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) 2012 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);
45 dmu_tx_stats_t dmu_tx_stats = {
46 { "dmu_tx_assigned", KSTAT_DATA_UINT64 },
47 { "dmu_tx_delay", KSTAT_DATA_UINT64 },
48 { "dmu_tx_error", KSTAT_DATA_UINT64 },
49 { "dmu_tx_suspended", KSTAT_DATA_UINT64 },
50 { "dmu_tx_group", KSTAT_DATA_UINT64 },
51 { "dmu_tx_how", KSTAT_DATA_UINT64 },
52 { "dmu_tx_memory_reserve", KSTAT_DATA_UINT64 },
53 { "dmu_tx_memory_reclaim", KSTAT_DATA_UINT64 },
54 { "dmu_tx_memory_inflight", KSTAT_DATA_UINT64 },
55 { "dmu_tx_dirty_throttle", KSTAT_DATA_UINT64 },
56 { "dmu_tx_write_limit", KSTAT_DATA_UINT64 },
57 { "dmu_tx_quota", KSTAT_DATA_UINT64 },
60 static kstat_t *dmu_tx_ksp;
63 dmu_tx_create_dd(dsl_dir_t *dd)
65 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_PUSHPAGE);
68 tx->tx_pool = dd->dd_pool;
69 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
70 offsetof(dmu_tx_hold_t, txh_node));
71 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
72 offsetof(dmu_tx_callback_t, dcb_node));
74 refcount_create(&tx->tx_space_written);
75 refcount_create(&tx->tx_space_freed);
81 dmu_tx_create(objset_t *os)
83 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
85 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
90 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
92 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
94 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
103 dmu_tx_is_syncing(dmu_tx_t *tx)
105 return (tx->tx_anyobj);
109 dmu_tx_private_ok(dmu_tx_t *tx)
111 return (tx->tx_anyobj);
114 static dmu_tx_hold_t *
115 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
116 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
122 if (object != DMU_NEW_OBJECT) {
123 err = dnode_hold(os, object, tx, &dn);
129 if (err == 0 && tx->tx_txg != 0) {
130 mutex_enter(&dn->dn_mtx);
132 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
133 * problem, but there's no way for it to happen (for
136 ASSERT(dn->dn_assigned_txg == 0);
137 dn->dn_assigned_txg = tx->tx_txg;
138 (void) refcount_add(&dn->dn_tx_holds, tx);
139 mutex_exit(&dn->dn_mtx);
143 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_PUSHPAGE);
147 txh->txh_type = type;
148 txh->txh_arg1 = arg1;
149 txh->txh_arg2 = arg2;
151 list_insert_tail(&tx->tx_holds, txh);
157 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
160 * If we're syncing, they can manipulate any object anyhow, and
161 * the hold on the dnode_t can cause problems.
163 if (!dmu_tx_is_syncing(tx)) {
164 (void) dmu_tx_hold_object_impl(tx, os,
165 object, THT_NEWOBJECT, 0, 0);
170 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
175 rw_enter(&dn->dn_struct_rwlock, RW_READER);
176 db = dbuf_hold_level(dn, level, blkid, FTAG);
177 rw_exit(&dn->dn_struct_rwlock);
180 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
186 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
187 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
189 objset_t *os = dn->dn_objset;
190 dsl_dataset_t *ds = os->os_dsl_dataset;
191 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
192 dmu_buf_impl_t *parent = NULL;
196 if (level >= dn->dn_nlevels || history[level] == blkid)
199 history[level] = blkid;
201 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
203 if (db == NULL || db == dn->dn_dbuf) {
207 ASSERT(DB_DNODE(db) == dn);
208 ASSERT(db->db_level == level);
209 ASSERT(db->db.db_size == space);
210 ASSERT(db->db_blkid == blkid);
212 parent = db->db_parent;
215 freeable = (bp && (freeable ||
216 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
219 txh->txh_space_tooverwrite += space;
221 txh->txh_space_towrite += space;
223 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
225 dmu_tx_count_twig(txh, dn, parent, level + 1,
226 blkid >> epbs, freeable, history);
231 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
233 dnode_t *dn = txh->txh_dnode;
234 uint64_t start, end, i;
235 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
242 min_bs = SPA_MINBLOCKSHIFT;
243 max_bs = SPA_MAXBLOCKSHIFT;
244 min_ibs = DN_MIN_INDBLKSHIFT;
245 max_ibs = DN_MAX_INDBLKSHIFT;
248 uint64_t history[DN_MAX_LEVELS];
249 int nlvls = dn->dn_nlevels;
253 * For i/o error checking, read the first and last level-0
254 * blocks (if they are not aligned), and all the level-1 blocks.
256 if (dn->dn_maxblkid == 0) {
257 delta = dn->dn_datablksz;
258 start = (off < dn->dn_datablksz) ? 0 : 1;
259 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
260 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
261 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
267 zio_t *zio = zio_root(dn->dn_objset->os_spa,
268 NULL, NULL, ZIO_FLAG_CANFAIL);
270 /* first level-0 block */
271 start = off >> dn->dn_datablkshift;
272 if (P2PHASE(off, dn->dn_datablksz) ||
273 len < dn->dn_datablksz) {
274 err = dmu_tx_check_ioerr(zio, dn, 0, start);
279 /* last level-0 block */
280 end = (off+len-1) >> dn->dn_datablkshift;
281 if (end != start && end <= dn->dn_maxblkid &&
282 P2PHASE(off+len, dn->dn_datablksz)) {
283 err = dmu_tx_check_ioerr(zio, dn, 0, end);
290 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
291 for (i = (start>>shft)+1; i < end>>shft; i++) {
292 err = dmu_tx_check_ioerr(zio, dn, 1, i);
301 delta = P2NPHASE(off, dn->dn_datablksz);
304 if (dn->dn_maxblkid > 0) {
306 * The blocksize can't change,
307 * so we can make a more precise estimate.
309 ASSERT(dn->dn_datablkshift != 0);
310 min_bs = max_bs = dn->dn_datablkshift;
311 min_ibs = max_ibs = dn->dn_indblkshift;
312 } else if (dn->dn_indblkshift > max_ibs) {
314 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
315 * the code will still work correctly on older pools.
317 min_ibs = max_ibs = dn->dn_indblkshift;
321 * If this write is not off the end of the file
322 * we need to account for overwrites/unref.
324 if (start <= dn->dn_maxblkid) {
325 for (l = 0; l < DN_MAX_LEVELS; l++)
328 while (start <= dn->dn_maxblkid) {
331 rw_enter(&dn->dn_struct_rwlock, RW_READER);
332 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
333 rw_exit(&dn->dn_struct_rwlock);
336 txh->txh_tx->tx_err = err;
340 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
345 * Account for new indirects appearing
346 * before this IO gets assigned into a txg.
349 epbs = min_ibs - SPA_BLKPTRSHIFT;
350 for (bits -= epbs * (nlvls - 1);
351 bits >= 0; bits -= epbs)
352 txh->txh_fudge += 1ULL << max_ibs;
358 delta = dn->dn_datablksz;
363 * 'end' is the last thing we will access, not one past.
364 * This way we won't overflow when accessing the last byte.
366 start = P2ALIGN(off, 1ULL << max_bs);
367 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
368 txh->txh_space_towrite += end - start + 1;
373 epbs = min_ibs - SPA_BLKPTRSHIFT;
376 * The object contains at most 2^(64 - min_bs) blocks,
377 * and each indirect level maps 2^epbs.
379 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
382 ASSERT3U(end, >=, start);
383 txh->txh_space_towrite += (end - start + 1) << max_ibs;
386 * We also need a new blkid=0 indirect block
387 * to reference any existing file data.
389 txh->txh_space_towrite += 1ULL << max_ibs;
394 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
399 txh->txh_tx->tx_err = err;
403 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
405 dnode_t *dn = txh->txh_dnode;
406 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
407 uint64_t space = mdn->dn_datablksz +
408 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
410 if (dn && dn->dn_dbuf->db_blkptr &&
411 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
412 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
413 txh->txh_space_tooverwrite += space;
414 txh->txh_space_tounref += space;
416 txh->txh_space_towrite += space;
417 if (dn && dn->dn_dbuf->db_blkptr)
418 txh->txh_space_tounref += space;
423 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
427 ASSERT(tx->tx_txg == 0);
428 ASSERT(len < DMU_MAX_ACCESS);
429 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
431 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
432 object, THT_WRITE, off, len);
436 dmu_tx_count_write(txh, off, len);
437 dmu_tx_count_dnode(txh);
441 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
443 uint64_t blkid, nblks, lastblk;
444 uint64_t space = 0, unref = 0, skipped = 0;
445 dnode_t *dn = txh->txh_dnode;
446 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
447 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
449 uint64_t l0span = 0, nl1blks = 0;
451 if (dn->dn_nlevels == 0)
455 * The struct_rwlock protects us against dn_nlevels
456 * changing, in case (against all odds) we manage to dirty &
457 * sync out the changes after we check for being dirty.
458 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
460 rw_enter(&dn->dn_struct_rwlock, RW_READER);
461 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
462 if (dn->dn_maxblkid == 0) {
463 if (off == 0 && len >= dn->dn_datablksz) {
467 rw_exit(&dn->dn_struct_rwlock);
471 blkid = off >> dn->dn_datablkshift;
472 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
474 if (blkid >= dn->dn_maxblkid) {
475 rw_exit(&dn->dn_struct_rwlock);
478 if (blkid + nblks > dn->dn_maxblkid)
479 nblks = dn->dn_maxblkid - blkid;
482 l0span = nblks; /* save for later use to calc level > 1 overhead */
483 if (dn->dn_nlevels == 1) {
485 for (i = 0; i < nblks; i++) {
486 blkptr_t *bp = dn->dn_phys->dn_blkptr;
487 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
489 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
490 dprintf_bp(bp, "can free old%s", "");
491 space += bp_get_dsize(spa, bp);
493 unref += BP_GET_ASIZE(bp);
499 lastblk = blkid + nblks - 1;
501 dmu_buf_impl_t *dbuf;
502 uint64_t ibyte, new_blkid;
504 int err, i, blkoff, tochk;
507 ibyte = blkid << dn->dn_datablkshift;
508 err = dnode_next_offset(dn,
509 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
510 new_blkid = ibyte >> dn->dn_datablkshift;
512 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
516 txh->txh_tx->tx_err = err;
519 if (new_blkid > lastblk) {
520 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
524 if (new_blkid > blkid) {
525 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
526 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
527 nblks -= new_blkid - blkid;
530 blkoff = P2PHASE(blkid, epb);
531 tochk = MIN(epb - blkoff, nblks);
533 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
535 txh->txh_tx->tx_err = err;
539 txh->txh_memory_tohold += dbuf->db.db_size;
542 * We don't check memory_tohold against DMU_MAX_ACCESS because
543 * memory_tohold is an over-estimation (especially the >L1
544 * indirect blocks), so it could fail. Callers should have
545 * already verified that they will not be holding too much
549 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
551 txh->txh_tx->tx_err = err;
552 dbuf_rele(dbuf, FTAG);
556 bp = dbuf->db.db_data;
559 for (i = 0; i < tochk; i++) {
560 if (dsl_dataset_block_freeable(ds, &bp[i],
562 dprintf_bp(&bp[i], "can free old%s", "");
563 space += bp_get_dsize(spa, &bp[i]);
565 unref += BP_GET_ASIZE(bp);
567 dbuf_rele(dbuf, FTAG);
573 rw_exit(&dn->dn_struct_rwlock);
576 * Add in memory requirements of higher-level indirects.
577 * This assumes a worst-possible scenario for dn_nlevels and a
578 * worst-possible distribution of l1-blocks over the region to free.
581 uint64_t blkcnt = 1 + ((l0span >> epbs) >> epbs);
584 * Here we don't use DN_MAX_LEVEL, but calculate it with the
585 * given datablkshift and indblkshift. This makes the
586 * difference between 19 and 8 on large files.
588 int maxlevel = 2 + (DN_MAX_OFFSET_SHIFT - dn->dn_datablkshift) /
589 (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
591 while (level++ < maxlevel) {
592 txh->txh_memory_tohold += MAX(MIN(blkcnt, nl1blks), 1)
593 << dn->dn_indblkshift;
594 blkcnt = 1 + (blkcnt >> epbs);
598 /* account for new level 1 indirect blocks that might show up */
600 txh->txh_fudge += skipped << dn->dn_indblkshift;
601 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
602 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
604 txh->txh_space_tofree += space;
605 txh->txh_space_tounref += unref;
609 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
613 uint64_t start, end, i;
617 ASSERT(tx->tx_txg == 0);
619 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
620 object, THT_FREE, off, len);
627 dmu_tx_count_write(txh, off, 1);
629 if (len != DMU_OBJECT_END)
630 dmu_tx_count_write(txh, off+len, 1);
632 dmu_tx_count_dnode(txh);
634 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
636 if (len == DMU_OBJECT_END)
637 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
640 * For i/o error checking, read the first and last level-0
641 * blocks, and all the level-1 blocks. The above count_write's
642 * have already taken care of the level-0 blocks.
644 if (dn->dn_nlevels > 1) {
645 shift = dn->dn_datablkshift + dn->dn_indblkshift -
647 start = off >> shift;
648 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
650 zio = zio_root(tx->tx_pool->dp_spa,
651 NULL, NULL, ZIO_FLAG_CANFAIL);
652 for (i = start; i <= end; i++) {
653 uint64_t ibyte = i << shift;
654 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
663 err = dmu_tx_check_ioerr(zio, dn, 1, i);
676 dmu_tx_count_free(txh, off, len);
680 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
687 ASSERT(tx->tx_txg == 0);
689 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
690 object, THT_ZAP, add, (uintptr_t)name);
695 dmu_tx_count_dnode(txh);
699 * We will be able to fit a new object's entries into one leaf
700 * block. So there will be at most 2 blocks total,
701 * including the header block.
703 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
707 ASSERT3U(DMU_OT_BYTESWAP(dn->dn_type), ==, DMU_BSWAP_ZAP);
709 if (dn->dn_maxblkid == 0 && !add) {
713 * If there is only one block (i.e. this is a micro-zap)
714 * and we are not adding anything, the accounting is simple.
716 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
723 * Use max block size here, since we don't know how much
724 * the size will change between now and the dbuf dirty call.
726 bp = &dn->dn_phys->dn_blkptr[0];
727 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
729 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
731 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
733 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
737 if (dn->dn_maxblkid > 0 && name) {
739 * access the name in this fat-zap so that we'll check
740 * for i/o errors to the leaf blocks, etc.
742 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
750 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
751 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
754 * If the modified blocks are scattered to the four winds,
755 * we'll have to modify an indirect twig for each.
757 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
758 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
759 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
760 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
762 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
766 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
770 ASSERT(tx->tx_txg == 0);
772 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
773 object, THT_BONUS, 0, 0);
775 dmu_tx_count_dnode(txh);
779 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
782 ASSERT(tx->tx_txg == 0);
784 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
785 DMU_NEW_OBJECT, THT_SPACE, space, 0);
787 txh->txh_space_towrite += space;
791 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
797 * By asserting that the tx is assigned, we're counting the
798 * number of dn_tx_holds, which is the same as the number of
799 * dn_holds. Otherwise, we'd be counting dn_holds, but
800 * dn_tx_holds could be 0.
802 ASSERT(tx->tx_txg != 0);
804 /* if (tx->tx_anyobj == TRUE) */
807 for (txh = list_head(&tx->tx_holds); txh;
808 txh = list_next(&tx->tx_holds, txh)) {
809 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
818 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
821 int match_object = FALSE, match_offset = FALSE;
827 ASSERT(tx->tx_txg != 0);
828 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
829 ASSERT3U(dn->dn_object, ==, db->db.db_object);
836 /* XXX No checking on the meta dnode for now */
837 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
842 for (txh = list_head(&tx->tx_holds); txh;
843 txh = list_next(&tx->tx_holds, txh)) {
844 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
845 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
847 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
848 int datablkshift = dn->dn_datablkshift ?
849 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
850 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
851 int shift = datablkshift + epbs * db->db_level;
852 uint64_t beginblk = shift >= 64 ? 0 :
853 (txh->txh_arg1 >> shift);
854 uint64_t endblk = shift >= 64 ? 0 :
855 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
856 uint64_t blkid = db->db_blkid;
858 /* XXX txh_arg2 better not be zero... */
860 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
861 txh->txh_type, beginblk, endblk);
863 switch (txh->txh_type) {
865 if (blkid >= beginblk && blkid <= endblk)
868 * We will let this hold work for the bonus
869 * or spill buffer so that we don't need to
870 * hold it when creating a new object.
872 if (blkid == DMU_BONUS_BLKID ||
873 blkid == DMU_SPILL_BLKID)
876 * They might have to increase nlevels,
877 * thus dirtying the new TLIBs. Or the
878 * might have to change the block size,
879 * thus dirying the new lvl=0 blk=0.
886 * We will dirty all the level 1 blocks in
887 * the free range and perhaps the first and
888 * last level 0 block.
890 if (blkid >= beginblk && (blkid <= endblk ||
891 txh->txh_arg2 == DMU_OBJECT_END))
895 if (blkid == DMU_SPILL_BLKID)
899 if (blkid == DMU_BONUS_BLKID)
909 ASSERT(!"bad txh_type");
912 if (match_object && match_offset) {
918 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
919 (u_longlong_t)db->db.db_object, db->db_level,
920 (u_longlong_t)db->db_blkid);
925 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
928 spa_t *spa = tx->tx_pool->dp_spa;
929 uint64_t memory, asize, fsize, usize;
930 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
932 ASSERT3U(tx->tx_txg, ==, 0);
935 DMU_TX_STAT_BUMP(dmu_tx_error);
939 if (spa_suspended(spa)) {
940 DMU_TX_STAT_BUMP(dmu_tx_suspended);
943 * If the user has indicated a blocking failure mode
944 * then return ERESTART which will block in dmu_tx_wait().
945 * Otherwise, return EIO so that an error can get
946 * propagated back to the VOP calls.
948 * Note that we always honor the txg_how flag regardless
949 * of the failuremode setting.
951 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
958 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
959 tx->tx_needassign_txh = NULL;
962 * NB: No error returns are allowed after txg_hold_open, but
963 * before processing the dnode holds, due to the
964 * dmu_tx_unassign() logic.
967 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
968 for (txh = list_head(&tx->tx_holds); txh;
969 txh = list_next(&tx->tx_holds, txh)) {
970 dnode_t *dn = txh->txh_dnode;
972 mutex_enter(&dn->dn_mtx);
973 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
974 mutex_exit(&dn->dn_mtx);
975 tx->tx_needassign_txh = txh;
976 DMU_TX_STAT_BUMP(dmu_tx_group);
979 if (dn->dn_assigned_txg == 0)
980 dn->dn_assigned_txg = tx->tx_txg;
981 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
982 (void) refcount_add(&dn->dn_tx_holds, tx);
983 mutex_exit(&dn->dn_mtx);
985 towrite += txh->txh_space_towrite;
986 tofree += txh->txh_space_tofree;
987 tooverwrite += txh->txh_space_tooverwrite;
988 tounref += txh->txh_space_tounref;
989 tohold += txh->txh_memory_tohold;
990 fudge += txh->txh_fudge;
994 * NB: This check must be after we've held the dnodes, so that
995 * the dmu_tx_unassign() logic will work properly
997 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg) {
998 DMU_TX_STAT_BUMP(dmu_tx_how);
1003 * If a snapshot has been taken since we made our estimates,
1004 * assume that we won't be able to free or overwrite anything.
1006 if (tx->tx_objset &&
1007 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
1008 tx->tx_lastsnap_txg) {
1009 towrite += tooverwrite;
1010 tooverwrite = tofree = 0;
1013 /* needed allocation: worst-case estimate of write space */
1014 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
1015 /* freed space estimate: worst-case overwrite + free estimate */
1016 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
1017 /* convert unrefd space to worst-case estimate */
1018 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
1019 /* calculate memory footprint estimate */
1020 memory = towrite + tooverwrite + tohold;
1024 * Add in 'tohold' to account for our dirty holds on this memory
1025 * XXX - the "fudge" factor is to account for skipped blocks that
1026 * we missed because dnode_next_offset() misses in-core-only blocks.
1028 tx->tx_space_towrite = asize +
1029 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
1030 tx->tx_space_tofree = tofree;
1031 tx->tx_space_tooverwrite = tooverwrite;
1032 tx->tx_space_tounref = tounref;
1035 if (tx->tx_dir && asize != 0) {
1036 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
1037 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1042 DMU_TX_STAT_BUMP(dmu_tx_assigned);
1048 dmu_tx_unassign(dmu_tx_t *tx)
1052 if (tx->tx_txg == 0)
1055 txg_rele_to_quiesce(&tx->tx_txgh);
1057 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1058 txh = list_next(&tx->tx_holds, txh)) {
1059 dnode_t *dn = txh->txh_dnode;
1063 mutex_enter(&dn->dn_mtx);
1064 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1066 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1067 dn->dn_assigned_txg = 0;
1068 cv_broadcast(&dn->dn_notxholds);
1070 mutex_exit(&dn->dn_mtx);
1073 txg_rele_to_sync(&tx->tx_txgh);
1075 tx->tx_lasttried_txg = tx->tx_txg;
1080 * Assign tx to a transaction group. txg_how can be one of:
1082 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1083 * a new one. This should be used when you're not holding locks.
1084 * If will only fail if we're truly out of space (or over quota).
1086 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1087 * blocking, returns immediately with ERESTART. This should be used
1088 * whenever you're holding locks. On an ERESTART error, the caller
1089 * should drop locks, do a dmu_tx_wait(tx), and try again.
1091 * (3) A specific txg. Use this if you need to ensure that multiple
1092 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1093 * returns ERESTART if it can't assign you into the requested txg.
1096 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1100 ASSERT(tx->tx_txg == 0);
1101 ASSERT(txg_how != 0);
1102 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1104 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1105 dmu_tx_unassign(tx);
1107 if (err != ERESTART || txg_how != TXG_WAIT)
1113 txg_rele_to_quiesce(&tx->tx_txgh);
1119 dmu_tx_wait(dmu_tx_t *tx)
1121 spa_t *spa = tx->tx_pool->dp_spa;
1123 ASSERT(tx->tx_txg == 0);
1126 * It's possible that the pool has become active after this thread
1127 * has tried to obtain a tx. If that's the case then his
1128 * tx_lasttried_txg would not have been assigned.
1130 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1131 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1132 } else if (tx->tx_needassign_txh) {
1133 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1135 mutex_enter(&dn->dn_mtx);
1136 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1137 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1138 mutex_exit(&dn->dn_mtx);
1139 tx->tx_needassign_txh = NULL;
1141 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1146 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1149 if (tx->tx_dir == NULL || delta == 0)
1153 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1154 tx->tx_space_towrite);
1155 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1157 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1163 dmu_tx_commit(dmu_tx_t *tx)
1167 ASSERT(tx->tx_txg != 0);
1169 while ((txh = list_head(&tx->tx_holds))) {
1170 dnode_t *dn = txh->txh_dnode;
1172 list_remove(&tx->tx_holds, txh);
1173 kmem_free(txh, sizeof (dmu_tx_hold_t));
1176 mutex_enter(&dn->dn_mtx);
1177 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1179 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1180 dn->dn_assigned_txg = 0;
1181 cv_broadcast(&dn->dn_notxholds);
1183 mutex_exit(&dn->dn_mtx);
1187 if (tx->tx_tempreserve_cookie)
1188 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1190 if (!list_is_empty(&tx->tx_callbacks))
1191 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1193 if (tx->tx_anyobj == FALSE)
1194 txg_rele_to_sync(&tx->tx_txgh);
1196 list_destroy(&tx->tx_callbacks);
1197 list_destroy(&tx->tx_holds);
1199 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1200 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1201 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1202 refcount_destroy_many(&tx->tx_space_written,
1203 refcount_count(&tx->tx_space_written));
1204 refcount_destroy_many(&tx->tx_space_freed,
1205 refcount_count(&tx->tx_space_freed));
1207 kmem_free(tx, sizeof (dmu_tx_t));
1211 dmu_tx_abort(dmu_tx_t *tx)
1215 ASSERT(tx->tx_txg == 0);
1217 while ((txh = list_head(&tx->tx_holds))) {
1218 dnode_t *dn = txh->txh_dnode;
1220 list_remove(&tx->tx_holds, txh);
1221 kmem_free(txh, sizeof (dmu_tx_hold_t));
1227 * Call any registered callbacks with an error code.
1229 if (!list_is_empty(&tx->tx_callbacks))
1230 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1232 list_destroy(&tx->tx_callbacks);
1233 list_destroy(&tx->tx_holds);
1235 refcount_destroy_many(&tx->tx_space_written,
1236 refcount_count(&tx->tx_space_written));
1237 refcount_destroy_many(&tx->tx_space_freed,
1238 refcount_count(&tx->tx_space_freed));
1240 kmem_free(tx, sizeof (dmu_tx_t));
1244 dmu_tx_get_txg(dmu_tx_t *tx)
1246 ASSERT(tx->tx_txg != 0);
1247 return (tx->tx_txg);
1251 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1253 dmu_tx_callback_t *dcb;
1255 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_PUSHPAGE);
1257 dcb->dcb_func = func;
1258 dcb->dcb_data = data;
1260 list_insert_tail(&tx->tx_callbacks, dcb);
1264 * Call all the commit callbacks on a list, with a given error code.
1267 dmu_tx_do_callbacks(list_t *cb_list, int error)
1269 dmu_tx_callback_t *dcb;
1271 while ((dcb = list_head(cb_list))) {
1272 list_remove(cb_list, dcb);
1273 dcb->dcb_func(dcb->dcb_data, error);
1274 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1279 * Interface to hold a bunch of attributes.
1280 * used for creating new files.
1281 * attrsize is the total size of all attributes
1282 * to be added during object creation
1284 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1288 * hold necessary attribute name for attribute registration.
1289 * should be a very rare case where this is needed. If it does
1290 * happen it would only happen on the first write to the file system.
1293 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1297 if (!sa->sa_need_attr_registration)
1300 for (i = 0; i != sa->sa_num_attrs; i++) {
1301 if (!sa->sa_attr_table[i].sa_registered) {
1302 if (sa->sa_reg_attr_obj)
1303 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1304 B_TRUE, sa->sa_attr_table[i].sa_name);
1306 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1307 B_TRUE, sa->sa_attr_table[i].sa_name);
1314 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1319 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1322 dn = txh->txh_dnode;
1327 /* If blkptr doesn't exist then add space to towrite */
1328 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
1329 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1333 bp = &dn->dn_phys->dn_spill;
1334 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1336 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1338 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1339 if (!BP_IS_HOLE(bp))
1340 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1345 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1347 sa_os_t *sa = tx->tx_objset->os_sa;
1349 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1351 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1354 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1355 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1357 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1358 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1359 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1360 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1363 dmu_tx_sa_registration_hold(sa, tx);
1365 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1368 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1375 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1377 * variable_size is the total size of all variable sized attributes
1378 * passed to this function. It is not the total size of all
1379 * variable size attributes that *may* exist on this object.
1382 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1385 sa_os_t *sa = tx->tx_objset->os_sa;
1387 ASSERT(hdl != NULL);
1389 object = sa_handle_object(hdl);
1391 dmu_tx_hold_bonus(tx, object);
1393 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1396 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1397 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
1398 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1399 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1400 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1401 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1404 dmu_tx_sa_registration_hold(sa, tx);
1406 if (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1407 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1409 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1410 ASSERT(tx->tx_txg == 0);
1411 dmu_tx_hold_spill(tx, object);
1413 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1418 if (dn->dn_have_spill) {
1419 ASSERT(tx->tx_txg == 0);
1420 dmu_tx_hold_spill(tx, object);
1429 dmu_tx_ksp = kstat_create("zfs", 0, "dmu_tx", "misc",
1430 KSTAT_TYPE_NAMED, sizeof (dmu_tx_stats) / sizeof (kstat_named_t),
1431 KSTAT_FLAG_VIRTUAL);
1433 if (dmu_tx_ksp != NULL) {
1434 dmu_tx_ksp->ks_data = &dmu_tx_stats;
1435 kstat_install(dmu_tx_ksp);
1442 if (dmu_tx_ksp != NULL) {
1443 kstat_delete(dmu_tx_ksp);
1448 #if defined(_KERNEL) && defined(HAVE_SPL)
1449 EXPORT_SYMBOL(dmu_tx_create);
1450 EXPORT_SYMBOL(dmu_tx_hold_write);
1451 EXPORT_SYMBOL(dmu_tx_hold_free);
1452 EXPORT_SYMBOL(dmu_tx_hold_zap);
1453 EXPORT_SYMBOL(dmu_tx_hold_bonus);
1454 EXPORT_SYMBOL(dmu_tx_abort);
1455 EXPORT_SYMBOL(dmu_tx_assign);
1456 EXPORT_SYMBOL(dmu_tx_wait);
1457 EXPORT_SYMBOL(dmu_tx_commit);
1458 EXPORT_SYMBOL(dmu_tx_get_txg);
1459 EXPORT_SYMBOL(dmu_tx_callback_register);
1460 EXPORT_SYMBOL(dmu_tx_do_callbacks);
1461 EXPORT_SYMBOL(dmu_tx_hold_spill);
1462 EXPORT_SYMBOL(dmu_tx_hold_sa_create);
1463 EXPORT_SYMBOL(dmu_tx_hold_sa);