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 (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
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>
40 #include <sys/range_tree.h>
42 static kmem_cache_t *dnode_cache;
44 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
45 * turned on when DEBUG is also defined.
52 #define DNODE_STAT_ADD(stat) ((stat)++)
54 #define DNODE_STAT_ADD(stat) /* nothing */
55 #endif /* DNODE_STATS */
57 static dnode_phys_t dnode_phys_zero;
59 int zfs_default_bs = SPA_MINBLOCKSHIFT;
60 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
62 SYSCTL_DECL(_vfs_zfs);
63 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_bs, CTLFLAG_RWTUN,
64 &zfs_default_bs, 0, "Default dnode block shift");
65 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_ibs, CTLFLAG_RWTUN,
66 &zfs_default_ibs, 0, "Default dnode indirect block shift");
69 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
73 dbuf_compare(const void *x1, const void *x2)
75 const dmu_buf_impl_t *d1 = x1;
76 const dmu_buf_impl_t *d2 = x2;
78 if (d1->db_level < d2->db_level) {
81 if (d1->db_level > d2->db_level) {
85 if (d1->db_blkid < d2->db_blkid) {
88 if (d1->db_blkid > d2->db_blkid) {
92 if (d1->db_state == DB_SEARCH) {
93 ASSERT3S(d2->db_state, !=, DB_SEARCH);
95 } else if (d2->db_state == DB_SEARCH) {
96 ASSERT3S(d1->db_state, !=, DB_SEARCH);
100 if ((uintptr_t)d1 < (uintptr_t)d2) {
103 if ((uintptr_t)d1 > (uintptr_t)d2) {
111 dnode_cons(void *arg, void *unused, int kmflag)
116 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
117 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
118 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
119 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
122 * Every dbuf has a reference, and dropping a tracked reference is
123 * O(number of references), so don't track dn_holds.
125 refcount_create_untracked(&dn->dn_holds);
126 refcount_create(&dn->dn_tx_holds);
127 list_link_init(&dn->dn_link);
129 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
130 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
131 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
132 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
133 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
134 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
135 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
137 for (i = 0; i < TXG_SIZE; i++) {
138 list_link_init(&dn->dn_dirty_link[i]);
139 dn->dn_free_ranges[i] = NULL;
140 list_create(&dn->dn_dirty_records[i],
141 sizeof (dbuf_dirty_record_t),
142 offsetof(dbuf_dirty_record_t, dr_dirty_node));
145 dn->dn_allocated_txg = 0;
147 dn->dn_assigned_txg = 0;
149 dn->dn_dirtyctx_firstset = NULL;
151 dn->dn_have_spill = B_FALSE;
161 dn->dn_dbufs_count = 0;
162 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
163 offsetof(dmu_buf_impl_t, db_link));
166 POINTER_INVALIDATE(&dn->dn_objset);
172 dnode_dest(void *arg, void *unused)
177 rw_destroy(&dn->dn_struct_rwlock);
178 mutex_destroy(&dn->dn_mtx);
179 mutex_destroy(&dn->dn_dbufs_mtx);
180 cv_destroy(&dn->dn_notxholds);
181 refcount_destroy(&dn->dn_holds);
182 refcount_destroy(&dn->dn_tx_holds);
183 ASSERT(!list_link_active(&dn->dn_link));
185 for (i = 0; i < TXG_SIZE; i++) {
186 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
187 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
188 list_destroy(&dn->dn_dirty_records[i]);
189 ASSERT0(dn->dn_next_nblkptr[i]);
190 ASSERT0(dn->dn_next_nlevels[i]);
191 ASSERT0(dn->dn_next_indblkshift[i]);
192 ASSERT0(dn->dn_next_bonustype[i]);
193 ASSERT0(dn->dn_rm_spillblk[i]);
194 ASSERT0(dn->dn_next_bonuslen[i]);
195 ASSERT0(dn->dn_next_blksz[i]);
198 ASSERT0(dn->dn_allocated_txg);
199 ASSERT0(dn->dn_free_txg);
200 ASSERT0(dn->dn_assigned_txg);
201 ASSERT0(dn->dn_dirtyctx);
202 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
203 ASSERT3P(dn->dn_bonus, ==, NULL);
204 ASSERT(!dn->dn_have_spill);
205 ASSERT3P(dn->dn_zio, ==, NULL);
206 ASSERT0(dn->dn_oldused);
207 ASSERT0(dn->dn_oldflags);
208 ASSERT0(dn->dn_olduid);
209 ASSERT0(dn->dn_oldgid);
210 ASSERT0(dn->dn_newuid);
211 ASSERT0(dn->dn_newgid);
212 ASSERT0(dn->dn_id_flags);
214 ASSERT0(dn->dn_dbufs_count);
215 avl_destroy(&dn->dn_dbufs);
221 ASSERT(dnode_cache == NULL);
222 dnode_cache = kmem_cache_create("dnode_t",
224 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
225 kmem_cache_set_move(dnode_cache, dnode_move);
231 kmem_cache_destroy(dnode_cache);
238 dnode_verify(dnode_t *dn)
240 int drop_struct_lock = FALSE;
243 ASSERT(dn->dn_objset);
244 ASSERT(dn->dn_handle->dnh_dnode == dn);
246 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
248 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
251 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
252 rw_enter(&dn->dn_struct_rwlock, RW_READER);
253 drop_struct_lock = TRUE;
255 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
257 ASSERT3U(dn->dn_indblkshift, >=, 0);
258 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
259 if (dn->dn_datablkshift) {
260 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
261 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
262 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
264 ASSERT3U(dn->dn_nlevels, <=, 30);
265 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
266 ASSERT3U(dn->dn_nblkptr, >=, 1);
267 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
268 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
269 ASSERT3U(dn->dn_datablksz, ==,
270 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
271 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
272 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
273 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
274 for (i = 0; i < TXG_SIZE; i++) {
275 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
278 if (dn->dn_phys->dn_type != DMU_OT_NONE)
279 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
280 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
281 if (dn->dn_dbuf != NULL) {
282 ASSERT3P(dn->dn_phys, ==,
283 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
284 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
286 if (drop_struct_lock)
287 rw_exit(&dn->dn_struct_rwlock);
292 dnode_byteswap(dnode_phys_t *dnp)
294 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
297 if (dnp->dn_type == DMU_OT_NONE) {
298 bzero(dnp, sizeof (dnode_phys_t));
302 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
303 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
304 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
305 dnp->dn_used = BSWAP_64(dnp->dn_used);
308 * dn_nblkptr is only one byte, so it's OK to read it in either
309 * byte order. We can't read dn_bouslen.
311 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
312 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
313 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
314 buf64[i] = BSWAP_64(buf64[i]);
317 * OK to check dn_bonuslen for zero, because it won't matter if
318 * we have the wrong byte order. This is necessary because the
319 * dnode dnode is smaller than a regular dnode.
321 if (dnp->dn_bonuslen != 0) {
323 * Note that the bonus length calculated here may be
324 * longer than the actual bonus buffer. This is because
325 * we always put the bonus buffer after the last block
326 * pointer (instead of packing it against the end of the
329 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
330 size_t len = DN_MAX_BONUSLEN - off;
331 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
332 dmu_object_byteswap_t byteswap =
333 DMU_OT_BYTESWAP(dnp->dn_bonustype);
334 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
337 /* Swap SPILL block if we have one */
338 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
339 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
344 dnode_buf_byteswap(void *vbuf, size_t size)
346 dnode_phys_t *buf = vbuf;
349 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
350 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
352 size >>= DNODE_SHIFT;
353 for (i = 0; i < size; i++) {
360 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
362 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
364 dnode_setdirty(dn, tx);
365 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
366 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
367 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
368 dn->dn_bonuslen = newsize;
370 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
372 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
373 rw_exit(&dn->dn_struct_rwlock);
377 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
379 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
380 dnode_setdirty(dn, tx);
381 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
382 dn->dn_bonustype = newtype;
383 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
384 rw_exit(&dn->dn_struct_rwlock);
388 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
390 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
391 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
392 dnode_setdirty(dn, tx);
393 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
394 dn->dn_have_spill = B_FALSE;
398 dnode_setdblksz(dnode_t *dn, int size)
400 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
401 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
402 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
403 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
404 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
405 dn->dn_datablksz = size;
406 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
407 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
411 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
412 uint64_t object, dnode_handle_t *dnh)
416 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
417 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
421 * Defer setting dn_objset until the dnode is ready to be a candidate
422 * for the dnode_move() callback.
424 dn->dn_object = object;
429 if (dnp->dn_datablkszsec) {
430 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
432 dn->dn_datablksz = 0;
433 dn->dn_datablkszsec = 0;
434 dn->dn_datablkshift = 0;
436 dn->dn_indblkshift = dnp->dn_indblkshift;
437 dn->dn_nlevels = dnp->dn_nlevels;
438 dn->dn_type = dnp->dn_type;
439 dn->dn_nblkptr = dnp->dn_nblkptr;
440 dn->dn_checksum = dnp->dn_checksum;
441 dn->dn_compress = dnp->dn_compress;
442 dn->dn_bonustype = dnp->dn_bonustype;
443 dn->dn_bonuslen = dnp->dn_bonuslen;
444 dn->dn_maxblkid = dnp->dn_maxblkid;
445 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
448 dmu_zfetch_init(&dn->dn_zfetch, dn);
450 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
452 mutex_enter(&os->os_lock);
453 if (dnh->dnh_dnode != NULL) {
454 /* Lost the allocation race. */
455 mutex_exit(&os->os_lock);
456 kmem_cache_free(dnode_cache, dn);
457 return (dnh->dnh_dnode);
461 * Exclude special dnodes from os_dnodes so an empty os_dnodes
462 * signifies that the special dnodes have no references from
463 * their children (the entries in os_dnodes). This allows
464 * dnode_destroy() to easily determine if the last child has
465 * been removed and then complete eviction of the objset.
467 if (!DMU_OBJECT_IS_SPECIAL(object))
468 list_insert_head(&os->os_dnodes, dn);
472 * Everything else must be valid before assigning dn_objset
473 * makes the dnode eligible for dnode_move().
478 mutex_exit(&os->os_lock);
480 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
485 * Caller must be holding the dnode handle, which is released upon return.
488 dnode_destroy(dnode_t *dn)
490 objset_t *os = dn->dn_objset;
491 boolean_t complete_os_eviction = B_FALSE;
493 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
495 mutex_enter(&os->os_lock);
496 POINTER_INVALIDATE(&dn->dn_objset);
497 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
498 list_remove(&os->os_dnodes, dn);
499 complete_os_eviction =
500 list_is_empty(&os->os_dnodes) &&
501 list_link_active(&os->os_evicting_node);
503 mutex_exit(&os->os_lock);
505 /* the dnode can no longer move, so we can release the handle */
506 zrl_remove(&dn->dn_handle->dnh_zrlock);
508 dn->dn_allocated_txg = 0;
510 dn->dn_assigned_txg = 0;
513 if (dn->dn_dirtyctx_firstset != NULL) {
514 kmem_free(dn->dn_dirtyctx_firstset, 1);
515 dn->dn_dirtyctx_firstset = NULL;
517 if (dn->dn_bonus != NULL) {
518 mutex_enter(&dn->dn_bonus->db_mtx);
519 dbuf_destroy(dn->dn_bonus);
524 dn->dn_have_spill = B_FALSE;
533 dmu_zfetch_fini(&dn->dn_zfetch);
534 kmem_cache_free(dnode_cache, dn);
535 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
537 if (complete_os_eviction)
538 dmu_objset_evict_done(os);
542 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
543 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
547 ASSERT3U(blocksize, <=,
548 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
550 blocksize = 1 << zfs_default_bs;
552 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
555 ibs = zfs_default_ibs;
557 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
559 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
560 dn->dn_object, tx->tx_txg, blocksize, ibs);
562 ASSERT(dn->dn_type == DMU_OT_NONE);
563 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
564 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
565 ASSERT(ot != DMU_OT_NONE);
566 ASSERT(DMU_OT_IS_VALID(ot));
567 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
568 (bonustype == DMU_OT_SA && bonuslen == 0) ||
569 (bonustype != DMU_OT_NONE && bonuslen != 0));
570 ASSERT(DMU_OT_IS_VALID(bonustype));
571 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
572 ASSERT(dn->dn_type == DMU_OT_NONE);
573 ASSERT0(dn->dn_maxblkid);
574 ASSERT0(dn->dn_allocated_txg);
575 ASSERT0(dn->dn_assigned_txg);
576 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
577 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
578 ASSERT(avl_is_empty(&dn->dn_dbufs));
580 for (i = 0; i < TXG_SIZE; i++) {
581 ASSERT0(dn->dn_next_nblkptr[i]);
582 ASSERT0(dn->dn_next_nlevels[i]);
583 ASSERT0(dn->dn_next_indblkshift[i]);
584 ASSERT0(dn->dn_next_bonuslen[i]);
585 ASSERT0(dn->dn_next_bonustype[i]);
586 ASSERT0(dn->dn_rm_spillblk[i]);
587 ASSERT0(dn->dn_next_blksz[i]);
588 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
589 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
590 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
594 dnode_setdblksz(dn, blocksize);
595 dn->dn_indblkshift = ibs;
597 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
601 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
602 dn->dn_bonustype = bonustype;
603 dn->dn_bonuslen = bonuslen;
604 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
605 dn->dn_compress = ZIO_COMPRESS_INHERIT;
609 if (dn->dn_dirtyctx_firstset) {
610 kmem_free(dn->dn_dirtyctx_firstset, 1);
611 dn->dn_dirtyctx_firstset = NULL;
614 dn->dn_allocated_txg = tx->tx_txg;
617 dnode_setdirty(dn, tx);
618 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
619 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
620 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
621 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
625 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
626 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
630 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
631 ASSERT3U(blocksize, <=,
632 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
633 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
634 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
635 ASSERT(tx->tx_txg != 0);
636 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
637 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
638 (bonustype == DMU_OT_SA && bonuslen == 0));
639 ASSERT(DMU_OT_IS_VALID(bonustype));
640 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
642 /* clean up any unreferenced dbufs */
643 dnode_evict_dbufs(dn);
647 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
648 dnode_setdirty(dn, tx);
649 if (dn->dn_datablksz != blocksize) {
650 /* change blocksize */
651 ASSERT(dn->dn_maxblkid == 0 &&
652 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
653 dnode_block_freed(dn, 0)));
654 dnode_setdblksz(dn, blocksize);
655 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
657 if (dn->dn_bonuslen != bonuslen)
658 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
660 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
663 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
664 if (dn->dn_bonustype != bonustype)
665 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
666 if (dn->dn_nblkptr != nblkptr)
667 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
668 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
669 dbuf_rm_spill(dn, tx);
670 dnode_rm_spill(dn, tx);
672 rw_exit(&dn->dn_struct_rwlock);
677 /* change bonus size and type */
678 mutex_enter(&dn->dn_mtx);
679 dn->dn_bonustype = bonustype;
680 dn->dn_bonuslen = bonuslen;
681 dn->dn_nblkptr = nblkptr;
682 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
683 dn->dn_compress = ZIO_COMPRESS_INHERIT;
684 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
686 /* fix up the bonus db_size */
688 dn->dn_bonus->db.db_size =
689 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
690 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
693 dn->dn_allocated_txg = tx->tx_txg;
694 mutex_exit(&dn->dn_mtx);
699 uint64_t dms_dnode_invalid;
700 uint64_t dms_dnode_recheck1;
701 uint64_t dms_dnode_recheck2;
702 uint64_t dms_dnode_special;
703 uint64_t dms_dnode_handle;
704 uint64_t dms_dnode_rwlock;
705 uint64_t dms_dnode_active;
707 #endif /* DNODE_STATS */
710 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
714 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
715 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
716 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
717 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
720 ndn->dn_objset = odn->dn_objset;
721 ndn->dn_object = odn->dn_object;
722 ndn->dn_dbuf = odn->dn_dbuf;
723 ndn->dn_handle = odn->dn_handle;
724 ndn->dn_phys = odn->dn_phys;
725 ndn->dn_type = odn->dn_type;
726 ndn->dn_bonuslen = odn->dn_bonuslen;
727 ndn->dn_bonustype = odn->dn_bonustype;
728 ndn->dn_nblkptr = odn->dn_nblkptr;
729 ndn->dn_checksum = odn->dn_checksum;
730 ndn->dn_compress = odn->dn_compress;
731 ndn->dn_nlevels = odn->dn_nlevels;
732 ndn->dn_indblkshift = odn->dn_indblkshift;
733 ndn->dn_datablkshift = odn->dn_datablkshift;
734 ndn->dn_datablkszsec = odn->dn_datablkszsec;
735 ndn->dn_datablksz = odn->dn_datablksz;
736 ndn->dn_maxblkid = odn->dn_maxblkid;
737 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
738 sizeof (odn->dn_next_nblkptr));
739 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
740 sizeof (odn->dn_next_nlevels));
741 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
742 sizeof (odn->dn_next_indblkshift));
743 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
744 sizeof (odn->dn_next_bonustype));
745 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
746 sizeof (odn->dn_rm_spillblk));
747 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
748 sizeof (odn->dn_next_bonuslen));
749 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
750 sizeof (odn->dn_next_blksz));
751 for (i = 0; i < TXG_SIZE; i++) {
752 list_move_tail(&ndn->dn_dirty_records[i],
753 &odn->dn_dirty_records[i]);
755 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
756 sizeof (odn->dn_free_ranges));
757 ndn->dn_allocated_txg = odn->dn_allocated_txg;
758 ndn->dn_free_txg = odn->dn_free_txg;
759 ndn->dn_assigned_txg = odn->dn_assigned_txg;
760 ndn->dn_dirtyctx = odn->dn_dirtyctx;
761 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
762 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
763 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
764 ASSERT(avl_is_empty(&ndn->dn_dbufs));
765 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
766 ndn->dn_dbufs_count = odn->dn_dbufs_count;
767 ndn->dn_bonus = odn->dn_bonus;
768 ndn->dn_have_spill = odn->dn_have_spill;
769 ndn->dn_zio = odn->dn_zio;
770 ndn->dn_oldused = odn->dn_oldused;
771 ndn->dn_oldflags = odn->dn_oldflags;
772 ndn->dn_olduid = odn->dn_olduid;
773 ndn->dn_oldgid = odn->dn_oldgid;
774 ndn->dn_newuid = odn->dn_newuid;
775 ndn->dn_newgid = odn->dn_newgid;
776 ndn->dn_id_flags = odn->dn_id_flags;
777 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
778 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
779 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
782 * Update back pointers. Updating the handle fixes the back pointer of
783 * every descendant dbuf as well as the bonus dbuf.
785 ASSERT(ndn->dn_handle->dnh_dnode == odn);
786 ndn->dn_handle->dnh_dnode = ndn;
787 if (ndn->dn_zfetch.zf_dnode == odn) {
788 ndn->dn_zfetch.zf_dnode = ndn;
792 * Invalidate the original dnode by clearing all of its back pointers.
795 odn->dn_handle = NULL;
796 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
797 offsetof(dmu_buf_impl_t, db_link));
798 odn->dn_dbufs_count = 0;
799 odn->dn_bonus = NULL;
800 odn->dn_zfetch.zf_dnode = NULL;
803 * Set the low bit of the objset pointer to ensure that dnode_move()
804 * recognizes the dnode as invalid in any subsequent callback.
806 POINTER_INVALIDATE(&odn->dn_objset);
809 * Satisfy the destructor.
811 for (i = 0; i < TXG_SIZE; i++) {
812 list_create(&odn->dn_dirty_records[i],
813 sizeof (dbuf_dirty_record_t),
814 offsetof(dbuf_dirty_record_t, dr_dirty_node));
815 odn->dn_free_ranges[i] = NULL;
816 odn->dn_next_nlevels[i] = 0;
817 odn->dn_next_indblkshift[i] = 0;
818 odn->dn_next_bonustype[i] = 0;
819 odn->dn_rm_spillblk[i] = 0;
820 odn->dn_next_bonuslen[i] = 0;
821 odn->dn_next_blksz[i] = 0;
823 odn->dn_allocated_txg = 0;
824 odn->dn_free_txg = 0;
825 odn->dn_assigned_txg = 0;
826 odn->dn_dirtyctx = 0;
827 odn->dn_dirtyctx_firstset = NULL;
828 odn->dn_have_spill = B_FALSE;
831 odn->dn_oldflags = 0;
836 odn->dn_id_flags = 0;
842 odn->dn_moved = (uint8_t)-1;
849 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
851 dnode_t *odn = buf, *ndn = newbuf;
857 * The dnode is on the objset's list of known dnodes if the objset
858 * pointer is valid. We set the low bit of the objset pointer when
859 * freeing the dnode to invalidate it, and the memory patterns written
860 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
861 * A newly created dnode sets the objset pointer last of all to indicate
862 * that the dnode is known and in a valid state to be moved by this
866 if (!POINTER_IS_VALID(os)) {
867 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
868 return (KMEM_CBRC_DONT_KNOW);
872 * Ensure that the objset does not go away during the move.
874 rw_enter(&os_lock, RW_WRITER);
875 if (os != odn->dn_objset) {
877 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
878 return (KMEM_CBRC_DONT_KNOW);
882 * If the dnode is still valid, then so is the objset. We know that no
883 * valid objset can be freed while we hold os_lock, so we can safely
884 * ensure that the objset remains in use.
886 mutex_enter(&os->os_lock);
889 * Recheck the objset pointer in case the dnode was removed just before
890 * acquiring the lock.
892 if (os != odn->dn_objset) {
893 mutex_exit(&os->os_lock);
895 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
896 return (KMEM_CBRC_DONT_KNOW);
900 * At this point we know that as long as we hold os->os_lock, the dnode
901 * cannot be freed and fields within the dnode can be safely accessed.
902 * The objset listing this dnode cannot go away as long as this dnode is
906 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
907 mutex_exit(&os->os_lock);
908 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
909 return (KMEM_CBRC_NO);
911 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
914 * Lock the dnode handle to prevent the dnode from obtaining any new
915 * holds. This also prevents the descendant dbufs and the bonus dbuf
916 * from accessing the dnode, so that we can discount their holds. The
917 * handle is safe to access because we know that while the dnode cannot
918 * go away, neither can its handle. Once we hold dnh_zrlock, we can
919 * safely move any dnode referenced only by dbufs.
921 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
922 mutex_exit(&os->os_lock);
923 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
924 return (KMEM_CBRC_LATER);
928 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
929 * We need to guarantee that there is a hold for every dbuf in order to
930 * determine whether the dnode is actively referenced. Falsely matching
931 * a dbuf to an active hold would lead to an unsafe move. It's possible
932 * that a thread already having an active dnode hold is about to add a
933 * dbuf, and we can't compare hold and dbuf counts while the add is in
936 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
937 zrl_exit(&odn->dn_handle->dnh_zrlock);
938 mutex_exit(&os->os_lock);
939 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
940 return (KMEM_CBRC_LATER);
944 * A dbuf may be removed (evicted) without an active dnode hold. In that
945 * case, the dbuf count is decremented under the handle lock before the
946 * dbuf's hold is released. This order ensures that if we count the hold
947 * after the dbuf is removed but before its hold is released, we will
948 * treat the unmatched hold as active and exit safely. If we count the
949 * hold before the dbuf is removed, the hold is discounted, and the
950 * removal is blocked until the move completes.
952 refcount = refcount_count(&odn->dn_holds);
953 ASSERT(refcount >= 0);
954 dbufs = odn->dn_dbufs_count;
956 /* We can't have more dbufs than dnode holds. */
957 ASSERT3U(dbufs, <=, refcount);
958 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
961 if (refcount > dbufs) {
962 rw_exit(&odn->dn_struct_rwlock);
963 zrl_exit(&odn->dn_handle->dnh_zrlock);
964 mutex_exit(&os->os_lock);
965 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
966 return (KMEM_CBRC_LATER);
969 rw_exit(&odn->dn_struct_rwlock);
972 * At this point we know that anyone with a hold on the dnode is not
973 * actively referencing it. The dnode is known and in a valid state to
974 * move. We're holding the locks needed to execute the critical section.
976 dnode_move_impl(odn, ndn);
978 list_link_replace(&odn->dn_link, &ndn->dn_link);
979 /* If the dnode was safe to move, the refcount cannot have changed. */
980 ASSERT(refcount == refcount_count(&ndn->dn_holds));
981 ASSERT(dbufs == ndn->dn_dbufs_count);
982 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
983 mutex_exit(&os->os_lock);
985 return (KMEM_CBRC_YES);
991 dnode_special_close(dnode_handle_t *dnh)
993 dnode_t *dn = dnh->dnh_dnode;
996 * Wait for final references to the dnode to clear. This can
997 * only happen if the arc is asyncronously evicting state that
998 * has a hold on this dnode while we are trying to evict this
1001 while (refcount_count(&dn->dn_holds) > 0)
1003 ASSERT(dn->dn_dbuf == NULL ||
1004 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1005 zrl_add(&dnh->dnh_zrlock);
1006 dnode_destroy(dn); /* implicit zrl_remove() */
1007 zrl_destroy(&dnh->dnh_zrlock);
1008 dnh->dnh_dnode = NULL;
1012 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1013 dnode_handle_t *dnh)
1017 dn = dnode_create(os, dnp, NULL, object, dnh);
1018 zrl_init(&dnh->dnh_zrlock);
1023 dnode_buf_evict_async(void *dbu)
1025 dnode_children_t *children_dnodes = dbu;
1028 for (i = 0; i < children_dnodes->dnc_count; i++) {
1029 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1033 * The dnode handle lock guards against the dnode moving to
1034 * another valid address, so there is no need here to guard
1035 * against changes to or from NULL.
1037 if (dnh->dnh_dnode == NULL) {
1038 zrl_destroy(&dnh->dnh_zrlock);
1042 zrl_add(&dnh->dnh_zrlock);
1043 dn = dnh->dnh_dnode;
1045 * If there are holds on this dnode, then there should
1046 * be holds on the dnode's containing dbuf as well; thus
1047 * it wouldn't be eligible for eviction and this function
1048 * would not have been called.
1050 ASSERT(refcount_is_zero(&dn->dn_holds));
1051 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1053 dnode_destroy(dn); /* implicit zrl_remove() */
1054 zrl_destroy(&dnh->dnh_zrlock);
1055 dnh->dnh_dnode = NULL;
1057 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1058 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1063 * EINVAL - invalid object number.
1065 * succeeds even for free dnodes.
1068 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1069 void *tag, dnode_t **dnp)
1072 int drop_struct_lock = FALSE;
1077 dnode_children_t *children_dnodes;
1078 dnode_handle_t *dnh;
1081 * If you are holding the spa config lock as writer, you shouldn't
1082 * be asking the DMU to do *anything* unless it's the root pool
1083 * which may require us to read from the root filesystem while
1084 * holding some (not all) of the locks as writer.
1086 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1087 (spa_is_root(os->os_spa) &&
1088 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1090 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1091 dn = (object == DMU_USERUSED_OBJECT) ?
1092 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1094 return (SET_ERROR(ENOENT));
1096 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1097 return (SET_ERROR(ENOENT));
1098 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1099 return (SET_ERROR(EEXIST));
1101 (void) refcount_add(&dn->dn_holds, tag);
1106 if (object == 0 || object >= DN_MAX_OBJECT)
1107 return (SET_ERROR(EINVAL));
1109 mdn = DMU_META_DNODE(os);
1110 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1114 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1115 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1116 drop_struct_lock = TRUE;
1119 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1121 db = dbuf_hold(mdn, blk, FTAG);
1122 if (drop_struct_lock)
1123 rw_exit(&mdn->dn_struct_rwlock);
1125 return (SET_ERROR(EIO));
1126 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1128 dbuf_rele(db, FTAG);
1132 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1133 epb = db->db.db_size >> DNODE_SHIFT;
1135 idx = object & (epb-1);
1137 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1138 children_dnodes = dmu_buf_get_user(&db->db);
1139 if (children_dnodes == NULL) {
1141 dnode_children_t *winner;
1142 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1143 epb * sizeof (dnode_handle_t), KM_SLEEP);
1144 children_dnodes->dnc_count = epb;
1145 dnh = &children_dnodes->dnc_children[0];
1146 for (i = 0; i < epb; i++) {
1147 zrl_init(&dnh[i].dnh_zrlock);
1149 dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1150 dnode_buf_evict_async, NULL);
1151 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1152 if (winner != NULL) {
1154 for (i = 0; i < epb; i++) {
1155 zrl_destroy(&dnh[i].dnh_zrlock);
1158 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1159 epb * sizeof (dnode_handle_t));
1160 children_dnodes = winner;
1163 ASSERT(children_dnodes->dnc_count == epb);
1165 dnh = &children_dnodes->dnc_children[idx];
1166 zrl_add(&dnh->dnh_zrlock);
1167 dn = dnh->dnh_dnode;
1169 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1171 dn = dnode_create(os, phys, db, object, dnh);
1174 mutex_enter(&dn->dn_mtx);
1176 if (dn->dn_free_txg ||
1177 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1178 ((flag & DNODE_MUST_BE_FREE) &&
1179 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1180 mutex_exit(&dn->dn_mtx);
1181 zrl_remove(&dnh->dnh_zrlock);
1182 dbuf_rele(db, FTAG);
1183 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1185 if (refcount_add(&dn->dn_holds, tag) == 1)
1186 dbuf_add_ref(db, dnh);
1187 mutex_exit(&dn->dn_mtx);
1189 /* Now we can rely on the hold to prevent the dnode from moving. */
1190 zrl_remove(&dnh->dnh_zrlock);
1193 ASSERT3P(dn->dn_dbuf, ==, db);
1194 ASSERT3U(dn->dn_object, ==, object);
1195 dbuf_rele(db, FTAG);
1202 * Return held dnode if the object is allocated, NULL if not.
1205 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1207 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1211 * Can only add a reference if there is already at least one
1212 * reference on the dnode. Returns FALSE if unable to add a
1216 dnode_add_ref(dnode_t *dn, void *tag)
1218 mutex_enter(&dn->dn_mtx);
1219 if (refcount_is_zero(&dn->dn_holds)) {
1220 mutex_exit(&dn->dn_mtx);
1223 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1224 mutex_exit(&dn->dn_mtx);
1229 dnode_rele(dnode_t *dn, void *tag)
1231 mutex_enter(&dn->dn_mtx);
1232 dnode_rele_and_unlock(dn, tag);
1236 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1239 /* Get while the hold prevents the dnode from moving. */
1240 dmu_buf_impl_t *db = dn->dn_dbuf;
1241 dnode_handle_t *dnh = dn->dn_handle;
1243 refs = refcount_remove(&dn->dn_holds, tag);
1244 mutex_exit(&dn->dn_mtx);
1247 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1248 * indirectly by dbuf_rele() while relying on the dnode handle to
1249 * prevent the dnode from moving, since releasing the last hold could
1250 * result in the dnode's parent dbuf evicting its dnode handles. For
1251 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1252 * other direct or indirect hold on the dnode must first drop the dnode
1255 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1257 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1258 if (refs == 0 && db != NULL) {
1260 * Another thread could add a hold to the dnode handle in
1261 * dnode_hold_impl() while holding the parent dbuf. Since the
1262 * hold on the parent dbuf prevents the handle from being
1263 * destroyed, the hold on the handle is OK. We can't yet assert
1264 * that the handle has zero references, but that will be
1265 * asserted anyway when the handle gets destroyed.
1272 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1274 objset_t *os = dn->dn_objset;
1275 uint64_t txg = tx->tx_txg;
1277 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1278 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1285 mutex_enter(&dn->dn_mtx);
1286 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1287 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1288 mutex_exit(&dn->dn_mtx);
1292 * Determine old uid/gid when necessary
1294 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1296 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1297 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1300 * If we are already marked dirty, we're done.
1302 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1303 multilist_sublist_unlock(mls);
1307 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1308 !avl_is_empty(&dn->dn_dbufs));
1309 ASSERT(dn->dn_datablksz != 0);
1310 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1311 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1312 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1314 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1315 dn->dn_object, txg);
1317 multilist_sublist_insert_head(mls, dn);
1319 multilist_sublist_unlock(mls);
1322 * The dnode maintains a hold on its containing dbuf as
1323 * long as there are holds on it. Each instantiated child
1324 * dbuf maintains a hold on the dnode. When the last child
1325 * drops its hold, the dnode will drop its hold on the
1326 * containing dbuf. We add a "dirty hold" here so that the
1327 * dnode will hang around after we finish processing its
1330 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1332 (void) dbuf_dirty(dn->dn_dbuf, tx);
1334 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1338 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1340 mutex_enter(&dn->dn_mtx);
1341 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1342 mutex_exit(&dn->dn_mtx);
1345 dn->dn_free_txg = tx->tx_txg;
1346 mutex_exit(&dn->dn_mtx);
1348 dnode_setdirty(dn, tx);
1352 * Try to change the block size for the indicated dnode. This can only
1353 * succeed if there are no blocks allocated or dirty beyond first block
1356 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1361 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1363 size = SPA_MINBLOCKSIZE;
1365 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1367 if (ibs == dn->dn_indblkshift)
1370 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1373 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1375 /* Check for any allocated blocks beyond the first */
1376 if (dn->dn_maxblkid != 0)
1379 mutex_enter(&dn->dn_dbufs_mtx);
1380 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1381 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1382 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1383 db->db_blkid != DMU_SPILL_BLKID) {
1384 mutex_exit(&dn->dn_dbufs_mtx);
1388 mutex_exit(&dn->dn_dbufs_mtx);
1390 if (ibs && dn->dn_nlevels != 1)
1393 /* resize the old block */
1394 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1396 dbuf_new_size(db, size, tx);
1397 else if (err != ENOENT)
1400 dnode_setdblksz(dn, size);
1401 dnode_setdirty(dn, tx);
1402 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1404 dn->dn_indblkshift = ibs;
1405 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1407 /* rele after we have fixed the blocksize in the dnode */
1409 dbuf_rele(db, FTAG);
1411 rw_exit(&dn->dn_struct_rwlock);
1415 rw_exit(&dn->dn_struct_rwlock);
1416 return (SET_ERROR(ENOTSUP));
1419 /* read-holding callers must not rely on the lock being continuously held */
1421 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1423 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1424 int epbs, new_nlevels;
1427 ASSERT(blkid != DMU_BONUS_BLKID);
1430 RW_READ_HELD(&dn->dn_struct_rwlock) :
1431 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1434 * if we have a read-lock, check to see if we need to do any work
1435 * before upgrading to a write-lock.
1438 if (blkid <= dn->dn_maxblkid)
1441 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1442 rw_exit(&dn->dn_struct_rwlock);
1443 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1447 if (blkid <= dn->dn_maxblkid)
1450 dn->dn_maxblkid = blkid;
1453 * Compute the number of levels necessary to support the new maxblkid.
1456 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1457 for (sz = dn->dn_nblkptr;
1458 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1461 if (new_nlevels > dn->dn_nlevels) {
1462 int old_nlevels = dn->dn_nlevels;
1465 dbuf_dirty_record_t *new, *dr, *dr_next;
1467 dn->dn_nlevels = new_nlevels;
1469 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1470 dn->dn_next_nlevels[txgoff] = new_nlevels;
1472 /* dirty the left indirects */
1473 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1475 new = dbuf_dirty(db, tx);
1476 dbuf_rele(db, FTAG);
1478 /* transfer the dirty records to the new indirect */
1479 mutex_enter(&dn->dn_mtx);
1480 mutex_enter(&new->dt.di.dr_mtx);
1481 list = &dn->dn_dirty_records[txgoff];
1482 for (dr = list_head(list); dr; dr = dr_next) {
1483 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1484 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1485 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1486 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1487 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1488 list_remove(&dn->dn_dirty_records[txgoff], dr);
1489 list_insert_tail(&new->dt.di.dr_children, dr);
1490 dr->dr_parent = new;
1493 mutex_exit(&new->dt.di.dr_mtx);
1494 mutex_exit(&dn->dn_mtx);
1499 rw_downgrade(&dn->dn_struct_rwlock);
1503 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1505 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1507 dmu_buf_will_dirty(&db->db, tx);
1508 dbuf_rele(db, FTAG);
1513 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1516 uint64_t blkoff, blkid, nblks;
1517 int blksz, blkshift, head, tail;
1521 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1522 blksz = dn->dn_datablksz;
1523 blkshift = dn->dn_datablkshift;
1524 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1526 if (len == DMU_OBJECT_END) {
1527 len = UINT64_MAX - off;
1532 * First, block align the region to free:
1535 head = P2NPHASE(off, blksz);
1536 blkoff = P2PHASE(off, blksz);
1537 if ((off >> blkshift) > dn->dn_maxblkid)
1540 ASSERT(dn->dn_maxblkid == 0);
1541 if (off == 0 && len >= blksz) {
1543 * Freeing the whole block; fast-track this request.
1544 * Note that we won't dirty any indirect blocks,
1545 * which is fine because we will be freeing the entire
1546 * file and thus all indirect blocks will be freed
1547 * by free_children().
1552 } else if (off >= blksz) {
1553 /* Freeing past end-of-data */
1556 /* Freeing part of the block. */
1558 ASSERT3U(head, >, 0);
1562 /* zero out any partial block data at the start of the range */
1564 ASSERT3U(blkoff + head, ==, blksz);
1567 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1568 TRUE, FALSE, FTAG, &db) == 0) {
1571 /* don't dirty if it isn't on disk and isn't dirty */
1572 if (db->db_last_dirty ||
1573 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1574 rw_exit(&dn->dn_struct_rwlock);
1575 dmu_buf_will_dirty(&db->db, tx);
1576 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1577 data = db->db.db_data;
1578 bzero(data + blkoff, head);
1580 dbuf_rele(db, FTAG);
1586 /* If the range was less than one block, we're done */
1590 /* If the remaining range is past end of file, we're done */
1591 if ((off >> blkshift) > dn->dn_maxblkid)
1594 ASSERT(ISP2(blksz));
1598 tail = P2PHASE(len, blksz);
1600 ASSERT0(P2PHASE(off, blksz));
1601 /* zero out any partial block data at the end of the range */
1605 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1606 TRUE, FALSE, FTAG, &db) == 0) {
1607 /* don't dirty if not on disk and not dirty */
1608 if (db->db_last_dirty ||
1609 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1610 rw_exit(&dn->dn_struct_rwlock);
1611 dmu_buf_will_dirty(&db->db, tx);
1612 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1613 bzero(db->db.db_data, tail);
1615 dbuf_rele(db, FTAG);
1620 /* If the range did not include a full block, we are done */
1624 ASSERT(IS_P2ALIGNED(off, blksz));
1625 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1626 blkid = off >> blkshift;
1627 nblks = len >> blkshift;
1632 * Dirty all the indirect blocks in this range. Note that only
1633 * the first and last indirect blocks can actually be written
1634 * (if they were partially freed) -- they must be dirtied, even if
1635 * they do not exist on disk yet. The interior blocks will
1636 * be freed by free_children(), so they will not actually be written.
1637 * Even though these interior blocks will not be written, we
1638 * dirty them for two reasons:
1640 * - It ensures that the indirect blocks remain in memory until
1641 * syncing context. (They have already been prefetched by
1642 * dmu_tx_hold_free(), so we don't have to worry about reading
1643 * them serially here.)
1645 * - The dirty space accounting will put pressure on the txg sync
1646 * mechanism to begin syncing, and to delay transactions if there
1647 * is a large amount of freeing. Even though these indirect
1648 * blocks will not be written, we could need to write the same
1649 * amount of space if we copy the freed BPs into deadlists.
1651 if (dn->dn_nlevels > 1) {
1652 uint64_t first, last;
1654 first = blkid >> epbs;
1655 dnode_dirty_l1(dn, first, tx);
1657 last = dn->dn_maxblkid >> epbs;
1659 last = (blkid + nblks - 1) >> epbs;
1661 dnode_dirty_l1(dn, last, tx);
1663 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1665 for (uint64_t i = first + 1; i < last; i++) {
1667 * Set i to the blockid of the next non-hole
1668 * level-1 indirect block at or after i. Note
1669 * that dnode_next_offset() operates in terms of
1670 * level-0-equivalent bytes.
1672 uint64_t ibyte = i << shift;
1673 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1680 * Normally we should not see an error, either
1681 * from dnode_next_offset() or dbuf_hold_level()
1682 * (except for ESRCH from dnode_next_offset).
1683 * If there is an i/o error, then when we read
1684 * this block in syncing context, it will use
1685 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1686 * to the "failmode" property. dnode_next_offset()
1687 * doesn't have a flag to indicate MUSTSUCCEED.
1692 dnode_dirty_l1(dn, i, tx);
1698 * Add this range to the dnode range list.
1699 * We will finish up this free operation in the syncing phase.
1701 mutex_enter(&dn->dn_mtx);
1702 int txgoff = tx->tx_txg & TXG_MASK;
1703 if (dn->dn_free_ranges[txgoff] == NULL) {
1704 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
1706 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1707 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1708 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1709 blkid, nblks, tx->tx_txg);
1710 mutex_exit(&dn->dn_mtx);
1712 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1713 dnode_setdirty(dn, tx);
1716 rw_exit(&dn->dn_struct_rwlock);
1720 dnode_spill_freed(dnode_t *dn)
1724 mutex_enter(&dn->dn_mtx);
1725 for (i = 0; i < TXG_SIZE; i++) {
1726 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1729 mutex_exit(&dn->dn_mtx);
1730 return (i < TXG_SIZE);
1733 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1735 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1737 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1740 if (blkid == DMU_BONUS_BLKID)
1744 * If we're in the process of opening the pool, dp will not be
1745 * set yet, but there shouldn't be anything dirty.
1750 if (dn->dn_free_txg)
1753 if (blkid == DMU_SPILL_BLKID)
1754 return (dnode_spill_freed(dn));
1756 mutex_enter(&dn->dn_mtx);
1757 for (i = 0; i < TXG_SIZE; i++) {
1758 if (dn->dn_free_ranges[i] != NULL &&
1759 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1762 mutex_exit(&dn->dn_mtx);
1763 return (i < TXG_SIZE);
1766 /* call from syncing context when we actually write/free space for this dnode */
1768 dnode_diduse_space(dnode_t *dn, int64_t delta)
1771 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1773 (u_longlong_t)dn->dn_phys->dn_used,
1776 mutex_enter(&dn->dn_mtx);
1777 space = DN_USED_BYTES(dn->dn_phys);
1779 ASSERT3U(space + delta, >=, space); /* no overflow */
1781 ASSERT3U(space, >=, -delta); /* no underflow */
1784 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1785 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1786 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1787 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1789 dn->dn_phys->dn_used = space;
1790 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1792 mutex_exit(&dn->dn_mtx);
1796 * Scans a block at the indicated "level" looking for a hole or data,
1797 * depending on 'flags'.
1799 * If level > 0, then we are scanning an indirect block looking at its
1800 * pointers. If level == 0, then we are looking at a block of dnodes.
1802 * If we don't find what we are looking for in the block, we return ESRCH.
1803 * Otherwise, return with *offset pointing to the beginning (if searching
1804 * forwards) or end (if searching backwards) of the range covered by the
1805 * block pointer we matched on (or dnode).
1807 * The basic search algorithm used below by dnode_next_offset() is to
1808 * use this function to search up the block tree (widen the search) until
1809 * we find something (i.e., we don't return ESRCH) and then search back
1810 * down the tree (narrow the search) until we reach our original search
1814 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1815 int lvl, uint64_t blkfill, uint64_t txg)
1817 dmu_buf_impl_t *db = NULL;
1819 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1820 uint64_t epb = 1ULL << epbs;
1821 uint64_t minfill, maxfill;
1823 int i, inc, error, span;
1825 dprintf("probing object %llu offset %llx level %d of %u\n",
1826 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1828 hole = ((flags & DNODE_FIND_HOLE) != 0);
1829 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1830 ASSERT(txg == 0 || !hole);
1832 if (lvl == dn->dn_phys->dn_nlevels) {
1834 epb = dn->dn_phys->dn_nblkptr;
1835 data = dn->dn_phys->dn_blkptr;
1837 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1838 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1840 if (error != ENOENT)
1845 * This can only happen when we are searching up
1846 * the block tree for data. We don't really need to
1847 * adjust the offset, as we will just end up looking
1848 * at the pointer to this block in its parent, and its
1849 * going to be unallocated, so we will skip over it.
1851 return (SET_ERROR(ESRCH));
1853 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1855 dbuf_rele(db, FTAG);
1858 data = db->db.db_data;
1862 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1863 db->db_blkptr->blk_birth <= txg ||
1864 BP_IS_HOLE(db->db_blkptr))) {
1866 * This can only happen when we are searching up the tree
1867 * and these conditions mean that we need to keep climbing.
1869 error = SET_ERROR(ESRCH);
1870 } else if (lvl == 0) {
1871 dnode_phys_t *dnp = data;
1873 ASSERT(dn->dn_type == DMU_OT_DNODE);
1875 for (i = (*offset >> span) & (blkfill - 1);
1876 i >= 0 && i < blkfill; i += inc) {
1877 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1879 *offset += (1ULL << span) * inc;
1881 if (i < 0 || i == blkfill)
1882 error = SET_ERROR(ESRCH);
1884 blkptr_t *bp = data;
1885 uint64_t start = *offset;
1886 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1888 maxfill = blkfill << ((lvl - 1) * epbs);
1895 *offset = *offset >> span;
1896 for (i = BF64_GET(*offset, 0, epbs);
1897 i >= 0 && i < epb; i += inc) {
1898 if (BP_GET_FILL(&bp[i]) >= minfill &&
1899 BP_GET_FILL(&bp[i]) <= maxfill &&
1900 (hole || bp[i].blk_birth > txg))
1902 if (inc > 0 || *offset > 0)
1905 *offset = *offset << span;
1907 /* traversing backwards; position offset at the end */
1908 ASSERT3U(*offset, <=, start);
1909 *offset = MIN(*offset + (1ULL << span) - 1, start);
1910 } else if (*offset < start) {
1913 if (i < 0 || i >= epb)
1914 error = SET_ERROR(ESRCH);
1918 dbuf_rele(db, FTAG);
1924 * Find the next hole, data, or sparse region at or after *offset.
1925 * The value 'blkfill' tells us how many items we expect to find
1926 * in an L0 data block; this value is 1 for normal objects,
1927 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1928 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1932 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1933 * Finds the next/previous hole/data in a file.
1934 * Used in dmu_offset_next().
1936 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1937 * Finds the next free/allocated dnode an objset's meta-dnode.
1938 * Only finds objects that have new contents since txg (ie.
1939 * bonus buffer changes and content removal are ignored).
1940 * Used in dmu_object_next().
1942 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1943 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1944 * Used in dmu_object_alloc().
1947 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1948 int minlvl, uint64_t blkfill, uint64_t txg)
1950 uint64_t initial_offset = *offset;
1954 if (!(flags & DNODE_FIND_HAVELOCK))
1955 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1957 if (dn->dn_phys->dn_nlevels == 0) {
1958 error = SET_ERROR(ESRCH);
1962 if (dn->dn_datablkshift == 0) {
1963 if (*offset < dn->dn_datablksz) {
1964 if (flags & DNODE_FIND_HOLE)
1965 *offset = dn->dn_datablksz;
1967 error = SET_ERROR(ESRCH);
1972 maxlvl = dn->dn_phys->dn_nlevels;
1974 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1975 error = dnode_next_offset_level(dn,
1976 flags, offset, lvl, blkfill, txg);
1981 while (error == 0 && --lvl >= minlvl) {
1982 error = dnode_next_offset_level(dn,
1983 flags, offset, lvl, blkfill, txg);
1987 * There's always a "virtual hole" at the end of the object, even
1988 * if all BP's which physically exist are non-holes.
1990 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1991 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1995 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1996 initial_offset < *offset : initial_offset > *offset))
1997 error = SET_ERROR(ESRCH);
1999 if (!(flags & DNODE_FIND_HAVELOCK))
2000 rw_exit(&dn->dn_struct_rwlock);