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 2007 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #pragma ident "%Z%%M% %I% %E% SMI"
28 #include <sys/zfs_context.h>
30 #include <sys/dnode.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
39 #include <sys/dmu_zfetch.h>
41 static int free_range_compar(const void *node1, const void *node2);
43 static kmem_cache_t *dnode_cache;
45 static dnode_phys_t dnode_phys_zero;
47 int zfs_default_bs = SPA_MINBLOCKSHIFT;
48 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
52 dnode_cons(void *arg, void *unused, int kmflag)
56 bzero(dn, sizeof (dnode_t));
58 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
59 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
60 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
61 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
62 refcount_create(&dn->dn_holds);
63 refcount_create(&dn->dn_tx_holds);
65 for (i = 0; i < TXG_SIZE; i++) {
66 avl_create(&dn->dn_ranges[i], free_range_compar,
67 sizeof (free_range_t),
68 offsetof(struct free_range, fr_node));
69 list_create(&dn->dn_dirty_records[i],
70 sizeof (dbuf_dirty_record_t),
71 offsetof(dbuf_dirty_record_t, dr_dirty_node));
74 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
75 offsetof(dmu_buf_impl_t, db_link));
82 dnode_dest(void *arg, void *unused)
87 cv_destroy(&dn->dn_notxholds);
88 rw_destroy(&dn->dn_struct_rwlock);
89 mutex_destroy(&dn->dn_mtx);
90 mutex_destroy(&dn->dn_dbufs_mtx);
91 refcount_destroy(&dn->dn_holds);
92 refcount_destroy(&dn->dn_tx_holds);
94 for (i = 0; i < TXG_SIZE; i++) {
95 avl_destroy(&dn->dn_ranges[i]);
96 list_destroy(&dn->dn_dirty_records[i]);
99 list_destroy(&dn->dn_dbufs);
105 dnode_cache = kmem_cache_create("dnode_t",
107 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
113 kmem_cache_destroy(dnode_cache);
119 dnode_verify(dnode_t *dn)
121 int drop_struct_lock = FALSE;
124 ASSERT(dn->dn_objset);
126 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
128 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
131 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
132 rw_enter(&dn->dn_struct_rwlock, RW_READER);
133 drop_struct_lock = TRUE;
135 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
137 ASSERT3U(dn->dn_indblkshift, >=, 0);
138 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
139 if (dn->dn_datablkshift) {
140 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
141 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
142 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
144 ASSERT3U(dn->dn_nlevels, <=, 30);
145 ASSERT3U(dn->dn_type, <=, DMU_OT_NUMTYPES);
146 ASSERT3U(dn->dn_nblkptr, >=, 1);
147 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
148 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
149 ASSERT3U(dn->dn_datablksz, ==,
150 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
151 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
152 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
153 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
154 for (i = 0; i < TXG_SIZE; i++) {
155 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
158 if (dn->dn_phys->dn_type != DMU_OT_NONE)
159 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
160 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || dn->dn_dbuf != NULL);
161 if (dn->dn_dbuf != NULL) {
162 ASSERT3P(dn->dn_phys, ==,
163 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
164 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
166 if (drop_struct_lock)
167 rw_exit(&dn->dn_struct_rwlock);
172 dnode_byteswap(dnode_phys_t *dnp)
174 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
177 if (dnp->dn_type == DMU_OT_NONE) {
178 bzero(dnp, sizeof (dnode_phys_t));
182 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
183 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
184 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
185 dnp->dn_used = BSWAP_64(dnp->dn_used);
188 * dn_nblkptr is only one byte, so it's OK to read it in either
189 * byte order. We can't read dn_bouslen.
191 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
192 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
193 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
194 buf64[i] = BSWAP_64(buf64[i]);
197 * OK to check dn_bonuslen for zero, because it won't matter if
198 * we have the wrong byte order. This is necessary because the
199 * dnode dnode is smaller than a regular dnode.
201 if (dnp->dn_bonuslen != 0) {
203 * Note that the bonus length calculated here may be
204 * longer than the actual bonus buffer. This is because
205 * we always put the bonus buffer after the last block
206 * pointer (instead of packing it against the end of the
209 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
210 size_t len = DN_MAX_BONUSLEN - off;
211 ASSERT3U(dnp->dn_bonustype, <, DMU_OT_NUMTYPES);
212 dmu_ot[dnp->dn_bonustype].ot_byteswap(dnp->dn_bonus + off, len);
217 dnode_buf_byteswap(void *vbuf, size_t size)
219 dnode_phys_t *buf = vbuf;
222 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
223 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
225 size >>= DNODE_SHIFT;
226 for (i = 0; i < size; i++) {
233 free_range_compar(const void *node1, const void *node2)
235 const free_range_t *rp1 = node1;
236 const free_range_t *rp2 = node2;
238 if (rp1->fr_blkid < rp2->fr_blkid)
240 else if (rp1->fr_blkid > rp2->fr_blkid)
246 dnode_setdblksz(dnode_t *dn, int size)
248 ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0);
249 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
250 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
251 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
252 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
253 dn->dn_datablksz = size;
254 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
255 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
259 dnode_create(objset_impl_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
262 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
265 dn->dn_object = object;
269 if (dnp->dn_datablkszsec)
270 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
271 dn->dn_indblkshift = dnp->dn_indblkshift;
272 dn->dn_nlevels = dnp->dn_nlevels;
273 dn->dn_type = dnp->dn_type;
274 dn->dn_nblkptr = dnp->dn_nblkptr;
275 dn->dn_checksum = dnp->dn_checksum;
276 dn->dn_compress = dnp->dn_compress;
277 dn->dn_bonustype = dnp->dn_bonustype;
278 dn->dn_bonuslen = dnp->dn_bonuslen;
279 dn->dn_maxblkid = dnp->dn_maxblkid;
281 dmu_zfetch_init(&dn->dn_zfetch, dn);
283 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
284 mutex_enter(&os->os_lock);
285 list_insert_head(&os->os_dnodes, dn);
286 mutex_exit(&os->os_lock);
292 dnode_destroy(dnode_t *dn)
294 objset_impl_t *os = dn->dn_objset;
299 for (i = 0; i < TXG_SIZE; i++) {
300 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
301 ASSERT(NULL == list_head(&dn->dn_dirty_records[i]));
302 ASSERT(0 == avl_numnodes(&dn->dn_ranges[i]));
304 ASSERT(NULL == list_head(&dn->dn_dbufs));
307 mutex_enter(&os->os_lock);
308 list_remove(&os->os_dnodes, dn);
309 mutex_exit(&os->os_lock);
311 if (dn->dn_dirtyctx_firstset) {
312 kmem_free(dn->dn_dirtyctx_firstset, 1);
313 dn->dn_dirtyctx_firstset = NULL;
315 dmu_zfetch_rele(&dn->dn_zfetch);
317 mutex_enter(&dn->dn_bonus->db_mtx);
318 dbuf_evict(dn->dn_bonus);
321 kmem_cache_free(dnode_cache, dn);
325 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
326 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
331 blocksize = 1 << zfs_default_bs;
332 else if (blocksize > SPA_MAXBLOCKSIZE)
333 blocksize = SPA_MAXBLOCKSIZE;
335 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
338 ibs = zfs_default_ibs;
340 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
342 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
343 dn->dn_object, tx->tx_txg, blocksize, ibs);
345 ASSERT(dn->dn_type == DMU_OT_NONE);
346 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
347 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
348 ASSERT(ot != DMU_OT_NONE);
349 ASSERT3U(ot, <, DMU_OT_NUMTYPES);
350 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
351 (bonustype != DMU_OT_NONE && bonuslen != 0));
352 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
353 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
354 ASSERT(dn->dn_type == DMU_OT_NONE);
355 ASSERT3U(dn->dn_maxblkid, ==, 0);
356 ASSERT3U(dn->dn_allocated_txg, ==, 0);
357 ASSERT3U(dn->dn_assigned_txg, ==, 0);
358 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
359 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
360 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
362 for (i = 0; i < TXG_SIZE; i++) {
363 ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
364 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
365 ASSERT3U(dn->dn_next_blksz[i], ==, 0);
366 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
367 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
368 ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0);
372 dnode_setdblksz(dn, blocksize);
373 dn->dn_indblkshift = ibs;
375 dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
376 dn->dn_bonustype = bonustype;
377 dn->dn_bonuslen = bonuslen;
378 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
379 dn->dn_compress = ZIO_COMPRESS_INHERIT;
383 if (dn->dn_dirtyctx_firstset) {
384 kmem_free(dn->dn_dirtyctx_firstset, 1);
385 dn->dn_dirtyctx_firstset = NULL;
388 dn->dn_allocated_txg = tx->tx_txg;
390 dnode_setdirty(dn, tx);
391 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
392 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
396 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
397 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
400 dmu_buf_impl_t *db = NULL;
402 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
403 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
404 ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0);
405 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
406 ASSERT(tx->tx_txg != 0);
407 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
408 (bonustype != DMU_OT_NONE && bonuslen != 0));
409 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
410 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
412 for (i = 0; i < TXG_SIZE; i++)
413 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
415 /* clean up any unreferenced dbufs */
416 (void) dnode_evict_dbufs(dn, 0);
417 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
420 * XXX I should really have a generation number to tell if we
423 if (blocksize != dn->dn_datablksz ||
424 dn->dn_bonustype != bonustype || dn->dn_bonuslen != bonuslen) {
425 /* free all old data */
426 dnode_free_range(dn, 0, -1ULL, tx);
429 /* change blocksize */
430 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
431 if (blocksize != dn->dn_datablksz &&
432 (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
433 list_head(&dn->dn_dbufs) != NULL)) {
434 db = dbuf_hold(dn, 0, FTAG);
435 dbuf_new_size(db, blocksize, tx);
437 dnode_setdblksz(dn, blocksize);
438 dnode_setdirty(dn, tx);
439 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
440 rw_exit(&dn->dn_struct_rwlock);
449 if (dn->dn_bonuslen != bonuslen) {
450 /* change bonus size */
452 bonuslen = 1; /* XXX */
453 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
454 if (dn->dn_bonus == NULL)
455 dn->dn_bonus = dbuf_create_bonus(dn);
457 rw_exit(&dn->dn_struct_rwlock);
458 if (refcount_add(&db->db_holds, FTAG) == 1)
459 dnode_add_ref(dn, db);
460 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED));
461 mutex_enter(&db->db_mtx);
462 ASSERT3U(db->db.db_size, ==, dn->dn_bonuslen);
463 ASSERT(db->db.db_data != NULL);
464 db->db.db_size = bonuslen;
465 mutex_exit(&db->db_mtx);
466 (void) dbuf_dirty(db, tx);
469 /* change bonus size and type */
470 mutex_enter(&dn->dn_mtx);
471 dn->dn_bonustype = bonustype;
472 dn->dn_bonuslen = bonuslen;
473 dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
474 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
475 dn->dn_compress = ZIO_COMPRESS_INHERIT;
476 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
479 * NB: we have to do the dbuf_rele after we've changed the
480 * dn_bonuslen, for the sake of dbuf_verify().
485 dn->dn_allocated_txg = tx->tx_txg;
486 mutex_exit(&dn->dn_mtx);
490 dnode_special_close(dnode_t *dn)
493 * Wait for final references to the dnode to clear. This can
494 * only happen if the arc is asyncronously evicting state that
495 * has a hold on this dnode while we are trying to evict this
498 while (refcount_count(&dn->dn_holds) > 0)
504 dnode_special_open(objset_impl_t *os, dnode_phys_t *dnp, uint64_t object)
506 dnode_t *dn = dnode_create(os, dnp, NULL, object);
512 dnode_buf_pageout(dmu_buf_t *db, void *arg)
514 dnode_t **children_dnodes = arg;
516 int epb = db->db_size >> DNODE_SHIFT;
518 for (i = 0; i < epb; i++) {
519 dnode_t *dn = children_dnodes[i];
526 * If there are holds on this dnode, then there should
527 * be holds on the dnode's containing dbuf as well; thus
528 * it wouldn't be eligable for eviction and this function
529 * would not have been called.
531 ASSERT(refcount_is_zero(&dn->dn_holds));
532 ASSERT(list_head(&dn->dn_dbufs) == NULL);
533 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
535 for (n = 0; n < TXG_SIZE; n++)
536 ASSERT(!list_link_active(&dn->dn_dirty_link[n]));
538 children_dnodes[i] = NULL;
541 kmem_free(children_dnodes, epb * sizeof (dnode_t *));
546 * EINVAL - invalid object number.
548 * succeeds even for free dnodes.
551 dnode_hold_impl(objset_impl_t *os, uint64_t object, int flag,
552 void *tag, dnode_t **dnp)
555 int drop_struct_lock = FALSE;
560 dnode_t **children_dnodes;
562 if (object == 0 || object >= DN_MAX_OBJECT)
565 mdn = os->os_meta_dnode;
569 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
570 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
571 drop_struct_lock = TRUE;
574 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
576 db = dbuf_hold(mdn, blk, FTAG);
577 if (drop_struct_lock)
578 rw_exit(&mdn->dn_struct_rwlock);
581 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
587 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
588 epb = db->db.db_size >> DNODE_SHIFT;
590 idx = object & (epb-1);
592 children_dnodes = dmu_buf_get_user(&db->db);
593 if (children_dnodes == NULL) {
595 children_dnodes = kmem_zalloc(epb * sizeof (dnode_t *),
597 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
598 dnode_buf_pageout)) {
599 kmem_free(children_dnodes, epb * sizeof (dnode_t *));
600 children_dnodes = winner;
604 if ((dn = children_dnodes[idx]) == NULL) {
606 dn = dnode_create(os, (dnode_phys_t *)db->db.db_data+idx,
608 winner = atomic_cas_ptr(&children_dnodes[idx], NULL, dn);
609 if (winner != NULL) {
615 mutex_enter(&dn->dn_mtx);
617 if (dn->dn_free_txg ||
618 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
619 ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)) {
620 mutex_exit(&dn->dn_mtx);
622 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
624 mutex_exit(&dn->dn_mtx);
626 if (refcount_add(&dn->dn_holds, tag) == 1)
627 dbuf_add_ref(db, dn);
630 ASSERT3P(dn->dn_dbuf, ==, db);
631 ASSERT3U(dn->dn_object, ==, object);
639 * Return held dnode if the object is allocated, NULL if not.
642 dnode_hold(objset_impl_t *os, uint64_t object, void *tag, dnode_t **dnp)
644 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
648 dnode_add_ref(dnode_t *dn, void *tag)
650 ASSERT(refcount_count(&dn->dn_holds) > 0);
651 (void) refcount_add(&dn->dn_holds, tag);
655 dnode_rele(dnode_t *dn, void *tag)
659 refs = refcount_remove(&dn->dn_holds, tag);
660 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
661 if (refs == 0 && dn->dn_dbuf)
662 dbuf_rele(dn->dn_dbuf, dn);
666 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
668 objset_impl_t *os = dn->dn_objset;
669 uint64_t txg = tx->tx_txg;
671 if (dn->dn_object == DMU_META_DNODE_OBJECT)
677 mutex_enter(&dn->dn_mtx);
678 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
679 /* ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); */
680 mutex_exit(&dn->dn_mtx);
683 mutex_enter(&os->os_lock);
686 * If we are already marked dirty, we're done.
688 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
689 mutex_exit(&os->os_lock);
693 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
694 ASSERT(dn->dn_datablksz != 0);
695 ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0);
697 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
700 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
701 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
703 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
706 mutex_exit(&os->os_lock);
709 * The dnode maintains a hold on its containing dbuf as
710 * long as there are holds on it. Each instantiated child
711 * dbuf maintaines a hold on the dnode. When the last child
712 * drops its hold, the dnode will drop its hold on the
713 * containing dbuf. We add a "dirty hold" here so that the
714 * dnode will hang around after we finish processing its
717 dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg);
719 (void) dbuf_dirty(dn->dn_dbuf, tx);
721 dsl_dataset_dirty(os->os_dsl_dataset, tx);
725 dnode_free(dnode_t *dn, dmu_tx_t *tx)
727 int txgoff = tx->tx_txg & TXG_MASK;
729 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
731 /* we should be the only holder... hopefully */
732 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
734 mutex_enter(&dn->dn_mtx);
735 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
736 mutex_exit(&dn->dn_mtx);
739 dn->dn_free_txg = tx->tx_txg;
740 mutex_exit(&dn->dn_mtx);
743 * If the dnode is already dirty, it needs to be moved from
744 * the dirty list to the free list.
746 mutex_enter(&dn->dn_objset->os_lock);
747 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
748 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
749 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
750 mutex_exit(&dn->dn_objset->os_lock);
752 mutex_exit(&dn->dn_objset->os_lock);
753 dnode_setdirty(dn, tx);
758 * Try to change the block size for the indicated dnode. This can only
759 * succeed if there are no blocks allocated or dirty beyond first block
762 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
764 dmu_buf_impl_t *db, *db_next;
765 int have_db0 = FALSE;
768 size = SPA_MINBLOCKSIZE;
769 if (size > SPA_MAXBLOCKSIZE)
770 size = SPA_MAXBLOCKSIZE;
772 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
774 if (ibs == dn->dn_indblkshift)
777 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
780 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
782 /* Check for any allocated blocks beyond the first */
783 if (dn->dn_phys->dn_maxblkid != 0)
786 mutex_enter(&dn->dn_dbufs_mtx);
787 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
788 db_next = list_next(&dn->dn_dbufs, db);
790 if (db->db_blkid == 0) {
792 } else if (db->db_blkid != DB_BONUS_BLKID) {
793 mutex_exit(&dn->dn_dbufs_mtx);
797 mutex_exit(&dn->dn_dbufs_mtx);
799 if (ibs && dn->dn_nlevels != 1)
803 if (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || have_db0) {
804 /* obtain the old block */
805 db = dbuf_hold(dn, 0, FTAG);
806 dbuf_new_size(db, size, tx);
809 dnode_setdblksz(dn, size);
810 dnode_setdirty(dn, tx);
811 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
813 dn->dn_indblkshift = ibs;
814 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
820 rw_exit(&dn->dn_struct_rwlock);
824 rw_exit(&dn->dn_struct_rwlock);
829 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx)
831 uint64_t txgoff = tx->tx_txg & TXG_MASK;
832 int drop_struct_lock = FALSE;
833 int epbs, new_nlevels;
836 ASSERT(blkid != DB_BONUS_BLKID);
838 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
839 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
840 drop_struct_lock = TRUE;
843 if (blkid <= dn->dn_maxblkid)
846 dn->dn_maxblkid = blkid;
849 * Compute the number of levels necessary to support the new maxblkid.
852 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
853 for (sz = dn->dn_nblkptr;
854 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
857 if (new_nlevels > dn->dn_nlevels) {
858 int old_nlevels = dn->dn_nlevels;
861 dbuf_dirty_record_t *new, *dr, *dr_next;
863 dn->dn_nlevels = new_nlevels;
865 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
866 dn->dn_next_nlevels[txgoff] = new_nlevels;
868 /* dirty the left indirects */
869 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
870 new = dbuf_dirty(db, tx);
873 /* transfer the dirty records to the new indirect */
874 mutex_enter(&dn->dn_mtx);
875 mutex_enter(&new->dt.di.dr_mtx);
876 list = &dn->dn_dirty_records[txgoff];
877 for (dr = list_head(list); dr; dr = dr_next) {
878 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
879 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
880 dr->dr_dbuf->db_blkid != DB_BONUS_BLKID) {
881 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
882 list_remove(&dn->dn_dirty_records[txgoff], dr);
883 list_insert_tail(&new->dt.di.dr_children, dr);
887 mutex_exit(&new->dt.di.dr_mtx);
888 mutex_exit(&dn->dn_mtx);
892 if (drop_struct_lock)
893 rw_exit(&dn->dn_struct_rwlock);
897 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
899 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
902 free_range_t rp_tofind;
903 uint64_t endblk = blkid + nblks;
905 ASSERT(MUTEX_HELD(&dn->dn_mtx));
906 ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
908 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
909 blkid, nblks, tx->tx_txg);
910 rp_tofind.fr_blkid = blkid;
911 rp = avl_find(tree, &rp_tofind, &where);
913 rp = avl_nearest(tree, where, AVL_BEFORE);
915 rp = avl_nearest(tree, where, AVL_AFTER);
917 while (rp && (rp->fr_blkid <= blkid + nblks)) {
918 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
919 free_range_t *nrp = AVL_NEXT(tree, rp);
921 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
922 /* clear this entire range */
923 avl_remove(tree, rp);
924 kmem_free(rp, sizeof (free_range_t));
925 } else if (blkid <= rp->fr_blkid &&
926 endblk > rp->fr_blkid && endblk < fr_endblk) {
927 /* clear the beginning of this range */
928 rp->fr_blkid = endblk;
929 rp->fr_nblks = fr_endblk - endblk;
930 } else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
931 endblk >= fr_endblk) {
932 /* clear the end of this range */
933 rp->fr_nblks = blkid - rp->fr_blkid;
934 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
935 /* clear a chunk out of this range */
936 free_range_t *new_rp =
937 kmem_alloc(sizeof (free_range_t), KM_SLEEP);
939 new_rp->fr_blkid = endblk;
940 new_rp->fr_nblks = fr_endblk - endblk;
941 avl_insert_here(tree, new_rp, rp, AVL_AFTER);
942 rp->fr_nblks = blkid - rp->fr_blkid;
944 /* there may be no overlap */
950 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
953 uint64_t blkoff, blkid, nblks;
957 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
958 blksz = dn->dn_datablksz;
960 /* If the range is past the end of the file, this is a no-op */
961 if (off >= blksz * (dn->dn_maxblkid+1))
964 len = UINT64_MAX - off;
969 * First, block align the region to free:
972 head = P2NPHASE(off, blksz);
973 blkoff = P2PHASE(off, blksz);
975 ASSERT(dn->dn_maxblkid == 0);
976 if (off == 0 && len >= blksz) {
977 /* Freeing the whole block; don't do any head. */
980 /* Freeing part of the block. */
982 ASSERT3U(head, >, 0);
986 /* zero out any partial block data at the start of the range */
988 ASSERT3U(blkoff + head, ==, blksz);
991 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
995 /* don't dirty if it isn't on disk and isn't dirty */
996 if (db->db_last_dirty ||
997 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
998 rw_exit(&dn->dn_struct_rwlock);
999 dbuf_will_dirty(db, tx);
1000 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1001 data = db->db.db_data;
1002 bzero(data + blkoff, head);
1004 dbuf_rele(db, FTAG);
1010 /* If the range was less than one block, we're done */
1011 if (len == 0 || off >= blksz * (dn->dn_maxblkid+1))
1016 * They are freeing the whole block of a
1017 * non-power-of-two blocksize file. Skip all the messy
1020 ASSERT3U(off, ==, 0);
1021 ASSERT3U(len, >=, blksz);
1026 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1027 int blkshift = dn->dn_datablkshift;
1029 /* If the remaining range is past end of file, we're done */
1030 if (off > dn->dn_maxblkid << blkshift)
1033 if (off + len == UINT64_MAX)
1036 tail = P2PHASE(len, blksz);
1038 ASSERT3U(P2PHASE(off, blksz), ==, 0);
1039 /* zero out any partial block data at the end of the range */
1043 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1044 TRUE, FTAG, &db) == 0) {
1045 /* don't dirty if not on disk and not dirty */
1046 if (db->db_last_dirty ||
1048 !BP_IS_HOLE(db->db_blkptr))) {
1049 rw_exit(&dn->dn_struct_rwlock);
1050 dbuf_will_dirty(db, tx);
1051 rw_enter(&dn->dn_struct_rwlock,
1053 bzero(db->db.db_data, tail);
1055 dbuf_rele(db, FTAG);
1059 /* If the range did not include a full block, we are done */
1063 /* dirty the left indirects */
1064 if (dn->dn_nlevels > 1 && off != 0) {
1065 db = dbuf_hold_level(dn, 1,
1066 (off - head) >> (blkshift + epbs), FTAG);
1067 dbuf_will_dirty(db, tx);
1068 dbuf_rele(db, FTAG);
1071 /* dirty the right indirects */
1072 if (dn->dn_nlevels > 1 && !trunc) {
1073 db = dbuf_hold_level(dn, 1,
1074 (off + len + tail - 1) >> (blkshift + epbs), FTAG);
1075 dbuf_will_dirty(db, tx);
1076 dbuf_rele(db, FTAG);
1080 * Finally, add this range to the dnode range list, we
1081 * will finish up this free operation in the syncing phase.
1083 ASSERT(IS_P2ALIGNED(off, 1<<blkshift));
1084 ASSERT(off + len == UINT64_MAX ||
1085 IS_P2ALIGNED(len, 1<<blkshift));
1086 blkid = off >> blkshift;
1087 nblks = len >> blkshift;
1090 dn->dn_maxblkid = (blkid ? blkid - 1 : 0);
1093 mutex_enter(&dn->dn_mtx);
1094 dnode_clear_range(dn, blkid, nblks, tx);
1096 free_range_t *rp, *found;
1098 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1100 /* Add new range to dn_ranges */
1101 rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1102 rp->fr_blkid = blkid;
1103 rp->fr_nblks = nblks;
1104 found = avl_find(tree, rp, &where);
1105 ASSERT(found == NULL);
1106 avl_insert(tree, rp, where);
1107 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1108 blkid, nblks, tx->tx_txg);
1110 mutex_exit(&dn->dn_mtx);
1112 dbuf_free_range(dn, blkid, nblks, tx);
1113 dnode_setdirty(dn, tx);
1115 rw_exit(&dn->dn_struct_rwlock);
1118 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1120 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1122 free_range_t range_tofind;
1123 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1126 if (blkid == DB_BONUS_BLKID)
1130 * If we're in the process of opening the pool, dp will not be
1131 * set yet, but there shouldn't be anything dirty.
1136 if (dn->dn_free_txg)
1140 * If dn_datablkshift is not set, then there's only a single
1141 * block, in which case there will never be a free range so it
1144 range_tofind.fr_blkid = blkid;
1145 mutex_enter(&dn->dn_mtx);
1146 for (i = 0; i < TXG_SIZE; i++) {
1147 free_range_t *range_found;
1150 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1152 ASSERT(range_found->fr_nblks > 0);
1155 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1157 range_found->fr_blkid + range_found->fr_nblks > blkid)
1160 mutex_exit(&dn->dn_mtx);
1161 return (i < TXG_SIZE);
1164 /* call from syncing context when we actually write/free space for this dnode */
1166 dnode_diduse_space(dnode_t *dn, int64_t delta)
1169 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1171 (u_longlong_t)dn->dn_phys->dn_used,
1174 mutex_enter(&dn->dn_mtx);
1175 space = DN_USED_BYTES(dn->dn_phys);
1177 ASSERT3U(space + delta, >=, space); /* no overflow */
1179 ASSERT3U(space, >=, -delta); /* no underflow */
1182 if (spa_version(dn->dn_objset->os_spa) < ZFS_VERSION_DNODE_BYTES) {
1183 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1184 ASSERT3U(P2PHASE(space, 1<<DEV_BSHIFT), ==, 0);
1185 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1187 dn->dn_phys->dn_used = space;
1188 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1190 mutex_exit(&dn->dn_mtx);
1194 * Call when we think we're going to write/free space in open context.
1195 * Be conservative (ie. OK to write less than this or free more than
1196 * this, but don't write more or free less).
1199 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1201 objset_impl_t *os = dn->dn_objset;
1202 dsl_dataset_t *ds = os->os_dsl_dataset;
1205 space = spa_get_asize(os->os_spa, space);
1208 dsl_dir_willuse_space(ds->ds_dir, space, tx);
1210 dmu_tx_willuse_space(tx, space);
1214 dnode_next_offset_level(dnode_t *dn, boolean_t hole, uint64_t *offset,
1215 int lvl, uint64_t blkfill, uint64_t txg)
1217 dmu_buf_impl_t *db = NULL;
1219 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1220 uint64_t epb = 1ULL << epbs;
1221 uint64_t minfill, maxfill;
1224 dprintf("probing object %llu offset %llx level %d of %u\n",
1225 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1227 if (lvl == dn->dn_phys->dn_nlevels) {
1229 epb = dn->dn_phys->dn_nblkptr;
1230 data = dn->dn_phys->dn_blkptr;
1232 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1233 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1235 if (error == ENOENT)
1236 return (hole ? 0 : ESRCH);
1239 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1241 dbuf_rele(db, FTAG);
1244 data = db->db.db_data;
1248 (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1250 } else if (lvl == 0) {
1251 dnode_phys_t *dnp = data;
1253 ASSERT(dn->dn_type == DMU_OT_DNODE);
1255 for (i = (*offset >> span) & (blkfill - 1); i < blkfill; i++) {
1256 boolean_t newcontents = B_TRUE;
1259 newcontents = B_FALSE;
1260 for (j = 0; j < dnp[i].dn_nblkptr; j++) {
1261 if (dnp[i].dn_blkptr[j].blk_birth > txg)
1262 newcontents = B_TRUE;
1265 if (!dnp[i].dn_type == hole && newcontents)
1267 *offset += 1ULL << span;
1272 blkptr_t *bp = data;
1273 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1275 maxfill = blkfill << ((lvl - 1) * epbs);
1282 for (i = (*offset >> span) & ((1ULL << epbs) - 1);
1284 if (bp[i].blk_fill >= minfill &&
1285 bp[i].blk_fill <= maxfill &&
1286 bp[i].blk_birth > txg)
1288 *offset += 1ULL << span;
1295 dbuf_rele(db, FTAG);
1301 * Find the next hole, data, or sparse region at or after *offset.
1302 * The value 'blkfill' tells us how many items we expect to find
1303 * in an L0 data block; this value is 1 for normal objects,
1304 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1305 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1309 * dnode_next_offset(dn, hole, offset, 1, 1, 0);
1310 * Finds the next hole/data in a file.
1311 * Used in dmu_offset_next().
1313 * dnode_next_offset(mdn, hole, offset, 0, DNODES_PER_BLOCK, txg);
1314 * Finds the next free/allocated dnode an objset's meta-dnode.
1315 * Only finds objects that have new contents since txg (ie.
1316 * bonus buffer changes and content removal are ignored).
1317 * Used in dmu_object_next().
1319 * dnode_next_offset(mdn, TRUE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1320 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1321 * Used in dmu_object_alloc().
1324 dnode_next_offset(dnode_t *dn, boolean_t hole, uint64_t *offset,
1325 int minlvl, uint64_t blkfill, uint64_t txg)
1329 uint64_t initial_offset = *offset;
1331 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1333 if (dn->dn_phys->dn_nlevels == 0) {
1334 rw_exit(&dn->dn_struct_rwlock);
1338 if (dn->dn_datablkshift == 0) {
1339 if (*offset < dn->dn_datablksz) {
1341 *offset = dn->dn_datablksz;
1345 rw_exit(&dn->dn_struct_rwlock);
1349 maxlvl = dn->dn_phys->dn_nlevels;
1351 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1352 error = dnode_next_offset_level(dn,
1353 hole, offset, lvl, blkfill, txg);
1358 while (--lvl >= minlvl && error == 0) {
1359 error = dnode_next_offset_level(dn,
1360 hole, offset, lvl, blkfill, txg);
1363 rw_exit(&dn->dn_struct_rwlock);
1365 if (error == 0 && initial_offset > *offset)