1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * This file and its contents are supplied under the terms of the
4 * Common Development and Distribution License ("CDDL"), version 1.0.
5 * You may only use this file in accordance with the terms of version
6 * 1.0 of the CDDL.
7 *
8 * A full copy of the text of the CDDL should have accompanied this
9 * source. A copy of the CDDL is also available via the Internet at
10 * https://opensource.org/license/CDDL-1.0.
11 */
12 /*
13 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
14 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
15 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
16 * Copyright (c) 2024, 2025, Klara, Inc.
17 */
18
19 #include <sys/dmu.h>
20 #include <sys/dmu_impl.h>
21 #include <sys/dbuf.h>
22 #include <sys/dmu_tx.h>
23 #include <sys/dmu_objset.h>
24 #include <sys/dsl_dataset.h>
25 #include <sys/dsl_dir.h>
26 #include <sys/dsl_pool.h>
27 #include <sys/zap_impl.h>
28 #include <sys/spa.h>
29 #include <sys/brt.h>
30 #include <sys/brt_impl.h>
31 #include <sys/sa.h>
32 #include <sys/sa_impl.h>
33 #include <sys/zfs_context.h>
34 #include <sys/trace_zfs.h>
35
36 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
37 uint64_t arg1, uint64_t arg2);
38
39 dmu_tx_stats_t dmu_tx_stats = {
40 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
41 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
42 { "dmu_tx_error", KSTAT_DATA_UINT64 },
43 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
44 { "dmu_tx_group", KSTAT_DATA_UINT64 },
45 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
46 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
47 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
48 { "dmu_tx_dirty_delay", KSTAT_DATA_UINT64 },
49 { "dmu_tx_dirty_over_max", KSTAT_DATA_UINT64 },
50 { "dmu_tx_dirty_frees_delay", KSTAT_DATA_UINT64 },
51 { "dmu_tx_wrlog_delay", KSTAT_DATA_UINT64 },
52 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
53 };
54
55 static kstat_t *dmu_tx_ksp;
56
57 dmu_tx_t *
dmu_tx_create_dd(dsl_dir_t * dd)58 dmu_tx_create_dd(dsl_dir_t *dd)
59 {
60 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
61 tx->tx_dir = dd;
62 if (dd != NULL)
63 tx->tx_pool = dd->dd_pool;
64 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
65 offsetof(dmu_tx_hold_t, txh_node));
66 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
67 offsetof(dmu_tx_callback_t, dcb_node));
68 tx->tx_start = gethrtime();
69 return (tx);
70 }
71
72 dmu_tx_t *
dmu_tx_create(objset_t * os)73 dmu_tx_create(objset_t *os)
74 {
75 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
76 tx->tx_objset = os;
77 return (tx);
78 }
79
80 dmu_tx_t *
dmu_tx_create_assigned(struct dsl_pool * dp,uint64_t txg)81 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
82 {
83 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
84
85 TXG_VERIFY(dp->dp_spa, txg);
86 tx->tx_pool = dp;
87 tx->tx_txg = txg;
88 tx->tx_anyobj = TRUE;
89
90 return (tx);
91 }
92
93 int
dmu_tx_is_syncing(dmu_tx_t * tx)94 dmu_tx_is_syncing(dmu_tx_t *tx)
95 {
96 return (tx->tx_anyobj);
97 }
98
99 int
dmu_tx_private_ok(dmu_tx_t * tx)100 dmu_tx_private_ok(dmu_tx_t *tx)
101 {
102 return (tx->tx_anyobj);
103 }
104
105 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)106 dmu_tx_hold_dnode_impl(dmu_tx_t *tx, dnode_t *dn, enum dmu_tx_hold_type type,
107 uint64_t arg1, uint64_t arg2)
108 {
109 dmu_tx_hold_t *txh;
110
111 if (dn != NULL) {
112 (void) zfs_refcount_add(&dn->dn_holds, tx);
113 if (tx->tx_txg != 0) {
114 mutex_enter(&dn->dn_mtx);
115 /*
116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 * problem, but there's no way for it to happen (for
118 * now, at least).
119 */
120 ASSERT0(dn->dn_assigned_txg);
121 dn->dn_assigned_txg = tx->tx_txg;
122 (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
123 mutex_exit(&dn->dn_mtx);
124 }
125 }
126
127 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
128 txh->txh_tx = tx;
129 txh->txh_dnode = dn;
130 zfs_refcount_create(&txh->txh_space_towrite);
131 zfs_refcount_create(&txh->txh_memory_tohold);
132 txh->txh_type = type;
133 txh->txh_arg1 = arg1;
134 txh->txh_arg2 = arg2;
135 list_insert_tail(&tx->tx_holds, txh);
136
137 return (txh);
138 }
139
140 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)141 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
142 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
143 {
144 dnode_t *dn = NULL;
145 dmu_tx_hold_t *txh;
146 int err;
147
148 if (object != DMU_NEW_OBJECT) {
149 err = dnode_hold(os, object, FTAG, &dn);
150 if (err != 0) {
151 tx->tx_err = err;
152 return (NULL);
153 }
154 }
155 txh = dmu_tx_hold_dnode_impl(tx, dn, type, arg1, arg2);
156 if (dn != NULL)
157 dnode_rele(dn, FTAG);
158 return (txh);
159 }
160
161 void
dmu_tx_add_new_object(dmu_tx_t * tx,dnode_t * dn)162 dmu_tx_add_new_object(dmu_tx_t *tx, dnode_t *dn)
163 {
164 /*
165 * If we're syncing, they can manipulate any object anyhow, and
166 * the hold on the dnode_t can cause problems.
167 */
168 if (!dmu_tx_is_syncing(tx))
169 (void) dmu_tx_hold_dnode_impl(tx, dn, THT_NEWOBJECT, 0, 0);
170 }
171
172 /*
173 * This function reads specified data from disk. The specified data will
174 * be needed to perform the transaction -- i.e, it will be read after
175 * we do dmu_tx_assign(). There are two reasons that we read the data now
176 * (before dmu_tx_assign()):
177 *
178 * 1. Reading it now has potentially better performance. The transaction
179 * has not yet been assigned, so the TXG is not held open, and also the
180 * caller typically has less locks held when calling dmu_tx_hold_*() than
181 * after the transaction has been assigned. This reduces the lock (and txg)
182 * hold times, thus reducing lock contention.
183 *
184 * 2. It is easier for callers (primarily the ZPL) to handle i/o errors
185 * that are detected before they start making changes to the DMU state
186 * (i.e. now). Once the transaction has been assigned, and some DMU
187 * state has been changed, it can be difficult to recover from an i/o
188 * error (e.g. to undo the changes already made in memory at the DMU
189 * layer). Typically code to do so does not exist in the caller -- it
190 * assumes that the data has already been cached and thus i/o errors are
191 * not possible.
192 *
193 * It has been observed that the i/o initiated here can be a performance
194 * problem, and it appears to be optional, because we don't look at the
195 * data which is read. However, removing this read would only serve to
196 * move the work elsewhere (after the dmu_tx_assign()), where it may
197 * have a greater impact on performance (in addition to the impact on
198 * fault tolerance noted above).
199 */
200 static int
dmu_tx_check_ioerr(zio_t * zio,dnode_t * dn,int level,uint64_t blkid)201 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
202 {
203 int err;
204 dmu_buf_impl_t *db;
205
206 rw_enter(&dn->dn_struct_rwlock, RW_READER);
207 err = dbuf_hold_impl(dn, level, blkid, TRUE, FALSE, FTAG, &db);
208 rw_exit(&dn->dn_struct_rwlock);
209 if (err == ENOENT)
210 return (0);
211 if (err != 0)
212 return (err);
213 /*
214 * DMU_IS_PREFETCH keeps the buffer temporarily in DBUF cache and ARC
215 * to avoid immediate eviction after the check. It will be promoted
216 * to demand access when dmu_buf_will_dirty() read it again.
217 */
218 err = dbuf_read(db, zio, DB_RF_CANFAIL | DMU_READ_NO_PREFETCH |
219 (level == 0 ? (DMU_KEEP_CACHING | DMU_IS_PREFETCH) : 0));
220 dbuf_rele(db, FTAG);
221 return (err);
222 }
223
224 static void
dmu_tx_count_write(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)225 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
226 {
227 dnode_t *dn = txh->txh_dnode;
228 int err = 0;
229
230 if (len == 0)
231 return;
232
233 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
234
235 if (dn == NULL)
236 return;
237
238 /*
239 * For i/o error checking, read the blocks that will be needed
240 * to perform the write: the first and last level-0 blocks (if
241 * they are not aligned, i.e. if they are partial-block writes),
242 * and all the level-1 blocks.
243 */
244 if (dn->dn_maxblkid == 0) {
245 if (off < dn->dn_datablksz &&
246 (off > 0 || len < dn->dn_datablksz)) {
247 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
248 if (err != 0) {
249 txh->txh_tx->tx_err = err;
250 }
251 }
252 } else {
253 zio_t *zio = zio_root(dn->dn_objset->os_spa,
254 NULL, NULL, ZIO_FLAG_CANFAIL);
255
256 /* first level-0 block */
257 uint64_t start = off >> dn->dn_datablkshift;
258 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
259 err = dmu_tx_check_ioerr(zio, dn, 0, start);
260 if (err != 0) {
261 txh->txh_tx->tx_err = err;
262 }
263 }
264
265 /* last level-0 block */
266 uint64_t end = (off + len - 1) >> dn->dn_datablkshift;
267 if (end != start && end <= dn->dn_maxblkid &&
268 P2PHASE(off + len, dn->dn_datablksz)) {
269 err = dmu_tx_check_ioerr(zio, dn, 0, end);
270 if (err != 0) {
271 txh->txh_tx->tx_err = err;
272 }
273 }
274
275 /* level-1 blocks */
276 if (dn->dn_nlevels > 1) {
277 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
278 for (uint64_t i = (start >> shft) + 1;
279 i < end >> shft; i++) {
280 err = dmu_tx_check_ioerr(zio, dn, 1, i);
281 if (err != 0) {
282 txh->txh_tx->tx_err = err;
283 }
284 }
285 }
286
287 err = zio_wait(zio);
288 if (err != 0) {
289 txh->txh_tx->tx_err = err;
290 }
291 }
292 }
293
294 static void
dmu_tx_count_append(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)295 dmu_tx_count_append(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
296 {
297 dnode_t *dn = txh->txh_dnode;
298 int err = 0;
299
300 if (len == 0)
301 return;
302
303 (void) zfs_refcount_add_many(&txh->txh_space_towrite, len, FTAG);
304
305 if (dn == NULL)
306 return;
307
308 /*
309 * For i/o error checking, read the blocks that will be needed
310 * to perform the append; first level-0 block (if not aligned, i.e.
311 * if they are partial-block writes), no additional blocks are read.
312 */
313 if (dn->dn_maxblkid == 0) {
314 if (off < dn->dn_datablksz &&
315 (off > 0 || len < dn->dn_datablksz)) {
316 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
317 if (err != 0) {
318 txh->txh_tx->tx_err = err;
319 }
320 }
321 } else {
322 zio_t *zio = zio_root(dn->dn_objset->os_spa,
323 NULL, NULL, ZIO_FLAG_CANFAIL);
324
325 /* first level-0 block */
326 uint64_t start = off >> dn->dn_datablkshift;
327 if (P2PHASE(off, dn->dn_datablksz) || len < dn->dn_datablksz) {
328 err = dmu_tx_check_ioerr(zio, dn, 0, start);
329 if (err != 0) {
330 txh->txh_tx->tx_err = err;
331 }
332 }
333
334 err = zio_wait(zio);
335 if (err != 0) {
336 txh->txh_tx->tx_err = err;
337 }
338 }
339 }
340
341 static void
dmu_tx_count_dnode(dmu_tx_hold_t * txh)342 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
343 {
344 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
345 DNODE_MIN_SIZE, FTAG);
346 }
347
348 void
dmu_tx_hold_write(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)349 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
350 {
351 dmu_tx_hold_t *txh;
352
353 ASSERT0(tx->tx_txg);
354 ASSERT3U(len, <=, DMU_MAX_ACCESS);
355 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
356
357 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
358 object, THT_WRITE, off, len);
359 if (txh != NULL) {
360 dmu_tx_count_write(txh, off, len);
361 dmu_tx_count_dnode(txh);
362 }
363 }
364
365 void
dmu_tx_hold_write_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)366 dmu_tx_hold_write_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
367 {
368 dmu_tx_hold_t *txh;
369
370 ASSERT0(tx->tx_txg);
371 ASSERT3U(len, <=, DMU_MAX_ACCESS);
372 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
373
374 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_WRITE, off, len);
375 if (txh != NULL) {
376 dmu_tx_count_write(txh, off, len);
377 dmu_tx_count_dnode(txh);
378 }
379 }
380
381 /*
382 * Should be used when appending to an object and the exact offset is unknown.
383 * The write must occur at or beyond the specified offset. Only the L0 block
384 * at provided offset will be prefetched.
385 */
386 void
dmu_tx_hold_append(dmu_tx_t * tx,uint64_t object,uint64_t off,int len)387 dmu_tx_hold_append(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
388 {
389 dmu_tx_hold_t *txh;
390
391 ASSERT0(tx->tx_txg);
392 ASSERT3U(len, <=, DMU_MAX_ACCESS);
393
394 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
395 object, THT_APPEND, off, DMU_OBJECT_END);
396 if (txh != NULL) {
397 dmu_tx_count_append(txh, off, len);
398 dmu_tx_count_dnode(txh);
399 }
400 }
401
402 void
dmu_tx_hold_append_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,int len)403 dmu_tx_hold_append_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, int len)
404 {
405 dmu_tx_hold_t *txh;
406
407 ASSERT0(tx->tx_txg);
408 ASSERT3U(len, <=, DMU_MAX_ACCESS);
409
410 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_APPEND, off, DMU_OBJECT_END);
411 if (txh != NULL) {
412 dmu_tx_count_append(txh, off, len);
413 dmu_tx_count_dnode(txh);
414 }
415 }
416
417 /*
418 * This function marks the transaction as being a "net free". The end
419 * result is that refquotas will be disabled for this transaction, and
420 * this transaction will be able to use half of the pool space overhead
421 * (see dsl_pool_adjustedsize()). Therefore this function should only
422 * be called for transactions that we expect will not cause a net increase
423 * in the amount of space used (but it's OK if that is occasionally not true).
424 */
425 void
dmu_tx_mark_netfree(dmu_tx_t * tx)426 dmu_tx_mark_netfree(dmu_tx_t *tx)
427 {
428 tx->tx_netfree = B_TRUE;
429 }
430
431 static void
dmu_tx_count_free(dmu_tx_hold_t * txh,uint64_t off,uint64_t len)432 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
433 {
434 dmu_tx_t *tx = txh->txh_tx;
435 dnode_t *dn = txh->txh_dnode;
436 int err;
437
438 ASSERT0(tx->tx_txg);
439
440 if (off >= (dn->dn_maxblkid + 1) * dn->dn_datablksz)
441 return;
442 if (len == DMU_OBJECT_END)
443 len = (dn->dn_maxblkid + 1) * dn->dn_datablksz - off;
444
445 /*
446 * For i/o error checking, we read the first and last level-0
447 * blocks if they are not aligned, and all the level-1 blocks.
448 *
449 * Note: dbuf_free_range() assumes that we have not instantiated
450 * any level-0 dbufs that will be completely freed. Therefore we must
451 * exercise care to not read or count the first and last blocks
452 * if they are blocksize-aligned.
453 */
454 if (dn->dn_datablkshift == 0) {
455 if (off != 0 || len < dn->dn_datablksz)
456 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
457 } else {
458 /* first block will be modified if it is not aligned */
459 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
460 dmu_tx_count_write(txh, off, 1);
461 /* last block will be modified if it is not aligned */
462 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
463 dmu_tx_count_write(txh, off + len, 1);
464 }
465
466 /*
467 * Check level-1 blocks.
468 */
469 if (dn->dn_nlevels > 1) {
470 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
471 SPA_BLKPTRSHIFT;
472 uint64_t start = off >> shift;
473 uint64_t end = (off + len) >> shift;
474
475 ASSERT(dn->dn_indblkshift != 0);
476
477 /*
478 * dnode_reallocate() can result in an object with indirect
479 * blocks having an odd data block size. In this case,
480 * just check the single block.
481 */
482 if (dn->dn_datablkshift == 0)
483 start = end = 0;
484
485 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
486 NULL, NULL, ZIO_FLAG_CANFAIL);
487 for (uint64_t i = start; i <= end; i++) {
488 uint64_t ibyte = i << shift;
489 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
490 i = ibyte >> shift;
491 if (err == ESRCH || i > end)
492 break;
493 if (err != 0) {
494 tx->tx_err = err;
495 (void) zio_wait(zio);
496 return;
497 }
498
499 (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
500 1 << dn->dn_indblkshift, FTAG);
501
502 err = dmu_tx_check_ioerr(zio, dn, 1, i);
503 if (err != 0) {
504 tx->tx_err = err;
505 (void) zio_wait(zio);
506 return;
507 }
508 }
509 err = zio_wait(zio);
510 if (err != 0) {
511 tx->tx_err = err;
512 return;
513 }
514 }
515 }
516
517 void
dmu_tx_hold_free(dmu_tx_t * tx,uint64_t object,uint64_t off,uint64_t len)518 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
519 {
520 dmu_tx_hold_t *txh;
521
522 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
523 object, THT_FREE, off, len);
524 if (txh != NULL) {
525 dmu_tx_count_dnode(txh);
526 dmu_tx_count_free(txh, off, len);
527 }
528 }
529
530 void
dmu_tx_hold_free_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len)531 dmu_tx_hold_free_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off, uint64_t len)
532 {
533 dmu_tx_hold_t *txh;
534
535 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_FREE, off, len);
536 if (txh != NULL) {
537 dmu_tx_count_dnode(txh);
538 dmu_tx_count_free(txh, off, len);
539 }
540 }
541
542 static void
dmu_tx_count_clone(dmu_tx_hold_t * txh,uint64_t off,uint64_t len,uint_t blksz)543 dmu_tx_count_clone(dmu_tx_hold_t *txh, uint64_t off, uint64_t len,
544 uint_t blksz)
545 {
546 dmu_tx_t *tx = txh->txh_tx;
547 dnode_t *dn = txh->txh_dnode;
548 int err;
549
550 ASSERT0(tx->tx_txg);
551 ASSERT(dn->dn_indblkshift != 0);
552 ASSERT(blksz != 0);
553 ASSERT0(off % blksz);
554
555 /*
556 * The last block of the range is cloned in full even if the range
557 * covers it only partially, in case it is the last block of the file.
558 */
559 len = P2ROUNDUP(len, (uint64_t)blksz);
560
561 (void) zfs_refcount_add_many(&txh->txh_memory_tohold,
562 len / blksz * sizeof (brt_entry_t), FTAG);
563
564 int shift = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
565 uint64_t start = off / blksz >> shift;
566 uint64_t end = (off + len) / blksz >> shift;
567
568 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
569 (end - start + 1) << dn->dn_indblkshift, FTAG);
570
571 zio_t *zio = zio_root(tx->tx_pool->dp_spa,
572 NULL, NULL, ZIO_FLAG_CANFAIL);
573 for (uint64_t i = start; i <= end; i++) {
574 err = dmu_tx_check_ioerr(zio, dn, 1, i);
575 if (err != 0) {
576 tx->tx_err = err;
577 (void) zio_wait(zio);
578 return;
579 }
580 }
581 err = zio_wait(zio);
582 if (err != 0) {
583 tx->tx_err = err;
584 return;
585 }
586
587 /*
588 * Each cloned block may dirty one BRT ZAP leaf in sync context.
589 */
590 tx->tx_sync_reserve += len / blksz *
591 brt_sync_dirty_est(tx->tx_pool->dp_spa);
592 }
593
594 void
dmu_tx_hold_clone_by_dnode(dmu_tx_t * tx,dnode_t * dn,uint64_t off,uint64_t len,uint_t blksz)595 dmu_tx_hold_clone_by_dnode(dmu_tx_t *tx, dnode_t *dn, uint64_t off,
596 uint64_t len, uint_t blksz)
597 {
598 dmu_tx_hold_t *txh;
599
600 ASSERT0(tx->tx_txg);
601 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
602
603 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_CLONE, off, len);
604 if (txh != NULL) {
605 dmu_tx_count_dnode(txh);
606 dmu_tx_count_clone(txh, off, len, blksz);
607 }
608 }
609
610 static void
dmu_tx_hold_zap_impl(dmu_tx_hold_t * txh,const char * name)611 dmu_tx_hold_zap_impl(dmu_tx_hold_t *txh, const char *name)
612 {
613 dmu_tx_t *tx = txh->txh_tx;
614 dnode_t *dn = txh->txh_dnode;
615 int err;
616
617 ASSERT0(tx->tx_txg);
618
619 dmu_tx_count_dnode(txh);
620
621 /*
622 * Modifying a almost-full microzap is around the worst case (128KB)
623 *
624 * If it is a fat zap, the worst case would be 7*16KB=112KB:
625 * - 3 blocks overwritten: target leaf, ptrtbl block, header block
626 * - 4 new blocks written if adding:
627 * - 2 blocks for possibly split leaves,
628 * - 2 grown ptrtbl blocks
629 */
630 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
631 zap_get_micro_max_size(tx->tx_pool->dp_spa), FTAG);
632
633 if (dn == NULL)
634 return;
635
636 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
637
638 if (dn->dn_maxblkid == 0 || name == NULL) {
639 /*
640 * This is a microzap (only one block), or we don't know
641 * the name. Check the first block for i/o errors.
642 */
643 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
644 if (err != 0) {
645 tx->tx_err = err;
646 }
647 } else {
648 /*
649 * Access the name so that we'll check for i/o errors to
650 * the leaf blocks, etc. We ignore ENOENT, as this name
651 * may not yet exist.
652 */
653 err = zap_lookup_by_dnode(dn, name, 8, 0, NULL);
654 if (err == EIO || err == ECKSUM || err == ENXIO) {
655 tx->tx_err = err;
656 }
657 }
658 }
659
660 void
dmu_tx_hold_zap(dmu_tx_t * tx,uint64_t object,int add,const char * name)661 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
662 {
663 dmu_tx_hold_t *txh;
664
665 ASSERT0(tx->tx_txg);
666
667 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
668 object, THT_ZAP, add, (uintptr_t)name);
669 if (txh != NULL)
670 dmu_tx_hold_zap_impl(txh, name);
671 }
672
673 void
dmu_tx_hold_zap_by_dnode(dmu_tx_t * tx,dnode_t * dn,int add,const char * name)674 dmu_tx_hold_zap_by_dnode(dmu_tx_t *tx, dnode_t *dn, int add, const char *name)
675 {
676 dmu_tx_hold_t *txh;
677
678 ASSERT0(tx->tx_txg);
679 ASSERT(dn != NULL);
680
681 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_ZAP, add, (uintptr_t)name);
682 if (txh != NULL)
683 dmu_tx_hold_zap_impl(txh, name);
684 }
685
686 void
dmu_tx_hold_bonus(dmu_tx_t * tx,uint64_t object)687 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
688 {
689 dmu_tx_hold_t *txh;
690
691 ASSERT0(tx->tx_txg);
692
693 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
694 object, THT_BONUS, 0, 0);
695 if (txh)
696 dmu_tx_count_dnode(txh);
697 }
698
699 void
dmu_tx_hold_bonus_by_dnode(dmu_tx_t * tx,dnode_t * dn)700 dmu_tx_hold_bonus_by_dnode(dmu_tx_t *tx, dnode_t *dn)
701 {
702 dmu_tx_hold_t *txh;
703
704 ASSERT0(tx->tx_txg);
705
706 txh = dmu_tx_hold_dnode_impl(tx, dn, THT_BONUS, 0, 0);
707 if (txh)
708 dmu_tx_count_dnode(txh);
709 }
710
711 void
dmu_tx_hold_space(dmu_tx_t * tx,uint64_t space)712 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
713 {
714 dmu_tx_hold_t *txh;
715
716 ASSERT0(tx->tx_txg);
717
718 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
719 DMU_NEW_OBJECT, THT_SPACE, space, 0);
720 if (txh) {
721 (void) zfs_refcount_add_many(
722 &txh->txh_space_towrite, space, FTAG);
723 }
724 }
725
726 #ifdef ZFS_DEBUG
727 void
dmu_tx_dirty_buf(dmu_tx_t * tx,dmu_buf_impl_t * db)728 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
729 {
730 boolean_t match_object = B_FALSE;
731 boolean_t match_offset = B_FALSE;
732
733 DB_DNODE_ENTER(db);
734 dnode_t *dn = DB_DNODE(db);
735 ASSERT(tx->tx_txg != 0);
736 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
737 ASSERT3U(dn->dn_object, ==, db->db.db_object);
738
739 if (tx->tx_anyobj) {
740 DB_DNODE_EXIT(db);
741 return;
742 }
743
744 /* XXX No checking on the meta dnode for now */
745 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
746 DB_DNODE_EXIT(db);
747 return;
748 }
749
750 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
751 txh = list_next(&tx->tx_holds, txh)) {
752 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
753 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
754 match_object = TRUE;
755 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
756 int datablkshift = dn->dn_datablkshift ?
757 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
758 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
759 int shift = datablkshift + epbs * db->db_level;
760 uint64_t beginblk = shift >= 64 ? 0 :
761 (txh->txh_arg1 >> shift);
762 uint64_t endblk = shift >= 64 ? 0 :
763 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
764 uint64_t blkid = db->db_blkid;
765
766 /* XXX txh_arg2 better not be zero... */
767
768 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
769 txh->txh_type, (u_longlong_t)beginblk,
770 (u_longlong_t)endblk);
771
772 switch (txh->txh_type) {
773 case THT_WRITE:
774 if (blkid >= beginblk && blkid <= endblk)
775 match_offset = TRUE;
776 /*
777 * We will let this hold work for the bonus
778 * or spill buffer so that we don't need to
779 * hold it when creating a new object.
780 */
781 if (blkid == DMU_BONUS_BLKID ||
782 blkid == DMU_SPILL_BLKID)
783 match_offset = TRUE;
784 /*
785 * They might have to increase nlevels,
786 * thus dirtying the new TLIBs. Or the
787 * might have to change the block size,
788 * thus dirying the new lvl=0 blk=0.
789 */
790 if (blkid == 0)
791 match_offset = TRUE;
792 break;
793 case THT_APPEND:
794 if (blkid >= beginblk && (blkid <= endblk ||
795 txh->txh_arg2 == DMU_OBJECT_END))
796 match_offset = TRUE;
797
798 /*
799 * THT_WRITE used for bonus and spill blocks.
800 */
801 ASSERT(blkid != DMU_BONUS_BLKID &&
802 blkid != DMU_SPILL_BLKID);
803
804 /*
805 * They might have to increase nlevels,
806 * thus dirtying the new TLIBs. Or the
807 * might have to change the block size,
808 * thus dirying the new lvl=0 blk=0.
809 */
810 if (blkid == 0)
811 match_offset = TRUE;
812 break;
813 case THT_FREE:
814 /*
815 * We will dirty all the level 1 blocks in
816 * the free range and perhaps the first and
817 * last level 0 block.
818 */
819 if (blkid >= beginblk && (blkid <= endblk ||
820 txh->txh_arg2 == DMU_OBJECT_END))
821 match_offset = TRUE;
822 break;
823 case THT_SPILL:
824 if (blkid == DMU_SPILL_BLKID)
825 match_offset = TRUE;
826 break;
827 case THT_BONUS:
828 if (blkid == DMU_BONUS_BLKID)
829 match_offset = TRUE;
830 break;
831 case THT_ZAP:
832 match_offset = TRUE;
833 break;
834 case THT_NEWOBJECT:
835 match_object = TRUE;
836 break;
837 case THT_CLONE:
838 if (blkid >= beginblk && blkid <= endblk)
839 match_offset = TRUE;
840 /*
841 * They might have to increase nlevels,
842 * thus dirtying the new TLIBs. Or the
843 * might have to change the block size,
844 * thus dirying the new lvl=0 blk=0.
845 */
846 if (blkid == 0)
847 match_offset = TRUE;
848 break;
849 default:
850 cmn_err(CE_PANIC, "bad txh_type %d",
851 txh->txh_type);
852 }
853 }
854 if (match_object && match_offset) {
855 DB_DNODE_EXIT(db);
856 return;
857 }
858 }
859 DB_DNODE_EXIT(db);
860 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
861 (u_longlong_t)db->db.db_object, db->db_level,
862 (u_longlong_t)db->db_blkid);
863 }
864 #endif
865
866 /*
867 * If we can't do 10 iops, something is wrong. Let us go ahead
868 * and hit zfs_dirty_data_max.
869 */
870 static const hrtime_t zfs_delay_max_ns = 100 * MICROSEC; /* 100 milliseconds */
871
872 /*
873 * We delay transactions when we've determined that the backend storage
874 * isn't able to accommodate the rate of incoming writes.
875 *
876 * If there is already a transaction waiting, we delay relative to when
877 * that transaction finishes waiting. This way the calculated min_time
878 * is independent of the number of threads concurrently executing
879 * transactions.
880 *
881 * If we are the only waiter, wait relative to when the transaction
882 * started, rather than the current time. This credits the transaction for
883 * "time already served", e.g. reading indirect blocks.
884 *
885 * The minimum time for a transaction to take is calculated as:
886 * min_time = scale * (dirty - min) / (max - dirty)
887 * min_time is then capped at zfs_delay_max_ns.
888 *
889 * The delay has two degrees of freedom that can be adjusted via tunables.
890 * The percentage of dirty data at which we start to delay is defined by
891 * zfs_delay_min_dirty_percent. This should typically be at or above
892 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
893 * delay after writing at full speed has failed to keep up with the incoming
894 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
895 * speaking, this variable determines the amount of delay at the midpoint of
896 * the curve.
897 *
898 * delay
899 * 10ms +-------------------------------------------------------------*+
900 * | *|
901 * 9ms + *+
902 * | *|
903 * 8ms + *+
904 * | * |
905 * 7ms + * +
906 * | * |
907 * 6ms + * +
908 * | * |
909 * 5ms + * +
910 * | * |
911 * 4ms + * +
912 * | * |
913 * 3ms + * +
914 * | * |
915 * 2ms + (midpoint) * +
916 * | | ** |
917 * 1ms + v *** +
918 * | zfs_delay_scale ----------> ******** |
919 * 0 +-------------------------------------*********----------------+
920 * 0% <- zfs_dirty_data_max -> 100%
921 *
922 * Note that since the delay is added to the outstanding time remaining on the
923 * most recent transaction, the delay is effectively the inverse of IOPS.
924 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
925 * was chosen such that small changes in the amount of accumulated dirty data
926 * in the first 3/4 of the curve yield relatively small differences in the
927 * amount of delay.
928 *
929 * The effects can be easier to understand when the amount of delay is
930 * represented on a log scale:
931 *
932 * delay
933 * 100ms +-------------------------------------------------------------++
934 * + +
935 * | |
936 * + *+
937 * 10ms + *+
938 * + ** +
939 * | (midpoint) ** |
940 * + | ** +
941 * 1ms + v **** +
942 * + zfs_delay_scale ----------> ***** +
943 * | **** |
944 * + **** +
945 * 100us + ** +
946 * + * +
947 * | * |
948 * + * +
949 * 10us + * +
950 * + +
951 * | |
952 * + +
953 * +--------------------------------------------------------------+
954 * 0% <- zfs_dirty_data_max -> 100%
955 *
956 * Note here that only as the amount of dirty data approaches its limit does
957 * the delay start to increase rapidly. The goal of a properly tuned system
958 * should be to keep the amount of dirty data out of that range by first
959 * ensuring that the appropriate limits are set for the I/O scheduler to reach
960 * optimal throughput on the backend storage, and then by changing the value
961 * of zfs_delay_scale to increase the steepness of the curve.
962 */
963 static void
dmu_tx_delay(dmu_tx_t * tx,uint64_t dirty)964 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
965 {
966 dsl_pool_t *dp = tx->tx_pool;
967 uint64_t delay_min_bytes, wrlog;
968 hrtime_t wakeup, tx_time = 0, now;
969
970 /* Calculate minimum transaction time for the dirty data amount. */
971 delay_min_bytes =
972 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
973 if (dirty >= zfs_dirty_data_max) {
974 /*
975 * The caller has already waited until the dirty data is
976 * under the max, but the sync context reservations added
977 * on top of it may push the sum beyond it.
978 */
979 tx_time = zfs_delay_max_ns;
980 } else if (dirty > delay_min_bytes) {
981 tx_time = zfs_delay_scale * (dirty - delay_min_bytes) /
982 (zfs_dirty_data_max - dirty);
983 }
984
985 /* Calculate minimum transaction time for the TX_WRITE log size. */
986 wrlog = aggsum_upper_bound(&dp->dp_wrlog_total);
987 delay_min_bytes =
988 zfs_wrlog_data_max * zfs_delay_min_dirty_percent / 100;
989 if (wrlog >= zfs_wrlog_data_max) {
990 tx_time = zfs_delay_max_ns;
991 } else if (wrlog > delay_min_bytes) {
992 tx_time = MAX(zfs_delay_scale * (wrlog - delay_min_bytes) /
993 (zfs_wrlog_data_max - wrlog), tx_time);
994 }
995
996 if (tx_time == 0)
997 return;
998
999 tx_time = MIN(tx_time, zfs_delay_max_ns);
1000 now = gethrtime();
1001 if (now > tx->tx_start + tx_time)
1002 return;
1003
1004 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1005 uint64_t, tx_time);
1006
1007 mutex_enter(&dp->dp_lock);
1008 wakeup = MAX(tx->tx_start + tx_time, dp->dp_last_wakeup + tx_time);
1009 dp->dp_last_wakeup = wakeup;
1010 mutex_exit(&dp->dp_lock);
1011
1012 zfs_sleep_until(wakeup);
1013 }
1014
1015 /*
1016 * This routine attempts to assign the transaction to a transaction group.
1017 * To do so, we must determine if there is sufficient free space on disk.
1018 *
1019 * If this is a "netfree" transaction (i.e. we called dmu_tx_mark_netfree()
1020 * on it), then it is assumed that there is sufficient free space,
1021 * unless there's insufficient slop space in the pool (see the comment
1022 * above spa_slop_shift in spa_misc.c).
1023 *
1024 * If it is not a "netfree" transaction, then if the data already on disk
1025 * is over the allowed usage (e.g. quota), this will fail with EDQUOT or
1026 * ENOSPC. Otherwise, if the current rough estimate of pending changes,
1027 * plus the rough estimate of this transaction's changes, may exceed the
1028 * allowed usage, then this will fail with ERESTART, which will cause the
1029 * caller to wait for the pending changes to be written to disk (by waiting
1030 * for the next TXG to open), and then check the space usage again.
1031 *
1032 * The rough estimate of pending changes is comprised of the sum of:
1033 *
1034 * - this transaction's holds' txh_space_towrite
1035 *
1036 * - dd_tempreserved[], which is the sum of in-flight transactions'
1037 * holds' txh_space_towrite (i.e. those transactions that have called
1038 * dmu_tx_assign() but not yet called dmu_tx_commit()).
1039 *
1040 * - dd_space_towrite[], which is the amount of dirtied dbufs.
1041 *
1042 * Note that all of these values are inflated by spa_get_worst_case_asize(),
1043 * which means that we may get ERESTART well before we are actually in danger
1044 * of running out of space, but this also mitigates any small inaccuracies
1045 * in the rough estimate (e.g. txh_space_towrite doesn't take into account
1046 * indirect blocks, and dd_space_towrite[] doesn't take into account changes
1047 * to the MOS).
1048 *
1049 * Note that due to this algorithm, it is possible to exceed the allowed
1050 * usage by one transaction. Also, as we approach the allowed usage,
1051 * we will allow a very limited amount of changes into each TXG, thus
1052 * decreasing performance.
1053 */
1054 static int
dmu_tx_try_assign(dmu_tx_t * tx)1055 dmu_tx_try_assign(dmu_tx_t *tx)
1056 {
1057 spa_t *spa = tx->tx_pool->dp_spa;
1058
1059 ASSERT0(tx->tx_txg);
1060
1061 if (tx->tx_err) {
1062 DMU_TX_STAT_BUMP(dmu_tx_error);
1063 return (SET_ERROR(EIO));
1064 }
1065
1066 if (spa_suspended(spa)) {
1067 DMU_TX_STAT_BUMP(dmu_tx_suspended);
1068
1069 /*
1070 * Let dmu_tx_assign() know specifically what happened, so
1071 * it can make the right choice based on the caller flags.
1072 */
1073 return (SET_ERROR(ESHUTDOWN));
1074 }
1075
1076 if (!tx->tx_dirty_delayed &&
1077 dsl_pool_need_wrlog_delay(tx->tx_pool)) {
1078 tx->tx_wait_dirty = B_TRUE;
1079 DMU_TX_STAT_BUMP(dmu_tx_wrlog_delay);
1080 return (SET_ERROR(ERESTART));
1081 }
1082
1083 if (!tx->tx_dirty_delayed &&
1084 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1085 tx->tx_wait_dirty = B_TRUE;
1086 DMU_TX_STAT_BUMP(dmu_tx_dirty_delay);
1087 return (SET_ERROR(ERESTART));
1088 }
1089
1090 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1091 tx->tx_needassign_txh = NULL;
1092
1093 /*
1094 * NB: No error returns are allowed after txg_hold_open, but
1095 * before processing the dnode holds, due to the
1096 * dmu_tx_unassign() logic.
1097 */
1098
1099 uint64_t towrite = 0;
1100 uint64_t tohold = 0;
1101 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1102 txh = list_next(&tx->tx_holds, txh)) {
1103 dnode_t *dn = txh->txh_dnode;
1104 if (dn != NULL) {
1105 /*
1106 * This thread can't hold the dn_struct_rwlock
1107 * while assigning the tx, because this can lead to
1108 * deadlock. Specifically, if this dnode is already
1109 * assigned to an earlier txg, this thread may need
1110 * to wait for that txg to sync (the ERESTART case
1111 * below). The other thread that has assigned this
1112 * dnode to an earlier txg prevents this txg from
1113 * syncing until its tx can complete (calling
1114 * dmu_tx_commit()), but it may need to acquire the
1115 * dn_struct_rwlock to do so (e.g. via
1116 * dmu_buf_hold*()).
1117 *
1118 * Note that this thread can't hold the lock for
1119 * read either, but the rwlock doesn't record
1120 * enough information to make that assertion.
1121 */
1122 ASSERT(!RW_WRITE_HELD(&dn->dn_struct_rwlock));
1123
1124 mutex_enter(&dn->dn_mtx);
1125 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1126 mutex_exit(&dn->dn_mtx);
1127 tx->tx_needassign_txh = txh;
1128 DMU_TX_STAT_BUMP(dmu_tx_group);
1129 return (SET_ERROR(ERESTART));
1130 }
1131 if (dn->dn_assigned_txg == 0)
1132 dn->dn_assigned_txg = tx->tx_txg;
1133 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1134 (void) zfs_refcount_add(&dn->dn_tx_holds, tx);
1135 mutex_exit(&dn->dn_mtx);
1136 }
1137 towrite += zfs_refcount_count(&txh->txh_space_towrite);
1138 tohold += zfs_refcount_count(&txh->txh_memory_tohold);
1139 }
1140
1141 /* needed allocation: worst-case estimate of write space */
1142 uint64_t asize = spa_get_worst_case_asize(tx->tx_pool->dp_spa, towrite);
1143 /* calculate memory footprint estimate */
1144 uint64_t memory = towrite + tohold;
1145
1146 if (tx->tx_dir != NULL && asize != 0) {
1147 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1148 asize, tx->tx_netfree, &tx->tx_tempreserve_cookie, tx);
1149 if (err != 0)
1150 return (err);
1151 }
1152
1153 dsl_pool_sync_reserve(tx->tx_pool, tx->tx_sync_reserve, tx);
1154
1155 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1156
1157 return (0);
1158 }
1159
1160 static void
dmu_tx_unassign(dmu_tx_t * tx)1161 dmu_tx_unassign(dmu_tx_t *tx)
1162 {
1163 if (tx->tx_txg == 0)
1164 return;
1165
1166 txg_rele_to_quiesce(&tx->tx_txgh);
1167
1168 /*
1169 * Walk the transaction's hold list, removing the hold on the
1170 * associated dnode, and notifying waiters if the refcount drops to 0.
1171 */
1172 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds);
1173 txh && txh != tx->tx_needassign_txh;
1174 txh = list_next(&tx->tx_holds, txh)) {
1175 dnode_t *dn = txh->txh_dnode;
1176
1177 if (dn == NULL)
1178 continue;
1179 mutex_enter(&dn->dn_mtx);
1180 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1181
1182 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1183 dn->dn_assigned_txg = 0;
1184 cv_broadcast(&dn->dn_notxholds);
1185 }
1186 mutex_exit(&dn->dn_mtx);
1187 }
1188
1189 txg_rele_to_sync(&tx->tx_txgh);
1190
1191 tx->tx_lasttried_txg = tx->tx_txg;
1192 tx->tx_txg = 0;
1193 }
1194
1195 /*
1196 * Assign tx to a transaction group; `flags` is a bitmask:
1197 *
1198 * If DMU_TX_WAIT is set and the currently open txg is full, this function
1199 * will wait until there's a new txg. This should be used when no locks
1200 * are being held. With this bit set, this function will only fail if
1201 * we're truly out of space (ENOSPC), over quota (EDQUOT), or required
1202 * data for the transaction could not be read from disk (EIO).
1203 *
1204 * If DMU_TX_WAIT is *not* set and we can't assign into the currently open
1205 * txg without blocking, this function will return immediately with
1206 * ERESTART. This should be used whenever locks are being held. On an
1207 * ERESTART error, the caller should drop all locks, call dmu_tx_wait(),
1208 * and try again.
1209 *
1210 * If DMU_TX_NOTHROTTLE is set, this indicates that this tx should not be
1211 * delayed due on the ZFS Write Throttle (see comments in dsl_pool.c for
1212 * details on the throttle). This is used by the VFS operations, after
1213 * they have already called dmu_tx_wait() (though most likely on a
1214 * different tx).
1215 *
1216 * If DMU_TX_SUSPEND is set, this indicates that this tx should ignore
1217 * the pool being or becoming suspending while it is in progress. This will
1218 * cause dmu_tx_assign() (and dmu_tx_wait()) to block until the pool resumes.
1219 * If this flag is not set and the pool suspends, the return will be either
1220 * ERESTART or EIO, depending on the value of the pool's failmode= property.
1221 *
1222 * It is guaranteed that subsequent successful calls to dmu_tx_assign()
1223 * will assign the tx to monotonically increasing txgs. Of course this is
1224 * not strong monotonicity, because the same txg can be returned multiple
1225 * times in a row. This guarantee holds both for subsequent calls from
1226 * one thread and for multiple threads. For example, it is impossible to
1227 * observe the following sequence of events:
1228 *
1229 * Thread 1 Thread 2
1230 *
1231 * dmu_tx_assign(T1, ...)
1232 * 1 <- dmu_tx_get_txg(T1)
1233 * dmu_tx_assign(T2, ...)
1234 * 2 <- dmu_tx_get_txg(T2)
1235 * dmu_tx_assign(T3, ...)
1236 * 1 <- dmu_tx_get_txg(T3)
1237 */
1238 int
dmu_tx_assign(dmu_tx_t * tx,dmu_tx_flag_t flags)1239 dmu_tx_assign(dmu_tx_t *tx, dmu_tx_flag_t flags)
1240 {
1241 int err;
1242
1243 ASSERT0(tx->tx_txg);
1244 ASSERT0(flags & ~(DMU_TX_WAIT | DMU_TX_NOTHROTTLE | DMU_TX_SUSPEND));
1245 IMPLY(flags & DMU_TX_SUSPEND, flags & DMU_TX_WAIT);
1246 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1247
1248 /* If we might wait, we must not hold the config lock. */
1249 IMPLY((flags & DMU_TX_WAIT), !dsl_pool_config_held(tx->tx_pool));
1250
1251 if ((flags & DMU_TX_NOTHROTTLE))
1252 tx->tx_dirty_delayed = B_TRUE;
1253
1254 if (!(flags & DMU_TX_SUSPEND))
1255 tx->tx_break_on_suspend = B_TRUE;
1256
1257 while ((err = dmu_tx_try_assign(tx)) != 0) {
1258 dmu_tx_unassign(tx);
1259
1260 boolean_t suspended = (err == ESHUTDOWN);
1261 if (suspended) {
1262 /*
1263 * Pool suspended. We need to decide whether to block
1264 * and retry, or return error, depending on the
1265 * caller's flags and the pool config.
1266 */
1267 if (flags & DMU_TX_SUSPEND)
1268 /*
1269 * The caller expressly does not care about
1270 * suspend, so treat it as a normal retry.
1271 */
1272 err = SET_ERROR(ERESTART);
1273 else if ((flags & DMU_TX_WAIT) &&
1274 spa_get_failmode(tx->tx_pool->dp_spa) ==
1275 ZIO_FAILURE_MODE_CONTINUE)
1276 /*
1277 * Caller wants to wait, but pool config is
1278 * overriding that, so return EIO to be
1279 * propagated back to userspace.
1280 */
1281 err = SET_ERROR(EIO);
1282 else
1283 /* Anything else, we should just block. */
1284 err = SET_ERROR(ERESTART);
1285 }
1286
1287 /*
1288 * Return unless we decided to retry, or the caller does not
1289 * want to block.
1290 */
1291 if (err != ERESTART || !(flags & DMU_TX_WAIT)) {
1292 ASSERT(err == EDQUOT || err == ENOSPC ||
1293 err == ERESTART || err == EIO);
1294 return (err);
1295 }
1296
1297 /*
1298 * Wait until there's room in this txg, or until it's been
1299 * synced out and a new one is available.
1300 *
1301 * If we're here because the pool suspended above, then we
1302 * unset tx_break_on_suspend to make sure that if dmu_tx_wait()
1303 * has to fall back to a txg_wait_synced_flags(), it doesn't
1304 * immediately return because the pool is suspended. That would
1305 * then immediately return here, and we'd end up in a busy loop
1306 * until the pool resumes.
1307 *
1308 * On the other hand, if the pool hasn't suspended yet, then it
1309 * should be allowed to break a txg wait if the pool does
1310 * suspend, so we can loop and reassess it in
1311 * dmu_tx_try_assign().
1312 */
1313 if (suspended)
1314 tx->tx_break_on_suspend = B_FALSE;
1315
1316 dmu_tx_wait(tx);
1317
1318 /*
1319 * Reset tx_break_on_suspend for DMU_TX_SUSPEND. We do this
1320 * here so that it's available if we return for some other
1321 * reason, and then the caller calls dmu_tx_wait().
1322 */
1323 if (!(flags & DMU_TX_SUSPEND))
1324 tx->tx_break_on_suspend = B_TRUE;
1325 }
1326
1327 txg_rele_to_quiesce(&tx->tx_txgh);
1328
1329 return (0);
1330 }
1331
1332 void
dmu_tx_wait(dmu_tx_t * tx)1333 dmu_tx_wait(dmu_tx_t *tx)
1334 {
1335 spa_t *spa = tx->tx_pool->dp_spa;
1336 dsl_pool_t *dp = tx->tx_pool;
1337 hrtime_t before;
1338
1339 ASSERT0(tx->tx_txg);
1340 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1341
1342 /*
1343 * Break on suspend according to whether or not DMU_TX_SUSPEND was
1344 * supplied to the previous dmu_tx_assign() call. For clients, this
1345 * ensures that after dmu_tx_assign() fails, the followup dmu_tx_wait()
1346 * gets the same behaviour wrt suspend. See also the comments in
1347 * dmu_tx_assign().
1348 */
1349 txg_wait_flag_t flags =
1350 (tx->tx_break_on_suspend ? TXG_WAIT_SUSPEND : TXG_WAIT_NONE);
1351
1352 before = gethrtime();
1353
1354 if (tx->tx_wait_dirty) {
1355 uint64_t dirty;
1356
1357 /*
1358 * dmu_tx_try_assign() has determined that we need to wait
1359 * because we've consumed much or all of the dirty buffer
1360 * space.
1361 */
1362 mutex_enter(&dp->dp_lock);
1363 if (dp->dp_dirty_total >= zfs_dirty_data_max)
1364 DMU_TX_STAT_BUMP(dmu_tx_dirty_over_max);
1365 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1366 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1367 dirty = dp->dp_dirty_total + dp->dp_sync_reserve_total;
1368 mutex_exit(&dp->dp_lock);
1369
1370 dmu_tx_delay(tx, dirty);
1371
1372 tx->tx_wait_dirty = B_FALSE;
1373
1374 /*
1375 * Note: setting tx_dirty_delayed only has effect if the
1376 * caller used DMU_TX_WAIT. Otherwise they are going to
1377 * destroy this tx and try again. The common case,
1378 * zfs_write(), uses DMU_TX_WAIT.
1379 */
1380 tx->tx_dirty_delayed = B_TRUE;
1381 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1382 /*
1383 * If the pool is suspended we need to wait until it
1384 * is resumed. Note that it's possible that the pool
1385 * has become active after this thread has tried to
1386 * obtain a tx. If that's the case then tx_lasttried_txg
1387 * would not have been set.
1388 */
1389 txg_wait_synced_flags(dp, spa_last_synced_txg(spa) + 1, flags);
1390 } else if (tx->tx_needassign_txh) {
1391 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1392
1393 mutex_enter(&dn->dn_mtx);
1394 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1395 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1396 mutex_exit(&dn->dn_mtx);
1397 tx->tx_needassign_txh = NULL;
1398 } else {
1399 /*
1400 * If we have a lot of dirty data just wait until we sync
1401 * out a TXG at which point we'll hopefully have synced
1402 * a portion of the changes.
1403 */
1404 txg_wait_synced_flags(dp, spa_last_synced_txg(spa) + 1, flags);
1405 }
1406
1407 spa_tx_assign_add_nsecs(spa, gethrtime() - before);
1408 }
1409
1410 static void
dmu_tx_destroy(dmu_tx_t * tx)1411 dmu_tx_destroy(dmu_tx_t *tx)
1412 {
1413 dmu_tx_hold_t *txh;
1414
1415 while ((txh = list_head(&tx->tx_holds)) != NULL) {
1416 dnode_t *dn = txh->txh_dnode;
1417
1418 list_remove(&tx->tx_holds, txh);
1419 zfs_refcount_destroy_many(&txh->txh_space_towrite,
1420 zfs_refcount_count(&txh->txh_space_towrite));
1421 zfs_refcount_destroy_many(&txh->txh_memory_tohold,
1422 zfs_refcount_count(&txh->txh_memory_tohold));
1423 kmem_free(txh, sizeof (dmu_tx_hold_t));
1424 if (dn != NULL)
1425 dnode_rele(dn, tx);
1426 }
1427
1428 list_destroy(&tx->tx_callbacks);
1429 list_destroy(&tx->tx_holds);
1430 kmem_free(tx, sizeof (dmu_tx_t));
1431 }
1432
1433 void
dmu_tx_commit(dmu_tx_t * tx)1434 dmu_tx_commit(dmu_tx_t *tx)
1435 {
1436 /* This function should only be used on assigned transactions. */
1437 ASSERT(tx->tx_txg != 0);
1438
1439 /*
1440 * Go through the transaction's hold list and remove holds on
1441 * associated dnodes, notifying waiters if no holds remain.
1442 */
1443 for (dmu_tx_hold_t *txh = list_head(&tx->tx_holds); txh != NULL;
1444 txh = list_next(&tx->tx_holds, txh)) {
1445 dnode_t *dn = txh->txh_dnode;
1446
1447 if (dn == NULL)
1448 continue;
1449
1450 mutex_enter(&dn->dn_mtx);
1451 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1452
1453 if (zfs_refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1454 dn->dn_assigned_txg = 0;
1455 cv_broadcast(&dn->dn_notxholds);
1456 }
1457 mutex_exit(&dn->dn_mtx);
1458 }
1459
1460 if (tx->tx_tempreserve_cookie)
1461 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1462
1463 if (!list_is_empty(&tx->tx_callbacks))
1464 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1465
1466 if (tx->tx_anyobj == FALSE)
1467 txg_rele_to_sync(&tx->tx_txgh);
1468
1469 dmu_tx_destroy(tx);
1470 }
1471
1472 void
dmu_tx_abort(dmu_tx_t * tx)1473 dmu_tx_abort(dmu_tx_t *tx)
1474 {
1475 /* This function should not be used on assigned transactions. */
1476 ASSERT0(tx->tx_txg);
1477
1478 /* Should not be needed, but better be safe than sorry. */
1479 if (tx->tx_tempreserve_cookie)
1480 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1481
1482 /*
1483 * Call any registered callbacks with an error code.
1484 */
1485 if (!list_is_empty(&tx->tx_callbacks))
1486 dmu_tx_do_callbacks(&tx->tx_callbacks, SET_ERROR(ECANCELED));
1487
1488 /* Should not be needed, but better be safe than sorry. */
1489 dmu_tx_unassign(tx);
1490
1491 dmu_tx_destroy(tx);
1492 }
1493
1494 uint64_t
dmu_tx_get_txg(dmu_tx_t * tx)1495 dmu_tx_get_txg(dmu_tx_t *tx)
1496 {
1497 ASSERT(tx->tx_txg != 0);
1498 return (tx->tx_txg);
1499 }
1500
1501 dsl_pool_t *
dmu_tx_pool(dmu_tx_t * tx)1502 dmu_tx_pool(dmu_tx_t *tx)
1503 {
1504 ASSERT(tx->tx_pool != NULL);
1505 return (tx->tx_pool);
1506 }
1507
1508 /*
1509 * Register a callback to be executed at the end of a TXG.
1510 *
1511 * Note: This currently exists for outside consumers, specifically the ZFS OSD
1512 * for Lustre. Please do not remove before checking that project. For examples
1513 * on how to use this see `ztest_commit_callback`.
1514 */
1515 void
dmu_tx_callback_register(dmu_tx_t * tx,dmu_tx_callback_func_t * func,void * data)1516 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1517 {
1518 dmu_tx_callback_t *dcb;
1519
1520 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1521
1522 dcb->dcb_func = func;
1523 dcb->dcb_data = data;
1524
1525 list_insert_tail(&tx->tx_callbacks, dcb);
1526 }
1527
1528 /*
1529 * Call all the commit callbacks on a list, with a given error code.
1530 */
1531 void
dmu_tx_do_callbacks(list_t * cb_list,int error)1532 dmu_tx_do_callbacks(list_t *cb_list, int error)
1533 {
1534 dmu_tx_callback_t *dcb;
1535
1536 while ((dcb = list_remove_tail(cb_list)) != NULL) {
1537 dcb->dcb_func(dcb->dcb_data, error);
1538 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1539 }
1540 }
1541
1542 /*
1543 * Interface to hold a bunch of attributes.
1544 * used for creating new files.
1545 * attrsize is the total size of all attributes
1546 * to be added during object creation
1547 *
1548 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1549 */
1550
1551 /*
1552 * hold necessary attribute name for attribute registration.
1553 * should be a very rare case where this is needed. If it does
1554 * happen it would only happen on the first write to the file system.
1555 */
1556 static void
dmu_tx_sa_registration_hold(sa_os_t * sa,dmu_tx_t * tx)1557 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1558 {
1559 if (!sa->sa_need_attr_registration)
1560 return;
1561
1562 for (int i = 0; i != sa->sa_num_attrs; i++) {
1563 if (!sa->sa_attr_table[i].sa_registered) {
1564 if (sa->sa_reg_attr_obj)
1565 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1566 B_TRUE, sa->sa_attr_table[i].sa_name);
1567 else
1568 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1569 B_TRUE, sa->sa_attr_table[i].sa_name);
1570 }
1571 }
1572 }
1573
1574 void
dmu_tx_hold_spill(dmu_tx_t * tx,uint64_t object)1575 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1576 {
1577 dmu_tx_hold_t *txh;
1578
1579 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1580 THT_SPILL, 0, 0);
1581 if (txh != NULL)
1582 (void) zfs_refcount_add_many(&txh->txh_space_towrite,
1583 SPA_OLD_MAXBLOCKSIZE, FTAG);
1584 }
1585
1586 void
dmu_tx_hold_sa_create(dmu_tx_t * tx,int attrsize)1587 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1588 {
1589 sa_os_t *sa = tx->tx_objset->os_sa;
1590
1591 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1592
1593 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1594 return;
1595
1596 if (tx->tx_objset->os_sa->sa_layout_attr_obj) {
1597 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1598 } else {
1599 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1600 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1601 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1602 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1603 }
1604
1605 dmu_tx_sa_registration_hold(sa, tx);
1606
1607 if (attrsize <= DN_OLD_MAX_BONUSLEN && !sa->sa_force_spill)
1608 return;
1609
1610 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1611 THT_SPILL, 0, 0);
1612 }
1613
1614 /*
1615 * Hold SA attribute
1616 *
1617 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1618 *
1619 * variable_size is the total size of all variable sized attributes
1620 * passed to this function. It is not the total size of all
1621 * variable size attributes that *may* exist on this object.
1622 */
1623 void
dmu_tx_hold_sa(dmu_tx_t * tx,sa_handle_t * hdl,boolean_t may_grow)1624 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1625 {
1626 uint64_t object;
1627 sa_os_t *sa = tx->tx_objset->os_sa;
1628
1629 ASSERT(hdl != NULL);
1630
1631 object = sa_handle_object(hdl);
1632
1633 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1634 DB_DNODE_ENTER(db);
1635 dmu_tx_hold_bonus_by_dnode(tx, DB_DNODE(db));
1636 DB_DNODE_EXIT(db);
1637
1638 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1639 return;
1640
1641 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1642 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1643 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1644 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1645 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1646 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1647 }
1648
1649 dmu_tx_sa_registration_hold(sa, tx);
1650
1651 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1652 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1653
1654 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1655 ASSERT0(tx->tx_txg);
1656 dmu_tx_hold_spill(tx, object);
1657 } else {
1658 DB_DNODE_ENTER(db);
1659 if (DB_DNODE(db)->dn_have_spill) {
1660 ASSERT0(tx->tx_txg);
1661 dmu_tx_hold_spill(tx, object);
1662 }
1663 DB_DNODE_EXIT(db);
1664 }
1665 }
1666
1667 void
dmu_tx_init(void)1668 dmu_tx_init(void)
1669 {
1670 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1671 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1672 KSTAT_FLAG_VIRTUAL);
1673
1674 if (dmu_tx_ksp != NULL) {
1675 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1676 kstat_install(dmu_tx_ksp);
1677 }
1678 }
1679
1680 void
dmu_tx_fini(void)1681 dmu_tx_fini(void)
1682 {
1683 if (dmu_tx_ksp != NULL) {
1684 kstat_delete(dmu_tx_ksp);
1685 dmu_tx_ksp = NULL;
1686 }
1687 }
1688
1689 #if defined(_KERNEL)
1690 EXPORT_SYMBOL(dmu_tx_create);
1691 EXPORT_SYMBOL(dmu_tx_hold_write);
1692 EXPORT_SYMBOL(dmu_tx_hold_write_by_dnode);
1693 EXPORT_SYMBOL(dmu_tx_hold_append);
1694 EXPORT_SYMBOL(dmu_tx_hold_append_by_dnode);
1695 EXPORT_SYMBOL(dmu_tx_hold_free);
1696 EXPORT_SYMBOL(dmu_tx_hold_free_by_dnode);
1697 EXPORT_SYMBOL(dmu_tx_hold_zap);
1698 EXPORT_SYMBOL(dmu_tx_hold_zap_by_dnode);
1699 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1700 EXPORT_SYMBOL(dmu_tx_hold_bonus_by_dnode);
1701 EXPORT_SYMBOL(dmu_tx_abort);
1702 EXPORT_SYMBOL(dmu_tx_assign);
1703 EXPORT_SYMBOL(dmu_tx_wait);
1704 EXPORT_SYMBOL(dmu_tx_commit);
1705 EXPORT_SYMBOL(dmu_tx_mark_netfree);
1706 EXPORT_SYMBOL(dmu_tx_get_txg);
1707 EXPORT_SYMBOL(dmu_tx_callback_register);
1708 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1709 EXPORT_SYMBOL(dmu_tx_hold_spill);
1710 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1711 EXPORT_SYMBOL(dmu_tx_hold_sa);
1712 #endif
1713