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.
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.
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]
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.
28 #include <sys/dmu_impl.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 */
38 #include <sys/sa_impl.h>
39 #include <sys/zfs_context.h>
40 #include <sys/varargs.h>
42 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
43 uint64_t arg1, uint64_t arg2);
47 dmu_tx_create_dd(dsl_dir_t *dd)
49 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
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));
58 refcount_create(&tx->tx_space_written);
59 refcount_create(&tx->tx_space_freed);
65 dmu_tx_create(objset_t *os)
67 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
69 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
74 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
76 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
78 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
87 dmu_tx_is_syncing(dmu_tx_t *tx)
89 return (tx->tx_anyobj);
93 dmu_tx_private_ok(dmu_tx_t *tx)
95 return (tx->tx_anyobj);
98 static dmu_tx_hold_t *
99 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
100 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
106 if (object != DMU_NEW_OBJECT) {
107 err = dnode_hold(os, object, tx, &dn);
113 if (err == 0 && tx->tx_txg != 0) {
114 mutex_enter(&dn->dn_mtx);
116 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
117 * problem, but there's no way for it to happen (for
120 ASSERT(dn->dn_assigned_txg == 0);
121 dn->dn_assigned_txg = tx->tx_txg;
122 (void) refcount_add(&dn->dn_tx_holds, tx);
123 mutex_exit(&dn->dn_mtx);
127 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
131 txh->txh_type = type;
132 txh->txh_arg1 = arg1;
133 txh->txh_arg2 = arg2;
135 list_insert_tail(&tx->tx_holds, txh);
141 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
144 * If we're syncing, they can manipulate any object anyhow, and
145 * the hold on the dnode_t can cause problems.
147 if (!dmu_tx_is_syncing(tx)) {
148 (void) dmu_tx_hold_object_impl(tx, os,
149 object, THT_NEWOBJECT, 0, 0);
154 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
159 rw_enter(&dn->dn_struct_rwlock, RW_READER);
160 db = dbuf_hold_level(dn, level, blkid, FTAG);
161 rw_exit(&dn->dn_struct_rwlock);
163 return (SET_ERROR(EIO));
164 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
170 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
171 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
173 objset_t *os = dn->dn_objset;
174 dsl_dataset_t *ds = os->os_dsl_dataset;
175 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
176 dmu_buf_impl_t *parent = NULL;
180 if (level >= dn->dn_nlevels || history[level] == blkid)
183 history[level] = blkid;
185 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
187 if (db == NULL || db == dn->dn_dbuf) {
191 ASSERT(DB_DNODE(db) == dn);
192 ASSERT(db->db_level == level);
193 ASSERT(db->db.db_size == space);
194 ASSERT(db->db_blkid == blkid);
196 parent = db->db_parent;
199 freeable = (bp && (freeable ||
200 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
203 txh->txh_space_tooverwrite += space;
205 txh->txh_space_towrite += space;
207 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
209 dmu_tx_count_twig(txh, dn, parent, level + 1,
210 blkid >> epbs, freeable, history);
215 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
217 dnode_t *dn = txh->txh_dnode;
218 uint64_t start, end, i;
219 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
225 min_bs = SPA_MINBLOCKSHIFT;
226 max_bs = SPA_MAXBLOCKSHIFT;
227 min_ibs = DN_MIN_INDBLKSHIFT;
228 max_ibs = DN_MAX_INDBLKSHIFT;
231 uint64_t history[DN_MAX_LEVELS];
232 int nlvls = dn->dn_nlevels;
236 * For i/o error checking, read the first and last level-0
237 * blocks (if they are not aligned), and all the level-1 blocks.
239 if (dn->dn_maxblkid == 0) {
240 delta = dn->dn_datablksz;
241 start = (off < dn->dn_datablksz) ? 0 : 1;
242 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
243 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
244 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
250 zio_t *zio = zio_root(dn->dn_objset->os_spa,
251 NULL, NULL, ZIO_FLAG_CANFAIL);
253 /* first level-0 block */
254 start = off >> dn->dn_datablkshift;
255 if (P2PHASE(off, dn->dn_datablksz) ||
256 len < dn->dn_datablksz) {
257 err = dmu_tx_check_ioerr(zio, dn, 0, start);
262 /* last level-0 block */
263 end = (off+len-1) >> dn->dn_datablkshift;
264 if (end != start && end <= dn->dn_maxblkid &&
265 P2PHASE(off+len, dn->dn_datablksz)) {
266 err = dmu_tx_check_ioerr(zio, dn, 0, end);
273 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
274 for (i = (start>>shft)+1; i < end>>shft; i++) {
275 err = dmu_tx_check_ioerr(zio, dn, 1, i);
284 delta = P2NPHASE(off, dn->dn_datablksz);
287 min_ibs = max_ibs = dn->dn_indblkshift;
288 if (dn->dn_maxblkid > 0) {
290 * The blocksize can't change,
291 * so we can make a more precise estimate.
293 ASSERT(dn->dn_datablkshift != 0);
294 min_bs = max_bs = dn->dn_datablkshift;
298 * If this write is not off the end of the file
299 * we need to account for overwrites/unref.
301 if (start <= dn->dn_maxblkid) {
302 for (int l = 0; l < DN_MAX_LEVELS; l++)
305 while (start <= dn->dn_maxblkid) {
308 rw_enter(&dn->dn_struct_rwlock, RW_READER);
309 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
310 rw_exit(&dn->dn_struct_rwlock);
313 txh->txh_tx->tx_err = err;
317 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
322 * Account for new indirects appearing
323 * before this IO gets assigned into a txg.
326 epbs = min_ibs - SPA_BLKPTRSHIFT;
327 for (bits -= epbs * (nlvls - 1);
328 bits >= 0; bits -= epbs)
329 txh->txh_fudge += 1ULL << max_ibs;
335 delta = dn->dn_datablksz;
340 * 'end' is the last thing we will access, not one past.
341 * This way we won't overflow when accessing the last byte.
343 start = P2ALIGN(off, 1ULL << max_bs);
344 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
345 txh->txh_space_towrite += end - start + 1;
350 epbs = min_ibs - SPA_BLKPTRSHIFT;
353 * The object contains at most 2^(64 - min_bs) blocks,
354 * and each indirect level maps 2^epbs.
356 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
359 ASSERT3U(end, >=, start);
360 txh->txh_space_towrite += (end - start + 1) << max_ibs;
363 * We also need a new blkid=0 indirect block
364 * to reference any existing file data.
366 txh->txh_space_towrite += 1ULL << max_ibs;
371 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 err = SET_ERROR(EFBIG);
376 txh->txh_tx->tx_err = err;
380 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
382 dnode_t *dn = txh->txh_dnode;
383 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
384 uint64_t space = mdn->dn_datablksz +
385 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
387 if (dn && dn->dn_dbuf->db_blkptr &&
388 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
389 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
390 txh->txh_space_tooverwrite += space;
391 txh->txh_space_tounref += space;
393 txh->txh_space_towrite += space;
394 if (dn && dn->dn_dbuf->db_blkptr)
395 txh->txh_space_tounref += space;
400 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
404 ASSERT(tx->tx_txg == 0);
405 ASSERT(len < DMU_MAX_ACCESS);
406 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
408 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
409 object, THT_WRITE, off, len);
413 dmu_tx_count_write(txh, off, len);
414 dmu_tx_count_dnode(txh);
418 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
420 uint64_t blkid, nblks, lastblk;
421 uint64_t space = 0, unref = 0, skipped = 0;
422 dnode_t *dn = txh->txh_dnode;
423 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
424 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
426 uint64_t l0span = 0, nl1blks = 0;
428 if (dn->dn_nlevels == 0)
432 * The struct_rwlock protects us against dn_nlevels
433 * changing, in case (against all odds) we manage to dirty &
434 * sync out the changes after we check for being dirty.
435 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
437 rw_enter(&dn->dn_struct_rwlock, RW_READER);
438 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
439 if (dn->dn_maxblkid == 0) {
440 if (off == 0 && len >= dn->dn_datablksz) {
444 rw_exit(&dn->dn_struct_rwlock);
448 blkid = off >> dn->dn_datablkshift;
449 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
451 if (blkid > dn->dn_maxblkid) {
452 rw_exit(&dn->dn_struct_rwlock);
455 if (blkid + nblks > dn->dn_maxblkid)
456 nblks = dn->dn_maxblkid - blkid + 1;
459 l0span = nblks; /* save for later use to calc level > 1 overhead */
460 if (dn->dn_nlevels == 1) {
462 for (i = 0; i < nblks; i++) {
463 blkptr_t *bp = dn->dn_phys->dn_blkptr;
464 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
466 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
467 dprintf_bp(bp, "can free old%s", "");
468 space += bp_get_dsize(spa, bp);
470 unref += BP_GET_ASIZE(bp);
476 lastblk = blkid + nblks - 1;
478 dmu_buf_impl_t *dbuf;
479 uint64_t ibyte, new_blkid;
481 int err, i, blkoff, tochk;
484 ibyte = blkid << dn->dn_datablkshift;
485 err = dnode_next_offset(dn,
486 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
487 new_blkid = ibyte >> dn->dn_datablkshift;
489 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
493 txh->txh_tx->tx_err = err;
496 if (new_blkid > lastblk) {
497 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
501 if (new_blkid > blkid) {
502 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
503 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
504 nblks -= new_blkid - blkid;
507 blkoff = P2PHASE(blkid, epb);
508 tochk = MIN(epb - blkoff, nblks);
510 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
512 txh->txh_tx->tx_err = err;
516 txh->txh_memory_tohold += dbuf->db.db_size;
519 * We don't check memory_tohold against DMU_MAX_ACCESS because
520 * memory_tohold is an over-estimation (especially the >L1
521 * indirect blocks), so it could fail. Callers should have
522 * already verified that they will not be holding too much
526 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
528 txh->txh_tx->tx_err = err;
529 dbuf_rele(dbuf, FTAG);
533 bp = dbuf->db.db_data;
536 for (i = 0; i < tochk; i++) {
537 if (dsl_dataset_block_freeable(ds, &bp[i],
539 dprintf_bp(&bp[i], "can free old%s", "");
540 space += bp_get_dsize(spa, &bp[i]);
542 unref += BP_GET_ASIZE(bp);
544 dbuf_rele(dbuf, FTAG);
550 rw_exit(&dn->dn_struct_rwlock);
553 * Add in memory requirements of higher-level indirects.
554 * This assumes a worst-possible scenario for dn_nlevels and a
555 * worst-possible distribution of l1-blocks over the region to free.
558 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
561 * Here we don't use DN_MAX_LEVEL, but calculate it with the
562 * given datablkshift and indblkshift. This makes the
563 * difference between 19 and 8 on large files.
565 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
566 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
568 while (level++ < maxlevel) {
569 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
570 << dn->dn_indblkshift;
571 blkcnt = 1 + (blkcnt >> epbs);
575 /* account for new level 1 indirect blocks that might show up */
577 txh->txh_fudge += skipped << dn->dn_indblkshift;
578 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
579 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
581 txh->txh_space_tofree += space;
582 txh->txh_space_tounref += unref;
586 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
593 ASSERT(tx->tx_txg == 0);
595 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
596 object, THT_FREE, off, len);
601 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
603 if (len == DMU_OBJECT_END)
604 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
606 dmu_tx_count_dnode(txh);
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.
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.
617 if (dn->dn_datablkshift == 0) {
618 if (off != 0 || len < dn->dn_datablksz)
619 dmu_tx_count_write(txh, 0, dn->dn_datablksz);
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);
630 * Check level-1 blocks.
632 if (dn->dn_nlevels > 1) {
633 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
635 uint64_t start = off >> shift;
636 uint64_t end = (off + len) >> shift;
638 ASSERT(dn->dn_indblkshift != 0);
641 * dnode_reallocate() can result in an object with indirect
642 * blocks having an odd data block size. In this case,
643 * just check the single block.
645 if (dn->dn_datablkshift == 0)
648 zio = zio_root(tx->tx_pool->dp_spa,
649 NULL, NULL, ZIO_FLAG_CANFAIL);
650 for (uint64_t i = start; i <= end; i++) {
651 uint64_t ibyte = i << shift;
652 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
661 err = dmu_tx_check_ioerr(zio, dn, 1, i);
674 dmu_tx_count_free(txh, off, len);
678 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
685 ASSERT(tx->tx_txg == 0);
687 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
688 object, THT_ZAP, add, (uintptr_t)name);
693 dmu_tx_count_dnode(txh);
697 * We will be able to fit a new object's entries into one leaf
698 * block. So there will be at most 2 blocks total,
699 * including the header block.
701 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
705 ASSERT3P(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
707 if (dn->dn_maxblkid == 0 && !add) {
711 * If there is only one block (i.e. this is a micro-zap)
712 * and we are not adding anything, the accounting is simple.
714 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
721 * Use max block size here, since we don't know how much
722 * the size will change between now and the dbuf dirty call.
724 bp = &dn->dn_phys->dn_blkptr[0];
725 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
727 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
729 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
731 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
735 if (dn->dn_maxblkid > 0 && name) {
737 * access the name in this fat-zap so that we'll check
738 * for i/o errors to the leaf blocks, etc.
740 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
748 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
749 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
752 * If the modified blocks are scattered to the four winds,
753 * we'll have to modify an indirect twig for each.
755 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
756 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
757 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
758 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
760 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
764 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
768 ASSERT(tx->tx_txg == 0);
770 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
771 object, THT_BONUS, 0, 0);
773 dmu_tx_count_dnode(txh);
777 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
780 ASSERT(tx->tx_txg == 0);
782 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
783 DMU_NEW_OBJECT, THT_SPACE, space, 0);
785 txh->txh_space_towrite += space;
789 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
795 * By asserting that the tx is assigned, we're counting the
796 * number of dn_tx_holds, which is the same as the number of
797 * dn_holds. Otherwise, we'd be counting dn_holds, but
798 * dn_tx_holds could be 0.
800 ASSERT(tx->tx_txg != 0);
802 /* if (tx->tx_anyobj == TRUE) */
805 for (txh = list_head(&tx->tx_holds); txh;
806 txh = list_next(&tx->tx_holds, txh)) {
807 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
816 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
819 int match_object = FALSE, match_offset = FALSE;
824 ASSERT(tx->tx_txg != 0);
825 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
826 ASSERT3U(dn->dn_object, ==, db->db.db_object);
833 /* XXX No checking on the meta dnode for now */
834 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
839 for (txh = list_head(&tx->tx_holds); txh;
840 txh = list_next(&tx->tx_holds, txh)) {
841 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
842 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
844 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
845 int datablkshift = dn->dn_datablkshift ?
846 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
847 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
848 int shift = datablkshift + epbs * db->db_level;
849 uint64_t beginblk = shift >= 64 ? 0 :
850 (txh->txh_arg1 >> shift);
851 uint64_t endblk = shift >= 64 ? 0 :
852 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
853 uint64_t blkid = db->db_blkid;
855 /* XXX txh_arg2 better not be zero... */
857 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
858 txh->txh_type, beginblk, endblk);
860 switch (txh->txh_type) {
862 if (blkid >= beginblk && blkid <= endblk)
865 * We will let this hold work for the bonus
866 * or spill buffer so that we don't need to
867 * hold it when creating a new object.
869 if (blkid == DMU_BONUS_BLKID ||
870 blkid == DMU_SPILL_BLKID)
873 * They might have to increase nlevels,
874 * thus dirtying the new TLIBs. Or the
875 * might have to change the block size,
876 * thus dirying the new lvl=0 blk=0.
883 * We will dirty all the level 1 blocks in
884 * the free range and perhaps the first and
885 * last level 0 block.
887 if (blkid >= beginblk && (blkid <= endblk ||
888 txh->txh_arg2 == DMU_OBJECT_END))
892 if (blkid == DMU_SPILL_BLKID)
896 if (blkid == DMU_BONUS_BLKID)
906 ASSERT(!"bad txh_type");
909 if (match_object && match_offset) {
915 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
916 (u_longlong_t)db->db.db_object, db->db_level,
917 (u_longlong_t)db->db_blkid);
922 dmu_tx_try_assign(dmu_tx_t *tx, txg_how_t txg_how)
925 spa_t *spa = tx->tx_pool->dp_spa;
926 uint64_t memory, asize, fsize, usize;
927 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
934 if (spa_suspended(spa)) {
936 * If the user has indicated a blocking failure mode
937 * then return ERESTART which will block in dmu_tx_wait().
938 * Otherwise, return EIO so that an error can get
939 * propagated back to the VOP calls.
941 * Note that we always honor the txg_how flag regardless
942 * of the failuremode setting.
944 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
946 return (SET_ERROR(EIO));
948 return (SET_ERROR(ERESTART));
951 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
952 tx->tx_needassign_txh = NULL;
955 * NB: No error returns are allowed after txg_hold_open, but
956 * before processing the dnode holds, due to the
957 * dmu_tx_unassign() logic.
960 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
961 for (txh = list_head(&tx->tx_holds); txh;
962 txh = list_next(&tx->tx_holds, txh)) {
963 dnode_t *dn = txh->txh_dnode;
965 mutex_enter(&dn->dn_mtx);
966 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
967 mutex_exit(&dn->dn_mtx);
968 tx->tx_needassign_txh = txh;
969 return (SET_ERROR(ERESTART));
971 if (dn->dn_assigned_txg == 0)
972 dn->dn_assigned_txg = tx->tx_txg;
973 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
974 (void) refcount_add(&dn->dn_tx_holds, tx);
975 mutex_exit(&dn->dn_mtx);
977 towrite += txh->txh_space_towrite;
978 tofree += txh->txh_space_tofree;
979 tooverwrite += txh->txh_space_tooverwrite;
980 tounref += txh->txh_space_tounref;
981 tohold += txh->txh_memory_tohold;
982 fudge += txh->txh_fudge;
986 * If a snapshot has been taken since we made our estimates,
987 * assume that we won't be able to free or overwrite anything.
990 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
991 tx->tx_lastsnap_txg) {
992 towrite += tooverwrite;
993 tooverwrite = tofree = 0;
996 /* needed allocation: worst-case estimate of write space */
997 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
998 /* freed space estimate: worst-case overwrite + free estimate */
999 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1000 /* convert unrefd space to worst-case estimate */
1001 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1002 /* calculate memory footprint estimate */
1003 memory = towrite + tooverwrite + tohold;
1007 * Add in 'tohold' to account for our dirty holds on this memory
1008 * XXX - the "fudge" factor is to account for skipped blocks that
1009 * we missed because dnode_next_offset() misses in-core-only blocks.
1011 tx->tx_space_towrite = asize +
1012 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1013 tx->tx_space_tofree = tofree;
1014 tx->tx_space_tooverwrite = tooverwrite;
1015 tx->tx_space_tounref = tounref;
1018 if (tx->tx_dir && asize != 0) {
1019 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1020 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1029 dmu_tx_unassign(dmu_tx_t *tx)
1033 if (tx->tx_txg == 0)
1036 txg_rele_to_quiesce(&tx->tx_txgh);
1039 * Walk the transaction's hold list, removing the hold on the
1040 * associated dnode, and notifying waiters if the refcount drops to 0.
1042 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1043 txh = list_next(&tx->tx_holds, txh)) {
1044 dnode_t *dn = txh->txh_dnode;
1048 mutex_enter(&dn->dn_mtx);
1049 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1051 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1052 dn->dn_assigned_txg = 0;
1053 cv_broadcast(&dn->dn_notxholds);
1055 mutex_exit(&dn->dn_mtx);
1058 txg_rele_to_sync(&tx->tx_txgh);
1060 tx->tx_lasttried_txg = tx->tx_txg;
1065 * Assign tx to a transaction group. txg_how can be one of:
1067 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1068 * a new one. This should be used when you're not holding locks.
1069 * It will only fail if we're truly out of space (or over quota).
1071 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1072 * blocking, returns immediately with ERESTART. This should be used
1073 * whenever you're holding locks. On an ERESTART error, the caller
1074 * should drop locks, do a dmu_tx_wait(tx), and try again.
1077 dmu_tx_assign(dmu_tx_t *tx, txg_how_t txg_how)
1081 ASSERT(tx->tx_txg == 0);
1082 ASSERT(txg_how == TXG_WAIT || txg_how == TXG_NOWAIT);
1083 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1085 /* If we might wait, we must not hold the config lock. */
1086 ASSERT(txg_how != TXG_WAIT || !dsl_pool_config_held(tx->tx_pool));
1088 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1089 dmu_tx_unassign(tx);
1091 if (err != ERESTART || txg_how != TXG_WAIT)
1097 txg_rele_to_quiesce(&tx->tx_txgh);
1103 dmu_tx_wait(dmu_tx_t *tx)
1105 spa_t *spa = tx->tx_pool->dp_spa;
1107 ASSERT(tx->tx_txg == 0);
1108 ASSERT(!dsl_pool_config_held(tx->tx_pool));
1111 * It's possible that the pool has become active after this thread
1112 * has tried to obtain a tx. If that's the case then his
1113 * tx_lasttried_txg would not have been assigned.
1115 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1116 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1117 } else if (tx->tx_needassign_txh) {
1118 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1120 mutex_enter(&dn->dn_mtx);
1121 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1122 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1123 mutex_exit(&dn->dn_mtx);
1124 tx->tx_needassign_txh = NULL;
1126 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1131 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1134 if (tx->tx_dir == NULL || delta == 0)
1138 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1139 tx->tx_space_towrite);
1140 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1142 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1148 dmu_tx_commit(dmu_tx_t *tx)
1152 ASSERT(tx->tx_txg != 0);
1155 * Go through the transaction's hold list and remove holds on
1156 * associated dnodes, notifying waiters if no holds remain.
1158 while (txh = list_head(&tx->tx_holds)) {
1159 dnode_t *dn = txh->txh_dnode;
1161 list_remove(&tx->tx_holds, txh);
1162 kmem_free(txh, sizeof (dmu_tx_hold_t));
1165 mutex_enter(&dn->dn_mtx);
1166 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1168 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1169 dn->dn_assigned_txg = 0;
1170 cv_broadcast(&dn->dn_notxholds);
1172 mutex_exit(&dn->dn_mtx);
1176 if (tx->tx_tempreserve_cookie)
1177 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1179 if (!list_is_empty(&tx->tx_callbacks))
1180 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1182 if (tx->tx_anyobj == FALSE)
1183 txg_rele_to_sync(&tx->tx_txgh);
1185 list_destroy(&tx->tx_callbacks);
1186 list_destroy(&tx->tx_holds);
1188 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1189 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1190 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1191 refcount_destroy_many(&tx->tx_space_written,
1192 refcount_count(&tx->tx_space_written));
1193 refcount_destroy_many(&tx->tx_space_freed,
1194 refcount_count(&tx->tx_space_freed));
1196 kmem_free(tx, sizeof (dmu_tx_t));
1200 dmu_tx_abort(dmu_tx_t *tx)
1204 ASSERT(tx->tx_txg == 0);
1206 while (txh = list_head(&tx->tx_holds)) {
1207 dnode_t *dn = txh->txh_dnode;
1209 list_remove(&tx->tx_holds, txh);
1210 kmem_free(txh, sizeof (dmu_tx_hold_t));
1216 * Call any registered callbacks with an error code.
1218 if (!list_is_empty(&tx->tx_callbacks))
1219 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1221 list_destroy(&tx->tx_callbacks);
1222 list_destroy(&tx->tx_holds);
1224 refcount_destroy_many(&tx->tx_space_written,
1225 refcount_count(&tx->tx_space_written));
1226 refcount_destroy_many(&tx->tx_space_freed,
1227 refcount_count(&tx->tx_space_freed));
1229 kmem_free(tx, sizeof (dmu_tx_t));
1233 dmu_tx_get_txg(dmu_tx_t *tx)
1235 ASSERT(tx->tx_txg != 0);
1236 return (tx->tx_txg);
1240 dmu_tx_pool(dmu_tx_t *tx)
1242 ASSERT(tx->tx_pool != NULL);
1243 return (tx->tx_pool);
1248 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1250 dmu_tx_callback_t *dcb;
1252 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1254 dcb->dcb_func = func;
1255 dcb->dcb_data = data;
1257 list_insert_tail(&tx->tx_callbacks, dcb);
1261 * Call all the commit callbacks on a list, with a given error code.
1264 dmu_tx_do_callbacks(list_t *cb_list, int error)
1266 dmu_tx_callback_t *dcb;
1268 while (dcb = list_head(cb_list)) {
1269 list_remove(cb_list, dcb);
1270 dcb->dcb_func(dcb->dcb_data, error);
1271 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1276 * Interface to hold a bunch of attributes.
1277 * used for creating new files.
1278 * attrsize is the total size of all attributes
1279 * to be added during object creation
1281 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1285 * hold necessary attribute name for attribute registration.
1286 * should be a very rare case where this is needed. If it does
1287 * happen it would only happen on the first write to the file system.
1290 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1294 if (!sa->sa_need_attr_registration)
1297 for (i = 0; i != sa->sa_num_attrs; i++) {
1298 if (!sa->sa_attr_table[i].sa_registered) {
1299 if (sa->sa_reg_attr_obj)
1300 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1301 B_TRUE, sa->sa_attr_table[i].sa_name);
1303 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1304 B_TRUE, sa->sa_attr_table[i].sa_name);
1311 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1316 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1319 dn = txh->txh_dnode;
1324 /* If blkptr doesn't exist then add space to towrite */
1325 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1326 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1330 bp = &dn->dn_phys->dn_spill;
1331 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1333 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1335 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1336 if (!BP_IS_HOLE(bp))
1337 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1342 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1344 sa_os_t *sa = tx->tx_objset->os_sa;
1346 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1348 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1351 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1352 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1354 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1355 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1356 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1357 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1360 dmu_tx_sa_registration_hold(sa, tx);
1362 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1365 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1372 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1374 * variable_size is the total size of all variable sized attributes
1375 * passed to this function. It is not the total size of all
1376 * variable size attributes that *may* exist on this object.
1379 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1382 sa_os_t *sa = tx->tx_objset->os_sa;
1384 ASSERT(hdl != NULL);
1386 object = sa_handle_object(hdl);
1388 dmu_tx_hold_bonus(tx, object);
1390 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1393 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1394 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1395 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1396 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1397 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1398 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1401 dmu_tx_sa_registration_hold(sa, tx);
1403 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1404 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1406 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1407 ASSERT(tx->tx_txg == 0);
1408 dmu_tx_hold_spill(tx, object);
1410 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1415 if (dn->dn_have_spill) {
1416 ASSERT(tx->tx_txg == 0);
1417 dmu_tx_hold_spill(tx, object);