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]
26 * Copyright 2017 RackTop Systems.
29 #include <sys/zfs_context.h>
31 #include <sys/dnode.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_dataset.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/range_tree.h>
43 dnode_stats_t dnode_stats = {
44 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64 },
45 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64 },
46 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64 },
47 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64 },
48 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64 },
49 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64 },
50 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64 },
51 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64 },
52 { "dnode_hold_free_hits", KSTAT_DATA_UINT64 },
53 { "dnode_hold_free_misses", KSTAT_DATA_UINT64 },
54 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64 },
55 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64 },
56 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64 },
57 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64 },
58 { "dnode_hold_free_txg", KSTAT_DATA_UINT64 },
59 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 },
60 { "dnode_allocate", KSTAT_DATA_UINT64 },
61 { "dnode_reallocate", KSTAT_DATA_UINT64 },
62 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
63 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
64 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
65 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
66 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
67 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
68 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
69 { "dnode_move_special", KSTAT_DATA_UINT64 },
70 { "dnode_move_handle", KSTAT_DATA_UINT64 },
71 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
72 { "dnode_move_active", KSTAT_DATA_UINT64 },
75 static kstat_t *dnode_ksp;
76 static kmem_cache_t *dnode_cache;
78 static dnode_phys_t dnode_phys_zero;
80 int zfs_default_bs = SPA_MINBLOCKSHIFT;
81 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
83 SYSCTL_DECL(_vfs_zfs);
84 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_bs, CTLFLAG_RWTUN,
85 &zfs_default_bs, 0, "Default dnode block shift");
86 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_ibs, CTLFLAG_RWTUN,
87 &zfs_default_ibs, 0, "Default dnode indirect block shift");
91 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
96 dbuf_compare(const void *x1, const void *x2)
98 const dmu_buf_impl_t *d1 = x1;
99 const dmu_buf_impl_t *d2 = x2;
101 int cmp = AVL_CMP(d1->db_level, d2->db_level);
105 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
109 if (d1->db_state == DB_SEARCH) {
110 ASSERT3S(d2->db_state, !=, DB_SEARCH);
112 } else if (d2->db_state == DB_SEARCH) {
113 ASSERT3S(d1->db_state, !=, DB_SEARCH);
117 return (AVL_PCMP(d1, d2));
122 dnode_cons(void *arg, void *unused, int kmflag)
127 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
128 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
129 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
130 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
133 * Every dbuf has a reference, and dropping a tracked reference is
134 * O(number of references), so don't track dn_holds.
136 zfs_refcount_create_untracked(&dn->dn_holds);
137 zfs_refcount_create(&dn->dn_tx_holds);
138 list_link_init(&dn->dn_link);
140 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
141 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
142 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
143 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
144 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
145 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
146 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
148 for (i = 0; i < TXG_SIZE; i++) {
149 multilist_link_init(&dn->dn_dirty_link[i]);
150 dn->dn_free_ranges[i] = NULL;
151 list_create(&dn->dn_dirty_records[i],
152 sizeof (dbuf_dirty_record_t),
153 offsetof(dbuf_dirty_record_t, dr_dirty_node));
156 dn->dn_allocated_txg = 0;
158 dn->dn_assigned_txg = 0;
159 dn->dn_dirty_txg = 0;
161 dn->dn_dirtyctx_firstset = NULL;
163 dn->dn_have_spill = B_FALSE;
173 dn->dn_dbufs_count = 0;
174 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
175 offsetof(dmu_buf_impl_t, db_link));
178 POINTER_INVALIDATE(&dn->dn_objset);
184 dnode_dest(void *arg, void *unused)
189 rw_destroy(&dn->dn_struct_rwlock);
190 mutex_destroy(&dn->dn_mtx);
191 mutex_destroy(&dn->dn_dbufs_mtx);
192 cv_destroy(&dn->dn_notxholds);
193 zfs_refcount_destroy(&dn->dn_holds);
194 zfs_refcount_destroy(&dn->dn_tx_holds);
195 ASSERT(!list_link_active(&dn->dn_link));
197 for (i = 0; i < TXG_SIZE; i++) {
198 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
199 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
200 list_destroy(&dn->dn_dirty_records[i]);
201 ASSERT0(dn->dn_next_nblkptr[i]);
202 ASSERT0(dn->dn_next_nlevels[i]);
203 ASSERT0(dn->dn_next_indblkshift[i]);
204 ASSERT0(dn->dn_next_bonustype[i]);
205 ASSERT0(dn->dn_rm_spillblk[i]);
206 ASSERT0(dn->dn_next_bonuslen[i]);
207 ASSERT0(dn->dn_next_blksz[i]);
210 ASSERT0(dn->dn_allocated_txg);
211 ASSERT0(dn->dn_free_txg);
212 ASSERT0(dn->dn_assigned_txg);
213 ASSERT0(dn->dn_dirty_txg);
214 ASSERT0(dn->dn_dirtyctx);
215 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
216 ASSERT3P(dn->dn_bonus, ==, NULL);
217 ASSERT(!dn->dn_have_spill);
218 ASSERT3P(dn->dn_zio, ==, NULL);
219 ASSERT0(dn->dn_oldused);
220 ASSERT0(dn->dn_oldflags);
221 ASSERT0(dn->dn_olduid);
222 ASSERT0(dn->dn_oldgid);
223 ASSERT0(dn->dn_newuid);
224 ASSERT0(dn->dn_newgid);
225 ASSERT0(dn->dn_id_flags);
227 ASSERT0(dn->dn_dbufs_count);
228 avl_destroy(&dn->dn_dbufs);
234 ASSERT(dnode_cache == NULL);
235 dnode_cache = kmem_cache_create("dnode_t",
237 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
239 kmem_cache_set_move(dnode_cache, dnode_move);
241 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
242 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
244 if (dnode_ksp != NULL) {
245 dnode_ksp->ks_data = &dnode_stats;
246 kstat_install(dnode_ksp);
254 if (dnode_ksp != NULL) {
255 kstat_delete(dnode_ksp);
259 kmem_cache_destroy(dnode_cache);
266 dnode_verify(dnode_t *dn)
268 int drop_struct_lock = FALSE;
271 ASSERT(dn->dn_objset);
272 ASSERT(dn->dn_handle->dnh_dnode == dn);
274 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
276 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
279 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
280 rw_enter(&dn->dn_struct_rwlock, RW_READER);
281 drop_struct_lock = TRUE;
283 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
285 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
286 ASSERT3U(dn->dn_indblkshift, >=, 0);
287 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
288 if (dn->dn_datablkshift) {
289 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
290 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
291 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
293 ASSERT3U(dn->dn_nlevels, <=, 30);
294 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
295 ASSERT3U(dn->dn_nblkptr, >=, 1);
296 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
297 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
298 ASSERT3U(dn->dn_datablksz, ==,
299 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
300 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
301 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
302 dn->dn_bonuslen, <=, max_bonuslen);
303 for (i = 0; i < TXG_SIZE; i++) {
304 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
307 if (dn->dn_phys->dn_type != DMU_OT_NONE)
308 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
309 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
310 if (dn->dn_dbuf != NULL) {
311 ASSERT3P(dn->dn_phys, ==,
312 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
313 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
315 if (drop_struct_lock)
316 rw_exit(&dn->dn_struct_rwlock);
321 dnode_byteswap(dnode_phys_t *dnp)
323 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
326 if (dnp->dn_type == DMU_OT_NONE) {
327 bzero(dnp, sizeof (dnode_phys_t));
331 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
332 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
333 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
334 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
335 dnp->dn_used = BSWAP_64(dnp->dn_used);
338 * dn_nblkptr is only one byte, so it's OK to read it in either
339 * byte order. We can't read dn_bouslen.
341 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
342 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
343 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
344 buf64[i] = BSWAP_64(buf64[i]);
347 * OK to check dn_bonuslen for zero, because it won't matter if
348 * we have the wrong byte order. This is necessary because the
349 * dnode dnode is smaller than a regular dnode.
351 if (dnp->dn_bonuslen != 0) {
353 * Note that the bonus length calculated here may be
354 * longer than the actual bonus buffer. This is because
355 * we always put the bonus buffer after the last block
356 * pointer (instead of packing it against the end of the
359 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
360 int slots = dnp->dn_extra_slots + 1;
361 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
362 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
363 dmu_object_byteswap_t byteswap =
364 DMU_OT_BYTESWAP(dnp->dn_bonustype);
365 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
368 /* Swap SPILL block if we have one */
369 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
370 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
375 dnode_buf_byteswap(void *vbuf, size_t size)
379 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
380 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
383 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
387 if (dnp->dn_type != DMU_OT_NONE)
388 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
393 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
395 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
397 dnode_setdirty(dn, tx);
398 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
399 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
400 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
401 dn->dn_bonuslen = newsize;
403 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
405 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
406 rw_exit(&dn->dn_struct_rwlock);
410 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
412 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
413 dnode_setdirty(dn, tx);
414 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
415 dn->dn_bonustype = newtype;
416 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
417 rw_exit(&dn->dn_struct_rwlock);
421 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
423 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
424 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
425 dnode_setdirty(dn, tx);
426 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
427 dn->dn_have_spill = B_FALSE;
431 dnode_setdblksz(dnode_t *dn, int size)
433 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
434 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
435 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
436 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
437 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
438 dn->dn_datablksz = size;
439 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
440 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
444 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
445 uint64_t object, dnode_handle_t *dnh)
449 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
451 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
456 * Defer setting dn_objset until the dnode is ready to be a candidate
457 * for the dnode_move() callback.
459 dn->dn_object = object;
464 if (dnp->dn_datablkszsec) {
465 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
467 dn->dn_datablksz = 0;
468 dn->dn_datablkszsec = 0;
469 dn->dn_datablkshift = 0;
471 dn->dn_indblkshift = dnp->dn_indblkshift;
472 dn->dn_nlevels = dnp->dn_nlevels;
473 dn->dn_type = dnp->dn_type;
474 dn->dn_nblkptr = dnp->dn_nblkptr;
475 dn->dn_checksum = dnp->dn_checksum;
476 dn->dn_compress = dnp->dn_compress;
477 dn->dn_bonustype = dnp->dn_bonustype;
478 dn->dn_bonuslen = dnp->dn_bonuslen;
479 dn->dn_num_slots = dnp->dn_extra_slots + 1;
480 dn->dn_maxblkid = dnp->dn_maxblkid;
481 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
484 dmu_zfetch_init(&dn->dn_zfetch, dn);
486 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
487 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
488 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
490 mutex_enter(&os->os_lock);
493 * Exclude special dnodes from os_dnodes so an empty os_dnodes
494 * signifies that the special dnodes have no references from
495 * their children (the entries in os_dnodes). This allows
496 * dnode_destroy() to easily determine if the last child has
497 * been removed and then complete eviction of the objset.
499 if (!DMU_OBJECT_IS_SPECIAL(object))
500 list_insert_head(&os->os_dnodes, dn);
504 * Everything else must be valid before assigning dn_objset
505 * makes the dnode eligible for dnode_move().
510 mutex_exit(&os->os_lock);
512 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
518 * Caller must be holding the dnode handle, which is released upon return.
521 dnode_destroy(dnode_t *dn)
523 objset_t *os = dn->dn_objset;
524 boolean_t complete_os_eviction = B_FALSE;
526 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
528 mutex_enter(&os->os_lock);
529 POINTER_INVALIDATE(&dn->dn_objset);
530 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
531 list_remove(&os->os_dnodes, dn);
532 complete_os_eviction =
533 list_is_empty(&os->os_dnodes) &&
534 list_link_active(&os->os_evicting_node);
536 mutex_exit(&os->os_lock);
538 /* the dnode can no longer move, so we can release the handle */
539 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
540 zrl_remove(&dn->dn_handle->dnh_zrlock);
542 dn->dn_allocated_txg = 0;
544 dn->dn_assigned_txg = 0;
545 dn->dn_dirty_txg = 0;
548 if (dn->dn_dirtyctx_firstset != NULL) {
549 kmem_free(dn->dn_dirtyctx_firstset, 1);
550 dn->dn_dirtyctx_firstset = NULL;
552 if (dn->dn_bonus != NULL) {
553 mutex_enter(&dn->dn_bonus->db_mtx);
554 dbuf_destroy(dn->dn_bonus);
559 dn->dn_have_spill = B_FALSE;
568 dmu_zfetch_fini(&dn->dn_zfetch);
569 kmem_cache_free(dnode_cache, dn);
570 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
572 if (complete_os_eviction)
573 dmu_objset_evict_done(os);
577 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
578 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
582 ASSERT3U(dn_slots, >, 0);
583 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
584 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
585 ASSERT3U(blocksize, <=,
586 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
588 blocksize = 1 << zfs_default_bs;
590 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
593 ibs = zfs_default_ibs;
595 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
597 dprintf("os=%p obj=%" PRIu64 " txg=%" PRIu64
598 " blocksize=%d ibs=%d dn_slots=%d\n",
599 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
600 DNODE_STAT_BUMP(dnode_allocate);
602 ASSERT(dn->dn_type == DMU_OT_NONE);
603 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
604 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
605 ASSERT(ot != DMU_OT_NONE);
606 ASSERT(DMU_OT_IS_VALID(ot));
607 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
608 (bonustype == DMU_OT_SA && bonuslen == 0) ||
609 (bonustype != DMU_OT_NONE && bonuslen != 0));
610 ASSERT(DMU_OT_IS_VALID(bonustype));
611 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
612 ASSERT(dn->dn_type == DMU_OT_NONE);
613 ASSERT0(dn->dn_maxblkid);
614 ASSERT0(dn->dn_allocated_txg);
615 ASSERT0(dn->dn_dirty_txg);
616 ASSERT0(dn->dn_assigned_txg);
617 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
618 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
619 ASSERT(avl_is_empty(&dn->dn_dbufs));
621 for (i = 0; i < TXG_SIZE; i++) {
622 ASSERT0(dn->dn_next_nblkptr[i]);
623 ASSERT0(dn->dn_next_nlevels[i]);
624 ASSERT0(dn->dn_next_indblkshift[i]);
625 ASSERT0(dn->dn_next_bonuslen[i]);
626 ASSERT0(dn->dn_next_bonustype[i]);
627 ASSERT0(dn->dn_rm_spillblk[i]);
628 ASSERT0(dn->dn_next_blksz[i]);
629 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
630 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
631 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
635 dnode_setdblksz(dn, blocksize);
636 dn->dn_indblkshift = ibs;
638 dn->dn_num_slots = dn_slots;
639 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
642 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
643 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
647 dn->dn_bonustype = bonustype;
648 dn->dn_bonuslen = bonuslen;
649 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
650 dn->dn_compress = ZIO_COMPRESS_INHERIT;
654 if (dn->dn_dirtyctx_firstset) {
655 kmem_free(dn->dn_dirtyctx_firstset, 1);
656 dn->dn_dirtyctx_firstset = NULL;
659 dn->dn_allocated_txg = tx->tx_txg;
662 dnode_setdirty(dn, tx);
663 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
664 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
665 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
666 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
670 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
671 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
675 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
676 ASSERT3U(blocksize, <=,
677 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
678 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
679 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
680 ASSERT(tx->tx_txg != 0);
681 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
682 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
683 (bonustype == DMU_OT_SA && bonuslen == 0));
684 ASSERT(DMU_OT_IS_VALID(bonustype));
685 ASSERT3U(bonuslen, <=,
686 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
687 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
689 dnode_free_interior_slots(dn);
690 DNODE_STAT_BUMP(dnode_reallocate);
692 /* clean up any unreferenced dbufs */
693 dnode_evict_dbufs(dn);
697 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
698 dnode_setdirty(dn, tx);
699 if (dn->dn_datablksz != blocksize) {
700 /* change blocksize */
701 ASSERT(dn->dn_maxblkid == 0 &&
702 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
703 dnode_block_freed(dn, 0)));
704 dnode_setdblksz(dn, blocksize);
705 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
707 if (dn->dn_bonuslen != bonuslen)
708 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
710 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
713 nblkptr = MIN(DN_MAX_NBLKPTR,
714 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
716 if (dn->dn_bonustype != bonustype)
717 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
718 if (dn->dn_nblkptr != nblkptr)
719 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
720 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
721 dbuf_rm_spill(dn, tx);
722 dnode_rm_spill(dn, tx);
724 rw_exit(&dn->dn_struct_rwlock);
729 /* change bonus size and type */
730 mutex_enter(&dn->dn_mtx);
731 dn->dn_bonustype = bonustype;
732 dn->dn_bonuslen = bonuslen;
733 dn->dn_num_slots = dn_slots;
734 dn->dn_nblkptr = nblkptr;
735 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
736 dn->dn_compress = ZIO_COMPRESS_INHERIT;
737 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
739 /* fix up the bonus db_size */
741 dn->dn_bonus->db.db_size =
742 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
743 (dn->dn_nblkptr - 1) * sizeof (blkptr_t);
744 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
747 dn->dn_allocated_txg = tx->tx_txg;
748 mutex_exit(&dn->dn_mtx);
753 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
757 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
758 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
759 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
760 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
763 ndn->dn_objset = odn->dn_objset;
764 ndn->dn_object = odn->dn_object;
765 ndn->dn_dbuf = odn->dn_dbuf;
766 ndn->dn_handle = odn->dn_handle;
767 ndn->dn_phys = odn->dn_phys;
768 ndn->dn_type = odn->dn_type;
769 ndn->dn_bonuslen = odn->dn_bonuslen;
770 ndn->dn_bonustype = odn->dn_bonustype;
771 ndn->dn_nblkptr = odn->dn_nblkptr;
772 ndn->dn_checksum = odn->dn_checksum;
773 ndn->dn_compress = odn->dn_compress;
774 ndn->dn_nlevels = odn->dn_nlevels;
775 ndn->dn_indblkshift = odn->dn_indblkshift;
776 ndn->dn_datablkshift = odn->dn_datablkshift;
777 ndn->dn_datablkszsec = odn->dn_datablkszsec;
778 ndn->dn_datablksz = odn->dn_datablksz;
779 ndn->dn_maxblkid = odn->dn_maxblkid;
780 ndn->dn_num_slots = odn->dn_num_slots;
781 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
782 sizeof (odn->dn_next_type));
783 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
784 sizeof (odn->dn_next_nblkptr));
785 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
786 sizeof (odn->dn_next_nlevels));
787 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
788 sizeof (odn->dn_next_indblkshift));
789 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
790 sizeof (odn->dn_next_bonustype));
791 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
792 sizeof (odn->dn_rm_spillblk));
793 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
794 sizeof (odn->dn_next_bonuslen));
795 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
796 sizeof (odn->dn_next_blksz));
797 for (i = 0; i < TXG_SIZE; i++) {
798 list_move_tail(&ndn->dn_dirty_records[i],
799 &odn->dn_dirty_records[i]);
801 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
802 sizeof (odn->dn_free_ranges));
803 ndn->dn_allocated_txg = odn->dn_allocated_txg;
804 ndn->dn_free_txg = odn->dn_free_txg;
805 ndn->dn_assigned_txg = odn->dn_assigned_txg;
806 ndn->dn_dirty_txg = odn->dn_dirty_txg;
807 ndn->dn_dirtyctx = odn->dn_dirtyctx;
808 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
809 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
810 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
811 ASSERT(avl_is_empty(&ndn->dn_dbufs));
812 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
813 ndn->dn_dbufs_count = odn->dn_dbufs_count;
814 ndn->dn_bonus = odn->dn_bonus;
815 ndn->dn_have_spill = odn->dn_have_spill;
816 ndn->dn_zio = odn->dn_zio;
817 ndn->dn_oldused = odn->dn_oldused;
818 ndn->dn_oldflags = odn->dn_oldflags;
819 ndn->dn_olduid = odn->dn_olduid;
820 ndn->dn_oldgid = odn->dn_oldgid;
821 ndn->dn_newuid = odn->dn_newuid;
822 ndn->dn_newgid = odn->dn_newgid;
823 ndn->dn_id_flags = odn->dn_id_flags;
824 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
825 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
826 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
829 * Update back pointers. Updating the handle fixes the back pointer of
830 * every descendant dbuf as well as the bonus dbuf.
832 ASSERT(ndn->dn_handle->dnh_dnode == odn);
833 ndn->dn_handle->dnh_dnode = ndn;
834 if (ndn->dn_zfetch.zf_dnode == odn) {
835 ndn->dn_zfetch.zf_dnode = ndn;
839 * Invalidate the original dnode by clearing all of its back pointers.
842 odn->dn_handle = NULL;
843 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
844 offsetof(dmu_buf_impl_t, db_link));
845 odn->dn_dbufs_count = 0;
846 odn->dn_bonus = NULL;
847 odn->dn_zfetch.zf_dnode = NULL;
850 * Set the low bit of the objset pointer to ensure that dnode_move()
851 * recognizes the dnode as invalid in any subsequent callback.
853 POINTER_INVALIDATE(&odn->dn_objset);
856 * Satisfy the destructor.
858 for (i = 0; i < TXG_SIZE; i++) {
859 list_create(&odn->dn_dirty_records[i],
860 sizeof (dbuf_dirty_record_t),
861 offsetof(dbuf_dirty_record_t, dr_dirty_node));
862 odn->dn_free_ranges[i] = NULL;
863 odn->dn_next_nlevels[i] = 0;
864 odn->dn_next_indblkshift[i] = 0;
865 odn->dn_next_bonustype[i] = 0;
866 odn->dn_rm_spillblk[i] = 0;
867 odn->dn_next_bonuslen[i] = 0;
868 odn->dn_next_blksz[i] = 0;
870 odn->dn_allocated_txg = 0;
871 odn->dn_free_txg = 0;
872 odn->dn_assigned_txg = 0;
873 odn->dn_dirty_txg = 0;
874 odn->dn_dirtyctx = 0;
875 odn->dn_dirtyctx_firstset = NULL;
876 odn->dn_have_spill = B_FALSE;
879 odn->dn_oldflags = 0;
884 odn->dn_id_flags = 0;
890 odn->dn_moved = (uint8_t)-1;
896 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
898 dnode_t *odn = buf, *ndn = newbuf;
904 * The dnode is on the objset's list of known dnodes if the objset
905 * pointer is valid. We set the low bit of the objset pointer when
906 * freeing the dnode to invalidate it, and the memory patterns written
907 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
908 * A newly created dnode sets the objset pointer last of all to indicate
909 * that the dnode is known and in a valid state to be moved by this
913 if (!POINTER_IS_VALID(os)) {
914 DNODE_STAT_BUMP(dnode_move_invalid);
915 return (KMEM_CBRC_DONT_KNOW);
919 * Ensure that the objset does not go away during the move.
921 rw_enter(&os_lock, RW_WRITER);
922 if (os != odn->dn_objset) {
924 DNODE_STAT_BUMP(dnode_move_recheck1);
925 return (KMEM_CBRC_DONT_KNOW);
929 * If the dnode is still valid, then so is the objset. We know that no
930 * valid objset can be freed while we hold os_lock, so we can safely
931 * ensure that the objset remains in use.
933 mutex_enter(&os->os_lock);
936 * Recheck the objset pointer in case the dnode was removed just before
937 * acquiring the lock.
939 if (os != odn->dn_objset) {
940 mutex_exit(&os->os_lock);
942 DNODE_STAT_BUMP(dnode_move_recheck2);
943 return (KMEM_CBRC_DONT_KNOW);
947 * At this point we know that as long as we hold os->os_lock, the dnode
948 * cannot be freed and fields within the dnode can be safely accessed.
949 * The objset listing this dnode cannot go away as long as this dnode is
953 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
954 mutex_exit(&os->os_lock);
955 DNODE_STAT_BUMP(dnode_move_special);
956 return (KMEM_CBRC_NO);
958 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
961 * Lock the dnode handle to prevent the dnode from obtaining any new
962 * holds. This also prevents the descendant dbufs and the bonus dbuf
963 * from accessing the dnode, so that we can discount their holds. The
964 * handle is safe to access because we know that while the dnode cannot
965 * go away, neither can its handle. Once we hold dnh_zrlock, we can
966 * safely move any dnode referenced only by dbufs.
968 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
969 mutex_exit(&os->os_lock);
970 DNODE_STAT_BUMP(dnode_move_handle);
971 return (KMEM_CBRC_LATER);
975 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
976 * We need to guarantee that there is a hold for every dbuf in order to
977 * determine whether the dnode is actively referenced. Falsely matching
978 * a dbuf to an active hold would lead to an unsafe move. It's possible
979 * that a thread already having an active dnode hold is about to add a
980 * dbuf, and we can't compare hold and dbuf counts while the add is in
983 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
984 zrl_exit(&odn->dn_handle->dnh_zrlock);
985 mutex_exit(&os->os_lock);
986 DNODE_STAT_BUMP(dnode_move_rwlock);
987 return (KMEM_CBRC_LATER);
991 * A dbuf may be removed (evicted) without an active dnode hold. In that
992 * case, the dbuf count is decremented under the handle lock before the
993 * dbuf's hold is released. This order ensures that if we count the hold
994 * after the dbuf is removed but before its hold is released, we will
995 * treat the unmatched hold as active and exit safely. If we count the
996 * hold before the dbuf is removed, the hold is discounted, and the
997 * removal is blocked until the move completes.
999 refcount = zfs_refcount_count(&odn->dn_holds);
1000 ASSERT(refcount >= 0);
1001 dbufs = DN_DBUFS_COUNT(odn);
1003 /* We can't have more dbufs than dnode holds. */
1004 ASSERT3U(dbufs, <=, refcount);
1005 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1008 if (refcount > dbufs) {
1009 rw_exit(&odn->dn_struct_rwlock);
1010 zrl_exit(&odn->dn_handle->dnh_zrlock);
1011 mutex_exit(&os->os_lock);
1012 DNODE_STAT_BUMP(dnode_move_active);
1013 return (KMEM_CBRC_LATER);
1016 rw_exit(&odn->dn_struct_rwlock);
1019 * At this point we know that anyone with a hold on the dnode is not
1020 * actively referencing it. The dnode is known and in a valid state to
1021 * move. We're holding the locks needed to execute the critical section.
1023 dnode_move_impl(odn, ndn);
1025 list_link_replace(&odn->dn_link, &ndn->dn_link);
1026 /* If the dnode was safe to move, the refcount cannot have changed. */
1027 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1028 ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1029 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1030 mutex_exit(&os->os_lock);
1032 return (KMEM_CBRC_YES);
1034 #endif /* illumos */
1035 #endif /* _KERNEL */
1038 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1040 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1042 for (int i = idx; i < idx + slots; i++) {
1043 dnode_handle_t *dnh = &children->dnc_children[i];
1044 zrl_add(&dnh->dnh_zrlock);
1049 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1051 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1053 for (int i = idx; i < idx + slots; i++) {
1054 dnode_handle_t *dnh = &children->dnc_children[i];
1056 if (zrl_is_locked(&dnh->dnh_zrlock))
1057 zrl_exit(&dnh->dnh_zrlock);
1059 zrl_remove(&dnh->dnh_zrlock);
1064 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1066 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1068 for (int i = idx; i < idx + slots; i++) {
1069 dnode_handle_t *dnh = &children->dnc_children[i];
1071 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1072 for (int j = idx; j < i; j++) {
1073 dnh = &children->dnc_children[j];
1074 zrl_exit(&dnh->dnh_zrlock);
1085 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1087 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1089 for (int i = idx; i < idx + slots; i++) {
1090 dnode_handle_t *dnh = &children->dnc_children[i];
1091 dnh->dnh_dnode = ptr;
1096 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1098 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1101 * If all dnode slots are either already free or
1102 * evictable return B_TRUE.
1104 for (int i = idx; i < idx + slots; i++) {
1105 dnode_handle_t *dnh = &children->dnc_children[i];
1106 dnode_t *dn = dnh->dnh_dnode;
1108 if (dn == DN_SLOT_FREE) {
1110 } else if (DN_SLOT_IS_PTR(dn)) {
1111 mutex_enter(&dn->dn_mtx);
1112 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1113 zfs_refcount_is_zero(&dn->dn_holds) &&
1114 !DNODE_IS_DIRTY(dn));
1115 mutex_exit(&dn->dn_mtx);
1130 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1132 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1134 for (int i = idx; i < idx + slots; i++) {
1135 dnode_handle_t *dnh = &children->dnc_children[i];
1137 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1139 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1140 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1141 dnode_destroy(dnh->dnh_dnode);
1142 dnh->dnh_dnode = DN_SLOT_FREE;
1148 dnode_free_interior_slots(dnode_t *dn)
1150 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1151 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1152 int idx = (dn->dn_object & (epb - 1)) + 1;
1153 int slots = dn->dn_num_slots - 1;
1158 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1160 while (!dnode_slots_tryenter(children, idx, slots))
1161 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1163 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1164 dnode_slots_rele(children, idx, slots);
1168 dnode_special_close(dnode_handle_t *dnh)
1170 dnode_t *dn = dnh->dnh_dnode;
1173 * Wait for final references to the dnode to clear. This can
1174 * only happen if the arc is asynchronously evicting state that
1175 * has a hold on this dnode while we are trying to evict this
1178 while (zfs_refcount_count(&dn->dn_holds) > 0)
1180 ASSERT(dn->dn_dbuf == NULL ||
1181 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1182 zrl_add(&dnh->dnh_zrlock);
1183 dnode_destroy(dn); /* implicit zrl_remove() */
1184 zrl_destroy(&dnh->dnh_zrlock);
1185 dnh->dnh_dnode = NULL;
1189 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1190 dnode_handle_t *dnh)
1194 zrl_init(&dnh->dnh_zrlock);
1195 zrl_tryenter(&dnh->dnh_zrlock);
1197 dn = dnode_create(os, dnp, NULL, object, dnh);
1200 zrl_exit(&dnh->dnh_zrlock);
1204 dnode_buf_evict_async(void *dbu)
1206 dnode_children_t *dnc = dbu;
1208 DNODE_STAT_BUMP(dnode_buf_evict);
1210 for (int i = 0; i < dnc->dnc_count; i++) {
1211 dnode_handle_t *dnh = &dnc->dnc_children[i];
1215 * The dnode handle lock guards against the dnode moving to
1216 * another valid address, so there is no need here to guard
1217 * against changes to or from NULL.
1219 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1220 zrl_destroy(&dnh->dnh_zrlock);
1221 dnh->dnh_dnode = DN_SLOT_UNINIT;
1225 zrl_add(&dnh->dnh_zrlock);
1226 dn = dnh->dnh_dnode;
1228 * If there are holds on this dnode, then there should
1229 * be holds on the dnode's containing dbuf as well; thus
1230 * it wouldn't be eligible for eviction and this function
1231 * would not have been called.
1233 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1234 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1236 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1237 zrl_destroy(&dnh->dnh_zrlock);
1238 dnh->dnh_dnode = DN_SLOT_UNINIT;
1240 kmem_free(dnc, sizeof (dnode_children_t) +
1241 dnc->dnc_count * sizeof (dnode_handle_t));
1245 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1246 * to ensure the hole at the specified object offset is large enough to
1247 * hold the dnode being created. The slots parameter is also used to ensure
1248 * a dnode does not span multiple dnode blocks. In both of these cases, if
1249 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1250 * are only possible when using DNODE_MUST_BE_FREE.
1252 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1253 * dnode_hold_impl() will check if the requested dnode is already consumed
1254 * as an extra dnode slot by an large dnode, in which case it returns
1258 * EINVAL - invalid object number or flags.
1259 * ENOSPC - hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1260 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1261 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1262 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1263 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1264 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1265 * EIO - i/o error error when reading the meta dnode dbuf.
1266 * succeeds even for free dnodes.
1269 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1270 void *tag, dnode_t **dnp)
1272 int epb, idx, err, i;
1273 int drop_struct_lock = FALSE;
1278 dnode_children_t *dnc;
1279 dnode_phys_t *dn_block;
1280 dnode_phys_t *dn_block_begin;
1281 dnode_handle_t *dnh;
1283 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1284 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1287 * If you are holding the spa config lock as writer, you shouldn't
1288 * be asking the DMU to do *anything* unless it's the root pool
1289 * which may require us to read from the root filesystem while
1290 * holding some (not all) of the locks as writer.
1292 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1293 (spa_is_root(os->os_spa) &&
1294 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1296 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1298 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1299 dn = (object == DMU_USERUSED_OBJECT) ?
1300 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1302 return (SET_ERROR(ENOENT));
1304 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1305 return (SET_ERROR(ENOENT));
1306 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1307 return (SET_ERROR(EEXIST));
1309 (void) zfs_refcount_add(&dn->dn_holds, tag);
1314 if (object == 0 || object >= DN_MAX_OBJECT)
1315 return (SET_ERROR(EINVAL));
1317 mdn = DMU_META_DNODE(os);
1318 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1322 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1323 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1324 drop_struct_lock = TRUE;
1327 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1329 db = dbuf_hold(mdn, blk, FTAG);
1330 if (drop_struct_lock)
1331 rw_exit(&mdn->dn_struct_rwlock);
1333 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1334 return (SET_ERROR(EIO));
1336 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1338 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1339 dbuf_rele(db, FTAG);
1343 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1344 epb = db->db.db_size >> DNODE_SHIFT;
1346 idx = object & (epb - 1);
1347 dn_block = (dnode_phys_t *)db->db.db_data;
1349 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1350 dnc = dmu_buf_get_user(&db->db);
1353 dnode_children_t *winner;
1356 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1357 epb * sizeof (dnode_handle_t), KM_SLEEP);
1358 dnc->dnc_count = epb;
1359 dnh = &dnc->dnc_children[0];
1361 /* Initialize dnode slot status from dnode_phys_t */
1362 for (int i = 0; i < epb; i++) {
1363 zrl_init(&dnh[i].dnh_zrlock);
1370 if (dn_block[i].dn_type != DMU_OT_NONE) {
1371 int interior = dn_block[i].dn_extra_slots;
1373 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1374 dnode_set_slots(dnc, i + 1, interior,
1378 dnh[i].dnh_dnode = DN_SLOT_FREE;
1383 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1384 dnode_buf_evict_async, NULL);
1385 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1386 if (winner != NULL) {
1388 for (int i = 0; i < epb; i++)
1389 zrl_destroy(&dnh[i].dnh_zrlock);
1391 kmem_free(dnc, sizeof (dnode_children_t) +
1392 epb * sizeof (dnode_handle_t));
1397 ASSERT(dnc->dnc_count == epb);
1398 dn = DN_SLOT_UNINIT;
1400 if (flag & DNODE_MUST_BE_ALLOCATED) {
1403 while (dn == DN_SLOT_UNINIT) {
1404 dnode_slots_hold(dnc, idx, slots);
1405 dnh = &dnc->dnc_children[idx];
1407 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1408 dn = dnh->dnh_dnode;
1410 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1411 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1412 dnode_slots_rele(dnc, idx, slots);
1413 dbuf_rele(db, FTAG);
1414 return (SET_ERROR(EEXIST));
1415 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1416 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1417 dnode_slots_rele(dnc, idx, slots);
1418 dbuf_rele(db, FTAG);
1419 return (SET_ERROR(ENOENT));
1422 dnode_slots_rele(dnc, idx, slots);
1423 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1424 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1429 * Someone else won the race and called dnode_create()
1430 * after we checked DN_SLOT_IS_PTR() above but before
1431 * we acquired the lock.
1433 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1434 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1435 dn = dnh->dnh_dnode;
1437 dn = dnode_create(os, dn_block + idx, db,
1442 mutex_enter(&dn->dn_mtx);
1443 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1444 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1445 mutex_exit(&dn->dn_mtx);
1446 dnode_slots_rele(dnc, idx, slots);
1447 dbuf_rele(db, FTAG);
1448 return (SET_ERROR(ENOENT));
1451 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1452 } else if (flag & DNODE_MUST_BE_FREE) {
1454 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1455 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1456 dbuf_rele(db, FTAG);
1457 return (SET_ERROR(ENOSPC));
1460 while (dn == DN_SLOT_UNINIT) {
1461 dnode_slots_hold(dnc, idx, slots);
1463 if (!dnode_check_slots_free(dnc, idx, slots)) {
1464 DNODE_STAT_BUMP(dnode_hold_free_misses);
1465 dnode_slots_rele(dnc, idx, slots);
1466 dbuf_rele(db, FTAG);
1467 return (SET_ERROR(ENOSPC));
1470 dnode_slots_rele(dnc, idx, slots);
1471 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1472 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1476 if (!dnode_check_slots_free(dnc, idx, slots)) {
1477 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1478 dnode_slots_rele(dnc, idx, slots);
1479 dbuf_rele(db, FTAG);
1480 return (SET_ERROR(ENOSPC));
1484 * Allocated but otherwise free dnodes which would
1485 * be in the interior of a multi-slot dnodes need
1486 * to be freed. Single slot dnodes can be safely
1487 * re-purposed as a performance optimization.
1490 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1492 dnh = &dnc->dnc_children[idx];
1493 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1494 dn = dnh->dnh_dnode;
1496 dn = dnode_create(os, dn_block + idx, db,
1501 mutex_enter(&dn->dn_mtx);
1502 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1503 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1504 mutex_exit(&dn->dn_mtx);
1505 dnode_slots_rele(dnc, idx, slots);
1506 dbuf_rele(db, FTAG);
1507 return (SET_ERROR(EEXIST));
1510 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1511 DNODE_STAT_BUMP(dnode_hold_free_hits);
1513 dbuf_rele(db, FTAG);
1514 return (SET_ERROR(EINVAL));
1517 if (dn->dn_free_txg) {
1518 DNODE_STAT_BUMP(dnode_hold_free_txg);
1520 mutex_exit(&dn->dn_mtx);
1521 dnode_slots_rele(dnc, idx, slots);
1522 dbuf_rele(db, FTAG);
1523 return (SET_ERROR((flag & DNODE_MUST_BE_ALLOCATED) ?
1527 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1528 dbuf_add_ref(db, dnh);
1530 mutex_exit(&dn->dn_mtx);
1532 /* Now we can rely on the hold to prevent the dnode from moving. */
1533 dnode_slots_rele(dnc, idx, slots);
1536 ASSERT3P(dn->dn_dbuf, ==, db);
1537 ASSERT3U(dn->dn_object, ==, object);
1538 dbuf_rele(db, FTAG);
1545 * Return held dnode if the object is allocated, NULL if not.
1548 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1550 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1555 * Can only add a reference if there is already at least one
1556 * reference on the dnode. Returns FALSE if unable to add a
1560 dnode_add_ref(dnode_t *dn, void *tag)
1562 mutex_enter(&dn->dn_mtx);
1563 if (zfs_refcount_is_zero(&dn->dn_holds)) {
1564 mutex_exit(&dn->dn_mtx);
1567 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1568 mutex_exit(&dn->dn_mtx);
1573 dnode_rele(dnode_t *dn, void *tag)
1575 mutex_enter(&dn->dn_mtx);
1576 dnode_rele_and_unlock(dn, tag, B_FALSE);
1580 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1583 /* Get while the hold prevents the dnode from moving. */
1584 dmu_buf_impl_t *db = dn->dn_dbuf;
1585 dnode_handle_t *dnh = dn->dn_handle;
1587 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1588 mutex_exit(&dn->dn_mtx);
1591 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1592 * indirectly by dbuf_rele() while relying on the dnode handle to
1593 * prevent the dnode from moving, since releasing the last hold could
1594 * result in the dnode's parent dbuf evicting its dnode handles. For
1595 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1596 * other direct or indirect hold on the dnode must first drop the dnode
1599 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1601 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1602 if (refs == 0 && db != NULL) {
1604 * Another thread could add a hold to the dnode handle in
1605 * dnode_hold_impl() while holding the parent dbuf. Since the
1606 * hold on the parent dbuf prevents the handle from being
1607 * destroyed, the hold on the handle is OK. We can't yet assert
1608 * that the handle has zero references, but that will be
1609 * asserted anyway when the handle gets destroyed.
1611 mutex_enter(&db->db_mtx);
1612 dbuf_rele_and_unlock(db, dnh, evicting);
1617 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1619 objset_t *os = dn->dn_objset;
1620 uint64_t txg = tx->tx_txg;
1622 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1623 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1630 mutex_enter(&dn->dn_mtx);
1631 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1632 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1633 mutex_exit(&dn->dn_mtx);
1637 * Determine old uid/gid when necessary
1639 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1641 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1642 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1645 * If we are already marked dirty, we're done.
1647 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1648 multilist_sublist_unlock(mls);
1652 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1653 !avl_is_empty(&dn->dn_dbufs));
1654 ASSERT(dn->dn_datablksz != 0);
1655 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1656 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1657 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1659 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1660 dn->dn_object, txg);
1662 multilist_sublist_insert_head(mls, dn);
1664 multilist_sublist_unlock(mls);
1667 * The dnode maintains a hold on its containing dbuf as
1668 * long as there are holds on it. Each instantiated child
1669 * dbuf maintains a hold on the dnode. When the last child
1670 * drops its hold, the dnode will drop its hold on the
1671 * containing dbuf. We add a "dirty hold" here so that the
1672 * dnode will hang around after we finish processing its
1675 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1677 (void) dbuf_dirty(dn->dn_dbuf, tx);
1679 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1683 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1685 mutex_enter(&dn->dn_mtx);
1686 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1687 mutex_exit(&dn->dn_mtx);
1690 dn->dn_free_txg = tx->tx_txg;
1691 mutex_exit(&dn->dn_mtx);
1693 dnode_setdirty(dn, tx);
1697 * Try to change the block size for the indicated dnode. This can only
1698 * succeed if there are no blocks allocated or dirty beyond first block
1701 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1706 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1708 size = SPA_MINBLOCKSIZE;
1710 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1712 if (ibs == dn->dn_indblkshift)
1715 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1718 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1720 /* Check for any allocated blocks beyond the first */
1721 if (dn->dn_maxblkid != 0)
1724 mutex_enter(&dn->dn_dbufs_mtx);
1725 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1726 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1727 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1728 db->db_blkid != DMU_SPILL_BLKID) {
1729 mutex_exit(&dn->dn_dbufs_mtx);
1733 mutex_exit(&dn->dn_dbufs_mtx);
1735 if (ibs && dn->dn_nlevels != 1)
1738 /* resize the old block */
1739 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1741 dbuf_new_size(db, size, tx);
1742 else if (err != ENOENT)
1745 dnode_setdblksz(dn, size);
1746 dnode_setdirty(dn, tx);
1747 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1749 dn->dn_indblkshift = ibs;
1750 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1752 /* rele after we have fixed the blocksize in the dnode */
1754 dbuf_rele(db, FTAG);
1756 rw_exit(&dn->dn_struct_rwlock);
1760 rw_exit(&dn->dn_struct_rwlock);
1761 return (SET_ERROR(ENOTSUP));
1764 /* read-holding callers must not rely on the lock being continuously held */
1766 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1768 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1769 int epbs, new_nlevels;
1772 ASSERT(blkid != DMU_BONUS_BLKID);
1775 RW_READ_HELD(&dn->dn_struct_rwlock) :
1776 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1779 * if we have a read-lock, check to see if we need to do any work
1780 * before upgrading to a write-lock.
1783 if (blkid <= dn->dn_maxblkid)
1786 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1787 rw_exit(&dn->dn_struct_rwlock);
1788 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1792 if (blkid <= dn->dn_maxblkid)
1795 dn->dn_maxblkid = blkid;
1798 * Compute the number of levels necessary to support the new maxblkid.
1801 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1802 for (sz = dn->dn_nblkptr;
1803 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1806 if (new_nlevels > dn->dn_nlevels) {
1807 int old_nlevels = dn->dn_nlevels;
1810 dbuf_dirty_record_t *new, *dr, *dr_next;
1812 dn->dn_nlevels = new_nlevels;
1814 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1815 dn->dn_next_nlevels[txgoff] = new_nlevels;
1817 /* dirty the left indirects */
1818 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1820 new = dbuf_dirty(db, tx);
1821 dbuf_rele(db, FTAG);
1823 /* transfer the dirty records to the new indirect */
1824 mutex_enter(&dn->dn_mtx);
1825 mutex_enter(&new->dt.di.dr_mtx);
1826 list = &dn->dn_dirty_records[txgoff];
1827 for (dr = list_head(list); dr; dr = dr_next) {
1828 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1829 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1830 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1831 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1832 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1833 list_remove(&dn->dn_dirty_records[txgoff], dr);
1834 list_insert_tail(&new->dt.di.dr_children, dr);
1835 dr->dr_parent = new;
1838 mutex_exit(&new->dt.di.dr_mtx);
1839 mutex_exit(&dn->dn_mtx);
1844 rw_downgrade(&dn->dn_struct_rwlock);
1848 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1850 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1852 dmu_buf_will_dirty(&db->db, tx);
1853 dbuf_rele(db, FTAG);
1858 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1862 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1865 dmu_buf_impl_t db_search;
1869 mutex_enter(&dn->dn_dbufs_mtx);
1871 db_search.db_level = 1;
1872 db_search.db_blkid = start_blkid + 1;
1873 db_search.db_state = DB_SEARCH;
1876 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1878 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1880 if (db == NULL || db->db_level != 1 ||
1881 db->db_blkid >= end_blkid) {
1886 * Setup the next blkid we want to search for.
1888 db_search.db_blkid = db->db_blkid + 1;
1889 ASSERT3U(db->db_blkid, >=, start_blkid);
1892 * If the dbuf transitions to DB_EVICTING while we're trying
1893 * to dirty it, then we will be unable to discover it in
1894 * the dbuf hash table. This will result in a call to
1895 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1896 * lock. To avoid a deadlock, we drop the lock before
1897 * dirtying the level-1 dbuf.
1899 mutex_exit(&dn->dn_dbufs_mtx);
1900 dnode_dirty_l1(dn, db->db_blkid, tx);
1901 mutex_enter(&dn->dn_dbufs_mtx);
1906 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1908 db_search.db_level = 1;
1909 db_search.db_blkid = start_blkid + 1;
1910 db_search.db_state = DB_SEARCH;
1911 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1913 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1914 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
1915 if (db->db_level != 1 || db->db_blkid >= end_blkid)
1917 ASSERT(db->db_dirtycnt > 0);
1920 mutex_exit(&dn->dn_dbufs_mtx);
1924 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1927 uint64_t blkoff, blkid, nblks;
1928 int blksz, blkshift, head, tail;
1932 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1933 blksz = dn->dn_datablksz;
1934 blkshift = dn->dn_datablkshift;
1935 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1937 if (len == DMU_OBJECT_END) {
1938 len = UINT64_MAX - off;
1943 * First, block align the region to free:
1946 head = P2NPHASE(off, blksz);
1947 blkoff = P2PHASE(off, blksz);
1948 if ((off >> blkshift) > dn->dn_maxblkid)
1951 ASSERT(dn->dn_maxblkid == 0);
1952 if (off == 0 && len >= blksz) {
1954 * Freeing the whole block; fast-track this request.
1958 if (dn->dn_nlevels > 1)
1959 dnode_dirty_l1(dn, 0, tx);
1961 } else if (off >= blksz) {
1962 /* Freeing past end-of-data */
1965 /* Freeing part of the block. */
1967 ASSERT3U(head, >, 0);
1971 /* zero out any partial block data at the start of the range */
1973 ASSERT3U(blkoff + head, ==, blksz);
1976 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1977 TRUE, FALSE, FTAG, &db) == 0) {
1980 /* don't dirty if it isn't on disk and isn't dirty */
1981 if (db->db_last_dirty ||
1982 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1983 rw_exit(&dn->dn_struct_rwlock);
1984 dmu_buf_will_dirty(&db->db, tx);
1985 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1986 data = db->db.db_data;
1987 bzero(data + blkoff, head);
1989 dbuf_rele(db, FTAG);
1995 /* If the range was less than one block, we're done */
1999 /* If the remaining range is past end of file, we're done */
2000 if ((off >> blkshift) > dn->dn_maxblkid)
2003 ASSERT(ISP2(blksz));
2007 tail = P2PHASE(len, blksz);
2009 ASSERT0(P2PHASE(off, blksz));
2010 /* zero out any partial block data at the end of the range */
2014 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2015 TRUE, FALSE, FTAG, &db) == 0) {
2016 /* don't dirty if not on disk and not dirty */
2017 if (db->db_last_dirty ||
2018 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2019 rw_exit(&dn->dn_struct_rwlock);
2020 dmu_buf_will_dirty(&db->db, tx);
2021 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2022 bzero(db->db.db_data, tail);
2024 dbuf_rele(db, FTAG);
2029 /* If the range did not include a full block, we are done */
2033 ASSERT(IS_P2ALIGNED(off, blksz));
2034 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2035 blkid = off >> blkshift;
2036 nblks = len >> blkshift;
2041 * Dirty all the indirect blocks in this range. Note that only
2042 * the first and last indirect blocks can actually be written
2043 * (if they were partially freed) -- they must be dirtied, even if
2044 * they do not exist on disk yet. The interior blocks will
2045 * be freed by free_children(), so they will not actually be written.
2046 * Even though these interior blocks will not be written, we
2047 * dirty them for two reasons:
2049 * - It ensures that the indirect blocks remain in memory until
2050 * syncing context. (They have already been prefetched by
2051 * dmu_tx_hold_free(), so we don't have to worry about reading
2052 * them serially here.)
2054 * - The dirty space accounting will put pressure on the txg sync
2055 * mechanism to begin syncing, and to delay transactions if there
2056 * is a large amount of freeing. Even though these indirect
2057 * blocks will not be written, we could need to write the same
2058 * amount of space if we copy the freed BPs into deadlists.
2060 if (dn->dn_nlevels > 1) {
2061 uint64_t first, last;
2063 first = blkid >> epbs;
2064 dnode_dirty_l1(dn, first, tx);
2066 last = dn->dn_maxblkid >> epbs;
2068 last = (blkid + nblks - 1) >> epbs;
2070 dnode_dirty_l1(dn, last, tx);
2072 dnode_dirty_l1range(dn, first, last, tx);
2074 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2076 for (uint64_t i = first + 1; i < last; i++) {
2078 * Set i to the blockid of the next non-hole
2079 * level-1 indirect block at or after i. Note
2080 * that dnode_next_offset() operates in terms of
2081 * level-0-equivalent bytes.
2083 uint64_t ibyte = i << shift;
2084 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2091 * Normally we should not see an error, either
2092 * from dnode_next_offset() or dbuf_hold_level()
2093 * (except for ESRCH from dnode_next_offset).
2094 * If there is an i/o error, then when we read
2095 * this block in syncing context, it will use
2096 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2097 * to the "failmode" property. dnode_next_offset()
2098 * doesn't have a flag to indicate MUSTSUCCEED.
2103 dnode_dirty_l1(dn, i, tx);
2109 * Add this range to the dnode range list.
2110 * We will finish up this free operation in the syncing phase.
2112 mutex_enter(&dn->dn_mtx);
2113 int txgoff = tx->tx_txg & TXG_MASK;
2114 if (dn->dn_free_ranges[txgoff] == NULL) {
2115 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2117 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2118 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2119 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2120 blkid, nblks, tx->tx_txg);
2121 mutex_exit(&dn->dn_mtx);
2123 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2124 dnode_setdirty(dn, tx);
2127 rw_exit(&dn->dn_struct_rwlock);
2131 dnode_spill_freed(dnode_t *dn)
2135 mutex_enter(&dn->dn_mtx);
2136 for (i = 0; i < TXG_SIZE; i++) {
2137 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2140 mutex_exit(&dn->dn_mtx);
2141 return (i < TXG_SIZE);
2144 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2146 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2148 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2151 if (blkid == DMU_BONUS_BLKID)
2155 * If we're in the process of opening the pool, dp will not be
2156 * set yet, but there shouldn't be anything dirty.
2161 if (dn->dn_free_txg)
2164 if (blkid == DMU_SPILL_BLKID)
2165 return (dnode_spill_freed(dn));
2167 mutex_enter(&dn->dn_mtx);
2168 for (i = 0; i < TXG_SIZE; i++) {
2169 if (dn->dn_free_ranges[i] != NULL &&
2170 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2173 mutex_exit(&dn->dn_mtx);
2174 return (i < TXG_SIZE);
2177 /* call from syncing context when we actually write/free space for this dnode */
2179 dnode_diduse_space(dnode_t *dn, int64_t delta)
2182 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2184 (u_longlong_t)dn->dn_phys->dn_used,
2187 mutex_enter(&dn->dn_mtx);
2188 space = DN_USED_BYTES(dn->dn_phys);
2190 ASSERT3U(space + delta, >=, space); /* no overflow */
2192 ASSERT3U(space, >=, -delta); /* no underflow */
2195 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2196 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2197 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2198 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2200 dn->dn_phys->dn_used = space;
2201 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2203 mutex_exit(&dn->dn_mtx);
2207 * Scans a block at the indicated "level" looking for a hole or data,
2208 * depending on 'flags'.
2210 * If level > 0, then we are scanning an indirect block looking at its
2211 * pointers. If level == 0, then we are looking at a block of dnodes.
2213 * If we don't find what we are looking for in the block, we return ESRCH.
2214 * Otherwise, return with *offset pointing to the beginning (if searching
2215 * forwards) or end (if searching backwards) of the range covered by the
2216 * block pointer we matched on (or dnode).
2218 * The basic search algorithm used below by dnode_next_offset() is to
2219 * use this function to search up the block tree (widen the search) until
2220 * we find something (i.e., we don't return ESRCH) and then search back
2221 * down the tree (narrow the search) until we reach our original search
2225 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2226 int lvl, uint64_t blkfill, uint64_t txg)
2228 dmu_buf_impl_t *db = NULL;
2230 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2231 uint64_t epb = 1ULL << epbs;
2232 uint64_t minfill, maxfill;
2234 int i, inc, error, span;
2236 dprintf("probing object %llu offset %llx level %d of %u\n",
2237 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
2239 hole = ((flags & DNODE_FIND_HOLE) != 0);
2240 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2241 ASSERT(txg == 0 || !hole);
2243 if (lvl == dn->dn_phys->dn_nlevels) {
2245 epb = dn->dn_phys->dn_nblkptr;
2246 data = dn->dn_phys->dn_blkptr;
2248 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2249 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2251 if (error != ENOENT)
2256 * This can only happen when we are searching up
2257 * the block tree for data. We don't really need to
2258 * adjust the offset, as we will just end up looking
2259 * at the pointer to this block in its parent, and its
2260 * going to be unallocated, so we will skip over it.
2262 return (SET_ERROR(ESRCH));
2264 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
2266 dbuf_rele(db, FTAG);
2269 data = db->db.db_data;
2273 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2274 db->db_blkptr->blk_birth <= txg ||
2275 BP_IS_HOLE(db->db_blkptr))) {
2277 * This can only happen when we are searching up the tree
2278 * and these conditions mean that we need to keep climbing.
2280 error = SET_ERROR(ESRCH);
2281 } else if (lvl == 0) {
2282 dnode_phys_t *dnp = data;
2284 ASSERT(dn->dn_type == DMU_OT_DNODE);
2285 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2287 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2288 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2289 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2294 error = SET_ERROR(ESRCH);
2296 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2299 blkptr_t *bp = data;
2300 uint64_t start = *offset;
2301 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2303 maxfill = blkfill << ((lvl - 1) * epbs);
2310 *offset = *offset >> span;
2311 for (i = BF64_GET(*offset, 0, epbs);
2312 i >= 0 && i < epb; i += inc) {
2313 if (BP_GET_FILL(&bp[i]) >= minfill &&
2314 BP_GET_FILL(&bp[i]) <= maxfill &&
2315 (hole || bp[i].blk_birth > txg))
2317 if (inc > 0 || *offset > 0)
2320 *offset = *offset << span;
2322 /* traversing backwards; position offset at the end */
2323 ASSERT3U(*offset, <=, start);
2324 *offset = MIN(*offset + (1ULL << span) - 1, start);
2325 } else if (*offset < start) {
2328 if (i < 0 || i >= epb)
2329 error = SET_ERROR(ESRCH);
2333 dbuf_rele(db, FTAG);
2339 * Find the next hole, data, or sparse region at or after *offset.
2340 * The value 'blkfill' tells us how many items we expect to find
2341 * in an L0 data block; this value is 1 for normal objects,
2342 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2343 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2347 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2348 * Finds the next/previous hole/data in a file.
2349 * Used in dmu_offset_next().
2351 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2352 * Finds the next free/allocated dnode an objset's meta-dnode.
2353 * Only finds objects that have new contents since txg (ie.
2354 * bonus buffer changes and content removal are ignored).
2355 * Used in dmu_object_next().
2357 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2358 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2359 * Used in dmu_object_alloc().
2362 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2363 int minlvl, uint64_t blkfill, uint64_t txg)
2365 uint64_t initial_offset = *offset;
2369 if (!(flags & DNODE_FIND_HAVELOCK))
2370 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2372 if (dn->dn_phys->dn_nlevels == 0) {
2373 error = SET_ERROR(ESRCH);
2377 if (dn->dn_datablkshift == 0) {
2378 if (*offset < dn->dn_datablksz) {
2379 if (flags & DNODE_FIND_HOLE)
2380 *offset = dn->dn_datablksz;
2382 error = SET_ERROR(ESRCH);
2387 maxlvl = dn->dn_phys->dn_nlevels;
2389 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2390 error = dnode_next_offset_level(dn,
2391 flags, offset, lvl, blkfill, txg);
2396 while (error == 0 && --lvl >= minlvl) {
2397 error = dnode_next_offset_level(dn,
2398 flags, offset, lvl, blkfill, txg);
2402 * There's always a "virtual hole" at the end of the object, even
2403 * if all BP's which physically exist are non-holes.
2405 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2406 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2410 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2411 initial_offset < *offset : initial_offset > *offset))
2412 error = SET_ERROR(ESRCH);
2414 if (!(flags & DNODE_FIND_HAVELOCK))
2415 rw_exit(&dn->dn_struct_rwlock);