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 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
27 #include <sys/dmu_impl.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dataset.h> /* for dsl_dataset_block_freeable() */
32 #include <sys/dsl_dir.h> /* for dsl_dir_tempreserve_*() */
33 #include <sys/dsl_pool.h>
34 #include <sys/zap_impl.h> /* for fzap_default_block_shift */
36 #include <sys/zfs_context.h>
38 typedef void (*dmu_tx_hold_func_t)(dmu_tx_t *tx, struct dnode *dn,
39 uint64_t arg1, uint64_t arg2);
43 dmu_tx_create_dd(dsl_dir_t *dd)
45 dmu_tx_t *tx = kmem_zalloc(sizeof (dmu_tx_t), KM_SLEEP);
48 tx->tx_pool = dd->dd_pool;
49 list_create(&tx->tx_holds, sizeof (dmu_tx_hold_t),
50 offsetof(dmu_tx_hold_t, txh_node));
52 refcount_create(&tx->tx_space_written);
53 refcount_create(&tx->tx_space_freed);
59 dmu_tx_create(objset_t *os)
61 dmu_tx_t *tx = dmu_tx_create_dd(os->os->os_dsl_dataset->ds_dir);
63 tx->tx_lastsnap_txg = dsl_dataset_prev_snap_txg(os->os->os_dsl_dataset);
68 dmu_tx_create_assigned(struct dsl_pool *dp, uint64_t txg)
70 dmu_tx_t *tx = dmu_tx_create_dd(NULL);
72 ASSERT3U(txg, <=, dp->dp_tx.tx_open_txg);
81 dmu_tx_is_syncing(dmu_tx_t *tx)
83 return (tx->tx_anyobj);
87 dmu_tx_private_ok(dmu_tx_t *tx)
89 return (tx->tx_anyobj);
92 static dmu_tx_hold_t *
93 dmu_tx_hold_object_impl(dmu_tx_t *tx, objset_t *os, uint64_t object,
94 enum dmu_tx_hold_type type, uint64_t arg1, uint64_t arg2)
100 if (object != DMU_NEW_OBJECT) {
101 err = dnode_hold(os->os, object, tx, &dn);
107 if (err == 0 && tx->tx_txg != 0) {
108 mutex_enter(&dn->dn_mtx);
110 * dn->dn_assigned_txg == tx->tx_txg doesn't pose a
111 * problem, but there's no way for it to happen (for
114 ASSERT(dn->dn_assigned_txg == 0);
115 dn->dn_assigned_txg = tx->tx_txg;
116 (void) refcount_add(&dn->dn_tx_holds, tx);
117 mutex_exit(&dn->dn_mtx);
121 txh = kmem_zalloc(sizeof (dmu_tx_hold_t), KM_SLEEP);
125 txh->txh_type = type;
126 txh->txh_arg1 = arg1;
127 txh->txh_arg2 = arg2;
129 list_insert_tail(&tx->tx_holds, txh);
135 dmu_tx_add_new_object(dmu_tx_t *tx, objset_t *os, uint64_t object)
138 * If we're syncing, they can manipulate any object anyhow, and
139 * the hold on the dnode_t can cause problems.
141 if (!dmu_tx_is_syncing(tx)) {
142 (void) dmu_tx_hold_object_impl(tx, os,
143 object, THT_NEWOBJECT, 0, 0);
148 dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
153 rw_enter(&dn->dn_struct_rwlock, RW_READER);
154 db = dbuf_hold_level(dn, level, blkid, FTAG);
155 rw_exit(&dn->dn_struct_rwlock);
158 err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
164 dmu_tx_count_indirects(dmu_tx_hold_t *txh, dmu_buf_impl_t *db,
165 boolean_t freeable, dmu_buf_impl_t **history)
167 int i = db->db_level + 1;
168 dnode_t *dn = db->db_dnode;
170 if (i >= dn->dn_nlevels)
175 uint64_t lvls = dn->dn_nlevels - i;
177 txh->txh_space_towrite += lvls << dn->dn_indblkshift;
181 if (db != history[i]) {
182 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
183 uint64_t space = 1ULL << dn->dn_indblkshift;
185 freeable = (db->db_blkptr && (freeable ||
186 dsl_dataset_block_freeable(ds, db->db_blkptr->blk_birth)));
188 txh->txh_space_tooverwrite += space;
190 txh->txh_space_towrite += space;
192 txh->txh_space_tounref += space;
194 dmu_tx_count_indirects(txh, db, freeable, history);
200 dmu_tx_count_write(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
202 dnode_t *dn = txh->txh_dnode;
203 uint64_t start, end, i;
204 int min_bs, max_bs, min_ibs, max_ibs, epbs, bits;
210 min_bs = SPA_MINBLOCKSHIFT;
211 max_bs = SPA_MAXBLOCKSHIFT;
212 min_ibs = DN_MIN_INDBLKSHIFT;
213 max_ibs = DN_MAX_INDBLKSHIFT;
216 dmu_buf_impl_t *last[DN_MAX_LEVELS];
217 int nlvls = dn->dn_nlevels;
221 * For i/o error checking, read the first and last level-0
222 * blocks (if they are not aligned), and all the level-1 blocks.
225 if (dn->dn_maxblkid == 0) {
226 delta = dn->dn_datablksz;
227 start = (off < dn->dn_datablksz) ? 0 : 1;
228 end = (off+len <= dn->dn_datablksz) ? 0 : 1;
229 if (start == 0 && (off > 0 || len < dn->dn_datablksz)) {
230 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
236 zio_t *zio = zio_root(dn->dn_objset->os_spa,
237 NULL, NULL, ZIO_FLAG_CANFAIL);
239 /* first level-0 block */
240 start = off >> dn->dn_datablkshift;
241 if (P2PHASE(off, dn->dn_datablksz) ||
242 len < dn->dn_datablksz) {
243 err = dmu_tx_check_ioerr(zio, dn, 0, start);
248 /* last level-0 block */
249 end = (off+len-1) >> dn->dn_datablkshift;
251 P2PHASE(off+len, dn->dn_datablksz)) {
252 err = dmu_tx_check_ioerr(zio, dn, 0, end);
259 int shft = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
260 for (i = (start>>shft)+1; i < end>>shft; i++) {
261 err = dmu_tx_check_ioerr(zio, dn, 1, i);
270 delta = P2NPHASE(off, dn->dn_datablksz);
273 if (dn->dn_maxblkid > 0) {
275 * The blocksize can't change,
276 * so we can make a more precise estimate.
278 ASSERT(dn->dn_datablkshift != 0);
279 min_bs = max_bs = dn->dn_datablkshift;
280 min_ibs = max_ibs = dn->dn_indblkshift;
281 } else if (dn->dn_indblkshift > max_ibs) {
283 * This ensures that if we reduce DN_MAX_INDBLKSHIFT,
284 * the code will still work correctly on older pools.
286 min_ibs = max_ibs = dn->dn_indblkshift;
290 * If this write is not off the end of the file
291 * we need to account for overwrites/unref.
293 if (start <= dn->dn_maxblkid)
294 bzero(last, sizeof (dmu_buf_impl_t *) * DN_MAX_LEVELS);
295 while (start <= dn->dn_maxblkid) {
296 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
297 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
300 rw_enter(&dn->dn_struct_rwlock, RW_READER);
301 db = dbuf_hold_level(dn, 0, start, FTAG);
302 rw_exit(&dn->dn_struct_rwlock);
303 if (db->db_blkptr && dsl_dataset_block_freeable(ds,
304 db->db_blkptr->blk_birth)) {
305 dprintf_bp(db->db_blkptr, "can free old%s", "");
306 txh->txh_space_tooverwrite += dn->dn_datablksz;
307 txh->txh_space_tounref += dn->dn_datablksz;
308 dmu_tx_count_indirects(txh, db, TRUE, last);
310 txh->txh_space_towrite += dn->dn_datablksz;
312 txh->txh_space_tounref +=
313 bp_get_dasize(spa, db->db_blkptr);
314 dmu_tx_count_indirects(txh, db, FALSE, last);
319 * Account for new indirects appearing
320 * before this IO gets assigned into a txg.
323 epbs = min_ibs - SPA_BLKPTRSHIFT;
324 for (bits -= epbs * (nlvls - 1);
325 bits >= 0; bits -= epbs)
326 txh->txh_fudge += 1ULL << max_ibs;
332 delta = dn->dn_datablksz;
337 * 'end' is the last thing we will access, not one past.
338 * This way we won't overflow when accessing the last byte.
340 start = P2ALIGN(off, 1ULL << max_bs);
341 end = P2ROUNDUP(off + len, 1ULL << max_bs) - 1;
342 txh->txh_space_towrite += end - start + 1;
347 epbs = min_ibs - SPA_BLKPTRSHIFT;
350 * The object contains at most 2^(64 - min_bs) blocks,
351 * and each indirect level maps 2^epbs.
353 for (bits = 64 - min_bs; bits >= 0; bits -= epbs) {
356 ASSERT3U(end, >=, start);
357 txh->txh_space_towrite += (end - start + 1) << max_ibs;
360 * We also need a new blkid=0 indirect block
361 * to reference any existing file data.
363 txh->txh_space_towrite += 1ULL << max_ibs;
368 if (txh->txh_space_towrite + txh->txh_space_tooverwrite >
373 txh->txh_tx->tx_err = err;
377 dmu_tx_count_dnode(dmu_tx_hold_t *txh)
379 dnode_t *dn = txh->txh_dnode;
380 dnode_t *mdn = txh->txh_tx->tx_objset->os->os_meta_dnode;
381 uint64_t space = mdn->dn_datablksz +
382 ((mdn->dn_nlevels-1) << mdn->dn_indblkshift);
384 if (dn && dn->dn_dbuf->db_blkptr &&
385 dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
386 dn->dn_dbuf->db_blkptr->blk_birth)) {
387 txh->txh_space_tooverwrite += space;
388 txh->txh_space_tounref += space;
390 txh->txh_space_towrite += space;
391 if (dn && dn->dn_dbuf->db_blkptr)
392 txh->txh_space_tounref += space;
397 dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len)
401 ASSERT(tx->tx_txg == 0);
402 ASSERT(len < DMU_MAX_ACCESS);
403 ASSERT(len == 0 || UINT64_MAX - off >= len - 1);
405 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
406 object, THT_WRITE, off, len);
410 dmu_tx_count_write(txh, off, len);
411 dmu_tx_count_dnode(txh);
415 dmu_tx_count_free(dmu_tx_hold_t *txh, uint64_t off, uint64_t len)
417 uint64_t blkid, nblks, lastblk;
418 uint64_t space = 0, unref = 0, skipped = 0;
419 dnode_t *dn = txh->txh_dnode;
420 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
421 spa_t *spa = txh->txh_tx->tx_pool->dp_spa;
424 if (dn->dn_nlevels == 0)
428 * The struct_rwlock protects us against dn_nlevels
429 * changing, in case (against all odds) we manage to dirty &
430 * sync out the changes after we check for being dirty.
431 * Also, dbuf_hold_level() wants us to have the struct_rwlock.
433 rw_enter(&dn->dn_struct_rwlock, RW_READER);
434 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
435 if (dn->dn_maxblkid == 0) {
436 if (off == 0 && len >= dn->dn_datablksz) {
440 rw_exit(&dn->dn_struct_rwlock);
444 blkid = off >> dn->dn_datablkshift;
445 nblks = (len + dn->dn_datablksz - 1) >> dn->dn_datablkshift;
447 if (blkid >= dn->dn_maxblkid) {
448 rw_exit(&dn->dn_struct_rwlock);
451 if (blkid + nblks > dn->dn_maxblkid)
452 nblks = dn->dn_maxblkid - blkid;
455 if (dn->dn_nlevels == 1) {
457 for (i = 0; i < nblks; i++) {
458 blkptr_t *bp = dn->dn_phys->dn_blkptr;
459 ASSERT3U(blkid + i, <, dn->dn_nblkptr);
461 if (dsl_dataset_block_freeable(ds, bp->blk_birth)) {
462 dprintf_bp(bp, "can free old%s", "");
463 space += bp_get_dasize(spa, bp);
465 unref += BP_GET_ASIZE(bp);
471 * Add in memory requirements of higher-level indirects.
472 * This assumes a worst-possible scenario for dn_nlevels.
475 uint64_t blkcnt = 1 + ((nblks >> epbs) >> epbs);
476 int level = (dn->dn_nlevels > 1) ? 2 : 1;
478 while (level++ < DN_MAX_LEVELS) {
479 txh->txh_memory_tohold += blkcnt << dn->dn_indblkshift;
480 blkcnt = 1 + (blkcnt >> epbs);
482 ASSERT(blkcnt <= dn->dn_nblkptr);
485 lastblk = blkid + nblks - 1;
487 dmu_buf_impl_t *dbuf;
488 uint64_t ibyte, new_blkid;
490 int err, i, blkoff, tochk;
493 ibyte = blkid << dn->dn_datablkshift;
494 err = dnode_next_offset(dn,
495 DNODE_FIND_HAVELOCK, &ibyte, 2, 1, 0);
496 new_blkid = ibyte >> dn->dn_datablkshift;
498 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
502 txh->txh_tx->tx_err = err;
505 if (new_blkid > lastblk) {
506 skipped += (lastblk >> epbs) - (blkid >> epbs) + 1;
510 if (new_blkid > blkid) {
511 ASSERT((new_blkid >> epbs) > (blkid >> epbs));
512 skipped += (new_blkid >> epbs) - (blkid >> epbs) - 1;
513 nblks -= new_blkid - blkid;
516 blkoff = P2PHASE(blkid, epb);
517 tochk = MIN(epb - blkoff, nblks);
519 dbuf = dbuf_hold_level(dn, 1, blkid >> epbs, FTAG);
521 txh->txh_memory_tohold += dbuf->db.db_size;
522 if (txh->txh_memory_tohold > DMU_MAX_ACCESS) {
523 txh->txh_tx->tx_err = E2BIG;
524 dbuf_rele(dbuf, FTAG);
527 err = dbuf_read(dbuf, NULL, DB_RF_HAVESTRUCT | DB_RF_CANFAIL);
529 txh->txh_tx->tx_err = err;
530 dbuf_rele(dbuf, FTAG);
534 bp = dbuf->db.db_data;
537 for (i = 0; i < tochk; i++) {
538 if (dsl_dataset_block_freeable(ds, bp[i].blk_birth)) {
539 dprintf_bp(&bp[i], "can free old%s", "");
540 space += bp_get_dasize(spa, &bp[i]);
542 unref += BP_GET_ASIZE(bp);
544 dbuf_rele(dbuf, FTAG);
549 rw_exit(&dn->dn_struct_rwlock);
551 /* account for new level 1 indirect blocks that might show up */
553 txh->txh_fudge += skipped << dn->dn_indblkshift;
554 skipped = MIN(skipped, DMU_MAX_DELETEBLKCNT >> epbs);
555 txh->txh_memory_tohold += skipped << dn->dn_indblkshift;
557 txh->txh_space_tofree += space;
558 txh->txh_space_tounref += unref;
562 dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off, uint64_t len)
566 uint64_t start, end, i;
570 ASSERT(tx->tx_txg == 0);
572 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
573 object, THT_FREE, off, len);
580 dmu_tx_count_write(txh, off, 1);
582 if (len != DMU_OBJECT_END)
583 dmu_tx_count_write(txh, off+len, 1);
585 if (off >= (dn->dn_maxblkid+1) * dn->dn_datablksz)
587 if (len == DMU_OBJECT_END)
588 len = (dn->dn_maxblkid+1) * dn->dn_datablksz - off;
591 * For i/o error checking, read the first and last level-0
592 * blocks, and all the level-1 blocks. The above count_write's
593 * have already taken care of the level-0 blocks.
595 if (dn->dn_nlevels > 1) {
596 shift = dn->dn_datablkshift + dn->dn_indblkshift -
598 start = off >> shift;
599 end = dn->dn_datablkshift ? ((off+len) >> shift) : 0;
601 zio = zio_root(tx->tx_pool->dp_spa,
602 NULL, NULL, ZIO_FLAG_CANFAIL);
603 for (i = start; i <= end; i++) {
604 uint64_t ibyte = i << shift;
605 err = dnode_next_offset(dn, 0, &ibyte, 2, 1, 0);
614 err = dmu_tx_check_ioerr(zio, dn, 1, i);
627 dmu_tx_count_dnode(txh);
628 dmu_tx_count_free(txh, off, len);
632 dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name)
639 ASSERT(tx->tx_txg == 0);
641 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
642 object, THT_ZAP, add, (uintptr_t)name);
647 dmu_tx_count_dnode(txh);
651 * We will be able to fit a new object's entries into one leaf
652 * block. So there will be at most 2 blocks total,
653 * including the header block.
655 dmu_tx_count_write(txh, 0, 2 << fzap_default_block_shift);
659 ASSERT3P(dmu_ot[dn->dn_type].ot_byteswap, ==, zap_byteswap);
661 if (dn->dn_maxblkid == 0 && !add) {
663 * If there is only one block (i.e. this is a micro-zap)
664 * and we are not adding anything, the accounting is simple.
666 err = dmu_tx_check_ioerr(NULL, dn, 0, 0);
673 * Use max block size here, since we don't know how much
674 * the size will change between now and the dbuf dirty call.
676 if (dsl_dataset_block_freeable(dn->dn_objset->os_dsl_dataset,
677 dn->dn_phys->dn_blkptr[0].blk_birth)) {
678 txh->txh_space_tooverwrite += SPA_MAXBLOCKSIZE;
680 txh->txh_space_towrite += SPA_MAXBLOCKSIZE;
682 if (dn->dn_phys->dn_blkptr[0].blk_birth)
683 txh->txh_space_tounref += SPA_MAXBLOCKSIZE;
687 if (dn->dn_maxblkid > 0 && name) {
689 * access the name in this fat-zap so that we'll check
690 * for i/o errors to the leaf blocks, etc.
692 err = zap_lookup(&dn->dn_objset->os, dn->dn_object, name,
700 err = zap_count_write(&dn->dn_objset->os, dn->dn_object, name, add,
701 &txh->txh_space_towrite, &txh->txh_space_tooverwrite);
704 * If the modified blocks are scattered to the four winds,
705 * we'll have to modify an indirect twig for each.
707 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
708 for (nblocks = dn->dn_maxblkid >> epbs; nblocks != 0; nblocks >>= epbs)
709 if (dn->dn_objset->os_dsl_dataset->ds_phys->ds_prev_snap_obj)
710 txh->txh_space_towrite += 3 << dn->dn_indblkshift;
712 txh->txh_space_tooverwrite += 3 << dn->dn_indblkshift;
716 dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object)
720 ASSERT(tx->tx_txg == 0);
722 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
723 object, THT_BONUS, 0, 0);
725 dmu_tx_count_dnode(txh);
729 dmu_tx_hold_space(dmu_tx_t *tx, uint64_t space)
732 ASSERT(tx->tx_txg == 0);
734 txh = dmu_tx_hold_object_impl(tx, tx->tx_objset,
735 DMU_NEW_OBJECT, THT_SPACE, space, 0);
737 txh->txh_space_towrite += space;
741 dmu_tx_holds(dmu_tx_t *tx, uint64_t object)
747 * By asserting that the tx is assigned, we're counting the
748 * number of dn_tx_holds, which is the same as the number of
749 * dn_holds. Otherwise, we'd be counting dn_holds, but
750 * dn_tx_holds could be 0.
752 ASSERT(tx->tx_txg != 0);
754 /* if (tx->tx_anyobj == TRUE) */
757 for (txh = list_head(&tx->tx_holds); txh;
758 txh = list_next(&tx->tx_holds, txh)) {
759 if (txh->txh_dnode && txh->txh_dnode->dn_object == object)
768 dmu_tx_dirty_buf(dmu_tx_t *tx, dmu_buf_impl_t *db)
771 int match_object = FALSE, match_offset = FALSE;
772 dnode_t *dn = db->db_dnode;
774 ASSERT(tx->tx_txg != 0);
775 ASSERT(tx->tx_objset == NULL || dn->dn_objset == tx->tx_objset->os);
776 ASSERT3U(dn->dn_object, ==, db->db.db_object);
781 /* XXX No checking on the meta dnode for now */
782 if (db->db.db_object == DMU_META_DNODE_OBJECT)
785 for (txh = list_head(&tx->tx_holds); txh;
786 txh = list_next(&tx->tx_holds, txh)) {
787 ASSERT(dn == NULL || dn->dn_assigned_txg == tx->tx_txg);
788 if (txh->txh_dnode == dn && txh->txh_type != THT_NEWOBJECT)
790 if (txh->txh_dnode == NULL || txh->txh_dnode == dn) {
791 int datablkshift = dn->dn_datablkshift ?
792 dn->dn_datablkshift : SPA_MAXBLOCKSHIFT;
793 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
794 int shift = datablkshift + epbs * db->db_level;
795 uint64_t beginblk = shift >= 64 ? 0 :
796 (txh->txh_arg1 >> shift);
797 uint64_t endblk = shift >= 64 ? 0 :
798 ((txh->txh_arg1 + txh->txh_arg2 - 1) >> shift);
799 uint64_t blkid = db->db_blkid;
801 /* XXX txh_arg2 better not be zero... */
803 dprintf("found txh type %x beginblk=%llx endblk=%llx\n",
804 txh->txh_type, beginblk, endblk);
806 switch (txh->txh_type) {
808 if (blkid >= beginblk && blkid <= endblk)
811 * We will let this hold work for the bonus
812 * buffer so that we don't need to hold it
813 * when creating a new object.
815 if (blkid == DB_BONUS_BLKID)
818 * They might have to increase nlevels,
819 * thus dirtying the new TLIBs. Or the
820 * might have to change the block size,
821 * thus dirying the new lvl=0 blk=0.
828 * We will dirty all the level 1 blocks in
829 * the free range and perhaps the first and
830 * last level 0 block.
832 if (blkid >= beginblk && (blkid <= endblk ||
833 txh->txh_arg2 == DMU_OBJECT_END))
837 if (blkid == DB_BONUS_BLKID)
847 ASSERT(!"bad txh_type");
850 if (match_object && match_offset)
853 panic("dirtying dbuf obj=%llx lvl=%u blkid=%llx but not tx_held\n",
854 (u_longlong_t)db->db.db_object, db->db_level,
855 (u_longlong_t)db->db_blkid);
860 dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
863 spa_t *spa = tx->tx_pool->dp_spa;
864 uint64_t memory, asize, fsize, usize;
865 uint64_t towrite, tofree, tooverwrite, tounref, tohold, fudge;
867 ASSERT3U(tx->tx_txg, ==, 0);
872 if (spa_suspended(spa)) {
874 * If the user has indicated a blocking failure mode
875 * then return ERESTART which will block in dmu_tx_wait().
876 * Otherwise, return EIO so that an error can get
877 * propagated back to the VOP calls.
879 * Note that we always honor the txg_how flag regardless
880 * of the failuremode setting.
882 if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_CONTINUE &&
889 tx->tx_txg = txg_hold_open(tx->tx_pool, &tx->tx_txgh);
890 tx->tx_needassign_txh = NULL;
893 * NB: No error returns are allowed after txg_hold_open, but
894 * before processing the dnode holds, due to the
895 * dmu_tx_unassign() logic.
898 towrite = tofree = tooverwrite = tounref = tohold = fudge = 0;
899 for (txh = list_head(&tx->tx_holds); txh;
900 txh = list_next(&tx->tx_holds, txh)) {
901 dnode_t *dn = txh->txh_dnode;
903 mutex_enter(&dn->dn_mtx);
904 if (dn->dn_assigned_txg == tx->tx_txg - 1) {
905 mutex_exit(&dn->dn_mtx);
906 tx->tx_needassign_txh = txh;
909 if (dn->dn_assigned_txg == 0)
910 dn->dn_assigned_txg = tx->tx_txg;
911 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
912 (void) refcount_add(&dn->dn_tx_holds, tx);
913 mutex_exit(&dn->dn_mtx);
915 towrite += txh->txh_space_towrite;
916 tofree += txh->txh_space_tofree;
917 tooverwrite += txh->txh_space_tooverwrite;
918 tounref += txh->txh_space_tounref;
919 tohold += txh->txh_memory_tohold;
920 fudge += txh->txh_fudge;
924 * NB: This check must be after we've held the dnodes, so that
925 * the dmu_tx_unassign() logic will work properly
927 if (txg_how >= TXG_INITIAL && txg_how != tx->tx_txg)
931 * If a snapshot has been taken since we made our estimates,
932 * assume that we won't be able to free or overwrite anything.
935 dsl_dataset_prev_snap_txg(tx->tx_objset->os->os_dsl_dataset) >
936 tx->tx_lastsnap_txg) {
937 towrite += tooverwrite;
938 tooverwrite = tofree = 0;
941 /* needed allocation: worst-case estimate of write space */
942 asize = spa_get_asize(tx->tx_pool->dp_spa, towrite + tooverwrite);
943 /* freed space estimate: worst-case overwrite + free estimate */
944 fsize = spa_get_asize(tx->tx_pool->dp_spa, tooverwrite) + tofree;
945 /* convert unrefd space to worst-case estimate */
946 usize = spa_get_asize(tx->tx_pool->dp_spa, tounref);
947 /* calculate memory footprint estimate */
948 memory = towrite + tooverwrite + tohold;
952 * Add in 'tohold' to account for our dirty holds on this memory
953 * XXX - the "fudge" factor is to account for skipped blocks that
954 * we missed because dnode_next_offset() misses in-core-only blocks.
956 tx->tx_space_towrite = asize +
957 spa_get_asize(tx->tx_pool->dp_spa, tohold + fudge);
958 tx->tx_space_tofree = tofree;
959 tx->tx_space_tooverwrite = tooverwrite;
960 tx->tx_space_tounref = tounref;
963 if (tx->tx_dir && asize != 0) {
964 int err = dsl_dir_tempreserve_space(tx->tx_dir, memory,
965 asize, fsize, usize, &tx->tx_tempreserve_cookie, tx);
974 dmu_tx_unassign(dmu_tx_t *tx)
981 txg_rele_to_quiesce(&tx->tx_txgh);
983 for (txh = list_head(&tx->tx_holds); txh != tx->tx_needassign_txh;
984 txh = list_next(&tx->tx_holds, txh)) {
985 dnode_t *dn = txh->txh_dnode;
989 mutex_enter(&dn->dn_mtx);
990 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
992 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
993 dn->dn_assigned_txg = 0;
994 cv_broadcast(&dn->dn_notxholds);
996 mutex_exit(&dn->dn_mtx);
999 txg_rele_to_sync(&tx->tx_txgh);
1001 tx->tx_lasttried_txg = tx->tx_txg;
1006 * Assign tx to a transaction group. txg_how can be one of:
1008 * (1) TXG_WAIT. If the current open txg is full, waits until there's
1009 * a new one. This should be used when you're not holding locks.
1010 * If will only fail if we're truly out of space (or over quota).
1012 * (2) TXG_NOWAIT. If we can't assign into the current open txg without
1013 * blocking, returns immediately with ERESTART. This should be used
1014 * whenever you're holding locks. On an ERESTART error, the caller
1015 * should drop locks, do a dmu_tx_wait(tx), and try again.
1017 * (3) A specific txg. Use this if you need to ensure that multiple
1018 * transactions all sync in the same txg. Like TXG_NOWAIT, it
1019 * returns ERESTART if it can't assign you into the requested txg.
1022 dmu_tx_assign(dmu_tx_t *tx, uint64_t txg_how)
1026 ASSERT(tx->tx_txg == 0);
1027 ASSERT(txg_how != 0);
1028 ASSERT(!dsl_pool_sync_context(tx->tx_pool));
1030 while ((err = dmu_tx_try_assign(tx, txg_how)) != 0) {
1031 dmu_tx_unassign(tx);
1033 if (err != ERESTART || txg_how != TXG_WAIT)
1039 txg_rele_to_quiesce(&tx->tx_txgh);
1045 dmu_tx_wait(dmu_tx_t *tx)
1047 spa_t *spa = tx->tx_pool->dp_spa;
1049 ASSERT(tx->tx_txg == 0);
1052 * It's possible that the pool has become active after this thread
1053 * has tried to obtain a tx. If that's the case then his
1054 * tx_lasttried_txg would not have been assigned.
1056 if (spa_suspended(spa) || tx->tx_lasttried_txg == 0) {
1057 txg_wait_synced(tx->tx_pool, spa_last_synced_txg(spa) + 1);
1058 } else if (tx->tx_needassign_txh) {
1059 dnode_t *dn = tx->tx_needassign_txh->txh_dnode;
1061 mutex_enter(&dn->dn_mtx);
1062 while (dn->dn_assigned_txg == tx->tx_lasttried_txg - 1)
1063 cv_wait(&dn->dn_notxholds, &dn->dn_mtx);
1064 mutex_exit(&dn->dn_mtx);
1065 tx->tx_needassign_txh = NULL;
1067 txg_wait_open(tx->tx_pool, tx->tx_lasttried_txg + 1);
1072 dmu_tx_willuse_space(dmu_tx_t *tx, int64_t delta)
1075 if (tx->tx_dir == NULL || delta == 0)
1079 ASSERT3U(refcount_count(&tx->tx_space_written) + delta, <=,
1080 tx->tx_space_towrite);
1081 (void) refcount_add_many(&tx->tx_space_written, delta, NULL);
1083 (void) refcount_add_many(&tx->tx_space_freed, -delta, NULL);
1089 dmu_tx_commit(dmu_tx_t *tx)
1093 ASSERT(tx->tx_txg != 0);
1095 while (txh = list_head(&tx->tx_holds)) {
1096 dnode_t *dn = txh->txh_dnode;
1098 list_remove(&tx->tx_holds, txh);
1099 kmem_free(txh, sizeof (dmu_tx_hold_t));
1102 mutex_enter(&dn->dn_mtx);
1103 ASSERT3U(dn->dn_assigned_txg, ==, tx->tx_txg);
1105 if (refcount_remove(&dn->dn_tx_holds, tx) == 0) {
1106 dn->dn_assigned_txg = 0;
1107 cv_broadcast(&dn->dn_notxholds);
1109 mutex_exit(&dn->dn_mtx);
1113 if (tx->tx_tempreserve_cookie)
1114 dsl_dir_tempreserve_clear(tx->tx_tempreserve_cookie, tx);
1116 if (tx->tx_anyobj == FALSE)
1117 txg_rele_to_sync(&tx->tx_txgh);
1118 list_destroy(&tx->tx_holds);
1120 dprintf("towrite=%llu written=%llu tofree=%llu freed=%llu\n",
1121 tx->tx_space_towrite, refcount_count(&tx->tx_space_written),
1122 tx->tx_space_tofree, refcount_count(&tx->tx_space_freed));
1123 refcount_destroy_many(&tx->tx_space_written,
1124 refcount_count(&tx->tx_space_written));
1125 refcount_destroy_many(&tx->tx_space_freed,
1126 refcount_count(&tx->tx_space_freed));
1128 kmem_free(tx, sizeof (dmu_tx_t));
1132 dmu_tx_abort(dmu_tx_t *tx)
1136 ASSERT(tx->tx_txg == 0);
1138 while (txh = list_head(&tx->tx_holds)) {
1139 dnode_t *dn = txh->txh_dnode;
1141 list_remove(&tx->tx_holds, txh);
1142 kmem_free(txh, sizeof (dmu_tx_hold_t));
1146 list_destroy(&tx->tx_holds);
1148 refcount_destroy_many(&tx->tx_space_written,
1149 refcount_count(&tx->tx_space_written));
1150 refcount_destroy_many(&tx->tx_space_freed,
1151 refcount_count(&tx->tx_space_freed));
1153 kmem_free(tx, sizeof (dmu_tx_t));
1157 dmu_tx_get_txg(dmu_tx_t *tx)
1159 ASSERT(tx->tx_txg != 0);
1160 return (tx->tx_txg);