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 */
24/*
25 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
26 */
27
28#include <sys/dmu.h>
29#include <sys/dmu_impl.h>
30#include <sys/dbuf.h>
31#include <sys/dmu_tx.h>
32#include <sys/dmu_objset.h>
33#include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
34#include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
35#include <sys/dsl_pool.h>
36#include <sys/zap_impl.h> /* for fzap_default_block_shift */
37#include <sys/spa.h>
38#include <sys/sa.h>
39#include <sys/sa_impl.h>
40#include <sys/zfs_context.h>
41#include <sys/varargs.h>
42
43typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 uint64_t arg1, uint64_t arg2);
45
46
47dmu_tx_t *
48dmu_tx_create_dd(dsl_dir_t *dd)
49{
50 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
51 tx->tx_dir = dd;
52 if (dd)
53 tx->tx_pool = dd->dd_pool;
54 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 offsetof(dmu_tx_hold_t, txh_node));
56 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 offsetof(dmu_tx_callback_t, dcb_node));
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
65dmu_tx_t *
66dmu_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
74dmu_tx_t *
75dmu_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
87int
88dmu_tx_is_syncing(dmu_tx_t *tx)
89{
90 return (tx->tx_anyobj);
91}
92
93int
94dmu_tx_private_ok(dmu_tx_t *tx)
95{
96 return (tx->tx_anyobj);
97}
98
99static dmu_tx_hold_t *
100dmu_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
141void
142dmu_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
154static int
155dmu_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 (EIO);
165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 dbuf_rele(db, FTAG);
167 return (err);
168}
169
170static void
171dmu_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 */
215static void
216dmu_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 if (dn->dn_maxblkid > 0) {
289 /*
290 * The blocksize can't change,
291 * so we can make a more precise estimate.
292 */
293 ASSERT(dn->dn_datablkshift != 0);
294 min_bs = max_bs = dn->dn_datablkshift;
295 min_ibs = max_ibs = dn->dn_indblkshift;
296 } else if (dn->dn_indblkshift > max_ibs) {
297 /*
298 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
299 * the code will still work correctly on older pools.
300 */
301 min_ibs = max_ibs = dn->dn_indblkshift;
302 }
303
304 /*
305 * If this write is not off the end of the file
306 * we need to account for overwrites/unref.
307 */
308 if (start <= dn->dn_maxblkid) {
309 for (int l = 0; l < DN_MAX_LEVELS; l++)
310 history[l] = -1ULL;
311 }
312 while (start <= dn->dn_maxblkid) {
313 dmu_buf_impl_t *db;
314
315 rw_enter(&dn->dn_struct_rwlock, RW_READER);
316 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
317 rw_exit(&dn->dn_struct_rwlock);
318
319 if (err) {
320 txh->txh_tx->tx_err = err;
321 return;
322 }
323
324 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
325 history);
326 dbuf_rele(db, FTAG);
327 if (++start > end) {
328 /*
329 * Account for new indirects appearing
330 * before this IO gets assigned into a txg.
331 */
332 bits = 64 - min_bs;
333 epbs = min_ibs - SPA_BLKPTRSHIFT;
334 for (bits -= epbs * (nlvls - 1);
335 bits >= 0; bits -= epbs)
336 txh->txh_fudge += 1ULL << max_ibs;
337 goto out;
338 }
339 off += delta;
340 if (len >= delta)
341 len -= delta;
342 delta = dn->dn_datablksz;
343 }
344 }
345
346 /*
347 * 'end' is the last thing we will access, not one past.
348 * This way we won't overflow when accessing the last byte.
349 */
350 start = P2ALIGN(off, 1ULL << max_bs);
351 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
352 txh->txh_space_towrite += end - start + 1;
353
354 start >>= min_bs;
355 end >>= min_bs;
356
357 epbs = min_ibs - SPA_BLKPTRSHIFT;
358
359 /*
360 * The object contains at most 2^(64 - min_bs) blocks,
361 * and each indirect level maps 2^epbs.
362 */
363 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
364 start >>= epbs;
365 end >>= epbs;
366 ASSERT3U(end, >=, start);
367 txh->txh_space_towrite += (end - start + 1) << max_ibs;
368 if (start != 0) {
369 /*
370 * We also need a new blkid=0 indirect block
371 * to reference any existing file data.
372 */
373 txh->txh_space_towrite += 1ULL << max_ibs;
374 }
375 }
376
377out:
378 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
379 2 * DMU_MAX_ACCESS)
380 err = EFBIG;
381
382 if (err)
383 txh->txh_tx->tx_err = err;
384}
385
386static void
387dmu_tx_count_dnode(dmu_tx_hold_t *txh)
388{
389 dnode_t *dn = txh->txh_dnode;
390 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
391 uint64_t space = mdn->dn_datablksz +
392 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
393
394 if (dn && dn->dn_dbuf->db_blkptr &&
395 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
396 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
397 txh->txh_space_tooverwrite += space;
398 txh->txh_space_tounref += space;
399 } else {
400 txh->txh_space_towrite += space;
401 if (dn && dn->dn_dbuf->db_blkptr)
402 txh->txh_space_tounref += space;
403 }
404}
405
406void
407dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
408{
409 dmu_tx_hold_t *txh;
410
411 ASSERT(tx->tx_txg == 0);
412 ASSERT(len < DMU_MAX_ACCESS);
413 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
414
415 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
416 object, THT_WRITE, off, len);
417 if (txh == NULL)
418 return;
419
420 dmu_tx_count_write(txh, off, len);
421 dmu_tx_count_dnode(txh);
422}
423
424static void
425dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
426{
427 uint64_t blkid, nblks, lastblk;
428 uint64_t space = 0, unref = 0, skipped = 0;
429 dnode_t *dn = txh->txh_dnode;
430 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
431 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
432 int epbs;
433
434 if (dn->dn_nlevels == 0)
435 return;
436
437 /*
438 * The struct_rwlock protects us against dn_nlevels
439 * changing, in case (against all odds) we manage to dirty &
440 * sync out the changes after we check for being dirty.
441 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
442 */
443 rw_enter(&dn->dn_struct_rwlock, RW_READER);
444 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
445 if (dn->dn_maxblkid == 0) {
446 if (off == 0 && len >= dn->dn_datablksz) {
447 blkid = 0;
448 nblks = 1;
449 } else {
450 rw_exit(&dn->dn_struct_rwlock);
451 return;
452 }
453 } else {
454 blkid = off >> dn->dn_datablkshift;
455 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
456
457 if (blkid >= dn->dn_maxblkid) {
458 rw_exit(&dn->dn_struct_rwlock);
459 return;
460 }
461 if (blkid + nblks > dn->dn_maxblkid)
462 nblks = dn->dn_maxblkid - blkid;
463
464 }
465 if (dn->dn_nlevels == 1) {
466 int i;
467 for (i = 0; i < nblks; i++) {
468 blkptr_t *bp = dn->dn_phys->dn_blkptr;
469 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
470 bp += blkid + i;
471 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
472 dprintf_bp(bp, "can free old%s", "");
473 space += bp_get_dsize(spa, bp);
474 }
475 unref += BP_GET_ASIZE(bp);
476 }
477 nblks = 0;
478 }
479
480 /*
481 * Add in memory requirements of higher-level indirects.
482 * This assumes a worst-possible scenario for dn_nlevels.
483 */
484 {
485 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
486 int level = (dn->dn_nlevels > 1) ? 2 : 1;
487
488 while (level++ < DN_MAX_LEVELS) {
489 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
490 blkcnt = 1 + (blkcnt >> epbs);
491 }
492 ASSERT(blkcnt <= dn->dn_nblkptr);
493 }
494
495 lastblk = blkid + nblks - 1;
496 while (nblks) {
497 dmu_buf_impl_t *dbuf;
498 uint64_t ibyte, new_blkid;
499 int epb = 1 << epbs;
500 int err, i, blkoff, tochk;
501 blkptr_t *bp;
502
503 ibyte = blkid << dn->dn_datablkshift;
504 err = dnode_next_offset(dn,
505 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
506 new_blkid = ibyte >> dn->dn_datablkshift;
507 if (err == ESRCH) {
508 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
509 break;
510 }
511 if (err) {
512 txh->txh_tx->tx_err = err;
513 break;
514 }
515 if (new_blkid > lastblk) {
516 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
517 break;
518 }
519
520 if (new_blkid > blkid) {
521 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
522 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
523 nblks -= new_blkid - blkid;
524 blkid = new_blkid;
525 }
526 blkoff = P2PHASE(blkid, epb);
527 tochk = MIN(epb - blkoff, nblks);
528
529 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
530 if (err) {
531 txh->txh_tx->tx_err = err;
532 break;
533 }
534
535 txh->txh_memory_tohold += dbuf->db.db_size;
536
537 /*
538 * We don't check memory_tohold against DMU_MAX_ACCESS because
539 * memory_tohold is an over-estimation (especially the >L1
540 * indirect blocks), so it could fail. Callers should have
541 * already verified that they will not be holding too much
542 * memory.
543 */
544
545 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
546 if (err != 0) {
547 txh->txh_tx->tx_err = err;
548 dbuf_rele(dbuf, FTAG);
549 break;
550 }
551
552 bp = dbuf->db.db_data;
553 bp += blkoff;
554
555 for (i = 0; i < tochk; i++) {
556 if (dsl_dataset_block_freeable(ds, &bp[i],
557 bp[i].blk_birth)) {
558 dprintf_bp(&bp[i], "can free old%s", "");
559 space += bp_get_dsize(spa, &bp[i]);
560 }
561 unref += BP_GET_ASIZE(bp);
562 }
563 dbuf_rele(dbuf, FTAG);
564
565 blkid += tochk;
566 nblks -= tochk;
567 }
568 rw_exit(&dn->dn_struct_rwlock);
569
570 /* account for new level 1 indirect blocks that might show up */
571 if (skipped > 0) {
572 txh->txh_fudge += skipped << dn->dn_indblkshift;
573 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
574 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
575 }
576 txh->txh_space_tofree += space;
577 txh->txh_space_tounref += unref;
578}
579
580void
581dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
582{
583 dmu_tx_hold_t *txh;
584 dnode_t *dn;
585 uint64_t start, end, i;
586 int err, shift;
587 zio_t *zio;
588
589 ASSERT(tx->tx_txg == 0);
590
591 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
592 object, THT_FREE, off, len);
593 if (txh == NULL)
594 return;
595 dn = txh->txh_dnode;
596
597 /* first block */
598 if (off != 0)
599 dmu_tx_count_write(txh, off, 1);
600 /* last block */
601 if (len != DMU_OBJECT_END)
602 dmu_tx_count_write(txh, off+len, 1);
603
604 dmu_tx_count_dnode(txh);
605
606 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
607 return;
608 if (len == DMU_OBJECT_END)
609 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
610
611 /*
612 * For i/o error checking, read the first and last level-0
613 * blocks, and all the level-1 blocks. The above count_write's
614 * have already taken care of the level-0 blocks.
615 */
616 if (dn->dn_nlevels > 1) {
617 shift = dn->dn_datablkshift + dn->dn_indblkshift -
618 SPA_BLKPTRSHIFT;
619 start = off >> shift;
620 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
621
622 zio = zio_root(tx->tx_pool->dp_spa,
623 NULL, NULL, ZIO_FLAG_CANFAIL);
624 for (i = start; i <= end; i++) {
625 uint64_t ibyte = i << shift;
626 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
627 i = ibyte >> shift;
628 if (err == ESRCH)
629 break;
630 if (err) {
631 tx->tx_err = err;
632 return;
633 }
634
635 err = dmu_tx_check_ioerr(zio, dn, 1, i);
636 if (err) {
637 tx->tx_err = err;
638 return;
639 }
640 }
641 err = zio_wait(zio);
642 if (err) {
643 tx->tx_err = err;
644 return;
645 }
646 }
647
648 dmu_tx_count_free(txh, off, len);
649}
650
651void
652dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
653{
654 dmu_tx_hold_t *txh;
655 dnode_t *dn;
656 uint64_t nblocks;
657 int epbs, err;
658
659 ASSERT(tx->tx_txg == 0);
660
661 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
662 object, THT_ZAP, add, (uintptr_t)name);
663 if (txh == NULL)
664 return;
665 dn = txh->txh_dnode;
666
667 dmu_tx_count_dnode(txh);
668
669 if (dn == NULL) {
670 /*
671 * We will be able to fit a new object's entries into one leaf
672 * block. So there will be at most 2 blocks total,
673 * including the header block.
674 */
675 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
676 return;
677 }
678
679 ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
680
681 if (dn->dn_maxblkid == 0 && !add) {
682 blkptr_t *bp;
683
684 /*
685 * If there is only one block (i.e. this is a micro-zap)
686 * and we are not adding anything, the accounting is simple.
687 */
688 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
689 if (err) {
690 tx->tx_err = err;
691 return;
692 }
693
694 /*
695 * Use max block size here, since we don't know how much
696 * the size will change between now and the dbuf dirty call.
697 */
698 bp = &dn->dn_phys->dn_blkptr[0];
699 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
700 bp, bp->blk_birth))
701 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
702 else
703 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
704 if (!BP_IS_HOLE(bp))
705 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
706 return;
707 }
708
709 if (dn->dn_maxblkid > 0 && name) {
710 /*
711 * access the name in this fat-zap so that we'll check
712 * for i/o errors to the leaf blocks, etc.
713 */
714 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
715 8, 0, NULL);
716 if (err == EIO) {
717 tx->tx_err = err;
718 return;
719 }
720 }
721
722 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
723 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
724
725 /*
726 * If the modified blocks are scattered to the four winds,
727 * we'll have to modify an indirect twig for each.
728 */
729 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
730 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
731 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
732 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
733 else
734 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
735}
736
737void
738dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
739{
740 dmu_tx_hold_t *txh;
741
742 ASSERT(tx->tx_txg == 0);
743
744 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
745 object, THT_BONUS, 0, 0);
746 if (txh)
747 dmu_tx_count_dnode(txh);
748}
749
750void
751dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
752{
753 dmu_tx_hold_t *txh;
754 ASSERT(tx->tx_txg == 0);
755
756 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
757 DMU_NEW_OBJECT, THT_SPACE, space, 0);
758
759 txh->txh_space_towrite += space;
760}
761
762int
763dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
764{
765 dmu_tx_hold_t *txh;
766 int holds = 0;
767
768 /*
769 * By asserting that the tx is assigned, we're counting the
770 * number of dn_tx_holds, which is the same as the number of
771 * dn_holds. Otherwise, we'd be counting dn_holds, but
772 * dn_tx_holds could be 0.
773 */
774 ASSERT(tx->tx_txg != 0);
775
776 /* if (tx->tx_anyobj == TRUE) */
777 /* return (0); */
778
779 for (txh = list_head(&tx->tx_holds); txh;
780 txh = list_next(&tx->tx_holds, txh)) {
781 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
782 holds++;
783 }
784
785 return (holds);
786}
787
788#ifdef ZFS_DEBUG
789void
790dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
791{
792 dmu_tx_hold_t *txh;
793 int match_object = FALSE, match_offset = FALSE;
794 dnode_t *dn;
795
796 DB_DNODE_ENTER(db);
797 dn = DB_DNODE(db);
798 ASSERT(tx->tx_txg != 0);
799 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
800 ASSERT3U(dn->dn_object, ==, db->db.db_object);
801
802 if (tx->tx_anyobj) {
803 DB_DNODE_EXIT(db);
804 return;
805 }
806
807 /* XXX No checking on the meta dnode for now */
808 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
809 DB_DNODE_EXIT(db);
810 return;
811 }
812
813 for (txh = list_head(&tx->tx_holds); txh;
814 txh = list_next(&tx->tx_holds, txh)) {
815 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
816 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
817 match_object = TRUE;
818 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
819 int datablkshift = dn->dn_datablkshift ?
820 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
821 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
822 int shift = datablkshift + epbs * db->db_level;
823 uint64_t beginblk = shift >= 64 ? 0 :
824 (txh->txh_arg1 >> shift);
825 uint64_t endblk = shift >= 64 ? 0 :
826 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
827 uint64_t blkid = db->db_blkid;
828
829 /* XXX txh_arg2 better not be zero... */
830
831 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
832 txh->txh_type, beginblk, endblk);
833
834 switch (txh->txh_type) {
835 case THT_WRITE:
836 if (blkid >= beginblk && blkid <= endblk)
837 match_offset = TRUE;
838 /*
839 * We will let this hold work for the bonus
840 * or spill buffer so that we don't need to
841 * hold it when creating a new object.
842 */
843 if (blkid == DMU_BONUS_BLKID ||
844 blkid == DMU_SPILL_BLKID)
845 match_offset = TRUE;
846 /*
847 * They might have to increase nlevels,
848 * thus dirtying the new TLIBs. Or the
849 * might have to change the block size,
850 * thus dirying the new lvl=0 blk=0.
851 */
852 if (blkid == 0)
853 match_offset = TRUE;
854 break;
855 case THT_FREE:
856 /*
857 * We will dirty all the level 1 blocks in
858 * the free range and perhaps the first and
859 * last level 0 block.
860 */
861 if (blkid >= beginblk && (blkid <= endblk ||
862 txh->txh_arg2 == DMU_OBJECT_END))
863 match_offset = TRUE;
864 break;
865 case THT_SPILL:
866 if (blkid == DMU_SPILL_BLKID)
867 match_offset = TRUE;
868 break;
869 case THT_BONUS:
870 if (blkid == DMU_BONUS_BLKID)
871 match_offset = TRUE;
872 break;
873 case THT_ZAP:
874 match_offset = TRUE;
875 break;
876 case THT_NEWOBJECT:
877 match_object = TRUE;
878 break;
879 default:
880 ASSERT(!"bad txh_type");
881 }
882 }
883 if (match_object && match_offset) {
884 DB_DNODE_EXIT(db);
885 return;
886 }
887 }
888 DB_DNODE_EXIT(db);
889 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
890 (u_longlong_t)db->db.db_object, db->db_level,
891 (u_longlong_t)db->db_blkid);
892}
893#endif
894
895static int
896dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
897{
898 dmu_tx_hold_t *txh;
899 spa_t *spa = tx->tx_pool->dp_spa;
900 uint64_t memory, asize, fsize, usize;
901 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
902
903 ASSERT3U(tx->tx_txg, ==, 0);
904
905 if (tx->tx_err)
906 return (tx->tx_err);
907
908 if (spa_suspended(spa)) {
909 /*
910 * If the user has indicated a blocking failure mode
911 * then return ERESTART which will block in dmu_tx_wait().
912 * Otherwise, return EIO so that an error can get
913 * propagated back to the VOP calls.
914 *
915 * Note that we always honor the txg_how flag regardless
916 * of the failuremode setting.
917 */
918 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
919 txg_how != TXG_WAIT)
920 return (EIO);
921
922 return (ERESTART);
923 }
924
925 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
926 tx->tx_needassign_txh = NULL;
927
928 /*
929 * NB: No error returns are allowed after txg_hold_open, but
930 * before processing the dnode holds, due to the
931 * dmu_tx_unassign() logic.
932 */
933
934 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
935 for (txh = list_head(&tx->tx_holds); txh;
936 txh = list_next(&tx->tx_holds, txh)) {
937 dnode_t *dn = txh->txh_dnode;
938 if (dn != NULL) {
939 mutex_enter(&dn->dn_mtx);
940 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
941 mutex_exit(&dn->dn_mtx);
942 tx->tx_needassign_txh = txh;
943 return (ERESTART);
944 }
945 if (dn->dn_assigned_txg == 0)
946 dn->dn_assigned_txg = tx->tx_txg;
947 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
948 (void) refcount_add(&dn->dn_tx_holds, tx);
949 mutex_exit(&dn->dn_mtx);
950 }
951 towrite += txh->txh_space_towrite;
952 tofree += txh->txh_space_tofree;
953 tooverwrite += txh->txh_space_tooverwrite;
954 tounref += txh->txh_space_tounref;
955 tohold += txh->txh_memory_tohold;
956 fudge += txh->txh_fudge;
957 }
958
959 /*
960 * NB: This check must be after we've held the dnodes, so that
961 * the dmu_tx_unassign() logic will work properly
962 */
963 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
964 return (ERESTART);
965
966 /*
967 * If a snapshot has been taken since we made our estimates,
968 * assume that we won't be able to free or overwrite anything.
969 */
970 if (tx->tx_objset &&
971 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
972 tx->tx_lastsnap_txg) {
973 towrite += tooverwrite;
974 tooverwrite = tofree = 0;
975 }
976
977 /* needed allocation: worst-case estimate of write space */
978 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
979 /* freed space estimate: worst-case overwrite + free estimate */
980 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
981 /* convert unrefd space to worst-case estimate */
982 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
983 /* calculate memory footprint estimate */
984 memory = towrite + tooverwrite + tohold;
985
986#ifdef ZFS_DEBUG
987 /*
988 * Add in 'tohold' to account for our dirty holds on this memory
989 * XXX - the "fudge" factor is to account for skipped blocks that
990 * we missed because dnode_next_offset() misses in-core-only blocks.
991 */
992 tx->tx_space_towrite = asize +
993 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
994 tx->tx_space_tofree = tofree;
995 tx->tx_space_tooverwrite = tooverwrite;
996 tx->tx_space_tounref = tounref;
997#endif
998
999 if (tx->tx_dir && asize != 0) {
1000 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1001 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1002 if (err)
1003 return (err);
1004 }
1005
1006 return (0);
1007}
1008
1009static void
1010dmu_tx_unassign(dmu_tx_t *tx)
1011{
1012 dmu_tx_hold_t *txh;
1013
1014 if (tx->tx_txg == 0)
1015 return;
1016
1017 txg_rele_to_quiesce(&tx->tx_txgh);
1018
1019 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1020 txh = list_next(&tx->tx_holds, txh)) {
1021 dnode_t *dn = txh->txh_dnode;
1022
1023 if (dn == NULL)
1024 continue;
1025 mutex_enter(&dn->dn_mtx);
1026 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1027
1028 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1029 dn->dn_assigned_txg = 0;
1030 cv_broadcast(&dn->dn_notxholds);
1031 }
1032 mutex_exit(&dn->dn_mtx);
1033 }
1034
1035 txg_rele_to_sync(&tx->tx_txgh);
1036
1037 tx->tx_lasttried_txg = tx->tx_txg;
1038 tx->tx_txg = 0;
1039}
1040
1041/*
1042 * Assign tx to a transaction group. txg_how can be one of:
1043 *
1044 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1045 * a new one. This should be used when you're not holding locks.
1046 * If will only fail if we're truly out of space (or over quota).
1047 *
1048 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1049 * blocking, returns immediately with ERESTART. This should be used
1050 * whenever you're holding locks. On an ERESTART error, the caller
1051 * should drop locks, do a dmu_tx_wait(tx), and try again.
1052 *
1053 * (3) A specific txg. Use this if you need to ensure that multiple
1054 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1055 * returns ERESTART if it can't assign you into the requested txg.
1056 */
1057int
1058dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1059{
1060 int err;
1061
1062 ASSERT(tx->tx_txg == 0);
1063 ASSERT(txg_how != 0);
1064 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1065
1066 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1067 dmu_tx_unassign(tx);
1068
1069 if (err != ERESTART || txg_how != TXG_WAIT)
1070 return (err);
1071
1072 dmu_tx_wait(tx);
1073 }
1074
1075 txg_rele_to_quiesce(&tx->tx_txgh);
1076
1077 return (0);
1078}
1079
1080void
1081dmu_tx_wait(dmu_tx_t *tx)
1082{
1083 spa_t *spa = tx->tx_pool->dp_spa;
1084
1085 ASSERT(tx->tx_txg == 0);
1086
1087 /*
1088 * It's possible that the pool has become active after this thread
1089 * has tried to obtain a tx. If that's the case then his
1090 * tx_lasttried_txg would not have been assigned.
1091 */
1092 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1093 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1094 } else if (tx->tx_needassign_txh) {
1095 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1096
1097 mutex_enter(&dn->dn_mtx);
1098 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1099 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1100 mutex_exit(&dn->dn_mtx);
1101 tx->tx_needassign_txh = NULL;
1102 } else {
1103 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1104 }
1105}
1106
1107void
1108dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1109{
1110#ifdef ZFS_DEBUG
1111 if (tx->tx_dir == NULL || delta == 0)
1112 return;
1113
1114 if (delta > 0) {
1115 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1116 tx->tx_space_towrite);
1117 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1118 } else {
1119 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1120 }
1121#endif
1122}
1123
1124void
1125dmu_tx_commit(dmu_tx_t *tx)
1126{
1127 dmu_tx_hold_t *txh;
1128
1129 ASSERT(tx->tx_txg != 0);
1130
1131 while (txh = list_head(&tx->tx_holds)) {
1132 dnode_t *dn = txh->txh_dnode;
1133
1134 list_remove(&tx->tx_holds, txh);
1135 kmem_free(txh, sizeof (dmu_tx_hold_t));
1136 if (dn == NULL)
1137 continue;
1138 mutex_enter(&dn->dn_mtx);
1139 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1140
1141 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1142 dn->dn_assigned_txg = 0;
1143 cv_broadcast(&dn->dn_notxholds);
1144 }
1145 mutex_exit(&dn->dn_mtx);
1146 dnode_rele(dn, tx);
1147 }
1148
1149 if (tx->tx_tempreserve_cookie)
1150 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1151
1152 if (!list_is_empty(&tx->tx_callbacks))
1153 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1154
1155 if (tx->tx_anyobj == FALSE)
1156 txg_rele_to_sync(&tx->tx_txgh);
1157
1158 list_destroy(&tx->tx_callbacks);
1159 list_destroy(&tx->tx_holds);
1160#ifdef ZFS_DEBUG
1161 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1162 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1163 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1164 refcount_destroy_many(&tx->tx_space_written,
1165 refcount_count(&tx->tx_space_written));
1166 refcount_destroy_many(&tx->tx_space_freed,
1167 refcount_count(&tx->tx_space_freed));
1168#endif
1169 kmem_free(tx, sizeof (dmu_tx_t));
1170}
1171
1172void
1173dmu_tx_abort(dmu_tx_t *tx)
1174{
1175 dmu_tx_hold_t *txh;
1176
1177 ASSERT(tx->tx_txg == 0);
1178
1179 while (txh = list_head(&tx->tx_holds)) {
1180 dnode_t *dn = txh->txh_dnode;
1181
1182 list_remove(&tx->tx_holds, txh);
1183 kmem_free(txh, sizeof (dmu_tx_hold_t));
1184 if (dn != NULL)
1185 dnode_rele(dn, tx);
1186 }
1187
1188 /*
1189 * Call any registered callbacks with an error code.
1190 */
1191 if (!list_is_empty(&tx->tx_callbacks))
1192 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1193
1194 list_destroy(&tx->tx_callbacks);
1195 list_destroy(&tx->tx_holds);
1196#ifdef ZFS_DEBUG
1197 refcount_destroy_many(&tx->tx_space_written,
1198 refcount_count(&tx->tx_space_written));
1199 refcount_destroy_many(&tx->tx_space_freed,
1200 refcount_count(&tx->tx_space_freed));
1201#endif
1202 kmem_free(tx, sizeof (dmu_tx_t));
1203}
1204
1205uint64_t
1206dmu_tx_get_txg(dmu_tx_t *tx)
1207{
1208 ASSERT(tx->tx_txg != 0);
1209 return (tx->tx_txg);
1210}
1211
1212void
1213dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1214{
1215 dmu_tx_callback_t *dcb;
1216
1217 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1218
1219 dcb->dcb_func = func;
1220 dcb->dcb_data = data;
1221
1222 list_insert_tail(&tx->tx_callbacks, dcb);
1223}
1224
1225/*
1226 * Call all the commit callbacks on a list, with a given error code.
1227 */
1228void
1229dmu_tx_do_callbacks(list_t *cb_list, int error)
1230{
1231 dmu_tx_callback_t *dcb;
1232
1233 while (dcb = list_head(cb_list)) {
1234 list_remove(cb_list, dcb);
1235 dcb->dcb_func(dcb->dcb_data, error);
1236 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1237 }
1238}
1239
1240/*
1241 * Interface to hold a bunch of attributes.
1242 * used for creating new files.
1243 * attrsize is the total size of all attributes
1244 * to be added during object creation
1245 *
1246 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1247 */
1248
1249/*
1250 * hold necessary attribute name for attribute registration.
1251 * should be a very rare case where this is needed. If it does
1252 * happen it would only happen on the first write to the file system.
1253 */
1254static void
1255dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1256{
1257 int i;
1258
1259 if (!sa->sa_need_attr_registration)
1260 return;
1261
1262 for (i = 0; i != sa->sa_num_attrs; i++) {
1263 if (!sa->sa_attr_table[i].sa_registered) {
1264 if (sa->sa_reg_attr_obj)
1265 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1266 B_TRUE, sa->sa_attr_table[i].sa_name);
1267 else
1268 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1269 B_TRUE, sa->sa_attr_table[i].sa_name);
1270 }
1271 }
1272}
1273
1274
1275void
1276dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1277{
1278 dnode_t *dn;
1279 dmu_tx_hold_t *txh;
1280
1281 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1282 THT_SPILL, 0, 0);
1283
1284 dn = txh->txh_dnode;
1285
1286 if (dn == NULL)
1287 return;
1288
1289 /* If blkptr doesn't exist then add space to towrite */
1290 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1291 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1292 } else {
1293 blkptr_t *bp;
1294
1295 bp = &dn->dn_phys->dn_spill;
1296 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1297 bp, bp->blk_birth))
1298 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1299 else
1300 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1301 if (!BP_IS_HOLE(bp))
1302 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1303 }
1304}
1305
1306void
1307dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1308{
1309 sa_os_t *sa = tx->tx_objset->os_sa;
1310
1311 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1312
1313 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1314 return;
1315
1316 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1317 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1318 else {
1319 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1320 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1321 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1323 }
1324
1325 dmu_tx_sa_registration_hold(sa, tx);
1326
1327 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1328 return;
1329
1330 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1331 THT_SPILL, 0, 0);
1332}
1333
1334/*
1335 * Hold SA attribute
1336 *
1337 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1338 *
1339 * variable_size is the total size of all variable sized attributes
1340 * passed to this function. It is not the total size of all
1341 * variable size attributes that *may* exist on this object.
1342 */
1343void
1344dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1345{
1346 uint64_t object;
1347 sa_os_t *sa = tx->tx_objset->os_sa;
1348
1349 ASSERT(hdl != NULL);
1350
1351 object = sa_handle_object(hdl);
1352
1353 dmu_tx_hold_bonus(tx, object);
1354
1355 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1356 return;
1357
1358 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1359 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1360 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1361 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1362 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1364 }
1365
1366 dmu_tx_sa_registration_hold(sa, tx);
1367
1368 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1369 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1370
1371 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1372 ASSERT(tx->tx_txg == 0);
1373 dmu_tx_hold_spill(tx, object);
1374 } else {
1375 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1376 dnode_t *dn;
1377
1378 DB_DNODE_ENTER(db);
1379 dn = DB_DNODE(db);
1380 if (dn->dn_have_spill) {
1381 ASSERT(tx->tx_txg == 0);
1382 dmu_tx_hold_spill(tx, object);
1383 }
1384 DB_DNODE_EXIT(db);
1385 }
1386}
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 */
24/*
25 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
26 */
27
28#include <sys/dmu.h>
29#include <sys/dmu_impl.h>
30#include <sys/dbuf.h>
31#include <sys/dmu_tx.h>
32#include <sys/dmu_objset.h>
33#include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
34#include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
35#include <sys/dsl_pool.h>
36#include <sys/zap_impl.h> /* for fzap_default_block_shift */
37#include <sys/spa.h>
38#include <sys/sa.h>
39#include <sys/sa_impl.h>
40#include <sys/zfs_context.h>
41#include <sys/varargs.h>
42
43typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
44 uint64_t arg1, uint64_t arg2);
45
46
47dmu_tx_t *
48dmu_tx_create_dd(dsl_dir_t *dd)
49{
50 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
51 tx->tx_dir = dd;
52 if (dd)
53 tx->tx_pool = dd->dd_pool;
54 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
55 offsetof(dmu_tx_hold_t, txh_node));
56 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
57 offsetof(dmu_tx_callback_t, dcb_node));
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
65dmu_tx_t *
66dmu_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
74dmu_tx_t *
75dmu_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
87int
88dmu_tx_is_syncing(dmu_tx_t *tx)
89{
90 return (tx->tx_anyobj);
91}
92
93int
94dmu_tx_private_ok(dmu_tx_t *tx)
95{
96 return (tx->tx_anyobj);
97}
98
99static dmu_tx_hold_t *
100dmu_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
141void
142dmu_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
154static int
155dmu_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 (EIO);
165 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
166 dbuf_rele(db, FTAG);
167 return (err);
168}
169
170static void
171dmu_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 */
215static void
216dmu_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 if (dn->dn_maxblkid > 0) {
289 /*
290 * The blocksize can't change,
291 * so we can make a more precise estimate.
292 */
293 ASSERT(dn->dn_datablkshift != 0);
294 min_bs = max_bs = dn->dn_datablkshift;
295 min_ibs = max_ibs = dn->dn_indblkshift;
296 } else if (dn->dn_indblkshift > max_ibs) {
297 /*
298 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
299 * the code will still work correctly on older pools.
300 */
301 min_ibs = max_ibs = dn->dn_indblkshift;
302 }
303
304 /*
305 * If this write is not off the end of the file
306 * we need to account for overwrites/unref.
307 */
308 if (start <= dn->dn_maxblkid) {
309 for (int l = 0; l < DN_MAX_LEVELS; l++)
310 history[l] = -1ULL;
311 }
312 while (start <= dn->dn_maxblkid) {
313 dmu_buf_impl_t *db;
314
315 rw_enter(&dn->dn_struct_rwlock, RW_READER);
316 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
317 rw_exit(&dn->dn_struct_rwlock);
318
319 if (err) {
320 txh->txh_tx->tx_err = err;
321 return;
322 }
323
324 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
325 history);
326 dbuf_rele(db, FTAG);
327 if (++start > end) {
328 /*
329 * Account for new indirects appearing
330 * before this IO gets assigned into a txg.
331 */
332 bits = 64 - min_bs;
333 epbs = min_ibs - SPA_BLKPTRSHIFT;
334 for (bits -= epbs * (nlvls - 1);
335 bits >= 0; bits -= epbs)
336 txh->txh_fudge += 1ULL << max_ibs;
337 goto out;
338 }
339 off += delta;
340 if (len >= delta)
341 len -= delta;
342 delta = dn->dn_datablksz;
343 }
344 }
345
346 /*
347 * 'end' is the last thing we will access, not one past.
348 * This way we won't overflow when accessing the last byte.
349 */
350 start = P2ALIGN(off, 1ULL << max_bs);
351 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
352 txh->txh_space_towrite += end - start + 1;
353
354 start >>= min_bs;
355 end >>= min_bs;
356
357 epbs = min_ibs - SPA_BLKPTRSHIFT;
358
359 /*
360 * The object contains at most 2^(64 - min_bs) blocks,
361 * and each indirect level maps 2^epbs.
362 */
363 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
364 start >>= epbs;
365 end >>= epbs;
366 ASSERT3U(end, >=, start);
367 txh->txh_space_towrite += (end - start + 1) << max_ibs;
368 if (start != 0) {
369 /*
370 * We also need a new blkid=0 indirect block
371 * to reference any existing file data.
372 */
373 txh->txh_space_towrite += 1ULL << max_ibs;
374 }
375 }
376
377out:
378 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
379 2 * DMU_MAX_ACCESS)
380 err = EFBIG;
381
382 if (err)
383 txh->txh_tx->tx_err = err;
384}
385
386static void
387dmu_tx_count_dnode(dmu_tx_hold_t *txh)
388{
389 dnode_t *dn = txh->txh_dnode;
390 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
391 uint64_t space = mdn->dn_datablksz +
392 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
393
394 if (dn && dn->dn_dbuf->db_blkptr &&
395 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
396 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
397 txh->txh_space_tooverwrite += space;
398 txh->txh_space_tounref += space;
399 } else {
400 txh->txh_space_towrite += space;
401 if (dn && dn->dn_dbuf->db_blkptr)
402 txh->txh_space_tounref += space;
403 }
404}
405
406void
407dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
408{
409 dmu_tx_hold_t *txh;
410
411 ASSERT(tx->tx_txg == 0);
412 ASSERT(len < DMU_MAX_ACCESS);
413 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
414
415 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
416 object, THT_WRITE, off, len);
417 if (txh == NULL)
418 return;
419
420 dmu_tx_count_write(txh, off, len);
421 dmu_tx_count_dnode(txh);
422}
423
424static void
425dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
426{
427 uint64_t blkid, nblks, lastblk;
428 uint64_t space = 0, unref = 0, skipped = 0;
429 dnode_t *dn = txh->txh_dnode;
430 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
431 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
432 int epbs;
433
434 if (dn->dn_nlevels == 0)
435 return;
436
437 /*
438 * The struct_rwlock protects us against dn_nlevels
439 * changing, in case (against all odds) we manage to dirty &
440 * sync out the changes after we check for being dirty.
441 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
442 */
443 rw_enter(&dn->dn_struct_rwlock, RW_READER);
444 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
445 if (dn->dn_maxblkid == 0) {
446 if (off == 0 && len >= dn->dn_datablksz) {
447 blkid = 0;
448 nblks = 1;
449 } else {
450 rw_exit(&dn->dn_struct_rwlock);
451 return;
452 }
453 } else {
454 blkid = off >> dn->dn_datablkshift;
455 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
456
457 if (blkid >= dn->dn_maxblkid) {
458 rw_exit(&dn->dn_struct_rwlock);
459 return;
460 }
461 if (blkid + nblks > dn->dn_maxblkid)
462 nblks = dn->dn_maxblkid - blkid;
463
464 }
465 if (dn->dn_nlevels == 1) {
466 int i;
467 for (i = 0; i < nblks; i++) {
468 blkptr_t *bp = dn->dn_phys->dn_blkptr;
469 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
470 bp += blkid + i;
471 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
472 dprintf_bp(bp, "can free old%s", "");
473 space += bp_get_dsize(spa, bp);
474 }
475 unref += BP_GET_ASIZE(bp);
476 }
477 nblks = 0;
478 }
479
480 /*
481 * Add in memory requirements of higher-level indirects.
482 * This assumes a worst-possible scenario for dn_nlevels.
483 */
484 {
485 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
486 int level = (dn->dn_nlevels > 1) ? 2 : 1;
487
488 while (level++ < DN_MAX_LEVELS) {
489 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
490 blkcnt = 1 + (blkcnt >> epbs);
491 }
492 ASSERT(blkcnt <= dn->dn_nblkptr);
493 }
494
495 lastblk = blkid + nblks - 1;
496 while (nblks) {
497 dmu_buf_impl_t *dbuf;
498 uint64_t ibyte, new_blkid;
499 int epb = 1 << epbs;
500 int err, i, blkoff, tochk;
501 blkptr_t *bp;
502
503 ibyte = blkid << dn->dn_datablkshift;
504 err = dnode_next_offset(dn,
505 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
506 new_blkid = ibyte >> dn->dn_datablkshift;
507 if (err == ESRCH) {
508 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
509 break;
510 }
511 if (err) {
512 txh->txh_tx->tx_err = err;
513 break;
514 }
515 if (new_blkid > lastblk) {
516 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
517 break;
518 }
519
520 if (new_blkid > blkid) {
521 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
522 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
523 nblks -= new_blkid - blkid;
524 blkid = new_blkid;
525 }
526 blkoff = P2PHASE(blkid, epb);
527 tochk = MIN(epb - blkoff, nblks);
528
529 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
530 if (err) {
531 txh->txh_tx->tx_err = err;
532 break;
533 }
534
535 txh->txh_memory_tohold += dbuf->db.db_size;
536
537 /*
538 * We don't check memory_tohold against DMU_MAX_ACCESS because
539 * memory_tohold is an over-estimation (especially the >L1
540 * indirect blocks), so it could fail. Callers should have
541 * already verified that they will not be holding too much
542 * memory.
543 */
544
545 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
546 if (err != 0) {
547 txh->txh_tx->tx_err = err;
548 dbuf_rele(dbuf, FTAG);
549 break;
550 }
551
552 bp = dbuf->db.db_data;
553 bp += blkoff;
554
555 for (i = 0; i < tochk; i++) {
556 if (dsl_dataset_block_freeable(ds, &bp[i],
557 bp[i].blk_birth)) {
558 dprintf_bp(&bp[i], "can free old%s", "");
559 space += bp_get_dsize(spa, &bp[i]);
560 }
561 unref += BP_GET_ASIZE(bp);
562 }
563 dbuf_rele(dbuf, FTAG);
564
565 blkid += tochk;
566 nblks -= tochk;
567 }
568 rw_exit(&dn->dn_struct_rwlock);
569
570 /* account for new level 1 indirect blocks that might show up */
571 if (skipped > 0) {
572 txh->txh_fudge += skipped << dn->dn_indblkshift;
573 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
574 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
575 }
576 txh->txh_space_tofree += space;
577 txh->txh_space_tounref += unref;
578}
579
580void
581dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
582{
583 dmu_tx_hold_t *txh;
584 dnode_t *dn;
585 uint64_t start, end, i;
586 int err, shift;
587 zio_t *zio;
588
589 ASSERT(tx->tx_txg == 0);
590
591 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
592 object, THT_FREE, off, len);
593 if (txh == NULL)
594 return;
595 dn = txh->txh_dnode;
596
597 /* first block */
598 if (off != 0)
599 dmu_tx_count_write(txh, off, 1);
600 /* last block */
601 if (len != DMU_OBJECT_END)
602 dmu_tx_count_write(txh, off+len, 1);
603
604 dmu_tx_count_dnode(txh);
605
606 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
607 return;
608 if (len == DMU_OBJECT_END)
609 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
610
611 /*
612 * For i/o error checking, read the first and last level-0
613 * blocks, and all the level-1 blocks. The above count_write's
614 * have already taken care of the level-0 blocks.
615 */
616 if (dn->dn_nlevels > 1) {
617 shift = dn->dn_datablkshift + dn->dn_indblkshift -
618 SPA_BLKPTRSHIFT;
619 start = off >> shift;
620 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
621
622 zio = zio_root(tx->tx_pool->dp_spa,
623 NULL, NULL, ZIO_FLAG_CANFAIL);
624 for (i = start; i <= end; i++) {
625 uint64_t ibyte = i << shift;
626 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
627 i = ibyte >> shift;
628 if (err == ESRCH)
629 break;
630 if (err) {
631 tx->tx_err = err;
632 return;
633 }
634
635 err = dmu_tx_check_ioerr(zio, dn, 1, i);
636 if (err) {
637 tx->tx_err = err;
638 return;
639 }
640 }
641 err = zio_wait(zio);
642 if (err) {
643 tx->tx_err = err;
644 return;
645 }
646 }
647
648 dmu_tx_count_free(txh, off, len);
649}
650
651void
652dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
653{
654 dmu_tx_hold_t *txh;
655 dnode_t *dn;
656 uint64_t nblocks;
657 int epbs, err;
658
659 ASSERT(tx->tx_txg == 0);
660
661 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
662 object, THT_ZAP, add, (uintptr_t)name);
663 if (txh == NULL)
664 return;
665 dn = txh->txh_dnode;
666
667 dmu_tx_count_dnode(txh);
668
669 if (dn == NULL) {
670 /*
671 * We will be able to fit a new object's entries into one leaf
672 * block. So there will be at most 2 blocks total,
673 * including the header block.
674 */
675 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
676 return;
677 }
678
679 ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
680
681 if (dn->dn_maxblkid == 0 && !add) {
682 blkptr_t *bp;
683
684 /*
685 * If there is only one block (i.e. this is a micro-zap)
686 * and we are not adding anything, the accounting is simple.
687 */
688 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
689 if (err) {
690 tx->tx_err = err;
691 return;
692 }
693
694 /*
695 * Use max block size here, since we don't know how much
696 * the size will change between now and the dbuf dirty call.
697 */
698 bp = &dn->dn_phys->dn_blkptr[0];
699 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
700 bp, bp->blk_birth))
701 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
702 else
703 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
704 if (!BP_IS_HOLE(bp))
705 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
706 return;
707 }
708
709 if (dn->dn_maxblkid > 0 && name) {
710 /*
711 * access the name in this fat-zap so that we'll check
712 * for i/o errors to the leaf blocks, etc.
713 */
714 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
715 8, 0, NULL);
716 if (err == EIO) {
717 tx->tx_err = err;
718 return;
719 }
720 }
721
722 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
723 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
724
725 /*
726 * If the modified blocks are scattered to the four winds,
727 * we'll have to modify an indirect twig for each.
728 */
729 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
730 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
731 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
732 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
733 else
734 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
735}
736
737void
738dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
739{
740 dmu_tx_hold_t *txh;
741
742 ASSERT(tx->tx_txg == 0);
743
744 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
745 object, THT_BONUS, 0, 0);
746 if (txh)
747 dmu_tx_count_dnode(txh);
748}
749
750void
751dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
752{
753 dmu_tx_hold_t *txh;
754 ASSERT(tx->tx_txg == 0);
755
756 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
757 DMU_NEW_OBJECT, THT_SPACE, space, 0);
758
759 txh->txh_space_towrite += space;
760}
761
762int
763dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
764{
765 dmu_tx_hold_t *txh;
766 int holds = 0;
767
768 /*
769 * By asserting that the tx is assigned, we're counting the
770 * number of dn_tx_holds, which is the same as the number of
771 * dn_holds. Otherwise, we'd be counting dn_holds, but
772 * dn_tx_holds could be 0.
773 */
774 ASSERT(tx->tx_txg != 0);
775
776 /* if (tx->tx_anyobj == TRUE) */
777 /* return (0); */
778
779 for (txh = list_head(&tx->tx_holds); txh;
780 txh = list_next(&tx->tx_holds, txh)) {
781 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
782 holds++;
783 }
784
785 return (holds);
786}
787
788#ifdef ZFS_DEBUG
789void
790dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
791{
792 dmu_tx_hold_t *txh;
793 int match_object = FALSE, match_offset = FALSE;
794 dnode_t *dn;
795
796 DB_DNODE_ENTER(db);
797 dn = DB_DNODE(db);
798 ASSERT(tx->tx_txg != 0);
799 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
800 ASSERT3U(dn->dn_object, ==, db->db.db_object);
801
802 if (tx->tx_anyobj) {
803 DB_DNODE_EXIT(db);
804 return;
805 }
806
807 /* XXX No checking on the meta dnode for now */
808 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
809 DB_DNODE_EXIT(db);
810 return;
811 }
812
813 for (txh = list_head(&tx->tx_holds); txh;
814 txh = list_next(&tx->tx_holds, txh)) {
815 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
816 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
817 match_object = TRUE;
818 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
819 int datablkshift = dn->dn_datablkshift ?
820 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
821 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
822 int shift = datablkshift + epbs * db->db_level;
823 uint64_t beginblk = shift >= 64 ? 0 :
824 (txh->txh_arg1 >> shift);
825 uint64_t endblk = shift >= 64 ? 0 :
826 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
827 uint64_t blkid = db->db_blkid;
828
829 /* XXX txh_arg2 better not be zero... */
830
831 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
832 txh->txh_type, beginblk, endblk);
833
834 switch (txh->txh_type) {
835 case THT_WRITE:
836 if (blkid >= beginblk && blkid <= endblk)
837 match_offset = TRUE;
838 /*
839 * We will let this hold work for the bonus
840 * or spill buffer so that we don't need to
841 * hold it when creating a new object.
842 */
843 if (blkid == DMU_BONUS_BLKID ||
844 blkid == DMU_SPILL_BLKID)
845 match_offset = TRUE;
846 /*
847 * They might have to increase nlevels,
848 * thus dirtying the new TLIBs. Or the
849 * might have to change the block size,
850 * thus dirying the new lvl=0 blk=0.
851 */
852 if (blkid == 0)
853 match_offset = TRUE;
854 break;
855 case THT_FREE:
856 /*
857 * We will dirty all the level 1 blocks in
858 * the free range and perhaps the first and
859 * last level 0 block.
860 */
861 if (blkid >= beginblk && (blkid <= endblk ||
862 txh->txh_arg2 == DMU_OBJECT_END))
863 match_offset = TRUE;
864 break;
865 case THT_SPILL:
866 if (blkid == DMU_SPILL_BLKID)
867 match_offset = TRUE;
868 break;
869 case THT_BONUS:
870 if (blkid == DMU_BONUS_BLKID)
871 match_offset = TRUE;
872 break;
873 case THT_ZAP:
874 match_offset = TRUE;
875 break;
876 case THT_NEWOBJECT:
877 match_object = TRUE;
878 break;
879 default:
880 ASSERT(!"bad txh_type");
881 }
882 }
883 if (match_object && match_offset) {
884 DB_DNODE_EXIT(db);
885 return;
886 }
887 }
888 DB_DNODE_EXIT(db);
889 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
890 (u_longlong_t)db->db.db_object, db->db_level,
891 (u_longlong_t)db->db_blkid);
892}
893#endif
894
895static int
896dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
897{
898 dmu_tx_hold_t *txh;
899 spa_t *spa = tx->tx_pool->dp_spa;
900 uint64_t memory, asize, fsize, usize;
901 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
902
903 ASSERT3U(tx->tx_txg, ==, 0);
904
905 if (tx->tx_err)
906 return (tx->tx_err);
907
908 if (spa_suspended(spa)) {
909 /*
910 * If the user has indicated a blocking failure mode
911 * then return ERESTART which will block in dmu_tx_wait().
912 * Otherwise, return EIO so that an error can get
913 * propagated back to the VOP calls.
914 *
915 * Note that we always honor the txg_how flag regardless
916 * of the failuremode setting.
917 */
918 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
919 txg_how != TXG_WAIT)
920 return (EIO);
921
922 return (ERESTART);
923 }
924
925 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
926 tx->tx_needassign_txh = NULL;
927
928 /*
929 * NB: No error returns are allowed after txg_hold_open, but
930 * before processing the dnode holds, due to the
931 * dmu_tx_unassign() logic.
932 */
933
934 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
935 for (txh = list_head(&tx->tx_holds); txh;
936 txh = list_next(&tx->tx_holds, txh)) {
937 dnode_t *dn = txh->txh_dnode;
938 if (dn != NULL) {
939 mutex_enter(&dn->dn_mtx);
940 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
941 mutex_exit(&dn->dn_mtx);
942 tx->tx_needassign_txh = txh;
943 return (ERESTART);
944 }
945 if (dn->dn_assigned_txg == 0)
946 dn->dn_assigned_txg = tx->tx_txg;
947 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
948 (void) refcount_add(&dn->dn_tx_holds, tx);
949 mutex_exit(&dn->dn_mtx);
950 }
951 towrite += txh->txh_space_towrite;
952 tofree += txh->txh_space_tofree;
953 tooverwrite += txh->txh_space_tooverwrite;
954 tounref += txh->txh_space_tounref;
955 tohold += txh->txh_memory_tohold;
956 fudge += txh->txh_fudge;
957 }
958
959 /*
960 * NB: This check must be after we've held the dnodes, so that
961 * the dmu_tx_unassign() logic will work properly
962 */
963 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
964 return (ERESTART);
965
966 /*
967 * If a snapshot has been taken since we made our estimates,
968 * assume that we won't be able to free or overwrite anything.
969 */
970 if (tx->tx_objset &&
971 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
972 tx->tx_lastsnap_txg) {
973 towrite += tooverwrite;
974 tooverwrite = tofree = 0;
975 }
976
977 /* needed allocation: worst-case estimate of write space */
978 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
979 /* freed space estimate: worst-case overwrite + free estimate */
980 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
981 /* convert unrefd space to worst-case estimate */
982 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
983 /* calculate memory footprint estimate */
984 memory = towrite + tooverwrite + tohold;
985
986#ifdef ZFS_DEBUG
987 /*
988 * Add in 'tohold' to account for our dirty holds on this memory
989 * XXX - the "fudge" factor is to account for skipped blocks that
990 * we missed because dnode_next_offset() misses in-core-only blocks.
991 */
992 tx->tx_space_towrite = asize +
993 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
994 tx->tx_space_tofree = tofree;
995 tx->tx_space_tooverwrite = tooverwrite;
996 tx->tx_space_tounref = tounref;
997#endif
998
999 if (tx->tx_dir && asize != 0) {
1000 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1001 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1002 if (err)
1003 return (err);
1004 }
1005
1006 return (0);
1007}
1008
1009static void
1010dmu_tx_unassign(dmu_tx_t *tx)
1011{
1012 dmu_tx_hold_t *txh;
1013
1014 if (tx->tx_txg == 0)
1015 return;
1016
1017 txg_rele_to_quiesce(&tx->tx_txgh);
1018
1019 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1020 txh = list_next(&tx->tx_holds, txh)) {
1021 dnode_t *dn = txh->txh_dnode;
1022
1023 if (dn == NULL)
1024 continue;
1025 mutex_enter(&dn->dn_mtx);
1026 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1027
1028 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1029 dn->dn_assigned_txg = 0;
1030 cv_broadcast(&dn->dn_notxholds);
1031 }
1032 mutex_exit(&dn->dn_mtx);
1033 }
1034
1035 txg_rele_to_sync(&tx->tx_txgh);
1036
1037 tx->tx_lasttried_txg = tx->tx_txg;
1038 tx->tx_txg = 0;
1039}
1040
1041/*
1042 * Assign tx to a transaction group. txg_how can be one of:
1043 *
1044 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1045 * a new one. This should be used when you're not holding locks.
1046 * If will only fail if we're truly out of space (or over quota).
1047 *
1048 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1049 * blocking, returns immediately with ERESTART. This should be used
1050 * whenever you're holding locks. On an ERESTART error, the caller
1051 * should drop locks, do a dmu_tx_wait(tx), and try again.
1052 *
1053 * (3) A specific txg. Use this if you need to ensure that multiple
1054 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1055 * returns ERESTART if it can't assign you into the requested txg.
1056 */
1057int
1058dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1059{
1060 int err;
1061
1062 ASSERT(tx->tx_txg == 0);
1063 ASSERT(txg_how != 0);
1064 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1065
1066 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1067 dmu_tx_unassign(tx);
1068
1069 if (err != ERESTART || txg_how != TXG_WAIT)
1070 return (err);
1071
1072 dmu_tx_wait(tx);
1073 }
1074
1075 txg_rele_to_quiesce(&tx->tx_txgh);
1076
1077 return (0);
1078}
1079
1080void
1081dmu_tx_wait(dmu_tx_t *tx)
1082{
1083 spa_t *spa = tx->tx_pool->dp_spa;
1084
1085 ASSERT(tx->tx_txg == 0);
1086
1087 /*
1088 * It's possible that the pool has become active after this thread
1089 * has tried to obtain a tx. If that's the case then his
1090 * tx_lasttried_txg would not have been assigned.
1091 */
1092 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1093 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1094 } else if (tx->tx_needassign_txh) {
1095 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1096
1097 mutex_enter(&dn->dn_mtx);
1098 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1099 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1100 mutex_exit(&dn->dn_mtx);
1101 tx->tx_needassign_txh = NULL;
1102 } else {
1103 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1104 }
1105}
1106
1107void
1108dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1109{
1110#ifdef ZFS_DEBUG
1111 if (tx->tx_dir == NULL || delta == 0)
1112 return;
1113
1114 if (delta > 0) {
1115 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1116 tx->tx_space_towrite);
1117 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1118 } else {
1119 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1120 }
1121#endif
1122}
1123
1124void
1125dmu_tx_commit(dmu_tx_t *tx)
1126{
1127 dmu_tx_hold_t *txh;
1128
1129 ASSERT(tx->tx_txg != 0);
1130
1131 while (txh = list_head(&tx->tx_holds)) {
1132 dnode_t *dn = txh->txh_dnode;
1133
1134 list_remove(&tx->tx_holds, txh);
1135 kmem_free(txh, sizeof (dmu_tx_hold_t));
1136 if (dn == NULL)
1137 continue;
1138 mutex_enter(&dn->dn_mtx);
1139 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1140
1141 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1142 dn->dn_assigned_txg = 0;
1143 cv_broadcast(&dn->dn_notxholds);
1144 }
1145 mutex_exit(&dn->dn_mtx);
1146 dnode_rele(dn, tx);
1147 }
1148
1149 if (tx->tx_tempreserve_cookie)
1150 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1151
1152 if (!list_is_empty(&tx->tx_callbacks))
1153 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1154
1155 if (tx->tx_anyobj == FALSE)
1156 txg_rele_to_sync(&tx->tx_txgh);
1157
1158 list_destroy(&tx->tx_callbacks);
1159 list_destroy(&tx->tx_holds);
1160#ifdef ZFS_DEBUG
1161 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1162 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1163 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1164 refcount_destroy_many(&tx->tx_space_written,
1165 refcount_count(&tx->tx_space_written));
1166 refcount_destroy_many(&tx->tx_space_freed,
1167 refcount_count(&tx->tx_space_freed));
1168#endif
1169 kmem_free(tx, sizeof (dmu_tx_t));
1170}
1171
1172void
1173dmu_tx_abort(dmu_tx_t *tx)
1174{
1175 dmu_tx_hold_t *txh;
1176
1177 ASSERT(tx->tx_txg == 0);
1178
1179 while (txh = list_head(&tx->tx_holds)) {
1180 dnode_t *dn = txh->txh_dnode;
1181
1182 list_remove(&tx->tx_holds, txh);
1183 kmem_free(txh, sizeof (dmu_tx_hold_t));
1184 if (dn != NULL)
1185 dnode_rele(dn, tx);
1186 }
1187
1188 /*
1189 * Call any registered callbacks with an error code.
1190 */
1191 if (!list_is_empty(&tx->tx_callbacks))
1192 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1193
1194 list_destroy(&tx->tx_callbacks);
1195 list_destroy(&tx->tx_holds);
1196#ifdef ZFS_DEBUG
1197 refcount_destroy_many(&tx->tx_space_written,
1198 refcount_count(&tx->tx_space_written));
1199 refcount_destroy_many(&tx->tx_space_freed,
1200 refcount_count(&tx->tx_space_freed));
1201#endif
1202 kmem_free(tx, sizeof (dmu_tx_t));
1203}
1204
1205uint64_t
1206dmu_tx_get_txg(dmu_tx_t *tx)
1207{
1208 ASSERT(tx->tx_txg != 0);
1209 return (tx->tx_txg);
1210}
1211
1212void
1213dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1214{
1215 dmu_tx_callback_t *dcb;
1216
1217 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1218
1219 dcb->dcb_func = func;
1220 dcb->dcb_data = data;
1221
1222 list_insert_tail(&tx->tx_callbacks, dcb);
1223}
1224
1225/*
1226 * Call all the commit callbacks on a list, with a given error code.
1227 */
1228void
1229dmu_tx_do_callbacks(list_t *cb_list, int error)
1230{
1231 dmu_tx_callback_t *dcb;
1232
1233 while (dcb = list_head(cb_list)) {
1234 list_remove(cb_list, dcb);
1235 dcb->dcb_func(dcb->dcb_data, error);
1236 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1237 }
1238}
1239
1240/*
1241 * Interface to hold a bunch of attributes.
1242 * used for creating new files.
1243 * attrsize is the total size of all attributes
1244 * to be added during object creation
1245 *
1246 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1247 */
1248
1249/*
1250 * hold necessary attribute name for attribute registration.
1251 * should be a very rare case where this is needed. If it does
1252 * happen it would only happen on the first write to the file system.
1253 */
1254static void
1255dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1256{
1257 int i;
1258
1259 if (!sa->sa_need_attr_registration)
1260 return;
1261
1262 for (i = 0; i != sa->sa_num_attrs; i++) {
1263 if (!sa->sa_attr_table[i].sa_registered) {
1264 if (sa->sa_reg_attr_obj)
1265 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1266 B_TRUE, sa->sa_attr_table[i].sa_name);
1267 else
1268 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1269 B_TRUE, sa->sa_attr_table[i].sa_name);
1270 }
1271 }
1272}
1273
1274
1275void
1276dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1277{
1278 dnode_t *dn;
1279 dmu_tx_hold_t *txh;
1280
1281 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1282 THT_SPILL, 0, 0);
1283
1284 dn = txh->txh_dnode;
1285
1286 if (dn == NULL)
1287 return;
1288
1289 /* If blkptr doesn't exist then add space to towrite */
1290 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1291 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1292 } else {
1293 blkptr_t *bp;
1294
1295 bp = &dn->dn_phys->dn_spill;
1296 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1297 bp, bp->blk_birth))
1298 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1299 else
1300 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1301 if (!BP_IS_HOLE(bp))
1302 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1303 }
1304}
1305
1306void
1307dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1308{
1309 sa_os_t *sa = tx->tx_objset->os_sa;
1310
1311 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1312
1313 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1314 return;
1315
1316 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1317 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1318 else {
1319 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1320 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1321 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1323 }
1324
1325 dmu_tx_sa_registration_hold(sa, tx);
1326
1327 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1328 return;
1329
1330 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1331 THT_SPILL, 0, 0);
1332}
1333
1334/*
1335 * Hold SA attribute
1336 *
1337 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1338 *
1339 * variable_size is the total size of all variable sized attributes
1340 * passed to this function. It is not the total size of all
1341 * variable size attributes that *may* exist on this object.
1342 */
1343void
1344dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1345{
1346 uint64_t object;
1347 sa_os_t *sa = tx->tx_objset->os_sa;
1348
1349 ASSERT(hdl != NULL);
1350
1351 object = sa_handle_object(hdl);
1352
1353 dmu_tx_hold_bonus(tx, object);
1354
1355 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1356 return;
1357
1358 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1359 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1360 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1361 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1362 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1364 }
1365
1366 dmu_tx_sa_registration_hold(sa, tx);
1367
1368 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1369 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1370
1371 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1372 ASSERT(tx->tx_txg == 0);
1373 dmu_tx_hold_spill(tx, object);
1374 } else {
1375 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1376 dnode_t *dn;
1377
1378 DB_DNODE_ENTER(db);
1379 dn = DB_DNODE(db);
1380 if (dn->dn_have_spill) {
1381 ASSERT(tx->tx_txg == 0);
1382 dmu_tx_hold_spill(tx, object);
1383 }
1384 DB_DNODE_EXIT(db);
1385 }
1386}