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.
26 #include <sys/dmu_impl.h>
28 #include <sys/dmu_tx.h>
29 #include <sys/dmu_objset.h>
30 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
31 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
32 #include <sys/dsl_pool.h>
33 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/sa_impl.h>
37 #include <sys/zfs_context.h>
38 #include <sys/varargs.h>
40 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
41 uint64_t arg1, uint64_t arg2);
45 dmu_tx_create_dd(dsl_dir_t *dd)
47 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
50 tx->tx_pool = dd->dd_pool;
51 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
52 offsetof(dmu_tx_hold_t, txh_node));
53 list_create(&tx->tx_callbacks, sizeof (dmu_tx_callback_t),
54 offsetof(dmu_tx_callback_t, dcb_node));
56 refcount_create(&tx->tx_space_written);
57 refcount_create(&tx->tx_space_freed);
63 dmu_tx_create(objset_t *os)
65 dmu_tx_t *tx = dmu_tx_create_dd(os->os_dsl_dataset->ds_dir);
67 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os_dsl_dataset);
72 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
74 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
76 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
85 dmu_tx_is_syncing(dmu_tx_t *tx)
87 return (tx->tx_anyobj);
91 dmu_tx_private_ok(dmu_tx_t *tx)
93 return (tx->tx_anyobj);
96 static dmu_tx_hold_t *
97 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
98 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
104 if (object != DMU_NEW_OBJECT) {
105 err = dnode_hold(os, object, tx, &dn);
111 if (err == 0 && tx->tx_txg != 0) {
112 mutex_enter(&dn->dn_mtx);
114 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
115 * problem, but there's no way for it to happen (for
118 ASSERT(dn->dn_assigned_txg == 0);
119 dn->dn_assigned_txg = tx->tx_txg;
120 (void) refcount_add(&dn->dn_tx_holds, tx);
121 mutex_exit(&dn->dn_mtx);
125 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
129 txh->txh_type = type;
130 txh->txh_arg1 = arg1;
131 txh->txh_arg2 = arg2;
133 list_insert_tail(&tx->tx_holds, txh);
139 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
142 * If we're syncing, they can manipulate any object anyhow, and
143 * the hold on the dnode_t can cause problems.
145 if (!dmu_tx_is_syncing(tx)) {
146 (void) dmu_tx_hold_object_impl(tx, os,
147 object, THT_NEWOBJECT, 0, 0);
152 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
157 rw_enter(&dn->dn_struct_rwlock, RW_READER);
158 db = dbuf_hold_level(dn, level, blkid, FTAG);
159 rw_exit(&dn->dn_struct_rwlock);
162 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
168 dmu_tx_count_twig(dmu_tx_hold_t *txh, dnode_t *dn, dmu_buf_impl_t *db,
169 int level, uint64_t blkid, boolean_t freeable, uint64_t *history)
171 objset_t *os = dn->dn_objset;
172 dsl_dataset_t *ds = os->os_dsl_dataset;
173 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
174 dmu_buf_impl_t *parent = NULL;
178 if (level >= dn->dn_nlevels || history[level] == blkid)
181 history[level] = blkid;
183 space = (level == 0) ? dn->dn_datablksz : (1ULL << dn->dn_indblkshift);
185 if (db == NULL || db == dn->dn_dbuf) {
189 ASSERT(DB_DNODE(db) == dn);
190 ASSERT(db->db_level == level);
191 ASSERT(db->db.db_size == space);
192 ASSERT(db->db_blkid == blkid);
194 parent = db->db_parent;
197 freeable = (bp && (freeable ||
198 dsl_dataset_block_freeable(ds, bp, bp->blk_birth)));
201 txh->txh_space_tooverwrite += space;
203 txh->txh_space_towrite += space;
205 txh->txh_space_tounref += bp_get_dsize(os->os_spa, bp);
207 dmu_tx_count_twig(txh, dn, parent, level + 1,
208 blkid >> epbs, freeable, history);
213 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
215 dnode_t *dn = txh->txh_dnode;
216 uint64_t start, end, i;
217 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
224 min_bs = SPA_MINBLOCKSHIFT;
225 max_bs = SPA_MAXBLOCKSHIFT;
226 min_ibs = DN_MIN_INDBLKSHIFT;
227 max_ibs = DN_MAX_INDBLKSHIFT;
230 uint64_t history[DN_MAX_LEVELS];
231 int nlvls = dn->dn_nlevels;
235 * For i/o error checking, read the first and last level-0
236 * blocks (if they are not aligned), and all the level-1 blocks.
238 if (dn->dn_maxblkid == 0) {
239 delta = dn->dn_datablksz;
240 start = (off < dn->dn_datablksz) ? 0 : 1;
241 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
242 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
243 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
249 zio_t *zio = zio_root(dn->dn_objset->os_spa,
250 NULL, NULL, ZIO_FLAG_CANFAIL);
252 /* first level-0 block */
253 start = off >> dn->dn_datablkshift;
254 if (P2PHASE(off, dn->dn_datablksz) ||
255 len < dn->dn_datablksz) {
256 err = dmu_tx_check_ioerr(zio, dn, 0, start);
261 /* last level-0 block */
262 end = (off+len-1) >> dn->dn_datablkshift;
263 if (end != start && end <= dn->dn_maxblkid &&
264 P2PHASE(off+len, dn->dn_datablksz)) {
265 err = dmu_tx_check_ioerr(zio, dn, 0, end);
272 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
273 for (i = (start>>shft)+1; i < end>>shft; i++) {
274 err = dmu_tx_check_ioerr(zio, dn, 1, i);
283 delta = P2NPHASE(off, dn->dn_datablksz);
286 if (dn->dn_maxblkid > 0) {
288 * The blocksize can't change,
289 * so we can make a more precise estimate.
291 ASSERT(dn->dn_datablkshift != 0);
292 min_bs = max_bs = dn->dn_datablkshift;
293 min_ibs = max_ibs = dn->dn_indblkshift;
294 } else if (dn->dn_indblkshift > max_ibs) {
296 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
297 * the code will still work correctly on older pools.
299 min_ibs = max_ibs = dn->dn_indblkshift;
303 * If this write is not off the end of the file
304 * we need to account for overwrites/unref.
306 if (start <= dn->dn_maxblkid) {
307 for (l = 0; l < DN_MAX_LEVELS; l++)
310 while (start <= dn->dn_maxblkid) {
313 rw_enter(&dn->dn_struct_rwlock, RW_READER);
314 err = dbuf_hold_impl(dn, 0, start, FALSE, FTAG, &db);
315 rw_exit(&dn->dn_struct_rwlock);
318 txh->txh_tx->tx_err = err;
322 dmu_tx_count_twig(txh, dn, db, 0, start, B_FALSE,
327 * Account for new indirects appearing
328 * before this IO gets assigned into a txg.
331 epbs = min_ibs - SPA_BLKPTRSHIFT;
332 for (bits -= epbs * (nlvls - 1);
333 bits >= 0; bits -= epbs)
334 txh->txh_fudge += 1ULL << max_ibs;
340 delta = dn->dn_datablksz;
345 * 'end' is the last thing we will access, not one past.
346 * This way we won't overflow when accessing the last byte.
348 start = P2ALIGN(off, 1ULL << max_bs);
349 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
350 txh->txh_space_towrite += end - start + 1;
355 epbs = min_ibs - SPA_BLKPTRSHIFT;
358 * The object contains at most 2^(64 - min_bs) blocks,
359 * and each indirect level maps 2^epbs.
361 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
364 ASSERT3U(end, >=, start);
365 txh->txh_space_towrite += (end - start + 1) << max_ibs;
368 * We also need a new blkid=0 indirect block
369 * to reference any existing file data.
371 txh->txh_space_towrite += 1ULL << max_ibs;
376 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
381 txh->txh_tx->tx_err = err;
385 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
387 dnode_t *dn = txh->txh_dnode;
388 dnode_t *mdn = DMU_META_DNODE(txh->txh_tx->tx_objset);
389 uint64_t space = mdn->dn_datablksz +
390 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
392 if (dn && dn->dn_dbuf->db_blkptr &&
393 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
394 dn->dn_dbuf->db_blkptr, dn->dn_dbuf->db_blkptr->blk_birth)) {
395 txh->txh_space_tooverwrite += space;
396 txh->txh_space_tounref += space;
398 txh->txh_space_towrite += space;
399 if (dn && dn->dn_dbuf->db_blkptr)
400 txh->txh_space_tounref += space;
405 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
409 ASSERT(tx->tx_txg == 0);
410 ASSERT(len < DMU_MAX_ACCESS);
411 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
413 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
414 object, THT_WRITE, off, len);
418 dmu_tx_count_write(txh, off, len);
419 dmu_tx_count_dnode(txh);
423 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
425 uint64_t blkid, nblks, lastblk;
426 uint64_t space = 0, unref = 0, skipped = 0;
427 dnode_t *dn = txh->txh_dnode;
428 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
429 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
432 if (dn->dn_nlevels == 0)
436 * The struct_rwlock protects us against dn_nlevels
437 * changing, in case (against all odds) we manage to dirty &
438 * sync out the changes after we check for being dirty.
439 * Also, dbuf_hold_impl() wants us to have the struct_rwlock.
441 rw_enter(&dn->dn_struct_rwlock, RW_READER);
442 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
443 if (dn->dn_maxblkid == 0) {
444 if (off == 0 && len >= dn->dn_datablksz) {
448 rw_exit(&dn->dn_struct_rwlock);
452 blkid = off >> dn->dn_datablkshift;
453 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
455 if (blkid >= dn->dn_maxblkid) {
456 rw_exit(&dn->dn_struct_rwlock);
459 if (blkid + nblks > dn->dn_maxblkid)
460 nblks = dn->dn_maxblkid - blkid;
463 if (dn->dn_nlevels == 1) {
465 for (i = 0; i < nblks; i++) {
466 blkptr_t *bp = dn->dn_phys->dn_blkptr;
467 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
469 if (dsl_dataset_block_freeable(ds, bp, bp->blk_birth)) {
470 dprintf_bp(bp, "can free old%s", "");
471 space += bp_get_dsize(spa, bp);
473 unref += BP_GET_ASIZE(bp);
479 * Add in memory requirements of higher-level indirects.
480 * This assumes a worst-possible scenario for dn_nlevels.
483 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
484 int level = (dn->dn_nlevels > 1) ? 2 : 1;
486 while (level++ < DN_MAX_LEVELS) {
487 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
488 blkcnt = 1 + (blkcnt >> epbs);
490 ASSERT(blkcnt <= dn->dn_nblkptr);
493 lastblk = blkid + nblks - 1;
495 dmu_buf_impl_t *dbuf;
496 uint64_t ibyte, new_blkid;
498 int err, i, blkoff, tochk;
501 ibyte = blkid << dn->dn_datablkshift;
502 err = dnode_next_offset(dn,
503 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
504 new_blkid = ibyte >> dn->dn_datablkshift;
506 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
510 txh->txh_tx->tx_err = err;
513 if (new_blkid > lastblk) {
514 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
518 if (new_blkid > blkid) {
519 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
520 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
521 nblks -= new_blkid - blkid;
524 blkoff = P2PHASE(blkid, epb);
525 tochk = MIN(epb - blkoff, nblks);
527 err = dbuf_hold_impl(dn, 1, blkid >> epbs, FALSE, FTAG, &dbuf);
529 txh->txh_tx->tx_err = err;
533 txh->txh_memory_tohold += dbuf->db.db_size;
536 * We don't check memory_tohold against DMU_MAX_ACCESS because
537 * memory_tohold is an over-estimation (especially the >L1
538 * indirect blocks), so it could fail. Callers should have
539 * already verified that they will not be holding too much
543 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
545 txh->txh_tx->tx_err = err;
546 dbuf_rele(dbuf, FTAG);
550 bp = dbuf->db.db_data;
553 for (i = 0; i < tochk; i++) {
554 if (dsl_dataset_block_freeable(ds, &bp[i],
556 dprintf_bp(&bp[i], "can free old%s", "");
557 space += bp_get_dsize(spa, &bp[i]);
559 unref += BP_GET_ASIZE(bp);
561 dbuf_rele(dbuf, FTAG);
566 rw_exit(&dn->dn_struct_rwlock);
568 /* account for new level 1 indirect blocks that might show up */
570 txh->txh_fudge += skipped << dn->dn_indblkshift;
571 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
572 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
574 txh->txh_space_tofree += space;
575 txh->txh_space_tounref += unref;
579 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
583 uint64_t start, end, i;
587 ASSERT(tx->tx_txg == 0);
589 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
590 object, THT_FREE, off, len);
597 dmu_tx_count_write(txh, off, 1);
599 if (len != DMU_OBJECT_END)
600 dmu_tx_count_write(txh, off+len, 1);
602 dmu_tx_count_dnode(txh);
604 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
606 if (len == DMU_OBJECT_END)
607 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
610 * For i/o error checking, read the first and last level-0
611 * blocks, and all the level-1 blocks. The above count_write's
612 * have already taken care of the level-0 blocks.
614 if (dn->dn_nlevels > 1) {
615 shift = dn->dn_datablkshift + dn->dn_indblkshift -
617 start = off >> shift;
618 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
620 zio = zio_root(tx->tx_pool->dp_spa,
621 NULL, NULL, ZIO_FLAG_CANFAIL);
622 for (i = start; i <= end; i++) {
623 uint64_t ibyte = i << shift;
624 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
633 err = dmu_tx_check_ioerr(zio, dn, 1, i);
646 dmu_tx_count_free(txh, off, len);
650 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
657 ASSERT(tx->tx_txg == 0);
659 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
660 object, THT_ZAP, add, (uintptr_t)name);
665 dmu_tx_count_dnode(txh);
669 * We will be able to fit a new object's entries into one leaf
670 * block. So there will be at most 2 blocks total,
671 * including the header block.
673 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
677 ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
679 if (dn->dn_maxblkid == 0 && !add) {
681 * If there is only one block (i.e. this is a micro-zap)
682 * and we are not adding anything, the accounting is simple.
684 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
691 * Use max block size here, since we don't know how much
692 * the size will change between now and the dbuf dirty call.
694 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
695 &dn->dn_phys->dn_blkptr[0],
696 dn->dn_phys->dn_blkptr[0].blk_birth)) {
697 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
699 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
701 if (dn->dn_phys->dn_blkptr[0].blk_birth)
702 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
706 if (dn->dn_maxblkid > 0 && name) {
708 * access the name in this fat-zap so that we'll check
709 * for i/o errors to the leaf blocks, etc.
711 err = zap_lookup(dn->dn_objset, dn->dn_object, name,
719 err = zap_count_write(dn->dn_objset, dn->dn_object, name, add,
720 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
723 * If the modified blocks are scattered to the four winds,
724 * we'll have to modify an indirect twig for each.
726 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
727 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
728 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
729 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
731 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
735 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
739 ASSERT(tx->tx_txg == 0);
741 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
742 object, THT_BONUS, 0, 0);
744 dmu_tx_count_dnode(txh);
748 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
751 ASSERT(tx->tx_txg == 0);
753 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
754 DMU_NEW_OBJECT, THT_SPACE, space, 0);
756 txh->txh_space_towrite += space;
760 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
766 * By asserting that the tx is assigned, we're counting the
767 * number of dn_tx_holds, which is the same as the number of
768 * dn_holds. Otherwise, we'd be counting dn_holds, but
769 * dn_tx_holds could be 0.
771 ASSERT(tx->tx_txg != 0);
773 /* if (tx->tx_anyobj == TRUE) */
776 for (txh = list_head(&tx->tx_holds); txh;
777 txh = list_next(&tx->tx_holds, txh)) {
778 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
787 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
790 int match_object = FALSE, match_offset = FALSE;
795 ASSERT(tx->tx_txg != 0);
796 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset);
797 ASSERT3U(dn->dn_object, ==, db->db.db_object);
804 /* XXX No checking on the meta dnode for now */
805 if (db->db.db_object == DMU_META_DNODE_OBJECT) {
810 for (txh = list_head(&tx->tx_holds); txh;
811 txh = list_next(&tx->tx_holds, txh)) {
812 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
813 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
815 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
816 int datablkshift = dn->dn_datablkshift ?
817 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
818 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
819 int shift = datablkshift + epbs * db->db_level;
820 uint64_t beginblk = shift >= 64 ? 0 :
821 (txh->txh_arg1 >> shift);
822 uint64_t endblk = shift >= 64 ? 0 :
823 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
824 uint64_t blkid = db->db_blkid;
826 /* XXX txh_arg2 better not be zero... */
828 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
829 txh->txh_type, beginblk, endblk);
831 switch (txh->txh_type) {
833 if (blkid >= beginblk && blkid <= endblk)
836 * We will let this hold work for the bonus
837 * or spill buffer so that we don't need to
838 * hold it when creating a new object.
840 if (blkid == DMU_BONUS_BLKID ||
841 blkid == DMU_SPILL_BLKID)
844 * They might have to increase nlevels,
845 * thus dirtying the new TLIBs. Or the
846 * might have to change the block size,
847 * thus dirying the new lvl=0 blk=0.
854 * We will dirty all the level 1 blocks in
855 * the free range and perhaps the first and
856 * last level 0 block.
858 if (blkid >= beginblk && (blkid <= endblk ||
859 txh->txh_arg2 == DMU_OBJECT_END))
863 if (blkid == DMU_SPILL_BLKID)
867 if (blkid == DMU_BONUS_BLKID)
877 ASSERT(!"bad txh_type");
880 if (match_object && match_offset) {
886 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
887 (u_longlong_t)db->db.db_object, db->db_level,
888 (u_longlong_t)db->db_blkid);
893 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
896 spa_t *spa = tx->tx_pool->dp_spa;
897 uint64_t memory, asize, fsize, usize;
898 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
900 ASSERT3U(tx->tx_txg, ==, 0);
905 if (spa_suspended(spa)) {
907 * If the user has indicated a blocking failure mode
908 * then return ERESTART which will block in dmu_tx_wait().
909 * Otherwise, return EIO so that an error can get
910 * propagated back to the VOP calls.
912 * Note that we always honor the txg_how flag regardless
913 * of the failuremode setting.
915 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
922 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
923 tx->tx_needassign_txh = NULL;
926 * NB: No error returns are allowed after txg_hold_open, but
927 * before processing the dnode holds, due to the
928 * dmu_tx_unassign() logic.
931 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
932 for (txh = list_head(&tx->tx_holds); txh;
933 txh = list_next(&tx->tx_holds, txh)) {
934 dnode_t *dn = txh->txh_dnode;
936 mutex_enter(&dn->dn_mtx);
937 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
938 mutex_exit(&dn->dn_mtx);
939 tx->tx_needassign_txh = txh;
942 if (dn->dn_assigned_txg == 0)
943 dn->dn_assigned_txg = tx->tx_txg;
944 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
945 (void) refcount_add(&dn->dn_tx_holds, tx);
946 mutex_exit(&dn->dn_mtx);
948 towrite += txh->txh_space_towrite;
949 tofree += txh->txh_space_tofree;
950 tooverwrite += txh->txh_space_tooverwrite;
951 tounref += txh->txh_space_tounref;
952 tohold += txh->txh_memory_tohold;
953 fudge += txh->txh_fudge;
957 * NB: This check must be after we've held the dnodes, so that
958 * the dmu_tx_unassign() logic will work properly
960 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
964 * If a snapshot has been taken since we made our estimates,
965 * assume that we won't be able to free or overwrite anything.
968 dsl_dataset_prev_snap_txg(tx->tx_objset->os_dsl_dataset) >
969 tx->tx_lastsnap_txg) {
970 towrite += tooverwrite;
971 tooverwrite = tofree = 0;
974 /* needed allocation: worst-case estimate of write space */
975 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
976 /* freed space estimate: worst-case overwrite + free estimate */
977 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
978 /* convert unrefd space to worst-case estimate */
979 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
980 /* calculate memory footprint estimate */
981 memory = towrite + tooverwrite + tohold;
985 * Add in 'tohold' to account for our dirty holds on this memory
986 * XXX - the "fudge" factor is to account for skipped blocks that
987 * we missed because dnode_next_offset() misses in-core-only blocks.
989 tx->tx_space_towrite = asize +
990 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
991 tx->tx_space_tofree = tofree;
992 tx->tx_space_tooverwrite = tooverwrite;
993 tx->tx_space_tounref = tounref;
996 if (tx->tx_dir && asize != 0) {
997 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
998 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
1007 dmu_tx_unassign(dmu_tx_t *tx)
1011 if (tx->tx_txg == 0)
1014 txg_rele_to_quiesce(&tx->tx_txgh);
1016 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
1017 txh = list_next(&tx->tx_holds, txh)) {
1018 dnode_t *dn = txh->txh_dnode;
1022 mutex_enter(&dn->dn_mtx);
1023 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1025 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1026 dn->dn_assigned_txg = 0;
1027 cv_broadcast(&dn->dn_notxholds);
1029 mutex_exit(&dn->dn_mtx);
1032 txg_rele_to_sync(&tx->tx_txgh);
1034 tx->tx_lasttried_txg = tx->tx_txg;
1039 * Assign tx to a transaction group. txg_how can be one of:
1041 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1042 * a new one. This should be used when you're not holding locks.
1043 * If will only fail if we're truly out of space (or over quota).
1045 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1046 * blocking, returns immediately with ERESTART. This should be used
1047 * whenever you're holding locks. On an ERESTART error, the caller
1048 * should drop locks, do a dmu_tx_wait(tx), and try again.
1050 * (3) A specific txg. Use this if you need to ensure that multiple
1051 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1052 * returns ERESTART if it can't assign you into the requested txg.
1055 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1059 ASSERT(tx->tx_txg == 0);
1060 ASSERT(txg_how != 0);
1061 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1063 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1064 dmu_tx_unassign(tx);
1066 if (err != ERESTART || txg_how != TXG_WAIT)
1072 txg_rele_to_quiesce(&tx->tx_txgh);
1078 dmu_tx_wait(dmu_tx_t *tx)
1080 spa_t *spa = tx->tx_pool->dp_spa;
1082 ASSERT(tx->tx_txg == 0);
1085 * It's possible that the pool has become active after this thread
1086 * has tried to obtain a tx. If that's the case then his
1087 * tx_lasttried_txg would not have been assigned.
1089 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1090 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1091 } else if (tx->tx_needassign_txh) {
1092 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1094 mutex_enter(&dn->dn_mtx);
1095 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1096 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1097 mutex_exit(&dn->dn_mtx);
1098 tx->tx_needassign_txh = NULL;
1100 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1105 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1108 if (tx->tx_dir == NULL || delta == 0)
1112 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1113 tx->tx_space_towrite);
1114 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1116 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1122 dmu_tx_commit(dmu_tx_t *tx)
1126 ASSERT(tx->tx_txg != 0);
1128 while ((txh = list_head(&tx->tx_holds))) {
1129 dnode_t *dn = txh->txh_dnode;
1131 list_remove(&tx->tx_holds, txh);
1132 kmem_free(txh, sizeof (dmu_tx_hold_t));
1135 mutex_enter(&dn->dn_mtx);
1136 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1138 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1139 dn->dn_assigned_txg = 0;
1140 cv_broadcast(&dn->dn_notxholds);
1142 mutex_exit(&dn->dn_mtx);
1146 if (tx->tx_tempreserve_cookie)
1147 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1149 if (!list_is_empty(&tx->tx_callbacks))
1150 txg_register_callbacks(&tx->tx_txgh, &tx->tx_callbacks);
1152 if (tx->tx_anyobj == FALSE)
1153 txg_rele_to_sync(&tx->tx_txgh);
1155 list_destroy(&tx->tx_callbacks);
1156 list_destroy(&tx->tx_holds);
1158 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1159 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1160 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1161 refcount_destroy_many(&tx->tx_space_written,
1162 refcount_count(&tx->tx_space_written));
1163 refcount_destroy_many(&tx->tx_space_freed,
1164 refcount_count(&tx->tx_space_freed));
1166 kmem_free(tx, sizeof (dmu_tx_t));
1170 dmu_tx_abort(dmu_tx_t *tx)
1174 ASSERT(tx->tx_txg == 0);
1176 while ((txh = list_head(&tx->tx_holds))) {
1177 dnode_t *dn = txh->txh_dnode;
1179 list_remove(&tx->tx_holds, txh);
1180 kmem_free(txh, sizeof (dmu_tx_hold_t));
1186 * Call any registered callbacks with an error code.
1188 if (!list_is_empty(&tx->tx_callbacks))
1189 dmu_tx_do_callbacks(&tx->tx_callbacks, ECANCELED);
1191 list_destroy(&tx->tx_callbacks);
1192 list_destroy(&tx->tx_holds);
1194 refcount_destroy_many(&tx->tx_space_written,
1195 refcount_count(&tx->tx_space_written));
1196 refcount_destroy_many(&tx->tx_space_freed,
1197 refcount_count(&tx->tx_space_freed));
1199 kmem_free(tx, sizeof (dmu_tx_t));
1203 dmu_tx_get_txg(dmu_tx_t *tx)
1205 ASSERT(tx->tx_txg != 0);
1206 return (tx->tx_txg);
1210 dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *func, void *data)
1212 dmu_tx_callback_t *dcb;
1214 dcb = kmem_alloc(sizeof (dmu_tx_callback_t), KM_SLEEP);
1216 dcb->dcb_func = func;
1217 dcb->dcb_data = data;
1219 list_insert_tail(&tx->tx_callbacks, dcb);
1223 * Call all the commit callbacks on a list, with a given error code.
1226 dmu_tx_do_callbacks(list_t *cb_list, int error)
1228 dmu_tx_callback_t *dcb;
1230 while ((dcb = list_head(cb_list))) {
1231 list_remove(cb_list, dcb);
1232 dcb->dcb_func(dcb->dcb_data, error);
1233 kmem_free(dcb, sizeof (dmu_tx_callback_t));
1238 * Interface to hold a bunch of attributes.
1239 * used for creating new files.
1240 * attrsize is the total size of all attributes
1241 * to be added during object creation
1243 * For updating/adding a single attribute dmu_tx_hold_sa() should be used.
1247 * hold necessary attribute name for attribute registration.
1248 * should be a very rare case where this is needed. If it does
1249 * happen it would only happen on the first write to the file system.
1252 dmu_tx_sa_registration_hold(sa_os_t *sa, dmu_tx_t *tx)
1256 if (!sa->sa_need_attr_registration)
1259 for (i = 0; i != sa->sa_num_attrs; i++) {
1260 if (!sa->sa_attr_table[i].sa_registered) {
1261 if (sa->sa_reg_attr_obj)
1262 dmu_tx_hold_zap(tx, sa->sa_reg_attr_obj,
1263 B_TRUE, sa->sa_attr_table[i].sa_name);
1265 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT,
1266 B_TRUE, sa->sa_attr_table[i].sa_name);
1273 dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object)
1279 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset, object,
1282 dn = txh->txh_dnode;
1287 /* If blkptr doesn't exist then add space to towrite */
1288 bp = &dn->dn_phys->dn_spill;
1289 if (BP_IS_HOLE(bp)) {
1290 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1291 txh->txh_space_tounref = 0;
1293 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
1295 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
1297 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
1299 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
1304 dmu_tx_hold_sa_create(dmu_tx_t *tx, int attrsize)
1306 sa_os_t *sa = tx->tx_objset->os_sa;
1308 dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT);
1310 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1313 if (tx->tx_objset->os_sa->sa_layout_attr_obj)
1314 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1316 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_LAYOUTS);
1317 dmu_tx_hold_zap(tx, sa->sa_master_obj, B_TRUE, SA_REGISTRY);
1318 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1319 dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, B_TRUE, NULL);
1322 dmu_tx_sa_registration_hold(sa, tx);
1324 if (attrsize <= DN_MAX_BONUSLEN && !sa->sa_force_spill)
1327 (void) dmu_tx_hold_object_impl(tx, tx->tx_objset, DMU_NEW_OBJECT,
1334 * dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *, attribute, add, size)
1336 * variable_size is the total size of all variable sized attributes
1337 * passed to this function. It is not the total size of all
1338 * variable size attributes that *may* exist on this object.
1341 dmu_tx_hold_sa(dmu_tx_t *tx, sa_handle_t *hdl, boolean_t may_grow)
1344 sa_os_t *sa = tx->tx_objset->os_sa;
1346 ASSERT(hdl != NULL);
1348 object = sa_handle_object(hdl);
1350 dmu_tx_hold_bonus(tx, object);
1352 if (tx->tx_objset->os_sa->sa_master_obj == 0)
1355 if (tx->tx_objset->os_sa->sa_reg_attr_obj == 0 ||
1356 tx->tx_objset->os_sa->sa_layout_attr_obj == 0) {
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 (may_grow && tx->tx_objset->os_sa->sa_layout_attr_obj)
1366 dmu_tx_hold_zap(tx, sa->sa_layout_attr_obj, B_TRUE, NULL);
1368 if (sa->sa_force_spill || may_grow || hdl->sa_spill) {
1369 ASSERT(tx->tx_txg == 0);
1370 dmu_tx_hold_spill(tx, object);
1372 dmu_buf_impl_t *db = (dmu_buf_impl_t *)hdl->sa_bonus;
1377 if (dn->dn_have_spill) {
1378 ASSERT(tx->tx_txg == 0);
1379 dmu_tx_hold_spill(tx, object);