xref: /freebsd-13-stable/sys/contrib/openzfs/module/zfs/dmu_tx.c (revision e6c1e181ba7f666e02b073be104eb3e241097d83)
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  */
26 
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h>
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/trace_zfs.h>
41 
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43     uint64_t arg1, uint64_t arg2);
44 
45 dmu_tx_stats_t dmu_tx_stats = {
46 	{ "dmu_tx_assigned",		KSTAT_DATA_UINT64 },
47 	{ "dmu_tx_delay",		KSTAT_DATA_UINT64 },
48 	{ "dmu_tx_error",		KSTAT_DATA_UINT64 },
49 	{ "dmu_tx_suspended",		KSTAT_DATA_UINT64 },
50 	{ "dmu_tx_group",		KSTAT_DATA_UINT64 },
51 	{ "dmu_tx_memory_reserve",	KSTAT_DATA_UINT64 },
52 	{ "dmu_tx_memory_reclaim",	KSTAT_DATA_UINT64 },
53 	{ "dmu_tx_dirty_throttle",	KSTAT_DATA_UINT64 },
54 	{ "dmu_tx_dirty_delay",		KSTAT_DATA_UINT64 },
55 	{ "dmu_tx_dirty_over_max",	KSTAT_DATA_UINT64 },
56 	{ "dmu_tx_dirty_frees_delay",	KSTAT_DATA_UINT64 },
57 	{ "dmu_tx_wrlog_delay",		KSTAT_DATA_UINT64 },
58 	{ "dmu_tx_quota",		KSTAT_DATA_UINT64 },
59 };
60 
61 static kstat_t *dmu_tx_ksp;
62 
63 dmu_tx_t *
dmu_tx_create_dd(dsl_dir_t * dd)64 dmu_tx_create_dd(dsl_dir_t *dd)
65 {
66 	dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
67 	tx->tx_dir = dd;
68 	if (dd != NULL)
69 		tx->tx_pool = dd->dd_pool;
70 	list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
71 	    offsetof(dmu_tx_hold_t, txh_node));
72 	list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
73 	    offsetof(dmu_tx_callback_t, dcb_node));
74 	tx->tx_start = gethrtime();
75 	return (tx);
76 }
77 
78 dmu_tx_t *
dmu_tx_create(objset_t * os)79 dmu_tx_create(objset_t *os)
80 {
81 	dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
82 	tx->tx_objset = os;
83 	return (tx);
84 }
85 
86 dmu_tx_t *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)87 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
88 {
89 	dmu_tx_t *tx = dmu_tx_create_dd(NULL);
90 
91 	TXG_VERIFY(dp->dp_spa, txg);
92 	tx->tx_pool = dp;
93 	tx->tx_txg = txg;
94 	tx->tx_anyobj = TRUE;
95 
96 	return (tx);
97 }
98 
99 int
dmu_tx_is_syncing(dmu_tx_t * tx)100 dmu_tx_is_syncing(dmu_tx_t *tx)
101 {
102 	return (tx->tx_anyobj);
103 }
104 
105 int
dmu_tx_private_ok(dmu_tx_t * tx)106 dmu_tx_private_ok(dmu_tx_t *tx)
107 {
108 	return (tx->tx_anyobj);
109 }
110 
111 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)112 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
113     uint64_t arg1, uint64_t arg2)
114 {
115 	dmu_tx_hold_t *txh;
116 
117 	if (dn != NULL) {
118 		(void) zfs_refcount_add(&dn->dn_holds, tx);
119 		if (tx->tx_txg != 0) {
120 			mutex_enter(&dn->dn_mtx);
121 			/*
122 			 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
123 			 * problem, but there's no way for it to happen (for
124 			 * now, at least).
125 			 */
126 			ASSERT(dn->dn_assigned_txg == 0);
127 			dn->dn_assigned_txg = tx->tx_txg;
128 			(void) zfs_refcount_add(&dn->dn_tx_holds, tx);
129 			mutex_exit(&dn->dn_mtx);
130 		}
131 	}
132 
133 	txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
134 	txh->txh_tx = tx;
135 	txh->txh_dnode = dn;
136 	zfs_refcount_create(&txh->txh_space_towrite);
137 	zfs_refcount_create(&txh->txh_memory_tohold);
138 	txh->txh_type = type;
139 	txh->txh_arg1 = arg1;
140 	txh->txh_arg2 = arg2;
141 	list_insert_tail(&tx->tx_holds, txh);
142 
143 	return (txh);
144 }
145 
146 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)147 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
148     enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
149 {
150 	dnode_t *dn = NULL;
151 	dmu_tx_hold_t *txh;
152 	int err;
153 
154 	if (object != DMU_NEW_OBJECT) {
155 		err = dnode_hold(os, object, FTAG, &dn);
156 		if (err != 0) {
157 			tx->tx_err = err;
158 			return (NULL);
159 		}
160 	}
161 	txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
162 	if (dn != NULL)
163 		dnode_rele(dn, FTAG);
164 	return (txh);
165 }
166 
167 void
dmu_tx_add_new_object(dmu_tx_t * tx,dnode_t * dn)168 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
169 {
170 	/*
171 	 * If we're syncing, they can manipulate any object anyhow, and
172 	 * the hold on the dnode_t can cause problems.
173 	 */
174 	if (!dmu_tx_is_syncing(tx))
175 		(void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
176 }
177 
178 /*
179  * This function reads specified data from disk.  The specified data will
180  * be needed to perform the transaction -- i.e, it will be read after
181  * we do dmu_tx_assign().  There are two reasons that we read the data now
182  * (before dmu_tx_assign()):
183  *
184  * 1. Reading it now has potentially better performance.  The transaction
185  * has not yet been assigned, so the TXG is not held open, and also the
186  * caller typically has less locks held when calling dmu_tx_hold_*() than
187  * after the transaction has been assigned.  This reduces the lock (and txg)
188  * hold times, thus reducing lock contention.
189  *
190  * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
191  * that are detected before they start making changes to the DMU state
192  * (i.e. now).  Once the transaction has been assigned, and some DMU
193  * state has been changed, it can be difficult to recover from an i/o
194  * error (e.g. to undo the changes already made in memory at the DMU
195  * layer).  Typically code to do so does not exist in the caller -- it
196  * assumes that the data has already been cached and thus i/o errors are
197  * not possible.
198  *
199  * It has been observed that the i/o initiated here can be a performance
200  * problem, and it appears to be optional, because we don't look at the
201  * data which is read.  However, removing this read would only serve to
202  * move the work elsewhere (after the dmu_tx_assign()), where it may
203  * have a greater impact on performance (in addition to the impact on
204  * fault tolerance noted above).
205  */
206 static int
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)207 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
208 {
209 	int err;
210 	dmu_buf_impl_t *db;
211 
212 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
213 	db = dbuf_hold_level(dn, level, blkid, FTAG);
214 	rw_exit(&dn->dn_struct_rwlock);
215 	if (db == NULL)
216 		return (SET_ERROR(EIO));
217 	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
218 	dbuf_rele(db, FTAG);
219 	return (err);
220 }
221 
222 /* ARGSUSED */
223 static void
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)224 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
225 {
226 	dnode_t *dn = txh->txh_dnode;
227 	int err = 0;
228 
229 	if (len == 0)
230 		return;
231 
232 	(void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
233 
234 	if (dn == NULL)
235 		return;
236 
237 	/*
238 	 * For i/o error checking, read the blocks that will be needed
239 	 * to perform the write: the first and last level-0 blocks (if
240 	 * they are not aligned, i.e. if they are partial-block writes),
241 	 * and all the level-1 blocks.
242 	 */
243 	if (dn->dn_maxblkid == 0) {
244 		if (off < dn->dn_datablksz &&
245 		    (off > 0 || len < dn->dn_datablksz)) {
246 			err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
247 			if (err != 0) {
248 				txh->txh_tx->tx_err = err;
249 			}
250 		}
251 	} else {
252 		zio_t *zio = zio_root(dn->dn_objset->os_spa,
253 		    NULL, NULL, ZIO_FLAG_CANFAIL);
254 
255 		/* first level-0 block */
256 		uint64_t start = off >> dn->dn_datablkshift;
257 		if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
258 			err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 			if (err != 0) {
260 				txh->txh_tx->tx_err = err;
261 			}
262 		}
263 
264 		/* last level-0 block */
265 		uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
266 		if (end != start && end <= dn->dn_maxblkid &&
267 		    P2PHASE(off + len, dn->dn_datablksz)) {
268 			err = dmu_tx_check_ioerr(zio, dn, 0, end);
269 			if (err != 0) {
270 				txh->txh_tx->tx_err = err;
271 			}
272 		}
273 
274 		/* level-1 blocks */
275 		if (dn->dn_nlevels > 1) {
276 			int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
277 			for (uint64_t i = (start >> shft) + 1;
278 			    i < end >> shft; i++) {
279 				err = dmu_tx_check_ioerr(zio, dn, 1, i);
280 				if (err != 0) {
281 					txh->txh_tx->tx_err = err;
282 				}
283 			}
284 		}
285 
286 		err = zio_wait(zio);
287 		if (err != 0) {
288 			txh->txh_tx->tx_err = err;
289 		}
290 	}
291 }
292 
293 static void
dmu_tx_count_append(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)294 dmu_tx_count_append(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
295 {
296 	dnode_t *dn = txh->txh_dnode;
297 	int err = 0;
298 
299 	if (len == 0)
300 		return;
301 
302 	(void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
303 
304 	if (dn == NULL)
305 		return;
306 
307 	/*
308 	 * For i/o error checking, read the blocks that will be needed
309 	 * to perform the append; first level-0 block (if not aligned, i.e.
310 	 * if they are partial-block writes), no additional blocks are read.
311 	 */
312 	if (dn->dn_maxblkid == 0) {
313 		if (off < dn->dn_datablksz &&
314 		    (off > 0 || len < dn->dn_datablksz)) {
315 			err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
316 			if (err != 0) {
317 				txh->txh_tx->tx_err = err;
318 			}
319 		}
320 	} else {
321 		zio_t *zio = zio_root(dn->dn_objset->os_spa,
322 		    NULL, NULL, ZIO_FLAG_CANFAIL);
323 
324 		/* first level-0 block */
325 		uint64_t start = off >> dn->dn_datablkshift;
326 		if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
327 			err = dmu_tx_check_ioerr(zio, dn, 0, start);
328 			if (err != 0) {
329 				txh->txh_tx->tx_err = err;
330 			}
331 		}
332 
333 		err = zio_wait(zio);
334 		if (err != 0) {
335 			txh->txh_tx->tx_err = err;
336 		}
337 	}
338 }
339 
340 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)341 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
342 {
343 	(void) zfs_refcount_add_many(&txh->txh_space_towrite,
344 	    DNODE_MIN_SIZE, FTAG);
345 }
346 
347 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)348 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
349 {
350 	dmu_tx_hold_t *txh;
351 
352 	ASSERT0(tx->tx_txg);
353 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
354 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
355 
356 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
357 	    object, THT_WRITE, off, len);
358 	if (txh != NULL) {
359 		dmu_tx_count_write(txh, off, len);
360 		dmu_tx_count_dnode(txh);
361 	}
362 }
363 
364 void
dmu_tx_hold_write_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)365 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
366 {
367 	dmu_tx_hold_t *txh;
368 
369 	ASSERT0(tx->tx_txg);
370 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
371 	ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
372 
373 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
374 	if (txh != NULL) {
375 		dmu_tx_count_write(txh, off, len);
376 		dmu_tx_count_dnode(txh);
377 	}
378 }
379 
380 /*
381  * Should be used when appending to an object and the exact offset is unknown.
382  * The write must occur at or beyond the specified offset.  Only the L0 block
383  * at provided offset will be prefetched.
384  */
385 void
dmu_tx_hold_append(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)386 dmu_tx_hold_append(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
387 {
388 	dmu_tx_hold_t *txh;
389 
390 	ASSERT0(tx->tx_txg);
391 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
392 
393 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
394 	    object, THT_APPEND, off, DMU_OBJECT_END);
395 	if (txh != NULL) {
396 		dmu_tx_count_append(txh, off, len);
397 		dmu_tx_count_dnode(txh);
398 	}
399 }
400 
401 void
dmu_tx_hold_append_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)402 dmu_tx_hold_append_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
403 {
404 	dmu_tx_hold_t *txh;
405 
406 	ASSERT0(tx->tx_txg);
407 	ASSERT3U(len, <=, DMU_MAX_ACCESS);
408 
409 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_APPEND, off, DMU_OBJECT_END);
410 	if (txh != NULL) {
411 		dmu_tx_count_append(txh, off, len);
412 		dmu_tx_count_dnode(txh);
413 	}
414 }
415 
416 /*
417  * This function marks the transaction as being a "net free".  The end
418  * result is that refquotas will be disabled for this transaction, and
419  * this transaction will be able to use half of the pool space overhead
420  * (see dsl_pool_adjustedsize()).  Therefore this function should only
421  * be called for transactions that we expect will not cause a net increase
422  * in the amount of space used (but it's OK if that is occasionally not true).
423  */
424 void
dmu_tx_mark_netfree(dmu_tx_t * tx)425 dmu_tx_mark_netfree(dmu_tx_t *tx)
426 {
427 	tx->tx_netfree = B_TRUE;
428 }
429 
430 static void
dmu_tx_hold_free_impl(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)431 dmu_tx_hold_free_impl(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
432 {
433 	dmu_tx_t *tx = txh->txh_tx;
434 	dnode_t *dn = txh->txh_dnode;
435 	int err;
436 
437 	ASSERT(tx->tx_txg == 0);
438 
439 	dmu_tx_count_dnode(txh);
440 
441 	if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
442 		return;
443 	if (len == DMU_OBJECT_END)
444 		len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
445 
446 	dmu_tx_count_dnode(txh);
447 
448 	/*
449 	 * For i/o error checking, we read the first and last level-0
450 	 * blocks if they are not aligned, and all the level-1 blocks.
451 	 *
452 	 * Note:  dbuf_free_range() assumes that we have not instantiated
453 	 * any level-0 dbufs that will be completely freed.  Therefore we must
454 	 * exercise care to not read or count the first and last blocks
455 	 * if they are blocksize-aligned.
456 	 */
457 	if (dn->dn_datablkshift == 0) {
458 		if (off != 0 || len < dn->dn_datablksz)
459 			dmu_tx_count_write(txh, 0, dn->dn_datablksz);
460 	} else {
461 		/* first block will be modified if it is not aligned */
462 		if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
463 			dmu_tx_count_write(txh, off, 1);
464 		/* last block will be modified if it is not aligned */
465 		if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
466 			dmu_tx_count_write(txh, off + len, 1);
467 	}
468 
469 	/*
470 	 * Check level-1 blocks.
471 	 */
472 	if (dn->dn_nlevels > 1) {
473 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
474 		    SPA_BLKPTRSHIFT;
475 		uint64_t start = off >> shift;
476 		uint64_t end = (off + len) >> shift;
477 
478 		ASSERT(dn->dn_indblkshift != 0);
479 
480 		/*
481 		 * dnode_reallocate() can result in an object with indirect
482 		 * blocks having an odd data block size.  In this case,
483 		 * just check the single block.
484 		 */
485 		if (dn->dn_datablkshift == 0)
486 			start = end = 0;
487 
488 		zio_t *zio = zio_root(tx->tx_pool->dp_spa,
489 		    NULL, NULL, ZIO_FLAG_CANFAIL);
490 		for (uint64_t i = start; i <= end; i++) {
491 			uint64_t ibyte = i << shift;
492 			err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
493 			i = ibyte >> shift;
494 			if (err == ESRCH || i > end)
495 				break;
496 			if (err != 0) {
497 				tx->tx_err = err;
498 				(void) zio_wait(zio);
499 				return;
500 			}
501 
502 			(void) zfs_refcount_add_many(&txh->txh_memory_tohold,
503 			    1 << dn->dn_indblkshift, FTAG);
504 
505 			err = dmu_tx_check_ioerr(zio, dn, 1, i);
506 			if (err != 0) {
507 				tx->tx_err = err;
508 				(void) zio_wait(zio);
509 				return;
510 			}
511 		}
512 		err = zio_wait(zio);
513 		if (err != 0) {
514 			tx->tx_err = err;
515 			return;
516 		}
517 	}
518 }
519 
520 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)521 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
522 {
523 	dmu_tx_hold_t *txh;
524 
525 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
526 	    object, THT_FREE, off, len);
527 	if (txh != NULL)
528 		(void) dmu_tx_hold_free_impl(txh, off, len);
529 }
530 
531 void
dmu_tx_hold_free_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len)532 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
533 {
534 	dmu_tx_hold_t *txh;
535 
536 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
537 	if (txh != NULL)
538 		(void) dmu_tx_hold_free_impl(txh, off, len);
539 }
540 
541 static void
dmu_tx_hold_zap_impl(dmu_tx_hold_t * txh,const char * name)542 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
543 {
544 	dmu_tx_t *tx = txh->txh_tx;
545 	dnode_t *dn = txh->txh_dnode;
546 	int err;
547 
548 	ASSERT(tx->tx_txg == 0);
549 
550 	dmu_tx_count_dnode(txh);
551 
552 	/*
553 	 * Modifying a almost-full microzap is around the worst case (128KB)
554 	 *
555 	 * If it is a fat zap, the worst case would be 7*16KB=112KB:
556 	 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
557 	 * - 4 new blocks written if adding:
558 	 *    - 2 blocks for possibly split leaves,
559 	 *    - 2 grown ptrtbl blocks
560 	 */
561 	(void) zfs_refcount_add_many(&txh->txh_space_towrite,
562 	    MZAP_MAX_BLKSZ, FTAG);
563 
564 	if (dn == NULL)
565 		return;
566 
567 	ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
568 
569 	if (dn->dn_maxblkid == 0 || name == NULL) {
570 		/*
571 		 * This is a microzap (only one block), or we don't know
572 		 * the name.  Check the first block for i/o errors.
573 		 */
574 		err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
575 		if (err != 0) {
576 			tx->tx_err = err;
577 		}
578 	} else {
579 		/*
580 		 * Access the name so that we'll check for i/o errors to
581 		 * the leaf blocks, etc.  We ignore ENOENT, as this name
582 		 * may not yet exist.
583 		 */
584 		err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
585 		if (err == EIO || err == ECKSUM || err == ENXIO) {
586 			tx->tx_err = err;
587 		}
588 	}
589 }
590 
591 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)592 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
593 {
594 	dmu_tx_hold_t *txh;
595 
596 	ASSERT0(tx->tx_txg);
597 
598 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
599 	    object, THT_ZAP, add, (uintptr_t)name);
600 	if (txh != NULL)
601 		dmu_tx_hold_zap_impl(txh, name);
602 }
603 
604 void
dmu_tx_hold_zap_by_dnode(dmu_tx_t * tx,dnode_t * dn,int add,const char * name)605 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
606 {
607 	dmu_tx_hold_t *txh;
608 
609 	ASSERT0(tx->tx_txg);
610 	ASSERT(dn != NULL);
611 
612 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
613 	if (txh != NULL)
614 		dmu_tx_hold_zap_impl(txh, name);
615 }
616 
617 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)618 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
619 {
620 	dmu_tx_hold_t *txh;
621 
622 	ASSERT(tx->tx_txg == 0);
623 
624 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
625 	    object, THT_BONUS, 0, 0);
626 	if (txh)
627 		dmu_tx_count_dnode(txh);
628 }
629 
630 void
dmu_tx_hold_bonus_by_dnode(dmu_tx_t * tx,dnode_t * dn)631 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
632 {
633 	dmu_tx_hold_t *txh;
634 
635 	ASSERT0(tx->tx_txg);
636 
637 	txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
638 	if (txh)
639 		dmu_tx_count_dnode(txh);
640 }
641 
642 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)643 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
644 {
645 	dmu_tx_hold_t *txh;
646 
647 	ASSERT(tx->tx_txg == 0);
648 
649 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
650 	    DMU_NEW_OBJECT, THT_SPACE, space, 0);
651 	if (txh) {
652 		(void) zfs_refcount_add_many(
653 		    &txh->txh_space_towrite, space, FTAG);
654 	}
655 }
656 
657 #ifdef ZFS_DEBUG
658 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)659 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
660 {
661 	boolean_t match_object = B_FALSE;
662 	boolean_t match_offset = B_FALSE;
663 
664 	DB_DNODE_ENTER(db);
665 	dnode_t *dn = DB_DNODE(db);
666 	ASSERT(tx->tx_txg != 0);
667 	ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
668 	ASSERT3U(dn->dn_object, ==, db->db.db_object);
669 
670 	if (tx->tx_anyobj) {
671 		DB_DNODE_EXIT(db);
672 		return;
673 	}
674 
675 	/* XXX No checking on the meta dnode for now */
676 	if (db->db.db_object == DMU_META_DNODE_OBJECT) {
677 		DB_DNODE_EXIT(db);
678 		return;
679 	}
680 
681 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
682 	    txh = list_next(&tx->tx_holds, txh)) {
683 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
684 		if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
685 			match_object = TRUE;
686 		if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
687 			int datablkshift = dn->dn_datablkshift ?
688 			    dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
689 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
690 			int shift = datablkshift + epbs * db->db_level;
691 			uint64_t beginblk = shift >= 64 ? 0 :
692 			    (txh->txh_arg1 >> shift);
693 			uint64_t endblk = shift >= 64 ? 0 :
694 			    ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
695 			uint64_t blkid = db->db_blkid;
696 
697 			/* XXX txh_arg2 better not be zero... */
698 
699 			dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
700 			    txh->txh_type, (u_longlong_t)beginblk,
701 			    (u_longlong_t)endblk);
702 
703 			switch (txh->txh_type) {
704 			case THT_WRITE:
705 				if (blkid >= beginblk && blkid <= endblk)
706 					match_offset = TRUE;
707 				/*
708 				 * We will let this hold work for the bonus
709 				 * or spill buffer so that we don't need to
710 				 * hold it when creating a new object.
711 				 */
712 				if (blkid == DMU_BONUS_BLKID ||
713 				    blkid == DMU_SPILL_BLKID)
714 					match_offset = TRUE;
715 				/*
716 				 * They might have to increase nlevels,
717 				 * thus dirtying the new TLIBs.  Or the
718 				 * might have to change the block size,
719 				 * thus dirying the new lvl=0 blk=0.
720 				 */
721 				if (blkid == 0)
722 					match_offset = TRUE;
723 				break;
724 			case THT_APPEND:
725 				if (blkid >= beginblk && (blkid <= endblk ||
726 				    txh->txh_arg2 == DMU_OBJECT_END))
727 					match_offset = TRUE;
728 
729 				/*
730 				 * THT_WRITE used for bonus and spill blocks.
731 				 */
732 				ASSERT(blkid != DMU_BONUS_BLKID &&
733 				    blkid != DMU_SPILL_BLKID);
734 
735 				/*
736 				 * They might have to increase nlevels,
737 				 * thus dirtying the new TLIBs.  Or the
738 				 * might have to change the block size,
739 				 * thus dirying the new lvl=0 blk=0.
740 				 */
741 				if (blkid == 0)
742 					match_offset = TRUE;
743 				break;
744 			case THT_FREE:
745 				/*
746 				 * We will dirty all the level 1 blocks in
747 				 * the free range and perhaps the first and
748 				 * last level 0 block.
749 				 */
750 				if (blkid >= beginblk && (blkid <= endblk ||
751 				    txh->txh_arg2 == DMU_OBJECT_END))
752 					match_offset = TRUE;
753 				break;
754 			case THT_SPILL:
755 				if (blkid == DMU_SPILL_BLKID)
756 					match_offset = TRUE;
757 				break;
758 			case THT_BONUS:
759 				if (blkid == DMU_BONUS_BLKID)
760 					match_offset = TRUE;
761 				break;
762 			case THT_ZAP:
763 				match_offset = TRUE;
764 				break;
765 			case THT_NEWOBJECT:
766 				match_object = TRUE;
767 				break;
768 			default:
769 				cmn_err(CE_PANIC, "bad txh_type %d",
770 				    txh->txh_type);
771 			}
772 		}
773 		if (match_object && match_offset) {
774 			DB_DNODE_EXIT(db);
775 			return;
776 		}
777 	}
778 	DB_DNODE_EXIT(db);
779 	panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
780 	    (u_longlong_t)db->db.db_object, db->db_level,
781 	    (u_longlong_t)db->db_blkid);
782 }
783 #endif
784 
785 /*
786  * If we can't do 10 iops, something is wrong.  Let us go ahead
787  * and hit zfs_dirty_data_max.
788  */
789 hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
790 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
791 
792 /*
793  * We delay transactions when we've determined that the backend storage
794  * isn't able to accommodate the rate of incoming writes.
795  *
796  * If there is already a transaction waiting, we delay relative to when
797  * that transaction finishes waiting.  This way the calculated min_time
798  * is independent of the number of threads concurrently executing
799  * transactions.
800  *
801  * If we are the only waiter, wait relative to when the transaction
802  * started, rather than the current time.  This credits the transaction for
803  * "time already served", e.g. reading indirect blocks.
804  *
805  * The minimum time for a transaction to take is calculated as:
806  *     min_time = scale * (dirty - min) / (max - dirty)
807  *     min_time is then capped at zfs_delay_max_ns.
808  *
809  * The delay has two degrees of freedom that can be adjusted via tunables.
810  * The percentage of dirty data at which we start to delay is defined by
811  * zfs_delay_min_dirty_percent. This should typically be at or above
812  * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
813  * delay after writing at full speed has failed to keep up with the incoming
814  * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
815  * speaking, this variable determines the amount of delay at the midpoint of
816  * the curve.
817  *
818  * delay
819  *  10ms +-------------------------------------------------------------*+
820  *       |                                                             *|
821  *   9ms +                                                             *+
822  *       |                                                             *|
823  *   8ms +                                                             *+
824  *       |                                                            * |
825  *   7ms +                                                            * +
826  *       |                                                            * |
827  *   6ms +                                                            * +
828  *       |                                                            * |
829  *   5ms +                                                           *  +
830  *       |                                                           *  |
831  *   4ms +                                                           *  +
832  *       |                                                           *  |
833  *   3ms +                                                          *   +
834  *       |                                                          *   |
835  *   2ms +                                              (midpoint) *    +
836  *       |                                                  |    **     |
837  *   1ms +                                                  v ***       +
838  *       |             zfs_delay_scale ---------->     ********         |
839  *     0 +-------------------------------------*********----------------+
840  *       0%                    <- zfs_dirty_data_max ->               100%
841  *
842  * Note that since the delay is added to the outstanding time remaining on the
843  * most recent transaction, the delay is effectively the inverse of IOPS.
844  * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
845  * was chosen such that small changes in the amount of accumulated dirty data
846  * in the first 3/4 of the curve yield relatively small differences in the
847  * amount of delay.
848  *
849  * The effects can be easier to understand when the amount of delay is
850  * represented on a log scale:
851  *
852  * delay
853  * 100ms +-------------------------------------------------------------++
854  *       +                                                              +
855  *       |                                                              |
856  *       +                                                             *+
857  *  10ms +                                                             *+
858  *       +                                                           ** +
859  *       |                                              (midpoint)  **  |
860  *       +                                                  |     **    +
861  *   1ms +                                                  v ****      +
862  *       +             zfs_delay_scale ---------->        *****         +
863  *       |                                             ****             |
864  *       +                                          ****                +
865  * 100us +                                        **                    +
866  *       +                                       *                      +
867  *       |                                      *                       |
868  *       +                                     *                        +
869  *  10us +                                     *                        +
870  *       +                                                              +
871  *       |                                                              |
872  *       +                                                              +
873  *       +--------------------------------------------------------------+
874  *       0%                    <- zfs_dirty_data_max ->               100%
875  *
876  * Note here that only as the amount of dirty data approaches its limit does
877  * the delay start to increase rapidly. The goal of a properly tuned system
878  * should be to keep the amount of dirty data out of that range by first
879  * ensuring that the appropriate limits are set for the I/O scheduler to reach
880  * optimal throughput on the backend storage, and then by changing the value
881  * of zfs_delay_scale to increase the steepness of the curve.
882  */
883 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)884 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
885 {
886 	dsl_pool_t *dp = tx->tx_pool;
887 	uint64_t delay_min_bytes, wrlog;
888 	hrtime_t wakeup, tx_time = 0, now;
889 
890 	/* Calculate minimum transaction time for the dirty data amount. */
891 	delay_min_bytes =
892 	    zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
893 	if (dirty > delay_min_bytes) {
894 		/*
895 		 * The caller has already waited until we are under the max.
896 		 * We make them pass us the amount of dirty data so we don't
897 		 * have to handle the case of it being >= the max, which
898 		 * could cause a divide-by-zero if it's == the max.
899 		 */
900 		ASSERT3U(dirty, <, zfs_dirty_data_max);
901 
902 		tx_time = zfs_delay_scale * (dirty - delay_min_bytes) /
903 		    (zfs_dirty_data_max - dirty);
904 	}
905 
906 	/* Calculate minimum transaction time for the TX_WRITE log size. */
907 	wrlog = aggsum_upper_bound(&dp->dp_wrlog_total);
908 	delay_min_bytes =
909 	    zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100;
910 	if (wrlog >= zfs_wrlog_data_max) {
911 		tx_time = zfs_delay_max_ns;
912 	} else if (wrlog > delay_min_bytes) {
913 		tx_time = MAX(zfs_delay_scale * (wrlog - delay_min_bytes) /
914 		    (zfs_wrlog_data_max - wrlog), tx_time);
915 	}
916 
917 	if (tx_time == 0)
918 		return;
919 
920 	tx_time = MIN(tx_time, zfs_delay_max_ns);
921 	now = gethrtime();
922 	if (now > tx->tx_start + tx_time)
923 		return;
924 
925 	DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
926 	    uint64_t, tx_time);
927 
928 	mutex_enter(&dp->dp_lock);
929 	wakeup = MAX(tx->tx_start + tx_time, dp->dp_last_wakeup + tx_time);
930 	dp->dp_last_wakeup = wakeup;
931 	mutex_exit(&dp->dp_lock);
932 
933 	zfs_sleep_until(wakeup);
934 }
935 
936 /*
937  * This routine attempts to assign the transaction to a transaction group.
938  * To do so, we must determine if there is sufficient free space on disk.
939  *
940  * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
941  * on it), then it is assumed that there is sufficient free space,
942  * unless there's insufficient slop space in the pool (see the comment
943  * above spa_slop_shift in spa_misc.c).
944  *
945  * If it is not a "netfree" transaction, then if the data already on disk
946  * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
947  * ENOSPC.  Otherwise, if the current rough estimate of pending changes,
948  * plus the rough estimate of this transaction's changes, may exceed the
949  * allowed usage, then this will fail with ERESTART, which will cause the
950  * caller to wait for the pending changes to be written to disk (by waiting
951  * for the next TXG to open), and then check the space usage again.
952  *
953  * The rough estimate of pending changes is comprised of the sum of:
954  *
955  *  - this transaction's holds' txh_space_towrite
956  *
957  *  - dd_tempreserved[], which is the sum of in-flight transactions'
958  *    holds' txh_space_towrite (i.e. those transactions that have called
959  *    dmu_tx_assign() but not yet called dmu_tx_commit()).
960  *
961  *  - dd_space_towrite[], which is the amount of dirtied dbufs.
962  *
963  * Note that all of these values are inflated by spa_get_worst_case_asize(),
964  * which means that we may get ERESTART well before we are actually in danger
965  * of running out of space, but this also mitigates any small inaccuracies
966  * in the rough estimate (e.g. txh_space_towrite doesn't take into account
967  * indirect blocks, and dd_space_towrite[] doesn't take into account changes
968  * to the MOS).
969  *
970  * Note that due to this algorithm, it is possible to exceed the allowed
971  * usage by one transaction.  Also, as we approach the allowed usage,
972  * we will allow a very limited amount of changes into each TXG, thus
973  * decreasing performance.
974  */
975 static int
dmu_tx_try_assign(dmu_tx_t * tx,uint64_t txg_how)976 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
977 {
978 	spa_t *spa = tx->tx_pool->dp_spa;
979 
980 	ASSERT0(tx->tx_txg);
981 
982 	if (tx->tx_err) {
983 		DMU_TX_STAT_BUMP(dmu_tx_error);
984 		return (tx->tx_err);
985 	}
986 
987 	if (spa_suspended(spa)) {
988 		DMU_TX_STAT_BUMP(dmu_tx_suspended);
989 
990 		/*
991 		 * If the user has indicated a blocking failure mode
992 		 * then return ERESTART which will block in dmu_tx_wait().
993 		 * Otherwise, return EIO so that an error can get
994 		 * propagated back to the VOP calls.
995 		 *
996 		 * Note that we always honor the txg_how flag regardless
997 		 * of the failuremode setting.
998 		 */
999 		if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1000 		    !(txg_how & TXG_WAIT))
1001 			return (SET_ERROR(EIO));
1002 
1003 		return (SET_ERROR(ERESTART));
1004 	}
1005 
1006 	if (!tx->tx_dirty_delayed &&
1007 	    dsl_pool_need_wrlog_delay(tx->tx_pool)) {
1008 		tx->tx_wait_dirty = B_TRUE;
1009 		DMU_TX_STAT_BUMP(dmu_tx_wrlog_delay);
1010 		return (SET_ERROR(ERESTART));
1011 	}
1012 
1013 	if (!tx->tx_dirty_delayed &&
1014 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
1015 		tx->tx_wait_dirty = B_TRUE;
1016 		DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1017 		return (SET_ERROR(ERESTART));
1018 	}
1019 
1020 	tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1021 	tx->tx_needassign_txh = NULL;
1022 
1023 	/*
1024 	 * NB: No error returns are allowed after txg_hold_open, but
1025 	 * before processing the dnode holds, due to the
1026 	 * dmu_tx_unassign() logic.
1027 	 */
1028 
1029 	uint64_t towrite = 0;
1030 	uint64_t tohold = 0;
1031 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1032 	    txh = list_next(&tx->tx_holds, txh)) {
1033 		dnode_t *dn = txh->txh_dnode;
1034 		if (dn != NULL) {
1035 			/*
1036 			 * This thread can't hold the dn_struct_rwlock
1037 			 * while assigning the tx, because this can lead to
1038 			 * deadlock. Specifically, if this dnode is already
1039 			 * assigned to an earlier txg, this thread may need
1040 			 * to wait for that txg to sync (the ERESTART case
1041 			 * below).  The other thread that has assigned this
1042 			 * dnode to an earlier txg prevents this txg from
1043 			 * syncing until its tx can complete (calling
1044 			 * dmu_tx_commit()), but it may need to acquire the
1045 			 * dn_struct_rwlock to do so (e.g. via
1046 			 * dmu_buf_hold*()).
1047 			 *
1048 			 * Note that this thread can't hold the lock for
1049 			 * read either, but the rwlock doesn't record
1050 			 * enough information to make that assertion.
1051 			 */
1052 			ASSERT(!RW_WRITE_HELD(&dn->dn_struct_rwlock));
1053 
1054 			mutex_enter(&dn->dn_mtx);
1055 			if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1056 				mutex_exit(&dn->dn_mtx);
1057 				tx->tx_needassign_txh = txh;
1058 				DMU_TX_STAT_BUMP(dmu_tx_group);
1059 				return (SET_ERROR(ERESTART));
1060 			}
1061 			if (dn->dn_assigned_txg == 0)
1062 				dn->dn_assigned_txg = tx->tx_txg;
1063 			ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1064 			(void) zfs_refcount_add(&dn->dn_tx_holds, tx);
1065 			mutex_exit(&dn->dn_mtx);
1066 		}
1067 		towrite += zfs_refcount_count(&txh->txh_space_towrite);
1068 		tohold += zfs_refcount_count(&txh->txh_memory_tohold);
1069 	}
1070 
1071 	/* needed allocation: worst-case estimate of write space */
1072 	uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
1073 	/* calculate memory footprint estimate */
1074 	uint64_t memory = towrite + tohold;
1075 
1076 	if (tx->tx_dir != NULL && asize != 0) {
1077 		int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1078 		    asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
1079 		if (err != 0)
1080 			return (err);
1081 	}
1082 
1083 	DMU_TX_STAT_BUMP(dmu_tx_assigned);
1084 
1085 	return (0);
1086 }
1087 
1088 static void
dmu_tx_unassign(dmu_tx_t * tx)1089 dmu_tx_unassign(dmu_tx_t *tx)
1090 {
1091 	if (tx->tx_txg == 0)
1092 		return;
1093 
1094 	txg_rele_to_quiesce(&tx->tx_txgh);
1095 
1096 	/*
1097 	 * Walk the transaction's hold list, removing the hold on the
1098 	 * associated dnode, and notifying waiters if the refcount drops to 0.
1099 	 */
1100 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
1101 	    txh && txh != tx->tx_needassign_txh;
1102 	    txh = list_next(&tx->tx_holds, txh)) {
1103 		dnode_t *dn = txh->txh_dnode;
1104 
1105 		if (dn == NULL)
1106 			continue;
1107 		mutex_enter(&dn->dn_mtx);
1108 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1109 
1110 		if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1111 			dn->dn_assigned_txg = 0;
1112 			cv_broadcast(&dn->dn_notxholds);
1113 		}
1114 		mutex_exit(&dn->dn_mtx);
1115 	}
1116 
1117 	txg_rele_to_sync(&tx->tx_txgh);
1118 
1119 	tx->tx_lasttried_txg = tx->tx_txg;
1120 	tx->tx_txg = 0;
1121 }
1122 
1123 /*
1124  * Assign tx to a transaction group; txg_how is a bitmask:
1125  *
1126  * If TXG_WAIT is set and the currently open txg is full, this function
1127  * will wait until there's a new txg. This should be used when no locks
1128  * are being held. With this bit set, this function will only fail if
1129  * we're truly out of space (or over quota).
1130  *
1131  * If TXG_WAIT is *not* set and we can't assign into the currently open
1132  * txg without blocking, this function will return immediately with
1133  * ERESTART. This should be used whenever locks are being held.  On an
1134  * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1135  * and try again.
1136  *
1137  * If TXG_NOTHROTTLE is set, this indicates that this tx should not be
1138  * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1139  * details on the throttle). This is used by the VFS operations, after
1140  * they have already called dmu_tx_wait() (though most likely on a
1141  * different tx).
1142  *
1143  * It is guaranteed that subsequent successful calls to dmu_tx_assign()
1144  * will assign the tx to monotonically increasing txgs. Of course this is
1145  * not strong monotonicity, because the same txg can be returned multiple
1146  * times in a row. This guarantee holds both for subsequent calls from
1147  * one thread and for multiple threads. For example, it is impossible to
1148  * observe the following sequence of events:
1149  *
1150  *          Thread 1                            Thread 2
1151  *
1152  *     dmu_tx_assign(T1, ...)
1153  *     1 <- dmu_tx_get_txg(T1)
1154  *                                       dmu_tx_assign(T2, ...)
1155  *                                       2 <- dmu_tx_get_txg(T2)
1156  *     dmu_tx_assign(T3, ...)
1157  *     1 <- dmu_tx_get_txg(T3)
1158  */
1159 int
dmu_tx_assign(dmu_tx_t * tx,uint64_t txg_how)1160 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1161 {
1162 	int err;
1163 
1164 	ASSERT(tx->tx_txg == 0);
1165 	ASSERT0(txg_how & ~(TXG_WAIT | TXG_NOTHROTTLE));
1166 	ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1167 
1168 	/* If we might wait, we must not hold the config lock. */
1169 	IMPLY((txg_how & TXG_WAIT), !dsl_pool_config_held(tx->tx_pool));
1170 
1171 	if ((txg_how & TXG_NOTHROTTLE))
1172 		tx->tx_dirty_delayed = B_TRUE;
1173 
1174 	while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1175 		dmu_tx_unassign(tx);
1176 
1177 		if (err != ERESTART || !(txg_how & TXG_WAIT))
1178 			return (err);
1179 
1180 		dmu_tx_wait(tx);
1181 	}
1182 
1183 	txg_rele_to_quiesce(&tx->tx_txgh);
1184 
1185 	return (0);
1186 }
1187 
1188 void
dmu_tx_wait(dmu_tx_t * tx)1189 dmu_tx_wait(dmu_tx_t *tx)
1190 {
1191 	spa_t *spa = tx->tx_pool->dp_spa;
1192 	dsl_pool_t *dp = tx->tx_pool;
1193 	hrtime_t before;
1194 
1195 	ASSERT(tx->tx_txg == 0);
1196 	ASSERT(!dsl_pool_config_held(tx->tx_pool));
1197 
1198 	before = gethrtime();
1199 
1200 	if (tx->tx_wait_dirty) {
1201 		uint64_t dirty;
1202 
1203 		/*
1204 		 * dmu_tx_try_assign() has determined that we need to wait
1205 		 * because we've consumed much or all of the dirty buffer
1206 		 * space.
1207 		 */
1208 		mutex_enter(&dp->dp_lock);
1209 		if (dp->dp_dirty_total >= zfs_dirty_data_max)
1210 			DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1211 		while (dp->dp_dirty_total >= zfs_dirty_data_max)
1212 			cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1213 		dirty = dp->dp_dirty_total;
1214 		mutex_exit(&dp->dp_lock);
1215 
1216 		dmu_tx_delay(tx, dirty);
1217 
1218 		tx->tx_wait_dirty = B_FALSE;
1219 
1220 		/*
1221 		 * Note: setting tx_dirty_delayed only has effect if the
1222 		 * caller used TX_WAIT.  Otherwise they are going to
1223 		 * destroy this tx and try again.  The common case,
1224 		 * zfs_write(), uses TX_WAIT.
1225 		 */
1226 		tx->tx_dirty_delayed = B_TRUE;
1227 	} else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1228 		/*
1229 		 * If the pool is suspended we need to wait until it
1230 		 * is resumed.  Note that it's possible that the pool
1231 		 * has become active after this thread has tried to
1232 		 * obtain a tx.  If that's the case then tx_lasttried_txg
1233 		 * would not have been set.
1234 		 */
1235 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1236 	} else if (tx->tx_needassign_txh) {
1237 		dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1238 
1239 		mutex_enter(&dn->dn_mtx);
1240 		while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1241 			cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1242 		mutex_exit(&dn->dn_mtx);
1243 		tx->tx_needassign_txh = NULL;
1244 	} else {
1245 		/*
1246 		 * If we have a lot of dirty data just wait until we sync
1247 		 * out a TXG at which point we'll hopefully have synced
1248 		 * a portion of the changes.
1249 		 */
1250 		txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1251 	}
1252 
1253 	spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1254 }
1255 
1256 static void
dmu_tx_destroy(dmu_tx_t * tx)1257 dmu_tx_destroy(dmu_tx_t *tx)
1258 {
1259 	dmu_tx_hold_t *txh;
1260 
1261 	while ((txh = list_head(&tx->tx_holds)) != NULL) {
1262 		dnode_t *dn = txh->txh_dnode;
1263 
1264 		list_remove(&tx->tx_holds, txh);
1265 		zfs_refcount_destroy_many(&txh->txh_space_towrite,
1266 		    zfs_refcount_count(&txh->txh_space_towrite));
1267 		zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1268 		    zfs_refcount_count(&txh->txh_memory_tohold));
1269 		kmem_free(txh, sizeof (dmu_tx_hold_t));
1270 		if (dn != NULL)
1271 			dnode_rele(dn, tx);
1272 	}
1273 
1274 	list_destroy(&tx->tx_callbacks);
1275 	list_destroy(&tx->tx_holds);
1276 	kmem_free(tx, sizeof (dmu_tx_t));
1277 }
1278 
1279 void
dmu_tx_commit(dmu_tx_t * tx)1280 dmu_tx_commit(dmu_tx_t *tx)
1281 {
1282 	ASSERT(tx->tx_txg != 0);
1283 
1284 	/*
1285 	 * Go through the transaction's hold list and remove holds on
1286 	 * associated dnodes, notifying waiters if no holds remain.
1287 	 */
1288 	for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1289 	    txh = list_next(&tx->tx_holds, txh)) {
1290 		dnode_t *dn = txh->txh_dnode;
1291 
1292 		if (dn == NULL)
1293 			continue;
1294 
1295 		mutex_enter(&dn->dn_mtx);
1296 		ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1297 
1298 		if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1299 			dn->dn_assigned_txg = 0;
1300 			cv_broadcast(&dn->dn_notxholds);
1301 		}
1302 		mutex_exit(&dn->dn_mtx);
1303 	}
1304 
1305 	if (tx->tx_tempreserve_cookie)
1306 		dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1307 
1308 	if (!list_is_empty(&tx->tx_callbacks))
1309 		txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1310 
1311 	if (tx->tx_anyobj == FALSE)
1312 		txg_rele_to_sync(&tx->tx_txgh);
1313 
1314 	dmu_tx_destroy(tx);
1315 }
1316 
1317 void
dmu_tx_abort(dmu_tx_t * tx)1318 dmu_tx_abort(dmu_tx_t *tx)
1319 {
1320 	ASSERT(tx->tx_txg == 0);
1321 
1322 	/*
1323 	 * Call any registered callbacks with an error code.
1324 	 */
1325 	if (!list_is_empty(&tx->tx_callbacks))
1326 		dmu_tx_do_callbacks(&tx->tx_callbacks, SET_ERROR(ECANCELED));
1327 
1328 	dmu_tx_destroy(tx);
1329 }
1330 
1331 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1332 dmu_tx_get_txg(dmu_tx_t *tx)
1333 {
1334 	ASSERT(tx->tx_txg != 0);
1335 	return (tx->tx_txg);
1336 }
1337 
1338 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1339 dmu_tx_pool(dmu_tx_t *tx)
1340 {
1341 	ASSERT(tx->tx_pool != NULL);
1342 	return (tx->tx_pool);
1343 }
1344 
1345 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1346 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1347 {
1348 	dmu_tx_callback_t *dcb;
1349 
1350 	dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1351 
1352 	dcb->dcb_func = func;
1353 	dcb->dcb_data = data;
1354 
1355 	list_insert_tail(&tx->tx_callbacks, dcb);
1356 }
1357 
1358 /*
1359  * Call all the commit callbacks on a list, with a given error code.
1360  */
1361 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1362 dmu_tx_do_callbacks(list_t *cb_list, int error)
1363 {
1364 	dmu_tx_callback_t *dcb;
1365 
1366 	while ((dcb = list_tail(cb_list)) != NULL) {
1367 		list_remove(cb_list, dcb);
1368 		dcb->dcb_func(dcb->dcb_data, error);
1369 		kmem_free(dcb, sizeof (dmu_tx_callback_t));
1370 	}
1371 }
1372 
1373 /*
1374  * Interface to hold a bunch of attributes.
1375  * used for creating new files.
1376  * attrsize is the total size of all attributes
1377  * to be added during object creation
1378  *
1379  * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1380  */
1381 
1382 /*
1383  * hold necessary attribute name for attribute registration.
1384  * should be a very rare case where this is needed.  If it does
1385  * happen it would only happen on the first write to the file system.
1386  */
1387 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1388 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1389 {
1390 	if (!sa->sa_need_attr_registration)
1391 		return;
1392 
1393 	for (int i = 0; i != sa->sa_num_attrs; i++) {
1394 		if (!sa->sa_attr_table[i].sa_registered) {
1395 			if (sa->sa_reg_attr_obj)
1396 				dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1397 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1398 			else
1399 				dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1400 				    B_TRUE, sa->sa_attr_table[i].sa_name);
1401 		}
1402 	}
1403 }
1404 
1405 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1406 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1407 {
1408 	dmu_tx_hold_t *txh;
1409 
1410 	txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1411 	    THT_SPILL, 0, 0);
1412 	if (txh != NULL)
1413 		(void) zfs_refcount_add_many(&txh->txh_space_towrite,
1414 		    SPA_OLD_MAXBLOCKSIZE, FTAG);
1415 }
1416 
1417 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1418 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1419 {
1420 	sa_os_t *sa = tx->tx_objset->os_sa;
1421 
1422 	dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1423 
1424 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1425 		return;
1426 
1427 	if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1428 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1429 	} else {
1430 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1431 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1432 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1433 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1434 	}
1435 
1436 	dmu_tx_sa_registration_hold(sa, tx);
1437 
1438 	if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1439 		return;
1440 
1441 	(void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1442 	    THT_SPILL, 0, 0);
1443 }
1444 
1445 /*
1446  * Hold SA attribute
1447  *
1448  * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1449  *
1450  * variable_size is the total size of all variable sized attributes
1451  * passed to this function.  It is not the total size of all
1452  * variable size attributes that *may* exist on this object.
1453  */
1454 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1455 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1456 {
1457 	uint64_t object;
1458 	sa_os_t *sa = tx->tx_objset->os_sa;
1459 
1460 	ASSERT(hdl != NULL);
1461 
1462 	object = sa_handle_object(hdl);
1463 
1464 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1465 	DB_DNODE_ENTER(db);
1466 	dmu_tx_hold_bonus_by_dnode(tx, DB_DNODE(db));
1467 	DB_DNODE_EXIT(db);
1468 
1469 	if (tx->tx_objset->os_sa->sa_master_obj == 0)
1470 		return;
1471 
1472 	if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1473 	    tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1474 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1475 		dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1476 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1477 		dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1478 	}
1479 
1480 	dmu_tx_sa_registration_hold(sa, tx);
1481 
1482 	if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1483 		dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1484 
1485 	if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1486 		ASSERT(tx->tx_txg == 0);
1487 		dmu_tx_hold_spill(tx, object);
1488 	} else {
1489 		dnode_t *dn;
1490 
1491 		DB_DNODE_ENTER(db);
1492 		dn = DB_DNODE(db);
1493 		if (dn->dn_have_spill) {
1494 			ASSERT(tx->tx_txg == 0);
1495 			dmu_tx_hold_spill(tx, object);
1496 		}
1497 		DB_DNODE_EXIT(db);
1498 	}
1499 }
1500 
1501 void
dmu_tx_init(void)1502 dmu_tx_init(void)
1503 {
1504 	dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1505 	    KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1506 	    KSTAT_FLAG_VIRTUAL);
1507 
1508 	if (dmu_tx_ksp != NULL) {
1509 		dmu_tx_ksp->ks_data = &dmu_tx_stats;
1510 		kstat_install(dmu_tx_ksp);
1511 	}
1512 }
1513 
1514 void
dmu_tx_fini(void)1515 dmu_tx_fini(void)
1516 {
1517 	if (dmu_tx_ksp != NULL) {
1518 		kstat_delete(dmu_tx_ksp);
1519 		dmu_tx_ksp = NULL;
1520 	}
1521 }
1522 
1523 #if defined(_KERNEL)
1524 EXPORT_SYMBOL(dmu_tx_create);
1525 EXPORT_SYMBOL(dmu_tx_hold_write);
1526 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode);
1527 EXPORT_SYMBOL(dmu_tx_hold_append);
1528 EXPORT_SYMBOL(dmu_tx_hold_append_by_dnode);
1529 EXPORT_SYMBOL(dmu_tx_hold_free);
1530 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode);
1531 EXPORT_SYMBOL(dmu_tx_hold_zap);
1532 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode);
1533 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1534 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode);
1535 EXPORT_SYMBOL(dmu_tx_abort);
1536 EXPORT_SYMBOL(dmu_tx_assign);
1537 EXPORT_SYMBOL(dmu_tx_wait);
1538 EXPORT_SYMBOL(dmu_tx_commit);
1539 EXPORT_SYMBOL(dmu_tx_mark_netfree);
1540 EXPORT_SYMBOL(dmu_tx_get_txg);
1541 EXPORT_SYMBOL(dmu_tx_callback_register);
1542 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1543 EXPORT_SYMBOL(dmu_tx_hold_spill);
1544 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1545 EXPORT_SYMBOL(dmu_tx_hold_sa);
1546 #endif
1547