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