1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2013 by Delphix. All rights reserved.
25 */
26
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
33 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
34 #include <sys/dsl_pool.h>
35 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/spa.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
41
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 uint64_t arg1, uint64_t arg2);
44
45
46 dmu_tx_t *
47 dmu_tx_create_dd(dsl_dir_t *dd)
48 {
49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 tx->tx_dir = dd;
51 if (dd != NULL)
52 tx->tx_pool = dd->dd_pool;
53 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
54 offsetof(dmu_tx_hold_t, txh_node));
55 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
56 offsetof(dmu_tx_callback_t, dcb_node));
57 tx->tx_start = gethrtime();
58 #ifdef ZFS_DEBUG
59 refcount_create(&tx->tx_space_written);
60 refcount_create(&tx->tx_space_freed);
61 #endif
62 return (tx);
63 }
64
65 dmu_tx_t *
66 dmu_tx_create(objset_t *os)
67 {
68 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 tx->tx_objset = os;
70 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
71 return (tx);
72 }
73
74 dmu_tx_t *
75 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 {
77 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78
79 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
80 tx->tx_pool = dp;
81 tx->tx_txg = txg;
82 tx->tx_anyobj = TRUE;
83
84 return (tx);
85 }
86
87 int
88 dmu_tx_is_syncing(dmu_tx_t *tx)
89 {
90 return (tx->tx_anyobj);
91 }
92
93 int
94 dmu_tx_private_ok(dmu_tx_t *tx)
95 {
96 return (tx->tx_anyobj);
97 }
98
99 static dmu_tx_hold_t *
100 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
101 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
102 {
103 dmu_tx_hold_t *txh;
104 dnode_t *dn = NULL;
105 int err;
106
107 if (object != DMU_NEW_OBJECT) {
108 err = dnode_hold(os, object, tx, &dn);
109 if (err) {
110 tx->tx_err = err;
111 return (NULL);
112 }
113
114 if (err == 0 && tx->tx_txg != 0) {
115 mutex_enter(&dn->dn_mtx);
116 /*
117 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
118 * problem, but there's no way for it to happen (for
119 * now, at least).
120 */
121 ASSERT(dn->dn_assigned_txg == 0);
122 dn->dn_assigned_txg = tx->tx_txg;
123 (void) refcount_add(&dn->dn_tx_holds, tx);
124 mutex_exit(&dn->dn_mtx);
125 }
126 }
127
128 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 txh->txh_tx = tx;
130 txh->txh_dnode = dn;
131 #ifdef ZFS_DEBUG
132 txh->txh_type = type;
133 txh->txh_arg1 = arg1;
134 txh->txh_arg2 = arg2;
135 #endif
136 list_insert_tail(&tx->tx_holds, txh);
137
138 return (txh);
139 }
140
141 void
142 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
143 {
144 /*
145 * If we're syncing, they can manipulate any object anyhow, and
146 * the hold on the dnode_t can cause problems.
147 */
148 if (!dmu_tx_is_syncing(tx)) {
149 (void) dmu_tx_hold_object_impl(tx, os,
150 object, THT_NEWOBJECT, 0, 0);
151 }
152 }
153
154 static int
155 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
156 {
157 int err;
158 dmu_buf_impl_t *db;
159
160 rw_enter(&dn->dn_struct_rwlock, RW_READER);
161 db = dbuf_hold_level(dn, level, blkid, FTAG);
162 rw_exit(&dn->dn_struct_rwlock);
163 if (db == NULL)
164 return (SET_ERROR(EIO));
165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 dbuf_rele(db, FTAG);
167 return (err);
168 }
169
170 static void
171 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
172 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 {
174 objset_t *os = dn->dn_objset;
175 dsl_dataset_t *ds = os->os_dsl_dataset;
176 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
177 dmu_buf_impl_t *parent = NULL;
178 blkptr_t *bp = NULL;
179 uint64_t space;
180
181 if (level >= dn->dn_nlevels || history[level] == blkid)
182 return;
183
184 history[level] = blkid;
185
186 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187
188 if (db == NULL || db == dn->dn_dbuf) {
189 ASSERT(level != 0);
190 db = NULL;
191 } else {
192 ASSERT(DB_DNODE(db) == dn);
193 ASSERT(db->db_level == level);
194 ASSERT(db->db.db_size == space);
195 ASSERT(db->db_blkid == blkid);
196 bp = db->db_blkptr;
197 parent = db->db_parent;
198 }
199
200 freeable = (bp && (freeable ||
201 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
202
203 if (freeable)
204 txh->txh_space_tooverwrite += space;
205 else
206 txh->txh_space_towrite += space;
207 if (bp)
208 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209
210 dmu_tx_count_twig(txh, dn, parent, level + 1,
211 blkid >> epbs, freeable, history);
212 }
213
214 /* ARGSUSED */
215 static void
216 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 {
218 dnode_t *dn = txh->txh_dnode;
219 uint64_t start, end, i;
220 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
221 int err = 0;
222
223 if (len == 0)
224 return;
225
226 min_bs = SPA_MINBLOCKSHIFT;
227 max_bs = SPA_MAXBLOCKSHIFT;
228 min_ibs = DN_MIN_INDBLKSHIFT;
229 max_ibs = DN_MAX_INDBLKSHIFT;
230
231 if (dn) {
232 uint64_t history[DN_MAX_LEVELS];
233 int nlvls = dn->dn_nlevels;
234 int delta;
235
236 /*
237 * For i/o error checking, read the first and last level-0
238 * blocks (if they are not aligned), and all the level-1 blocks.
239 */
240 if (dn->dn_maxblkid == 0) {
241 delta = dn->dn_datablksz;
242 start = (off < dn->dn_datablksz) ? 0 : 1;
243 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
244 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
245 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
246 if (err)
247 goto out;
248 delta -= off;
249 }
250 } else {
251 zio_t *zio = zio_root(dn->dn_objset->os_spa,
252 NULL, NULL, ZIO_FLAG_CANFAIL);
253
254 /* first level-0 block */
255 start = off >> dn->dn_datablkshift;
256 if (P2PHASE(off, dn->dn_datablksz) ||
257 len < dn->dn_datablksz) {
258 err = dmu_tx_check_ioerr(zio, dn, 0, start);
259 if (err)
260 goto out;
261 }
262
263 /* last level-0 block */
264 end = (off+len-1) >> dn->dn_datablkshift;
265 if (end != start && end <= dn->dn_maxblkid &&
266 P2PHASE(off+len, dn->dn_datablksz)) {
267 err = dmu_tx_check_ioerr(zio, dn, 0, end);
268 if (err)
269 goto out;
270 }
271
272 /* level-1 blocks */
273 if (nlvls > 1) {
274 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
275 for (i = (start>>shft)+1; i < end>>shft; i++) {
276 err = dmu_tx_check_ioerr(zio, dn, 1, i);
277 if (err)
278 goto out;
279 }
280 }
281
282 err = zio_wait(zio);
283 if (err)
284 goto out;
285 delta = P2NPHASE(off, dn->dn_datablksz);
286 }
287
288 min_ibs = max_ibs = dn->dn_indblkshift;
289 if (dn->dn_maxblkid > 0) {
290 /*
291 * The blocksize can't change,
292 * so we can make a more precise estimate.
293 */
294 ASSERT(dn->dn_datablkshift != 0);
295 min_bs = max_bs = dn->dn_datablkshift;
296 }
297
298 /*
299 * If this write is not off the end of the file
300 * we need to account for overwrites/unref.
301 */
302 if (start <= dn->dn_maxblkid) {
303 for (int l = 0; l < DN_MAX_LEVELS; l++)
304 history[l] = -1ULL;
305 }
306 while (start <= dn->dn_maxblkid) {
307 dmu_buf_impl_t *db;
308
309 rw_enter(&dn->dn_struct_rwlock, RW_READER);
310 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
311 rw_exit(&dn->dn_struct_rwlock);
312
313 if (err) {
314 txh->txh_tx->tx_err = err;
315 return;
316 }
317
318 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
319 history);
320 dbuf_rele(db, FTAG);
321 if (++start > end) {
322 /*
323 * Account for new indirects appearing
324 * before this IO gets assigned into a txg.
325 */
326 bits = 64 - min_bs;
327 epbs = min_ibs - SPA_BLKPTRSHIFT;
328 for (bits -= epbs * (nlvls - 1);
329 bits >= 0; bits -= epbs)
330 txh->txh_fudge += 1ULL << max_ibs;
331 goto out;
332 }
333 off += delta;
334 if (len >= delta)
335 len -= delta;
336 delta = dn->dn_datablksz;
337 }
338 }
339
340 /*
341 * 'end' is the last thing we will access, not one past.
342 * This way we won't overflow when accessing the last byte.
343 */
344 start = P2ALIGN(off, 1ULL << max_bs);
345 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
346 txh->txh_space_towrite += end - start + 1;
347
348 start >>= min_bs;
349 end >>= min_bs;
350
351 epbs = min_ibs - SPA_BLKPTRSHIFT;
352
353 /*
354 * The object contains at most 2^(64 - min_bs) blocks,
355 * and each indirect level maps 2^epbs.
356 */
357 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
358 start >>= epbs;
359 end >>= epbs;
360 ASSERT3U(end, >=, start);
361 txh->txh_space_towrite += (end - start + 1) << max_ibs;
362 if (start != 0) {
363 /*
364 * We also need a new blkid=0 indirect block
365 * to reference any existing file data.
366 */
367 txh->txh_space_towrite += 1ULL << max_ibs;
368 }
369 }
370
371 out:
372 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 2 * DMU_MAX_ACCESS)
374 err = SET_ERROR(EFBIG);
375
376 if (err)
377 txh->txh_tx->tx_err = err;
378 }
379
380 static void
381 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
382 {
383 dnode_t *dn = txh->txh_dnode;
384 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
385 uint64_t space = mdn->dn_datablksz +
386 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
387
388 if (dn && dn->dn_dbuf->db_blkptr &&
389 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
390 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
391 txh->txh_space_tooverwrite += space;
392 txh->txh_space_tounref += space;
393 } else {
394 txh->txh_space_towrite += space;
395 if (dn && dn->dn_dbuf->db_blkptr)
396 txh->txh_space_tounref += space;
397 }
398 }
399
400 void
401 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
402 {
403 dmu_tx_hold_t *txh;
404
405 ASSERT(tx->tx_txg == 0);
406 ASSERT(len < DMU_MAX_ACCESS);
407 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
408
409 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
410 object, THT_WRITE, off, len);
411 if (txh == NULL)
412 return;
413
414 dmu_tx_count_write(txh, off, len);
415 dmu_tx_count_dnode(txh);
416 }
417
418 static void
419 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
420 {
421 uint64_t blkid, nblks, lastblk;
422 uint64_t space = 0, unref = 0, skipped = 0;
423 dnode_t *dn = txh->txh_dnode;
424 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
425 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
426 int epbs;
427 uint64_t l0span = 0, nl1blks = 0;
428
429 if (dn->dn_nlevels == 0)
430 return;
431
432 /*
433 * The struct_rwlock protects us against dn_nlevels
434 * changing, in case (against all odds) we manage to dirty &
435 * sync out the changes after we check for being dirty.
436 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 */
438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
440 if (dn->dn_maxblkid == 0) {
441 if (off == 0 && len >= dn->dn_datablksz) {
442 blkid = 0;
443 nblks = 1;
444 } else {
445 rw_exit(&dn->dn_struct_rwlock);
446 return;
447 }
448 } else {
449 blkid = off >> dn->dn_datablkshift;
450 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
451
452 if (blkid > dn->dn_maxblkid) {
453 rw_exit(&dn->dn_struct_rwlock);
454 return;
455 }
456 if (blkid + nblks > dn->dn_maxblkid)
457 nblks = dn->dn_maxblkid - blkid + 1;
458
459 }
460 l0span = nblks; /* save for later use to calc level > 1 overhead */
461 if (dn->dn_nlevels == 1) {
462 int i;
463 for (i = 0; i < nblks; i++) {
464 blkptr_t *bp = dn->dn_phys->dn_blkptr;
465 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
466 bp += blkid + i;
467 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
468 dprintf_bp(bp, "can free old%s", "");
469 space += bp_get_dsize(spa, bp);
470 }
471 unref += BP_GET_ASIZE(bp);
472 }
473 nl1blks = 1;
474 nblks = 0;
475 }
476
477 lastblk = blkid + nblks - 1;
478 while (nblks) {
479 dmu_buf_impl_t *dbuf;
480 uint64_t ibyte, new_blkid;
481 int epb = 1 << epbs;
482 int err, i, blkoff, tochk;
483 blkptr_t *bp;
484
485 ibyte = blkid << dn->dn_datablkshift;
486 err = dnode_next_offset(dn,
487 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
488 new_blkid = ibyte >> dn->dn_datablkshift;
489 if (err == ESRCH) {
490 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
491 break;
492 }
493 if (err) {
494 txh->txh_tx->tx_err = err;
495 break;
496 }
497 if (new_blkid > lastblk) {
498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
499 break;
500 }
501
502 if (new_blkid > blkid) {
503 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
504 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
505 nblks -= new_blkid - blkid;
506 blkid = new_blkid;
507 }
508 blkoff = P2PHASE(blkid, epb);
509 tochk = MIN(epb - blkoff, nblks);
510
511 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
512 if (err) {
513 txh->txh_tx->tx_err = err;
514 break;
515 }
516
517 txh->txh_memory_tohold += dbuf->db.db_size;
518
519 /*
520 * We don't check memory_tohold against DMU_MAX_ACCESS because
521 * memory_tohold is an over-estimation (especially the >L1
522 * indirect blocks), so it could fail. Callers should have
523 * already verified that they will not be holding too much
524 * memory.
525 */
526
527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
528 if (err != 0) {
529 txh->txh_tx->tx_err = err;
530 dbuf_rele(dbuf, FTAG);
531 break;
532 }
533
534 bp = dbuf->db.db_data;
535 bp += blkoff;
536
537 for (i = 0; i < tochk; i++) {
538 if (dsl_dataset_block_freeable(ds, &bp[i],
539 bp[i].blk_birth)) {
540 dprintf_bp(&bp[i], "can free old%s", "");
541 space += bp_get_dsize(spa, &bp[i]);
542 }
543 unref += BP_GET_ASIZE(bp);
544 }
545 dbuf_rele(dbuf, FTAG);
546
547 ++nl1blks;
548 blkid += tochk;
549 nblks -= tochk;
550 }
551 rw_exit(&dn->dn_struct_rwlock);
552
553 /*
554 * Add in memory requirements of higher-level indirects.
555 * This assumes a worst-possible scenario for dn_nlevels and a
556 * worst-possible distribution of l1-blocks over the region to free.
557 */
558 {
559 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
560 int level = 2;
561 /*
562 * Here we don't use DN_MAX_LEVEL, but calculate it with the
563 * given datablkshift and indblkshift. This makes the
564 * difference between 19 and 8 on large files.
565 */
566 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
567 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
568
569 while (level++ < maxlevel) {
570 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
571 << dn->dn_indblkshift;
572 blkcnt = 1 + (blkcnt >> epbs);
573 }
574 }
575
576 /* account for new level 1 indirect blocks that might show up */
577 if (skipped > 0) {
578 txh->txh_fudge += skipped << dn->dn_indblkshift;
579 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
580 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
581 }
582 txh->txh_space_tofree += space;
583 txh->txh_space_tounref += unref;
584 }
585
586 void
587 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
588 {
589 dmu_tx_hold_t *txh;
590 dnode_t *dn;
591 int err;
592 zio_t *zio;
593
594 ASSERT(tx->tx_txg == 0);
595
596 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
597 object, THT_FREE, off, len);
598 if (txh == NULL)
599 return;
600 dn = txh->txh_dnode;
601 dmu_tx_count_dnode(txh);
602
603 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
604 return;
605 if (len == DMU_OBJECT_END)
606 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
607
608 /*
609 * For i/o error checking, we read the first and last level-0
610 * blocks if they are not aligned, and all the level-1 blocks.
611 *
612 * Note: dbuf_free_range() assumes that we have not instantiated
613 * any level-0 dbufs that will be completely freed. Therefore we must
614 * exercise care to not read or count the first and last blocks
615 * if they are blocksize-aligned.
616 */
617 if (dn->dn_datablkshift == 0) {
618 if (off != 0 || len < dn->dn_datablksz)
619 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
620 } else {
621 /* first block will be modified if it is not aligned */
622 if (!IS_P2ALIGNED(off, 1 << dn->dn_datablkshift))
623 dmu_tx_count_write(txh, off, 1);
624 /* last block will be modified if it is not aligned */
625 if (!IS_P2ALIGNED(off + len, 1 << dn->dn_datablkshift))
626 dmu_tx_count_write(txh, off+len, 1);
627 }
628
629 /*
630 * Check level-1 blocks.
631 */
632 if (dn->dn_nlevels > 1) {
633 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
634 SPA_BLKPTRSHIFT;
635 uint64_t start = off >> shift;
636 uint64_t end = (off + len) >> shift;
637
638 ASSERT(dn->dn_datablkshift != 0);
639 ASSERT(dn->dn_indblkshift != 0);
640
641 zio = zio_root(tx->tx_pool->dp_spa,
642 NULL, NULL, ZIO_FLAG_CANFAIL);
643 for (uint64_t i = start; i <= end; i++) {
644 uint64_t ibyte = i << shift;
645 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
646 i = ibyte >> shift;
647 if (err == ESRCH)
648 break;
649 if (err) {
650 tx->tx_err = err;
651 return;
652 }
653
654 err = dmu_tx_check_ioerr(zio, dn, 1, i);
655 if (err) {
656 tx->tx_err = err;
657 return;
658 }
659 }
660 err = zio_wait(zio);
661 if (err) {
662 tx->tx_err = err;
663 return;
664 }
665 }
666
667 dmu_tx_count_free(txh, off, len);
668 }
669
670 void
671 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
672 {
673 dmu_tx_hold_t *txh;
674 dnode_t *dn;
675 uint64_t nblocks;
676 int epbs, err;
677
678 ASSERT(tx->tx_txg == 0);
679
680 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
681 object, THT_ZAP, add, (uintptr_t)name);
682 if (txh == NULL)
683 return;
684 dn = txh->txh_dnode;
685
686 dmu_tx_count_dnode(txh);
687
688 if (dn == NULL) {
689 /*
690 * We will be able to fit a new object's entries into one leaf
691 * block. So there will be at most 2 blocks total,
692 * including the header block.
693 */
694 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
695 return;
696 }
697
698 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
699
700 if (dn->dn_maxblkid == 0 && !add) {
701 blkptr_t *bp;
702
703 /*
704 * If there is only one block (i.e. this is a micro-zap)
705 * and we are not adding anything, the accounting is simple.
706 */
707 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
708 if (err) {
709 tx->tx_err = err;
710 return;
711 }
712
713 /*
714 * Use max block size here, since we don't know how much
715 * the size will change between now and the dbuf dirty call.
716 */
717 bp = &dn->dn_phys->dn_blkptr[0];
718 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
719 bp, bp->blk_birth))
720 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
721 else
722 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
723 if (!BP_IS_HOLE(bp))
724 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
725 return;
726 }
727
728 if (dn->dn_maxblkid > 0 && name) {
729 /*
730 * access the name in this fat-zap so that we'll check
731 * for i/o errors to the leaf blocks, etc.
732 */
733 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
734 8, 0, NULL);
735 if (err == EIO) {
736 tx->tx_err = err;
737 return;
738 }
739 }
740
741 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
742 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
743
744 /*
745 * If the modified blocks are scattered to the four winds,
746 * we'll have to modify an indirect twig for each.
747 */
748 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
749 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
750 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
751 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
752 else
753 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
754 }
755
756 void
757 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
758 {
759 dmu_tx_hold_t *txh;
760
761 ASSERT(tx->tx_txg == 0);
762
763 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
764 object, THT_BONUS, 0, 0);
765 if (txh)
766 dmu_tx_count_dnode(txh);
767 }
768
769 void
770 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
771 {
772 dmu_tx_hold_t *txh;
773 ASSERT(tx->tx_txg == 0);
774
775 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
776 DMU_NEW_OBJECT, THT_SPACE, space, 0);
777
778 txh->txh_space_towrite += space;
779 }
780
781 int
782 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
783 {
784 dmu_tx_hold_t *txh;
785 int holds = 0;
786
787 /*
788 * By asserting that the tx is assigned, we're counting the
789 * number of dn_tx_holds, which is the same as the number of
790 * dn_holds. Otherwise, we'd be counting dn_holds, but
791 * dn_tx_holds could be 0.
792 */
793 ASSERT(tx->tx_txg != 0);
794
795 /* if (tx->tx_anyobj == TRUE) */
796 /* return (0); */
797
798 for (txh = list_head(&tx->tx_holds); txh;
799 txh = list_next(&tx->tx_holds, txh)) {
800 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
801 holds++;
802 }
803
804 return (holds);
805 }
806
807 #ifdef ZFS_DEBUG
808 void
809 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
810 {
811 dmu_tx_hold_t *txh;
812 int match_object = FALSE, match_offset = FALSE;
813 dnode_t *dn;
814
815 DB_DNODE_ENTER(db);
816 dn = DB_DNODE(db);
817 ASSERT(tx->tx_txg != 0);
818 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
819 ASSERT3U(dn->dn_object, ==, db->db.db_object);
820
821 if (tx->tx_anyobj) {
822 DB_DNODE_EXIT(db);
823 return;
824 }
825
826 /* XXX No checking on the meta dnode for now */
827 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
828 DB_DNODE_EXIT(db);
829 return;
830 }
831
832 for (txh = list_head(&tx->tx_holds); txh;
833 txh = list_next(&tx->tx_holds, txh)) {
834 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
835 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
836 match_object = TRUE;
837 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
838 int datablkshift = dn->dn_datablkshift ?
839 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
840 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
841 int shift = datablkshift + epbs * db->db_level;
842 uint64_t beginblk = shift >= 64 ? 0 :
843 (txh->txh_arg1 >> shift);
844 uint64_t endblk = shift >= 64 ? 0 :
845 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
846 uint64_t blkid = db->db_blkid;
847
848 /* XXX txh_arg2 better not be zero... */
849
850 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
851 txh->txh_type, beginblk, endblk);
852
853 switch (txh->txh_type) {
854 case THT_WRITE:
855 if (blkid >= beginblk && blkid <= endblk)
856 match_offset = TRUE;
857 /*
858 * We will let this hold work for the bonus
859 * or spill buffer so that we don't need to
860 * hold it when creating a new object.
861 */
862 if (blkid == DMU_BONUS_BLKID ||
863 blkid == DMU_SPILL_BLKID)
864 match_offset = TRUE;
865 /*
866 * They might have to increase nlevels,
867 * thus dirtying the new TLIBs. Or the
868 * might have to change the block size,
869 * thus dirying the new lvl=0 blk=0.
870 */
871 if (blkid == 0)
872 match_offset = TRUE;
873 break;
874 case THT_FREE:
875 /*
876 * We will dirty all the level 1 blocks in
877 * the free range and perhaps the first and
878 * last level 0 block.
879 */
880 if (blkid >= beginblk && (blkid <= endblk ||
881 txh->txh_arg2 == DMU_OBJECT_END))
882 match_offset = TRUE;
883 break;
884 case THT_SPILL:
885 if (blkid == DMU_SPILL_BLKID)
886 match_offset = TRUE;
887 break;
888 case THT_BONUS:
889 if (blkid == DMU_BONUS_BLKID)
890 match_offset = TRUE;
891 break;
892 case THT_ZAP:
893 match_offset = TRUE;
894 break;
895 case THT_NEWOBJECT:
896 match_object = TRUE;
897 break;
898 default:
899 ASSERT(!"bad txh_type");
900 }
901 }
902 if (match_object && match_offset) {
903 DB_DNODE_EXIT(db);
904 return;
905 }
906 }
907 DB_DNODE_EXIT(db);
908 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
909 (u_longlong_t)db->db.db_object, db->db_level,
910 (u_longlong_t)db->db_blkid);
911 }
912 #endif
913
914 /*
915 * If we can't do 10 iops, something is wrong. Let us go ahead
916 * and hit zfs_dirty_data_max.
917 */
918 hrtime_t zfs_delay_max_ns = MSEC2NSEC(100);
919 int zfs_delay_resolution_ns = 100 * 1000; /* 100 microseconds */
920
921 /*
922 * We delay transactions when we've determined that the backend storage
923 * isn't able to accommodate the rate of incoming writes.
924 *
925 * If there is already a transaction waiting, we delay relative to when
926 * that transaction finishes waiting. This way the calculated min_time
927 * is independent of the number of threads concurrently executing
928 * transactions.
929 *
930 * If we are the only waiter, wait relative to when the transaction
931 * started, rather than the current time. This credits the transaction for
932 * "time already served", e.g. reading indirect blocks.
933 *
934 * The minimum time for a transaction to take is calculated as:
935 * min_time = scale * (dirty - min) / (max - dirty)
936 * min_time is then capped at zfs_delay_max_ns.
937 *
938 * The delay has two degrees of freedom that can be adjusted via tunables.
939 * The percentage of dirty data at which we start to delay is defined by
940 * zfs_delay_min_dirty_percent. This should typically be at or above
941 * zfs_vdev_async_write_active_max_dirty_percent so that we only start to
942 * delay after writing at full speed has failed to keep up with the incoming
943 * write rate. The scale of the curve is defined by zfs_delay_scale. Roughly
944 * speaking, this variable determines the amount of delay at the midpoint of
945 * the curve.
946 *
947 * delay
948 * 10ms +-------------------------------------------------------------*+
949 * | *|
950 * 9ms + *+
951 * | *|
952 * 8ms + *+
953 * | * |
954 * 7ms + * +
955 * | * |
956 * 6ms + * +
957 * | * |
958 * 5ms + * +
959 * | * |
960 * 4ms + * +
961 * | * |
962 * 3ms + * +
963 * | * |
964 * 2ms + (midpoint) * +
965 * | | ** |
966 * 1ms + v *** +
967 * | zfs_delay_scale ----------> ******** |
968 * 0 +-------------------------------------*********----------------+
969 * 0% <- zfs_dirty_data_max -> 100%
970 *
971 * Note that since the delay is added to the outstanding time remaining on the
972 * most recent transaction, the delay is effectively the inverse of IOPS.
973 * Here the midpoint of 500us translates to 2000 IOPS. The shape of the curve
974 * was chosen such that small changes in the amount of accumulated dirty data
975 * in the first 3/4 of the curve yield relatively small differences in the
976 * amount of delay.
977 *
978 * The effects can be easier to understand when the amount of delay is
979 * represented on a log scale:
980 *
981 * delay
982 * 100ms +-------------------------------------------------------------++
983 * + +
984 * | |
985 * + *+
986 * 10ms + *+
987 * + ** +
988 * | (midpoint) ** |
989 * + | ** +
990 * 1ms + v **** +
991 * + zfs_delay_scale ----------> ***** +
992 * | **** |
993 * + **** +
994 * 100us + ** +
995 * + * +
996 * | * |
997 * + * +
998 * 10us + * +
999 * + +
1000 * | |
1001 * + +
1002 * +--------------------------------------------------------------+
1003 * 0% <- zfs_dirty_data_max -> 100%
1004 *
1005 * Note here that only as the amount of dirty data approaches its limit does
1006 * the delay start to increase rapidly. The goal of a properly tuned system
1007 * should be to keep the amount of dirty data out of that range by first
1008 * ensuring that the appropriate limits are set for the I/O scheduler to reach
1009 * optimal throughput on the backend storage, and then by changing the value
1010 * of zfs_delay_scale to increase the steepness of the curve.
1011 */
1012 static void
1013 dmu_tx_delay(dmu_tx_t *tx, uint64_t dirty)
1014 {
1015 dsl_pool_t *dp = tx->tx_pool;
1016 uint64_t delay_min_bytes =
1017 zfs_dirty_data_max * zfs_delay_min_dirty_percent / 100;
1018 hrtime_t wakeup, min_tx_time, now;
1019
1020 if (dirty <= delay_min_bytes)
1021 return;
1022
1023 /*
1024 * The caller has already waited until we are under the max.
1025 * We make them pass us the amount of dirty data so we don't
1026 * have to handle the case of it being >= the max, which could
1027 * cause a divide-by-zero if it's == the max.
1028 */
1029 ASSERT3U(dirty, <, zfs_dirty_data_max);
1030
1031 now = gethrtime();
1032 min_tx_time = zfs_delay_scale *
1033 (dirty - delay_min_bytes) / (zfs_dirty_data_max - dirty);
1034 if (now > tx->tx_start + min_tx_time)
1035 return;
1036
1037 min_tx_time = MIN(min_tx_time, zfs_delay_max_ns);
1038
1039 DTRACE_PROBE3(delay__mintime, dmu_tx_t *, tx, uint64_t, dirty,
1040 uint64_t, min_tx_time);
1041
1042 mutex_enter(&dp->dp_lock);
1043 wakeup = MAX(tx->tx_start + min_tx_time,
1044 dp->dp_last_wakeup + min_tx_time);
1045 dp->dp_last_wakeup = wakeup;
1046 mutex_exit(&dp->dp_lock);
1047
1048 #ifdef _KERNEL
1049 mutex_enter(&curthread->t_delay_lock);
1050 while (cv_timedwait_hires(&curthread->t_delay_cv,
1051 &curthread->t_delay_lock, wakeup, zfs_delay_resolution_ns,
1052 CALLOUT_FLAG_ABSOLUTE | CALLOUT_FLAG_ROUNDUP) > 0)
1053 continue;
1054 mutex_exit(&curthread->t_delay_lock);
1055 #else
1056 hrtime_t delta = wakeup - gethrtime();
1057 struct timespec ts;
1058 ts.tv_sec = delta / NANOSEC;
1059 ts.tv_nsec = delta % NANOSEC;
1060 (void) nanosleep(&ts, NULL);
1061 #endif
1062 }
1063
1064 static int
1065 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
1066 {
1067 dmu_tx_hold_t *txh;
1068 spa_t *spa = tx->tx_pool->dp_spa;
1069 uint64_t memory, asize, fsize, usize;
1070 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
1071
1072 ASSERT0(tx->tx_txg);
1073
1074 if (tx->tx_err)
1075 return (tx->tx_err);
1076
1077 if (spa_suspended(spa)) {
1078 /*
1079 * If the user has indicated a blocking failure mode
1080 * then return ERESTART which will block in dmu_tx_wait().
1081 * Otherwise, return EIO so that an error can get
1082 * propagated back to the VOP calls.
1083 *
1084 * Note that we always honor the txg_how flag regardless
1085 * of the failuremode setting.
1086 */
1087 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
1088 txg_how != TXG_WAIT)
1089 return (SET_ERROR(EIO));
1090
1091 return (SET_ERROR(ERESTART));
1092 }
1093
1094 if (!tx->tx_waited &&
1095 dsl_pool_need_dirty_delay(tx->tx_pool)) {
1096 tx->tx_wait_dirty = B_TRUE;
1097 return (SET_ERROR(ERESTART));
1098 }
1099
1100 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
1101 tx->tx_needassign_txh = NULL;
1102
1103 /*
1104 * NB: No error returns are allowed after txg_hold_open, but
1105 * before processing the dnode holds, due to the
1106 * dmu_tx_unassign() logic.
1107 */
1108
1109 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
1110 for (txh = list_head(&tx->tx_holds); txh;
1111 txh = list_next(&tx->tx_holds, txh)) {
1112 dnode_t *dn = txh->txh_dnode;
1113 if (dn != NULL) {
1114 mutex_enter(&dn->dn_mtx);
1115 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
1116 mutex_exit(&dn->dn_mtx);
1117 tx->tx_needassign_txh = txh;
1118 return (SET_ERROR(ERESTART));
1119 }
1120 if (dn->dn_assigned_txg == 0)
1121 dn->dn_assigned_txg = tx->tx_txg;
1122 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1123 (void) refcount_add(&dn->dn_tx_holds, tx);
1124 mutex_exit(&dn->dn_mtx);
1125 }
1126 towrite += txh->txh_space_towrite;
1127 tofree += txh->txh_space_tofree;
1128 tooverwrite += txh->txh_space_tooverwrite;
1129 tounref += txh->txh_space_tounref;
1130 tohold += txh->txh_memory_tohold;
1131 fudge += txh->txh_fudge;
1132 }
1133
1134 /*
1135 * If a snapshot has been taken since we made our estimates,
1136 * assume that we won't be able to free or overwrite anything.
1137 */
1138 if (tx->tx_objset &&
1139 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1140 tx->tx_lastsnap_txg) {
1141 towrite += tooverwrite;
1142 tooverwrite = tofree = 0;
1143 }
1144
1145 /* needed allocation: worst-case estimate of write space */
1146 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1147 /* freed space estimate: worst-case overwrite + free estimate */
1148 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1149 /* convert unrefd space to worst-case estimate */
1150 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1151 /* calculate memory footprint estimate */
1152 memory = towrite + tooverwrite + tohold;
1153
1154 #ifdef ZFS_DEBUG
1155 /*
1156 * Add in 'tohold' to account for our dirty holds on this memory
1157 * XXX - the "fudge" factor is to account for skipped blocks that
1158 * we missed because dnode_next_offset() misses in-core-only blocks.
1159 */
1160 tx->tx_space_towrite = asize +
1161 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1162 tx->tx_space_tofree = tofree;
1163 tx->tx_space_tooverwrite = tooverwrite;
1164 tx->tx_space_tounref = tounref;
1165 #endif
1166
1167 if (tx->tx_dir && asize != 0) {
1168 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1169 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1170 if (err)
1171 return (err);
1172 }
1173
1174 return (0);
1175 }
1176
1177 static void
1178 dmu_tx_unassign(dmu_tx_t *tx)
1179 {
1180 dmu_tx_hold_t *txh;
1181
1182 if (tx->tx_txg == 0)
1183 return;
1184
1185 txg_rele_to_quiesce(&tx->tx_txgh);
1186
1187 /*
1188 * Walk the transaction's hold list, removing the hold on the
1189 * associated dnode, and notifying waiters if the refcount drops to 0.
1190 */
1191 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1192 txh = list_next(&tx->tx_holds, txh)) {
1193 dnode_t *dn = txh->txh_dnode;
1194
1195 if (dn == NULL)
1196 continue;
1197 mutex_enter(&dn->dn_mtx);
1198 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1199
1200 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1201 dn->dn_assigned_txg = 0;
1202 cv_broadcast(&dn->dn_notxholds);
1203 }
1204 mutex_exit(&dn->dn_mtx);
1205 }
1206
1207 txg_rele_to_sync(&tx->tx_txgh);
1208
1209 tx->tx_lasttried_txg = tx->tx_txg;
1210 tx->tx_txg = 0;
1211 }
1212
1213 /*
1214 * Assign tx to a transaction group. txg_how can be one of:
1215 *
1216 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1217 * a new one. This should be used when you're not holding locks.
1218 * It will only fail if we're truly out of space (or over quota).
1219 *
1220 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1221 * blocking, returns immediately with ERESTART. This should be used
1222 * whenever you're holding locks. On an ERESTART error, the caller
1223 * should drop locks, do a dmu_tx_wait(tx), and try again.
1224 *
1225 * (3) TXG_WAITED. Like TXG_NOWAIT, but indicates that dmu_tx_wait()
1226 * has already been called on behalf of this operation (though
1227 * most likely on a different tx).
1228 */
1229 int
1230 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1231 {
1232 int err;
1233
1234 ASSERT(tx->tx_txg == 0);
1235 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT ||
1236 txg_how == TXG_WAITED);
1237 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1238
1239 /* If we might wait, we must not hold the config lock. */
1240 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1241
1242 if (txg_how == TXG_WAITED)
1243 tx->tx_waited = B_TRUE;
1244
1245 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1246 dmu_tx_unassign(tx);
1247
1248 if (err != ERESTART || txg_how != TXG_WAIT)
1249 return (err);
1250
1251 dmu_tx_wait(tx);
1252 }
1253
1254 txg_rele_to_quiesce(&tx->tx_txgh);
1255
1256 return (0);
1257 }
1258
1259 void
1260 dmu_tx_wait(dmu_tx_t *tx)
1261 {
1262 spa_t *spa = tx->tx_pool->dp_spa;
1263 dsl_pool_t *dp = tx->tx_pool;
1264
1265 ASSERT(tx->tx_txg == 0);
1266 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1267
1268 if (tx->tx_wait_dirty) {
1269 /*
1270 * dmu_tx_try_assign() has determined that we need to wait
1271 * because we've consumed much or all of the dirty buffer
1272 * space.
1273 */
1274 mutex_enter(&dp->dp_lock);
1275 while (dp->dp_dirty_total >= zfs_dirty_data_max)
1276 cv_wait(&dp->dp_spaceavail_cv, &dp->dp_lock);
1277 uint64_t dirty = dp->dp_dirty_total;
1278 mutex_exit(&dp->dp_lock);
1279
1280 dmu_tx_delay(tx, dirty);
1281
1282 tx->tx_wait_dirty = B_FALSE;
1283
1284 /*
1285 * Note: setting tx_waited only has effect if the caller
1286 * used TX_WAIT. Otherwise they are going to destroy
1287 * this tx and try again. The common case, zfs_write(),
1288 * uses TX_WAIT.
1289 */
1290 tx->tx_waited = B_TRUE;
1291 } else if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1292 /*
1293 * If the pool is suspended we need to wait until it
1294 * is resumed. Note that it's possible that the pool
1295 * has become active after this thread has tried to
1296 * obtain a tx. If that's the case then tx_lasttried_txg
1297 * would not have been set.
1298 */
1299 txg_wait_synced(dp, spa_last_synced_txg(spa) + 1);
1300 } else if (tx->tx_needassign_txh) {
1301 /*
1302 * A dnode is assigned to the quiescing txg. Wait for its
1303 * transaction to complete.
1304 */
1305 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1306
1307 mutex_enter(&dn->dn_mtx);
1308 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1309 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1310 mutex_exit(&dn->dn_mtx);
1311 tx->tx_needassign_txh = NULL;
1312 } else {
1313 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1314 }
1315 }
1316
1317 void
1318 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1319 {
1320 #ifdef ZFS_DEBUG
1321 if (tx->tx_dir == NULL || delta == 0)
1322 return;
1323
1324 if (delta > 0) {
1325 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1326 tx->tx_space_towrite);
1327 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1328 } else {
1329 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1330 }
1331 #endif
1332 }
1333
1334 void
1335 dmu_tx_commit(dmu_tx_t *tx)
1336 {
1337 dmu_tx_hold_t *txh;
1338
1339 ASSERT(tx->tx_txg != 0);
1340
1341 /*
1342 * Go through the transaction's hold list and remove holds on
1343 * associated dnodes, notifying waiters if no holds remain.
1344 */
1345 while (txh = list_head(&tx->tx_holds)) {
1346 dnode_t *dn = txh->txh_dnode;
1347
1348 list_remove(&tx->tx_holds, txh);
1349 kmem_free(txh, sizeof (dmu_tx_hold_t));
1350 if (dn == NULL)
1351 continue;
1352 mutex_enter(&dn->dn_mtx);
1353 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1354
1355 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1356 dn->dn_assigned_txg = 0;
1357 cv_broadcast(&dn->dn_notxholds);
1358 }
1359 mutex_exit(&dn->dn_mtx);
1360 dnode_rele(dn, tx);
1361 }
1362
1363 if (tx->tx_tempreserve_cookie)
1364 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1365
1366 if (!list_is_empty(&tx->tx_callbacks))
1367 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1368
1369 if (tx->tx_anyobj == FALSE)
1370 txg_rele_to_sync(&tx->tx_txgh);
1371
1372 list_destroy(&tx->tx_callbacks);
1373 list_destroy(&tx->tx_holds);
1374 #ifdef ZFS_DEBUG
1375 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1376 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1377 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1378 refcount_destroy_many(&tx->tx_space_written,
1379 refcount_count(&tx->tx_space_written));
1380 refcount_destroy_many(&tx->tx_space_freed,
1381 refcount_count(&tx->tx_space_freed));
1382 #endif
1383 kmem_free(tx, sizeof (dmu_tx_t));
1384 }
1385
1386 void
1387 dmu_tx_abort(dmu_tx_t *tx)
1388 {
1389 dmu_tx_hold_t *txh;
1390
1391 ASSERT(tx->tx_txg == 0);
1392
1393 while (txh = list_head(&tx->tx_holds)) {
1394 dnode_t *dn = txh->txh_dnode;
1395
1396 list_remove(&tx->tx_holds, txh);
1397 kmem_free(txh, sizeof (dmu_tx_hold_t));
1398 if (dn != NULL)
1399 dnode_rele(dn, tx);
1400 }
1401
1402 /*
1403 * Call any registered callbacks with an error code.
1404 */
1405 if (!list_is_empty(&tx->tx_callbacks))
1406 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1407
1408 list_destroy(&tx->tx_callbacks);
1409 list_destroy(&tx->tx_holds);
1410 #ifdef ZFS_DEBUG
1411 refcount_destroy_many(&tx->tx_space_written,
1412 refcount_count(&tx->tx_space_written));
1413 refcount_destroy_many(&tx->tx_space_freed,
1414 refcount_count(&tx->tx_space_freed));
1415 #endif
1416 kmem_free(tx, sizeof (dmu_tx_t));
1417 }
1418
1419 uint64_t
1420 dmu_tx_get_txg(dmu_tx_t *tx)
1421 {
1422 ASSERT(tx->tx_txg != 0);
1423 return (tx->tx_txg);
1424 }
1425
1426 dsl_pool_t *
1427 dmu_tx_pool(dmu_tx_t *tx)
1428 {
1429 ASSERT(tx->tx_pool != NULL);
1430 return (tx->tx_pool);
1431 }
1432
1433
1434 void
1435 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1436 {
1437 dmu_tx_callback_t *dcb;
1438
1439 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1440
1441 dcb->dcb_func = func;
1442 dcb->dcb_data = data;
1443
1444 list_insert_tail(&tx->tx_callbacks, dcb);
1445 }
1446
1447 /*
1448 * Call all the commit callbacks on a list, with a given error code.
1449 */
1450 void
1451 dmu_tx_do_callbacks(list_t *cb_list, int error)
1452 {
1453 dmu_tx_callback_t *dcb;
1454
1455 while (dcb = list_head(cb_list)) {
1456 list_remove(cb_list, dcb);
1457 dcb->dcb_func(dcb->dcb_data, error);
1458 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1459 }
1460 }
1461
1462 /*
1463 * Interface to hold a bunch of attributes.
1464 * used for creating new files.
1465 * attrsize is the total size of all attributes
1466 * to be added during object creation
1467 *
1468 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1469 */
1470
1471 /*
1472 * hold necessary attribute name for attribute registration.
1473 * should be a very rare case where this is needed. If it does
1474 * happen it would only happen on the first write to the file system.
1475 */
1476 static void
1477 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1478 {
1479 int i;
1480
1481 if (!sa->sa_need_attr_registration)
1482 return;
1483
1484 for (i = 0; i != sa->sa_num_attrs; i++) {
1485 if (!sa->sa_attr_table[i].sa_registered) {
1486 if (sa->sa_reg_attr_obj)
1487 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1488 B_TRUE, sa->sa_attr_table[i].sa_name);
1489 else
1490 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1491 B_TRUE, sa->sa_attr_table[i].sa_name);
1492 }
1493 }
1494 }
1495
1496
1497 void
1498 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1499 {
1500 dnode_t *dn;
1501 dmu_tx_hold_t *txh;
1502
1503 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1504 THT_SPILL, 0, 0);
1505
1506 dn = txh->txh_dnode;
1507
1508 if (dn == NULL)
1509 return;
1510
1511 /* If blkptr doesn't exist then add space to towrite */
1512 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1513 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1514 } else {
1515 blkptr_t *bp;
1516
1517 bp = &dn->dn_phys->dn_spill;
1518 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1519 bp, bp->blk_birth))
1520 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1521 else
1522 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1523 if (!BP_IS_HOLE(bp))
1524 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1525 }
1526 }
1527
1528 void
1529 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1530 {
1531 sa_os_t *sa = tx->tx_objset->os_sa;
1532
1533 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1534
1535 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1536 return;
1537
1538 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1539 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1540 else {
1541 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1542 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1543 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1544 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1545 }
1546
1547 dmu_tx_sa_registration_hold(sa, tx);
1548
1549 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1550 return;
1551
1552 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1553 THT_SPILL, 0, 0);
1554 }
1555
1556 /*
1557 * Hold SA attribute
1558 *
1559 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1560 *
1561 * variable_size is the total size of all variable sized attributes
1562 * passed to this function. It is not the total size of all
1563 * variable size attributes that *may* exist on this object.
1564 */
1565 void
1566 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1567 {
1568 uint64_t object;
1569 sa_os_t *sa = tx->tx_objset->os_sa;
1570
1571 ASSERT(hdl != NULL);
1572
1573 object = sa_handle_object(hdl);
1574
1575 dmu_tx_hold_bonus(tx, object);
1576
1577 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1578 return;
1579
1580 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1581 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1582 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1583 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1584 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1585 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1586 }
1587
1588 dmu_tx_sa_registration_hold(sa, tx);
1589
1590 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1591 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1592
1593 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1594 ASSERT(tx->tx_txg == 0);
1595 dmu_tx_hold_spill(tx, object);
1596 } else {
1597 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1598 dnode_t *dn;
1599
1600 DB_DNODE_ENTER(db);
1601 dn = DB_DNODE(db);
1602 if (dn->dn_have_spill) {
1603 ASSERT(tx->tx_txg == 0);
1604 dmu_tx_hold_spill(tx, object);
1605 }
1606 DB_DNODE_EXIT(db);
1607 }
1608 }