1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * CDDL HEADER START
4 *
5 * The contents of this file are subject to the terms of the
6 * Common Development and Distribution License (the "License").
7 * You may not use this file except in compliance with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or https://opensource.org/licenses/CDDL-1.0.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22 /*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
25 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
26 * Copyright (c) 2024, 2025, Klara, Inc.
27 */
28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dbuf.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_pool.h>
37 #include <sys/zap_impl.h>
38 #include <sys/spa.h>
39 #include <sys/brt_impl.h>
40 #include <sys/sa.h>
41 #include <sys/sa_impl.h>
42 #include <sys/zfs_context.h>
43 #include <sys/trace_zfs.h>
44
45 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
46 uint64_t arg1, uint64_t arg2);
47
48 dmu_tx_stats_t dmu_tx_stats = {
49 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
50 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
51 { "dmu_tx_error", KSTAT_DATA_UINT64 },
52 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
53 { "dmu_tx_group", KSTAT_DATA_UINT64 },
54 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
55 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
56 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
57 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
58 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
59 { "dmu_tx_dirty_frees_delay", KSTAT_DATA_UINT64 },
60 { "dmu_tx_wrlog_delay", KSTAT_DATA_UINT64 },
61 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
62 };
63
64 static kstat_t *dmu_tx_ksp;
65
66 dmu_tx_t *
dmu_tx_create_dd(dsl_dir_t * dd)67 dmu_tx_create_dd(dsl_dir_t *dd)
68 {
69 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
70 tx->tx_dir = dd;
71 if (dd != NULL)
72 tx->tx_pool = dd->dd_pool;
73 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
74 offsetof(dmu_tx_hold_t, txh_node));
75 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
76 offsetof(dmu_tx_callback_t, dcb_node));
77 tx->tx_start = gethrtime();
78 return (tx);
79 }
80
81 dmu_tx_t *
dmu_tx_create(objset_t * os)82 dmu_tx_create(objset_t *os)
83 {
84 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
85 tx->tx_objset = os;
86 return (tx);
87 }
88
89 dmu_tx_t *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)90 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
91 {
92 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
93
94 TXG_VERIFY(dp->dp_spa, txg);
95 tx->tx_pool = dp;
96 tx->tx_txg = txg;
97 tx->tx_anyobj = TRUE;
98
99 return (tx);
100 }
101
102 int
dmu_tx_is_syncing(dmu_tx_t * tx)103 dmu_tx_is_syncing(dmu_tx_t *tx)
104 {
105 return (tx->tx_anyobj);
106 }
107
108 int
dmu_tx_private_ok(dmu_tx_t * tx)109 dmu_tx_private_ok(dmu_tx_t *tx)
110 {
111 return (tx->tx_anyobj);
112 }
113
114 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)115 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
116 uint64_t arg1, uint64_t arg2)
117 {
118 dmu_tx_hold_t *txh;
119
120 if (dn != NULL) {
121 (void) zfs_refcount_add(&dn->dn_holds, tx);
122 if (tx->tx_txg != 0) {
123 mutex_enter(&dn->dn_mtx);
124 /*
125 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
126 * problem, but there's no way for it to happen (for
127 * now, at least).
128 */
129 ASSERT(dn->dn_assigned_txg == 0);
130 dn->dn_assigned_txg = tx->tx_txg;
131 (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
132 mutex_exit(&dn->dn_mtx);
133 }
134 }
135
136 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
137 txh->txh_tx = tx;
138 txh->txh_dnode = dn;
139 zfs_refcount_create(&txh->txh_space_towrite);
140 zfs_refcount_create(&txh->txh_memory_tohold);
141 txh->txh_type = type;
142 txh->txh_arg1 = arg1;
143 txh->txh_arg2 = arg2;
144 list_insert_tail(&tx->tx_holds, txh);
145
146 return (txh);
147 }
148
149 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)150 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
151 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
152 {
153 dnode_t *dn = NULL;
154 dmu_tx_hold_t *txh;
155 int err;
156
157 if (object != DMU_NEW_OBJECT) {
158 err = dnode_hold(os, object, FTAG, &dn);
159 if (err != 0) {
160 tx->tx_err = err;
161 return (NULL);
162 }
163 }
164 txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
165 if (dn != NULL)
166 dnode_rele(dn, FTAG);
167 return (txh);
168 }
169
170 void
dmu_tx_add_new_object(dmu_tx_t * tx,dnode_t * dn)171 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
172 {
173 /*
174 * If we're syncing, they can manipulate any object anyhow, and
175 * the hold on the dnode_t can cause problems.
176 */
177 if (!dmu_tx_is_syncing(tx))
178 (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
179 }
180
181 /*
182 * This function reads specified data from disk. The specified data will
183 * be needed to perform the transaction -- i.e, it will be read after
184 * we do dmu_tx_assign(). There are two reasons that we read the data now
185 * (before dmu_tx_assign()):
186 *
187 * 1. Reading it now has potentially better performance. The transaction
188 * has not yet been assigned, so the TXG is not held open, and also the
189 * caller typically has less locks held when calling dmu_tx_hold_*() than
190 * after the transaction has been assigned. This reduces the lock (and txg)
191 * hold times, thus reducing lock contention.
192 *
193 * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
194 * that are detected before they start making changes to the DMU state
195 * (i.e. now). Once the transaction has been assigned, and some DMU
196 * state has been changed, it can be difficult to recover from an i/o
197 * error (e.g. to undo the changes already made in memory at the DMU
198 * layer). Typically code to do so does not exist in the caller -- it
199 * assumes that the data has already been cached and thus i/o errors are
200 * not possible.
201 *
202 * It has been observed that the i/o initiated here can be a performance
203 * problem, and it appears to be optional, because we don't look at the
204 * data which is read. However, removing this read would only serve to
205 * move the work elsewhere (after the dmu_tx_assign()), where it may
206 * have a greater impact on performance (in addition to the impact on
207 * fault tolerance noted above).
208 */
209 static int
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)210 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
211 {
212 int err;
213 dmu_buf_impl_t *db;
214
215 rw_enter(&dn->dn_struct_rwlock, RW_READER);
216 err = dbuf_hold_impl(dn, level, blkid, TRUE, FALSE, FTAG, &db);
217 rw_exit(&dn->dn_struct_rwlock);
218 if (err == ENOENT)
219 return (0);
220 if (err != 0)
221 return (err);
222 /*
223 * PARTIAL_FIRST allows caching for uncacheable blocks. It will
224 * be cleared after dmu_buf_will_dirty() call dbuf_read() again.
225 */
226 err = dbuf_read(db, zio, DB_RF_CANFAIL | DMU_READ_NO_PREFETCH |
227 (level == 0 ? (DMU_UNCACHEDIO | DMU_PARTIAL_FIRST) : 0));
228 dbuf_rele(db, FTAG);
229 return (err);
230 }
231
232 static void
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)233 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
234 {
235 dnode_t *dn = txh->txh_dnode;
236 int err = 0;
237
238 if (len == 0)
239 return;
240
241 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
242
243 if (dn == NULL)
244 return;
245
246 /*
247 * For i/o error checking, read the blocks that will be needed
248 * to perform the write: the first and last level-0 blocks (if
249 * they are not aligned, i.e. if they are partial-block writes),
250 * and all the level-1 blocks.
251 */
252 if (dn->dn_maxblkid == 0) {
253 if (off < dn->dn_datablksz &&
254 (off > 0 || len < dn->dn_datablksz)) {
255 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
256 if (err != 0) {
257 txh->txh_tx->tx_err = err;
258 }
259 }
260 } else {
261 zio_t *zio = zio_root(dn->dn_objset->os_spa,
262 NULL, NULL, ZIO_FLAG_CANFAIL);
263
264 /* first level-0 block */
265 uint64_t start = off >> dn->dn_datablkshift;
266 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
267 err = dmu_tx_check_ioerr(zio, dn, 0, start);
268 if (err != 0) {
269 txh->txh_tx->tx_err = err;
270 }
271 }
272
273 /* last level-0 block */
274 uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
275 if (end != start && end <= dn->dn_maxblkid &&
276 P2PHASE(off + len, dn->dn_datablksz)) {
277 err = dmu_tx_check_ioerr(zio, dn, 0, end);
278 if (err != 0) {
279 txh->txh_tx->tx_err = err;
280 }
281 }
282
283 /* level-1 blocks */
284 if (dn->dn_nlevels > 1) {
285 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
286 for (uint64_t i = (start >> shft) + 1;
287 i < end >> shft; i++) {
288 err = dmu_tx_check_ioerr(zio, dn, 1, i);
289 if (err != 0) {
290 txh->txh_tx->tx_err = err;
291 }
292 }
293 }
294
295 err = zio_wait(zio);
296 if (err != 0) {
297 txh->txh_tx->tx_err = err;
298 }
299 }
300 }
301
302 static void
dmu_tx_count_append(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)303 dmu_tx_count_append(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
304 {
305 dnode_t *dn = txh->txh_dnode;
306 int err = 0;
307
308 if (len == 0)
309 return;
310
311 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
312
313 if (dn == NULL)
314 return;
315
316 /*
317 * For i/o error checking, read the blocks that will be needed
318 * to perform the append; first level-0 block (if not aligned, i.e.
319 * if they are partial-block writes), no additional blocks are read.
320 */
321 if (dn->dn_maxblkid == 0) {
322 if (off < dn->dn_datablksz &&
323 (off > 0 || len < dn->dn_datablksz)) {
324 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
325 if (err != 0) {
326 txh->txh_tx->tx_err = err;
327 }
328 }
329 } else {
330 zio_t *zio = zio_root(dn->dn_objset->os_spa,
331 NULL, NULL, ZIO_FLAG_CANFAIL);
332
333 /* first level-0 block */
334 uint64_t start = off >> dn->dn_datablkshift;
335 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
336 err = dmu_tx_check_ioerr(zio, dn, 0, start);
337 if (err != 0) {
338 txh->txh_tx->tx_err = err;
339 }
340 }
341
342 err = zio_wait(zio);
343 if (err != 0) {
344 txh->txh_tx->tx_err = err;
345 }
346 }
347 }
348
349 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)350 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
351 {
352 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
353 DNODE_MIN_SIZE, FTAG);
354 }
355
356 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)357 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
358 {
359 dmu_tx_hold_t *txh;
360
361 ASSERT0(tx->tx_txg);
362 ASSERT3U(len, <=, DMU_MAX_ACCESS);
363 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
364
365 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
366 object, THT_WRITE, off, len);
367 if (txh != NULL) {
368 dmu_tx_count_write(txh, off, len);
369 dmu_tx_count_dnode(txh);
370 }
371 }
372
373 void
dmu_tx_hold_write_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)374 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
375 {
376 dmu_tx_hold_t *txh;
377
378 ASSERT0(tx->tx_txg);
379 ASSERT3U(len, <=, DMU_MAX_ACCESS);
380 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
381
382 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
383 if (txh != NULL) {
384 dmu_tx_count_write(txh, off, len);
385 dmu_tx_count_dnode(txh);
386 }
387 }
388
389 /*
390 * Should be used when appending to an object and the exact offset is unknown.
391 * The write must occur at or beyond the specified offset. Only the L0 block
392 * at provided offset will be prefetched.
393 */
394 void
dmu_tx_hold_append(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)395 dmu_tx_hold_append(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
396 {
397 dmu_tx_hold_t *txh;
398
399 ASSERT0(tx->tx_txg);
400 ASSERT3U(len, <=, DMU_MAX_ACCESS);
401
402 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
403 object, THT_APPEND, off, DMU_OBJECT_END);
404 if (txh != NULL) {
405 dmu_tx_count_append(txh, off, len);
406 dmu_tx_count_dnode(txh);
407 }
408 }
409
410 void
dmu_tx_hold_append_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)411 dmu_tx_hold_append_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
412 {
413 dmu_tx_hold_t *txh;
414
415 ASSERT0(tx->tx_txg);
416 ASSERT3U(len, <=, DMU_MAX_ACCESS);
417
418 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_APPEND, off, DMU_OBJECT_END);
419 if (txh != NULL) {
420 dmu_tx_count_append(txh, off, len);
421 dmu_tx_count_dnode(txh);
422 }
423 }
424
425 /*
426 * This function marks the transaction as being a "net free". The end
427 * result is that refquotas will be disabled for this transaction, and
428 * this transaction will be able to use half of the pool space overhead
429 * (see dsl_pool_adjustedsize()). Therefore this function should only
430 * be called for transactions that we expect will not cause a net increase
431 * in the amount of space used (but it's OK if that is occasionally not true).
432 */
433 void
dmu_tx_mark_netfree(dmu_tx_t * tx)434 dmu_tx_mark_netfree(dmu_tx_t *tx)
435 {
436 tx->tx_netfree = B_TRUE;
437 }
438
439 static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)440 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
441 {
442 dmu_tx_t *tx = txh->txh_tx;
443 dnode_t *dn = txh->txh_dnode;
444 int err;
445
446 ASSERT(tx->tx_txg == 0);
447
448 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
449 return;
450 if (len == DMU_OBJECT_END)
451 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
452
453 /*
454 * For i/o error checking, we read the first and last level-0
455 * blocks if they are not aligned, and all the level-1 blocks.
456 *
457 * Note: dbuf_free_range() assumes that we have not instantiated
458 * any level-0 dbufs that will be completely freed. Therefore we must
459 * exercise care to not read or count the first and last blocks
460 * if they are blocksize-aligned.
461 */
462 if (dn->dn_datablkshift == 0) {
463 if (off != 0 || len < dn->dn_datablksz)
464 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
465 } else {
466 /* first block will be modified if it is not aligned */
467 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
468 dmu_tx_count_write(txh, off, 1);
469 /* last block will be modified if it is not aligned */
470 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
471 dmu_tx_count_write(txh, off + len, 1);
472 }
473
474 /*
475 * Check level-1 blocks.
476 */
477 if (dn->dn_nlevels > 1) {
478 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
479 SPA_BLKPTRSHIFT;
480 uint64_t start = off >> shift;
481 uint64_t end = (off + len) >> shift;
482
483 ASSERT(dn->dn_indblkshift != 0);
484
485 /*
486 * dnode_reallocate() can result in an object with indirect
487 * blocks having an odd data block size. In this case,
488 * just check the single block.
489 */
490 if (dn->dn_datablkshift == 0)
491 start = end = 0;
492
493 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
494 NULL, NULL, ZIO_FLAG_CANFAIL);
495 for (uint64_t i = start; i <= end; i++) {
496 uint64_t ibyte = i << shift;
497 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
498 i = ibyte >> shift;
499 if (err == ESRCH || i > end)
500 break;
501 if (err != 0) {
502 tx->tx_err = err;
503 (void) zio_wait(zio);
504 return;
505 }
506
507 (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
508 1 << dn->dn_indblkshift, FTAG);
509
510 err = dmu_tx_check_ioerr(zio, dn, 1, i);
511 if (err != 0) {
512 tx->tx_err = err;
513 (void) zio_wait(zio);
514 return;
515 }
516 }
517 err = zio_wait(zio);
518 if (err != 0) {
519 tx->tx_err = err;
520 return;
521 }
522 }
523 }
524
525 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)526 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
527 {
528 dmu_tx_hold_t *txh;
529
530 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
531 object, THT_FREE, off, len);
532 if (txh != NULL) {
533 dmu_tx_count_dnode(txh);
534 dmu_tx_count_free(txh, off, len);
535 }
536 }
537
538 void
dmu_tx_hold_free_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len)539 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
540 {
541 dmu_tx_hold_t *txh;
542
543 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
544 if (txh != NULL) {
545 dmu_tx_count_dnode(txh);
546 dmu_tx_count_free(txh, off, len);
547 }
548 }
549
550 static void
dmu_tx_count_clone(dmu_tx_hold_t * txh,uint64_t off,uint64_t len,uint_t blksz)551 dmu_tx_count_clone(dmu_tx_hold_t *txh, uint64_t off, uint64_t len,
552 uint_t blksz)
553 {
554 dmu_tx_t *tx = txh->txh_tx;
555 dnode_t *dn = txh->txh_dnode;
556 int err;
557
558 ASSERT0(tx->tx_txg);
559 ASSERT(dn->dn_indblkshift != 0);
560 ASSERT(blksz != 0);
561 ASSERT0(off % blksz);
562
563 (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
564 len / blksz * sizeof (brt_entry_t), FTAG);
565
566 int shift = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
567 uint64_t start = off / blksz >> shift;
568 uint64_t end = (off + len) / blksz >> shift;
569
570 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
571 (end - start + 1) << dn->dn_indblkshift, FTAG);
572
573 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
574 NULL, NULL, ZIO_FLAG_CANFAIL);
575 for (uint64_t i = start; i <= end; i++) {
576 err = dmu_tx_check_ioerr(zio, dn, 1, i);
577 if (err != 0) {
578 tx->tx_err = err;
579 break;
580 }
581 }
582 err = zio_wait(zio);
583 if (err != 0)
584 tx->tx_err = err;
585 }
586
587 void
dmu_tx_hold_clone_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len,uint_t blksz)588 dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off,
589 uint64_t len, uint_t blksz)
590 {
591 dmu_tx_hold_t *txh;
592
593 ASSERT0(tx->tx_txg);
594 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
595
596 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_CLONE, off, len);
597 if (txh != NULL) {
598 dmu_tx_count_dnode(txh);
599 dmu_tx_count_clone(txh, off, len, blksz);
600 }
601 }
602
603 static void
dmu_tx_hold_zap_impl(dmu_tx_hold_t * txh,const char * name)604 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
605 {
606 dmu_tx_t *tx = txh->txh_tx;
607 dnode_t *dn = txh->txh_dnode;
608 int err;
609
610 ASSERT(tx->tx_txg == 0);
611
612 dmu_tx_count_dnode(txh);
613
614 /*
615 * Modifying a almost-full microzap is around the worst case (128KB)
616 *
617 * If it is a fat zap, the worst case would be 7*16KB=112KB:
618 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
619 * - 4 new blocks written if adding:
620 * - 2 blocks for possibly split leaves,
621 * - 2 grown ptrtbl blocks
622 */
623 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
624 zap_get_micro_max_size(tx->tx_pool->dp_spa), FTAG);
625
626 if (dn == NULL)
627 return;
628
629 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
630
631 if (dn->dn_maxblkid == 0 || name == NULL) {
632 /*
633 * This is a microzap (only one block), or we don't know
634 * the name. Check the first block for i/o errors.
635 */
636 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
637 if (err != 0) {
638 tx->tx_err = err;
639 }
640 } else {
641 /*
642 * Access the name so that we'll check for i/o errors to
643 * the leaf blocks, etc. We ignore ENOENT, as this name
644 * may not yet exist.
645 */
646 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
647 if (err == EIO || err == ECKSUM || err == ENXIO) {
648 tx->tx_err = err;
649 }
650 }
651 }
652
653 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)654 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
655 {
656 dmu_tx_hold_t *txh;
657
658 ASSERT0(tx->tx_txg);
659
660 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
661 object, THT_ZAP, add, (uintptr_t)name);
662 if (txh != NULL)
663 dmu_tx_hold_zap_impl(txh, name);
664 }
665
666 void
dmu_tx_hold_zap_by_dnode(dmu_tx_t * tx,dnode_t * dn,int add,const char * name)667 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
668 {
669 dmu_tx_hold_t *txh;
670
671 ASSERT0(tx->tx_txg);
672 ASSERT(dn != NULL);
673
674 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
675 if (txh != NULL)
676 dmu_tx_hold_zap_impl(txh, name);
677 }
678
679 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)680 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
681 {
682 dmu_tx_hold_t *txh;
683
684 ASSERT(tx->tx_txg == 0);
685
686 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
687 object, THT_BONUS, 0, 0);
688 if (txh)
689 dmu_tx_count_dnode(txh);
690 }
691
692 void
dmu_tx_hold_bonus_by_dnode(dmu_tx_t * tx,dnode_t * dn)693 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
694 {
695 dmu_tx_hold_t *txh;
696
697 ASSERT0(tx->tx_txg);
698
699 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
700 if (txh)
701 dmu_tx_count_dnode(txh);
702 }
703
704 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)705 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
706 {
707 dmu_tx_hold_t *txh;
708
709 ASSERT(tx->tx_txg == 0);
710
711 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
712 DMU_NEW_OBJECT, THT_SPACE, space, 0);
713 if (txh) {
714 (void) zfs_refcount_add_many(
715 &txh->txh_space_towrite, space, FTAG);
716 }
717 }
718
719 #ifdef ZFS_DEBUG
720 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)721 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
722 {
723 boolean_t match_object = B_FALSE;
724 boolean_t match_offset = B_FALSE;
725
726 DB_DNODE_ENTER(db);
727 dnode_t *dn = DB_DNODE(db);
728 ASSERT(tx->tx_txg != 0);
729 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
730 ASSERT3U(dn->dn_object, ==, db->db.db_object);
731
732 if (tx->tx_anyobj) {
733 DB_DNODE_EXIT(db);
734 return;
735 }
736
737 /* XXX No checking on the meta dnode for now */
738 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
739 DB_DNODE_EXIT(db);
740 return;
741 }
742
743 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
744 txh = list_next(&tx->tx_holds, txh)) {
745 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
746 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
747 match_object = TRUE;
748 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
749 int datablkshift = dn->dn_datablkshift ?
750 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
751 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
752 int shift = datablkshift + epbs * db->db_level;
753 uint64_t beginblk = shift >= 64 ? 0 :
754 (txh->txh_arg1 >> shift);
755 uint64_t endblk = shift >= 64 ? 0 :
756 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
757 uint64_t blkid = db->db_blkid;
758
759 /* XXX txh_arg2 better not be zero... */
760
761 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
762 txh->txh_type, (u_longlong_t)beginblk,
763 (u_longlong_t)endblk);
764
765 switch (txh->txh_type) {
766 case THT_WRITE:
767 if (blkid >= beginblk && blkid <= endblk)
768 match_offset = TRUE;
769 /*
770 * We will let this hold work for the bonus
771 * or spill buffer so that we don't need to
772 * hold it when creating a new object.
773 */
774 if (blkid == DMU_BONUS_BLKID ||
775 blkid == DMU_SPILL_BLKID)
776 match_offset = TRUE;
777 /*
778 * They might have to increase nlevels,
779 * thus dirtying the new TLIBs. Or the
780 * might have to change the block size,
781 * thus dirying the new lvl=0 blk=0.
782 */
783 if (blkid == 0)
784 match_offset = TRUE;
785 break;
786 case THT_APPEND:
787 if (blkid >= beginblk && (blkid <= endblk ||
788 txh->txh_arg2 == DMU_OBJECT_END))
789 match_offset = TRUE;
790
791 /*
792 * THT_WRITE used for bonus and spill blocks.
793 */
794 ASSERT(blkid != DMU_BONUS_BLKID &&
795 blkid != DMU_SPILL_BLKID);
796
797 /*
798 * They might have to increase nlevels,
799 * thus dirtying the new TLIBs. Or the
800 * might have to change the block size,
801 * thus dirying the new lvl=0 blk=0.
802 */
803 if (blkid == 0)
804 match_offset = TRUE;
805 break;
806 case THT_FREE:
807 /*
808 * We will dirty all the level 1 blocks in
809 * the free range and perhaps the first and
810 * last level 0 block.
811 */
812 if (blkid >= beginblk && (blkid <= endblk ||
813 txh->txh_arg2 == DMU_OBJECT_END))
814 match_offset = TRUE;
815 break;
816 case THT_SPILL:
817 if (blkid == DMU_SPILL_BLKID)
818 match_offset = TRUE;
819 break;
820 case THT_BONUS:
821 if (blkid == DMU_BONUS_BLKID)
822 match_offset = TRUE;
823 break;
824 case THT_ZAP:
825 match_offset = TRUE;
826 break;
827 case THT_NEWOBJECT:
828 match_object = TRUE;
829 break;
830 case THT_CLONE:
831 if (blkid >= beginblk && blkid <= endblk)
832 match_offset = TRUE;
833 /*
834 * They might have to increase nlevels,
835 * thus dirtying the new TLIBs. Or the
836 * might have to change the block size,
837 * thus dirying the new lvl=0 blk=0.
838 */
839 if (blkid == 0)
840 match_offset = TRUE;
841 break;
842 default:
843 cmn_err(CE_PANIC, "bad txh_type %d",
844 txh->txh_type);
845 }
846 }
847 if (match_object && match_offset) {
848 DB_DNODE_EXIT(db);
849 return;
850 }
851 }
852 DB_DNODE_EXIT(db);
853 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
854 (u_longlong_t)db->db.db_object, db->db_level,
855 (u_longlong_t)db->db_blkid);
856 }
857 #endif
858
859 /*
860 * If we can't do 10 iops, something is wrong. Let us go ahead
861 * and hit zfs_dirty_data_max.
862 */
863 static const hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
864
865 /*
866 * We delay transactions when we've determined that the backend storage
867 * isn't able to accommodate the rate of incoming writes.
868 *
869 * If there is already a transaction waiting, we delay relative to when
870 * that transaction finishes waiting. This way the calculated min_time
871 * is independent of the number of threads concurrently executing
872 * transactions.
873 *
874 * If we are the only waiter, wait relative to when the transaction
875 * started, rather than the current time. This credits the transaction for
876 * "time already served", e.g. reading indirect blocks.
877 *
878 * The minimum time for a transaction to take is calculated as:
879 * min_time = scale * (dirty - min) / (max - dirty)
880 * min_time is then capped at zfs_delay_max_ns.
881 *
882 * The delay has two degrees of freedom that can be adjusted via tunables.
883 * The percentage of dirty data at which we start to delay is defined by
884 * zfs_delay_min_dirty_percent. This should typically be at or above
885 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
886 * delay after writing at full speed has failed to keep up with the incoming
887 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
888 * speaking, this variable determines the amount of delay at the midpoint of
889 * the curve.
890 *
891 * delay
892 * 10ms +-------------------------------------------------------------*+
893 * | *|
894 * 9ms + *+
895 * | *|
896 * 8ms + *+
897 * | * |
898 * 7ms + * +
899 * | * |
900 * 6ms + * +
901 * | * |
902 * 5ms + * +
903 * | * |
904 * 4ms + * +
905 * | * |
906 * 3ms + * +
907 * | * |
908 * 2ms + (midpoint) * +
909 * | | ** |
910 * 1ms + v *** +
911 * | zfs_delay_scale ----------> ******** |
912 * 0 +-------------------------------------*********----------------+
913 * 0% <- zfs_dirty_data_max -> 100%
914 *
915 * Note that since the delay is added to the outstanding time remaining on the
916 * most recent transaction, the delay is effectively the inverse of IOPS.
917 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
918 * was chosen such that small changes in the amount of accumulated dirty data
919 * in the first 3/4 of the curve yield relatively small differences in the
920 * amount of delay.
921 *
922 * The effects can be easier to understand when the amount of delay is
923 * represented on a log scale:
924 *
925 * delay
926 * 100ms +-------------------------------------------------------------++
927 * + +
928 * | |
929 * + *+
930 * 10ms + *+
931 * + ** +
932 * | (midpoint) ** |
933 * + | ** +
934 * 1ms + v **** +
935 * + zfs_delay_scale ----------> ***** +
936 * | **** |
937 * + **** +
938 * 100us + ** +
939 * + * +
940 * | * |
941 * + * +
942 * 10us + * +
943 * + +
944 * | |
945 * + +
946 * +--------------------------------------------------------------+
947 * 0% <- zfs_dirty_data_max -> 100%
948 *
949 * Note here that only as the amount of dirty data approaches its limit does
950 * the delay start to increase rapidly. The goal of a properly tuned system
951 * should be to keep the amount of dirty data out of that range by first
952 * ensuring that the appropriate limits are set for the I/O scheduler to reach
953 * optimal throughput on the backend storage, and then by changing the value
954 * of zfs_delay_scale to increase the steepness of the curve.
955 */
956 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)957 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
958 {
959 dsl_pool_t *dp = tx->tx_pool;
960 uint64_t delay_min_bytes, wrlog;
961 hrtime_t wakeup, tx_time = 0, now;
962
963 /* Calculate minimum transaction time for the dirty data amount. */
964 delay_min_bytes =
965 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
966 if (dirty > delay_min_bytes) {
967 /*
968 * The caller has already waited until we are under the max.
969 * We make them pass us the amount of dirty data so we don't
970 * have to handle the case of it being >= the max, which
971 * could cause a divide-by-zero if it's == the max.
972 */
973 ASSERT3U(dirty, <, zfs_dirty_data_max);
974
975 tx_time = zfs_delay_scale * (dirty - delay_min_bytes) /
976 (zfs_dirty_data_max - dirty);
977 }
978
979 /* Calculate minimum transaction time for the TX_WRITE log size. */
980 wrlog = aggsum_upper_bound(&dp->dp_wrlog_total);
981 delay_min_bytes =
982 zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100;
983 if (wrlog >= zfs_wrlog_data_max) {
984 tx_time = zfs_delay_max_ns;
985 } else if (wrlog > delay_min_bytes) {
986 tx_time = MAX(zfs_delay_scale * (wrlog - delay_min_bytes) /
987 (zfs_wrlog_data_max - wrlog), tx_time);
988 }
989
990 if (tx_time == 0)
991 return;
992
993 tx_time = MIN(tx_time, zfs_delay_max_ns);
994 now = gethrtime();
995 if (now > tx->tx_start + tx_time)
996 return;
997
998 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
999 uint64_t, tx_time);
1000
1001 mutex_enter(&dp->dp_lock);
1002 wakeup = MAX(tx->tx_start + tx_time, dp->dp_last_wakeup + tx_time);
1003 dp->dp_last_wakeup = wakeup;
1004 mutex_exit(&dp->dp_lock);
1005
1006 zfs_sleep_until(wakeup);
1007 }
1008
1009 /*
1010 * This routine attempts to assign the transaction to a transaction group.
1011 * To do so, we must determine if there is sufficient free space on disk.
1012 *
1013 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
1014 * on it), then it is assumed that there is sufficient free space,
1015 * unless there's insufficient slop space in the pool (see the comment
1016 * above spa_slop_shift in spa_misc.c).
1017 *
1018 * If it is not a "netfree" transaction, then if the data already on disk
1019 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
1020 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
1021 * plus the rough estimate of this transaction's changes, may exceed the
1022 * allowed usage, then this will fail with ERESTART, which will cause the
1023 * caller to wait for the pending changes to be written to disk (by waiting
1024 * for the next TXG to open), and then check the space usage again.
1025 *
1026 * The rough estimate of pending changes is comprised of the sum of:
1027 *
1028 * - this transaction's holds' txh_space_towrite
1029 *
1030 * - dd_tempreserved[], which is the sum of in-flight transactions'
1031 * holds' txh_space_towrite (i.e. those transactions that have called
1032 * dmu_tx_assign() but not yet called dmu_tx_commit()).
1033 *
1034 * - dd_space_towrite[], which is the amount of dirtied dbufs.
1035 *
1036 * Note that all of these values are inflated by spa_get_worst_case_asize(),
1037 * which means that we may get ERESTART well before we are actually in danger
1038 * of running out of space, but this also mitigates any small inaccuracies
1039 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
1040 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
1041 * to the MOS).
1042 *
1043 * Note that due to this algorithm, it is possible to exceed the allowed
1044 * usage by one transaction. Also, as we approach the allowed usage,
1045 * we will allow a very limited amount of changes into each TXG, thus
1046 * decreasing performance.
1047 */
1048 static int
dmu_tx_try_assign(dmu_tx_t * tx)1049 dmu_tx_try_assign(dmu_tx_t *tx)
1050 {
1051 spa_t *spa = tx->tx_pool->dp_spa;
1052
1053 ASSERT0(tx->tx_txg);
1054
1055 if (tx->tx_err) {
1056 DMU_TX_STAT_BUMP(dmu_tx_error);
1057 return (tx->tx_err);
1058 }
1059
1060 if (spa_suspended(spa)) {
1061 DMU_TX_STAT_BUMP(dmu_tx_suspended);
1062
1063 /*
1064 * Let dmu_tx_assign() know specifically what happened, so
1065 * it can make the right choice based on the caller flags.
1066 */
1067 return (SET_ERROR(ESHUTDOWN));
1068 }
1069
1070 if (!tx->tx_dirty_delayed &&
1071 dsl_pool_need_wrlog_delay(tx->tx_pool)) {
1072 tx->tx_wait_dirty = B_TRUE;
1073 DMU_TX_STAT_BUMP(dmu_tx_wrlog_delay);
1074 return (SET_ERROR(ERESTART));
1075 }
1076
1077 if (!tx->tx_dirty_delayed &&
1078 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1079 tx->tx_wait_dirty = B_TRUE;
1080 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1081 return (SET_ERROR(ERESTART));
1082 }
1083
1084 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1085 tx->tx_needassign_txh = NULL;
1086
1087 /*
1088 * NB: No error returns are allowed after txg_hold_open, but
1089 * before processing the dnode holds, due to the
1090 * dmu_tx_unassign() logic.
1091 */
1092
1093 uint64_t towrite = 0;
1094 uint64_t tohold = 0;
1095 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1096 txh = list_next(&tx->tx_holds, txh)) {
1097 dnode_t *dn = txh->txh_dnode;
1098 if (dn != NULL) {
1099 /*
1100 * This thread can't hold the dn_struct_rwlock
1101 * while assigning the tx, because this can lead to
1102 * deadlock. Specifically, if this dnode is already
1103 * assigned to an earlier txg, this thread may need
1104 * to wait for that txg to sync (the ERESTART case
1105 * below). The other thread that has assigned this
1106 * dnode to an earlier txg prevents this txg from
1107 * syncing until its tx can complete (calling
1108 * dmu_tx_commit()), but it may need to acquire the
1109 * dn_struct_rwlock to do so (e.g. via
1110 * dmu_buf_hold*()).
1111 *
1112 * Note that this thread can't hold the lock for
1113 * read either, but the rwlock doesn't record
1114 * enough information to make that assertion.
1115 */
1116 ASSERT(!RW_WRITE_HELD(&dn->dn_struct_rwlock));
1117
1118 mutex_enter(&dn->dn_mtx);
1119 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1120 mutex_exit(&dn->dn_mtx);
1121 tx->tx_needassign_txh = txh;
1122 DMU_TX_STAT_BUMP(dmu_tx_group);
1123 return (SET_ERROR(ERESTART));
1124 }
1125 if (dn->dn_assigned_txg == 0)
1126 dn->dn_assigned_txg = tx->tx_txg;
1127 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1128 (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
1129 mutex_exit(&dn->dn_mtx);
1130 }
1131 towrite += zfs_refcount_count(&txh->txh_space_towrite);
1132 tohold += zfs_refcount_count(&txh->txh_memory_tohold);
1133 }
1134
1135 /* needed allocation: worst-case estimate of write space */
1136 uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
1137 /* calculate memory footprint estimate */
1138 uint64_t memory = towrite + tohold;
1139
1140 if (tx->tx_dir != NULL && asize != 0) {
1141 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1142 asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
1143 if (err != 0)
1144 return (err);
1145 }
1146
1147 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1148
1149 return (0);
1150 }
1151
1152 static void
dmu_tx_unassign(dmu_tx_t * tx)1153 dmu_tx_unassign(dmu_tx_t *tx)
1154 {
1155 if (tx->tx_txg == 0)
1156 return;
1157
1158 txg_rele_to_quiesce(&tx->tx_txgh);
1159
1160 /*
1161 * Walk the transaction's hold list, removing the hold on the
1162 * associated dnode, and notifying waiters if the refcount drops to 0.
1163 */
1164 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
1165 txh && txh != tx->tx_needassign_txh;
1166 txh = list_next(&tx->tx_holds, txh)) {
1167 dnode_t *dn = txh->txh_dnode;
1168
1169 if (dn == NULL)
1170 continue;
1171 mutex_enter(&dn->dn_mtx);
1172 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1173
1174 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1175 dn->dn_assigned_txg = 0;
1176 cv_broadcast(&dn->dn_notxholds);
1177 }
1178 mutex_exit(&dn->dn_mtx);
1179 }
1180
1181 txg_rele_to_sync(&tx->tx_txgh);
1182
1183 tx->tx_lasttried_txg = tx->tx_txg;
1184 tx->tx_txg = 0;
1185 }
1186
1187 /*
1188 * Assign tx to a transaction group; `flags` is a bitmask:
1189 *
1190 * If DMU_TX_WAIT is set and the currently open txg is full, this function
1191 * will wait until there's a new txg. This should be used when no locks
1192 * are being held. With this bit set, this function will only fail if
1193 * we're truly out of space (or over quota).
1194 *
1195 * If DMU_TX_WAIT is *not* set and we can't assign into the currently open
1196 * txg without blocking, this function will return immediately with
1197 * ERESTART. This should be used whenever locks are being held. On an
1198 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1199 * and try again.
1200 *
1201 * If DMU_TX_NOTHROTTLE is set, this indicates that this tx should not be
1202 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1203 * details on the throttle). This is used by the VFS operations, after
1204 * they have already called dmu_tx_wait() (though most likely on a
1205 * different tx).
1206 *
1207 * If DMU_TX_SUSPEND is set, this indicates that this tx should ignore
1208 * the pool being or becoming suspending while it is in progress. This will
1209 * cause dmu_tx_assign() (and dmu_tx_wait()) to block until the pool resumes.
1210 * If this flag is not set and the pool suspends, the return will be either
1211 * ERESTART or EIO, depending on the value of the pool's failmode= property.
1212 *
1213 * It is guaranteed that subsequent successful calls to dmu_tx_assign()
1214 * will assign the tx to monotonically increasing txgs. Of course this is
1215 * not strong monotonicity, because the same txg can be returned multiple
1216 * times in a row. This guarantee holds both for subsequent calls from
1217 * one thread and for multiple threads. For example, it is impossible to
1218 * observe the following sequence of events:
1219 *
1220 * Thread 1 Thread 2
1221 *
1222 * dmu_tx_assign(T1, ...)
1223 * 1 <- dmu_tx_get_txg(T1)
1224 * dmu_tx_assign(T2, ...)
1225 * 2 <- dmu_tx_get_txg(T2)
1226 * dmu_tx_assign(T3, ...)
1227 * 1 <- dmu_tx_get_txg(T3)
1228 */
1229 int
dmu_tx_assign(dmu_tx_t * tx,dmu_tx_flag_t flags)1230 dmu_tx_assign(dmu_tx_t *tx, dmu_tx_flag_t flags)
1231 {
1232 int err;
1233
1234 ASSERT(tx->tx_txg == 0);
1235 ASSERT0(flags & ~(DMU_TX_WAIT | DMU_TX_NOTHROTTLE | DMU_TX_SUSPEND));
1236 IMPLY(flags & DMU_TX_SUSPEND, flags & DMU_TX_WAIT);
1237 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1238
1239 /* If we might wait, we must not hold the config lock. */
1240 IMPLY((flags & DMU_TX_WAIT), !dsl_pool_config_held(tx->tx_pool));
1241
1242 if ((flags & DMU_TX_NOTHROTTLE))
1243 tx->tx_dirty_delayed = B_TRUE;
1244
1245 if (!(flags & DMU_TX_SUSPEND))
1246 tx->tx_break_on_suspend = B_TRUE;
1247
1248 while ((err = dmu_tx_try_assign(tx)) != 0) {
1249 dmu_tx_unassign(tx);
1250
1251 boolean_t suspended = (err == ESHUTDOWN);
1252 if (suspended) {
1253 /*
1254 * Pool suspended. We need to decide whether to block
1255 * and retry, or return error, depending on the
1256 * caller's flags and the pool config.
1257 */
1258 if (flags & DMU_TX_SUSPEND)
1259 /*
1260 * The caller expressly does not care about
1261 * suspend, so treat it as a normal retry.
1262 */
1263 err = SET_ERROR(ERESTART);
1264 else if ((flags & DMU_TX_WAIT) &&
1265 spa_get_failmode(tx->tx_pool->dp_spa) ==
1266 ZIO_FAILURE_MODE_CONTINUE)
1267 /*
1268 * Caller wants to wait, but pool config is
1269 * overriding that, so return EIO to be
1270 * propagated back to userspace.
1271 */
1272 err = SET_ERROR(EIO);
1273 else
1274 /* Anything else, we should just block. */
1275 err = SET_ERROR(ERESTART);
1276 }
1277
1278 /*
1279 * Return unless we decided to retry, or the caller does not
1280 * want to block.
1281 */
1282 if (err != ERESTART || !(flags & DMU_TX_WAIT))
1283 return (err);
1284
1285 /*
1286 * Wait until there's room in this txg, or until it's been
1287 * synced out and a new one is available.
1288 *
1289 * If we're here because the pool suspended above, then we
1290 * unset tx_break_on_suspend to make sure that if dmu_tx_wait()
1291 * has to fall back to a txg_wait_synced_flags(), it doesn't
1292 * immediately return because the pool is suspended. That would
1293 * then immediately return here, and we'd end up in a busy loop
1294 * until the pool resumes.
1295 *
1296 * On the other hand, if the pool hasn't suspended yet, then it
1297 * should be allowed to break a txg wait if the pool does
1298 * suspend, so we can loop and reassess it in
1299 * dmu_tx_try_assign().
1300 */
1301 if (suspended)
1302 tx->tx_break_on_suspend = B_FALSE;
1303
1304 dmu_tx_wait(tx);
1305
1306 /*
1307 * Reset tx_break_on_suspend for DMU_TX_SUSPEND. We do this
1308 * here so that it's available if we return for some other
1309 * reason, and then the caller calls dmu_tx_wait().
1310 */
1311 if (!(flags & DMU_TX_SUSPEND))
1312 tx->tx_break_on_suspend = B_TRUE;
1313 }
1314
1315 txg_rele_to_quiesce(&tx->tx_txgh);
1316
1317 return (0);
1318 }
1319
1320 void
dmu_tx_wait(dmu_tx_t * tx)1321 dmu_tx_wait(dmu_tx_t *tx)
1322 {
1323 spa_t *spa = tx->tx_pool->dp_spa;
1324 dsl_pool_t *dp = tx->tx_pool;
1325 hrtime_t before;
1326
1327 ASSERT(tx->tx_txg == 0);
1328 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1329
1330 /*
1331 * Break on suspend according to whether or not DMU_TX_SUSPEND was
1332 * supplied to the previous dmu_tx_assign() call. For clients, this
1333 * ensures that after dmu_tx_assign() fails, the followup dmu_tx_wait()
1334 * gets the same behaviour wrt suspend. See also the comments in
1335 * dmu_tx_assign().
1336 */
1337 txg_wait_flag_t flags =
1338 (tx->tx_break_on_suspend ? TXG_WAIT_SUSPEND : TXG_WAIT_NONE);
1339
1340 before = gethrtime();
1341
1342 if (tx->tx_wait_dirty) {
1343 uint64_t dirty;
1344
1345 /*
1346 * dmu_tx_try_assign() has determined that we need to wait
1347 * because we've consumed much or all of the dirty buffer
1348 * space.
1349 */
1350 mutex_enter(&dp->dp_lock);
1351 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1352 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1353 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1354 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1355 dirty = dp->dp_dirty_total;
1356 mutex_exit(&dp->dp_lock);
1357
1358 dmu_tx_delay(tx, dirty);
1359
1360 tx->tx_wait_dirty = B_FALSE;
1361
1362 /*
1363 * Note: setting tx_dirty_delayed only has effect if the
1364 * caller used DMU_TX_WAIT. Otherwise they are going to
1365 * destroy this tx and try again. The common case,
1366 * zfs_write(), uses DMU_TX_WAIT.
1367 */
1368 tx->tx_dirty_delayed = B_TRUE;
1369 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1370 /*
1371 * If the pool is suspended we need to wait until it
1372 * is resumed. Note that it's possible that the pool
1373 * has become active after this thread has tried to
1374 * obtain a tx. If that's the case then tx_lasttried_txg
1375 * would not have been set.
1376 */
1377 txg_wait_synced_flags(dp, spa_last_synced_txg(spa) + 1, flags);
1378 } else if (tx->tx_needassign_txh) {
1379 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1380
1381 mutex_enter(&dn->dn_mtx);
1382 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1383 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1384 mutex_exit(&dn->dn_mtx);
1385 tx->tx_needassign_txh = NULL;
1386 } else {
1387 /*
1388 * If we have a lot of dirty data just wait until we sync
1389 * out a TXG at which point we'll hopefully have synced
1390 * a portion of the changes.
1391 */
1392 txg_wait_synced_flags(dp, spa_last_synced_txg(spa) + 1, flags);
1393 }
1394
1395 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1396 }
1397
1398 static void
dmu_tx_destroy(dmu_tx_t * tx)1399 dmu_tx_destroy(dmu_tx_t *tx)
1400 {
1401 dmu_tx_hold_t *txh;
1402
1403 while ((txh = list_head(&tx->tx_holds)) != NULL) {
1404 dnode_t *dn = txh->txh_dnode;
1405
1406 list_remove(&tx->tx_holds, txh);
1407 zfs_refcount_destroy_many(&txh->txh_space_towrite,
1408 zfs_refcount_count(&txh->txh_space_towrite));
1409 zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1410 zfs_refcount_count(&txh->txh_memory_tohold));
1411 kmem_free(txh, sizeof (dmu_tx_hold_t));
1412 if (dn != NULL)
1413 dnode_rele(dn, tx);
1414 }
1415
1416 list_destroy(&tx->tx_callbacks);
1417 list_destroy(&tx->tx_holds);
1418 kmem_free(tx, sizeof (dmu_tx_t));
1419 }
1420
1421 void
dmu_tx_commit(dmu_tx_t * tx)1422 dmu_tx_commit(dmu_tx_t *tx)
1423 {
1424 /* This function should only be used on assigned transactions. */
1425 ASSERT(tx->tx_txg != 0);
1426
1427 /*
1428 * Go through the transaction's hold list and remove holds on
1429 * associated dnodes, notifying waiters if no holds remain.
1430 */
1431 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1432 txh = list_next(&tx->tx_holds, txh)) {
1433 dnode_t *dn = txh->txh_dnode;
1434
1435 if (dn == NULL)
1436 continue;
1437
1438 mutex_enter(&dn->dn_mtx);
1439 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1440
1441 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1442 dn->dn_assigned_txg = 0;
1443 cv_broadcast(&dn->dn_notxholds);
1444 }
1445 mutex_exit(&dn->dn_mtx);
1446 }
1447
1448 if (tx->tx_tempreserve_cookie)
1449 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1450
1451 if (!list_is_empty(&tx->tx_callbacks))
1452 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1453
1454 if (tx->tx_anyobj == FALSE)
1455 txg_rele_to_sync(&tx->tx_txgh);
1456
1457 dmu_tx_destroy(tx);
1458 }
1459
1460 void
dmu_tx_abort(dmu_tx_t * tx)1461 dmu_tx_abort(dmu_tx_t *tx)
1462 {
1463 /* This function should not be used on assigned transactions. */
1464 ASSERT0(tx->tx_txg);
1465
1466 /* Should not be needed, but better be safe than sorry. */
1467 if (tx->tx_tempreserve_cookie)
1468 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1469
1470 /*
1471 * Call any registered callbacks with an error code.
1472 */
1473 if (!list_is_empty(&tx->tx_callbacks))
1474 dmu_tx_do_callbacks(&tx->tx_callbacks, SET_ERROR(ECANCELED));
1475
1476 /* Should not be needed, but better be safe than sorry. */
1477 dmu_tx_unassign(tx);
1478
1479 dmu_tx_destroy(tx);
1480 }
1481
1482 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1483 dmu_tx_get_txg(dmu_tx_t *tx)
1484 {
1485 ASSERT(tx->tx_txg != 0);
1486 return (tx->tx_txg);
1487 }
1488
1489 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1490 dmu_tx_pool(dmu_tx_t *tx)
1491 {
1492 ASSERT(tx->tx_pool != NULL);
1493 return (tx->tx_pool);
1494 }
1495
1496 /*
1497 * Register a callback to be executed at the end of a TXG.
1498 *
1499 * Note: This currently exists for outside consumers, specifically the ZFS OSD
1500 * for Lustre. Please do not remove before checking that project. For examples
1501 * on how to use this see `ztest_commit_callback`.
1502 */
1503 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1504 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1505 {
1506 dmu_tx_callback_t *dcb;
1507
1508 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1509
1510 dcb->dcb_func = func;
1511 dcb->dcb_data = data;
1512
1513 list_insert_tail(&tx->tx_callbacks, dcb);
1514 }
1515
1516 /*
1517 * Call all the commit callbacks on a list, with a given error code.
1518 */
1519 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1520 dmu_tx_do_callbacks(list_t *cb_list, int error)
1521 {
1522 dmu_tx_callback_t *dcb;
1523
1524 while ((dcb = list_remove_tail(cb_list)) != NULL) {
1525 dcb->dcb_func(dcb->dcb_data, error);
1526 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1527 }
1528 }
1529
1530 /*
1531 * Interface to hold a bunch of attributes.
1532 * used for creating new files.
1533 * attrsize is the total size of all attributes
1534 * to be added during object creation
1535 *
1536 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1537 */
1538
1539 /*
1540 * hold necessary attribute name for attribute registration.
1541 * should be a very rare case where this is needed. If it does
1542 * happen it would only happen on the first write to the file system.
1543 */
1544 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1545 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1546 {
1547 if (!sa->sa_need_attr_registration)
1548 return;
1549
1550 for (int i = 0; i != sa->sa_num_attrs; i++) {
1551 if (!sa->sa_attr_table[i].sa_registered) {
1552 if (sa->sa_reg_attr_obj)
1553 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1554 B_TRUE, sa->sa_attr_table[i].sa_name);
1555 else
1556 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1557 B_TRUE, sa->sa_attr_table[i].sa_name);
1558 }
1559 }
1560 }
1561
1562 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1563 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1564 {
1565 dmu_tx_hold_t *txh;
1566
1567 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1568 THT_SPILL, 0, 0);
1569 if (txh != NULL)
1570 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
1571 SPA_OLD_MAXBLOCKSIZE, FTAG);
1572 }
1573
1574 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1575 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1576 {
1577 sa_os_t *sa = tx->tx_objset->os_sa;
1578
1579 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1580
1581 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1582 return;
1583
1584 if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1585 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1586 } else {
1587 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1588 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1589 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1590 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1591 }
1592
1593 dmu_tx_sa_registration_hold(sa, tx);
1594
1595 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1596 return;
1597
1598 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1599 THT_SPILL, 0, 0);
1600 }
1601
1602 /*
1603 * Hold SA attribute
1604 *
1605 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1606 *
1607 * variable_size is the total size of all variable sized attributes
1608 * passed to this function. It is not the total size of all
1609 * variable size attributes that *may* exist on this object.
1610 */
1611 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1612 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1613 {
1614 uint64_t object;
1615 sa_os_t *sa = tx->tx_objset->os_sa;
1616
1617 ASSERT(hdl != NULL);
1618
1619 object = sa_handle_object(hdl);
1620
1621 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1622 DB_DNODE_ENTER(db);
1623 dmu_tx_hold_bonus_by_dnode(tx, DB_DNODE(db));
1624 DB_DNODE_EXIT(db);
1625
1626 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1627 return;
1628
1629 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1630 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1631 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1632 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1633 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1634 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1635 }
1636
1637 dmu_tx_sa_registration_hold(sa, tx);
1638
1639 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1640 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1641
1642 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1643 ASSERT(tx->tx_txg == 0);
1644 dmu_tx_hold_spill(tx, object);
1645 } else {
1646 DB_DNODE_ENTER(db);
1647 if (DB_DNODE(db)->dn_have_spill) {
1648 ASSERT(tx->tx_txg == 0);
1649 dmu_tx_hold_spill(tx, object);
1650 }
1651 DB_DNODE_EXIT(db);
1652 }
1653 }
1654
1655 void
dmu_tx_init(void)1656 dmu_tx_init(void)
1657 {
1658 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1659 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1660 KSTAT_FLAG_VIRTUAL);
1661
1662 if (dmu_tx_ksp != NULL) {
1663 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1664 kstat_install(dmu_tx_ksp);
1665 }
1666 }
1667
1668 void
dmu_tx_fini(void)1669 dmu_tx_fini(void)
1670 {
1671 if (dmu_tx_ksp != NULL) {
1672 kstat_delete(dmu_tx_ksp);
1673 dmu_tx_ksp = NULL;
1674 }
1675 }
1676
1677 #if defined(_KERNEL)
1678 EXPORT_SYMBOL(dmu_tx_create);
1679 EXPORT_SYMBOL(dmu_tx_hold_write);
1680 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode);
1681 EXPORT_SYMBOL(dmu_tx_hold_append);
1682 EXPORT_SYMBOL(dmu_tx_hold_append_by_dnode);
1683 EXPORT_SYMBOL(dmu_tx_hold_free);
1684 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode);
1685 EXPORT_SYMBOL(dmu_tx_hold_zap);
1686 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode);
1687 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1688 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode);
1689 EXPORT_SYMBOL(dmu_tx_abort);
1690 EXPORT_SYMBOL(dmu_tx_assign);
1691 EXPORT_SYMBOL(dmu_tx_wait);
1692 EXPORT_SYMBOL(dmu_tx_commit);
1693 EXPORT_SYMBOL(dmu_tx_mark_netfree);
1694 EXPORT_SYMBOL(dmu_tx_get_txg);
1695 EXPORT_SYMBOL(dmu_tx_callback_register);
1696 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1697 EXPORT_SYMBOL(dmu_tx_hold_spill);
1698 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1699 EXPORT_SYMBOL(dmu_tx_hold_sa);
1700 #endif
1701