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, 2019 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27 #include <sys/zfs_context.h>
29 #include <sys/dnode.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 #include <sys/trace_zfs.h>
41 #include <sys/zfs_project.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_free_interior_lock_retry", KSTAT_DATA_UINT64 },
59 { "dnode_allocate", KSTAT_DATA_UINT64 },
60 { "dnode_reallocate", KSTAT_DATA_UINT64 },
61 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
62 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
63 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
64 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
65 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
66 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
67 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
68 { "dnode_move_special", KSTAT_DATA_UINT64 },
69 { "dnode_move_handle", KSTAT_DATA_UINT64 },
70 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
71 { "dnode_move_active", KSTAT_DATA_UINT64 },
74 static kstat_t *dnode_ksp;
75 static kmem_cache_t *dnode_cache;
77 static dnode_phys_t dnode_phys_zero __maybe_unused;
79 int zfs_default_bs = SPA_MINBLOCKSHIFT;
80 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
83 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
87 dbuf_compare(const void *x1, const void *x2)
89 const dmu_buf_impl_t *d1 = x1;
90 const dmu_buf_impl_t *d2 = x2;
92 int cmp = TREE_CMP(d1->db_level, d2->db_level);
96 cmp = TREE_CMP(d1->db_blkid, d2->db_blkid);
100 if (d1->db_state == DB_SEARCH) {
101 ASSERT3S(d2->db_state, !=, DB_SEARCH);
103 } else if (d2->db_state == DB_SEARCH) {
104 ASSERT3S(d1->db_state, !=, DB_SEARCH);
108 return (TREE_PCMP(d1, d2));
113 dnode_cons(void *arg, void *unused, int kmflag)
118 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
119 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
120 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
121 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
122 cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
125 * Every dbuf has a reference, and dropping a tracked reference is
126 * O(number of references), so don't track dn_holds.
128 zfs_refcount_create_untracked(&dn->dn_holds);
129 zfs_refcount_create(&dn->dn_tx_holds);
130 list_link_init(&dn->dn_link);
132 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
133 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
134 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
135 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
136 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
137 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
138 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
139 bzero(&dn->dn_next_maxblkid[0], sizeof (dn->dn_next_maxblkid));
141 for (i = 0; i < TXG_SIZE; i++) {
142 multilist_link_init(&dn->dn_dirty_link[i]);
143 dn->dn_free_ranges[i] = NULL;
144 list_create(&dn->dn_dirty_records[i],
145 sizeof (dbuf_dirty_record_t),
146 offsetof(dbuf_dirty_record_t, dr_dirty_node));
149 dn->dn_allocated_txg = 0;
151 dn->dn_assigned_txg = 0;
152 dn->dn_dirty_txg = 0;
154 dn->dn_dirtyctx_firstset = NULL;
156 dn->dn_have_spill = B_FALSE;
162 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
165 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
168 dn->dn_dbufs_count = 0;
169 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
170 offsetof(dmu_buf_impl_t, db_link));
178 dnode_dest(void *arg, void *unused)
183 rw_destroy(&dn->dn_struct_rwlock);
184 mutex_destroy(&dn->dn_mtx);
185 mutex_destroy(&dn->dn_dbufs_mtx);
186 cv_destroy(&dn->dn_notxholds);
187 cv_destroy(&dn->dn_nodnholds);
188 zfs_refcount_destroy(&dn->dn_holds);
189 zfs_refcount_destroy(&dn->dn_tx_holds);
190 ASSERT(!list_link_active(&dn->dn_link));
192 for (i = 0; i < TXG_SIZE; i++) {
193 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
194 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
195 list_destroy(&dn->dn_dirty_records[i]);
196 ASSERT0(dn->dn_next_nblkptr[i]);
197 ASSERT0(dn->dn_next_nlevels[i]);
198 ASSERT0(dn->dn_next_indblkshift[i]);
199 ASSERT0(dn->dn_next_bonustype[i]);
200 ASSERT0(dn->dn_rm_spillblk[i]);
201 ASSERT0(dn->dn_next_bonuslen[i]);
202 ASSERT0(dn->dn_next_blksz[i]);
203 ASSERT0(dn->dn_next_maxblkid[i]);
206 ASSERT0(dn->dn_allocated_txg);
207 ASSERT0(dn->dn_free_txg);
208 ASSERT0(dn->dn_assigned_txg);
209 ASSERT0(dn->dn_dirty_txg);
210 ASSERT0(dn->dn_dirtyctx);
211 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
212 ASSERT3P(dn->dn_bonus, ==, NULL);
213 ASSERT(!dn->dn_have_spill);
214 ASSERT3P(dn->dn_zio, ==, NULL);
215 ASSERT0(dn->dn_oldused);
216 ASSERT0(dn->dn_oldflags);
217 ASSERT0(dn->dn_olduid);
218 ASSERT0(dn->dn_oldgid);
219 ASSERT0(dn->dn_oldprojid);
220 ASSERT0(dn->dn_newuid);
221 ASSERT0(dn->dn_newgid);
222 ASSERT0(dn->dn_newprojid);
223 ASSERT0(dn->dn_id_flags);
225 ASSERT0(dn->dn_dbufs_count);
226 avl_destroy(&dn->dn_dbufs);
232 ASSERT(dnode_cache == NULL);
233 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
234 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
235 kmem_cache_set_move(dnode_cache, dnode_move);
237 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
238 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
240 if (dnode_ksp != NULL) {
241 dnode_ksp->ks_data = &dnode_stats;
242 kstat_install(dnode_ksp);
249 if (dnode_ksp != NULL) {
250 kstat_delete(dnode_ksp);
254 kmem_cache_destroy(dnode_cache);
261 dnode_verify(dnode_t *dn)
263 int drop_struct_lock = FALSE;
266 ASSERT(dn->dn_objset);
267 ASSERT(dn->dn_handle->dnh_dnode == dn);
269 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
271 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
274 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
275 rw_enter(&dn->dn_struct_rwlock, RW_READER);
276 drop_struct_lock = TRUE;
278 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
280 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
281 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
282 if (dn->dn_datablkshift) {
283 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
284 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
285 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
287 ASSERT3U(dn->dn_nlevels, <=, 30);
288 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
289 ASSERT3U(dn->dn_nblkptr, >=, 1);
290 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
291 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
292 ASSERT3U(dn->dn_datablksz, ==,
293 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
294 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
295 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
296 dn->dn_bonuslen, <=, max_bonuslen);
297 for (i = 0; i < TXG_SIZE; i++) {
298 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
301 if (dn->dn_phys->dn_type != DMU_OT_NONE)
302 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
303 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
304 if (dn->dn_dbuf != NULL) {
305 ASSERT3P(dn->dn_phys, ==,
306 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
307 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
309 if (drop_struct_lock)
310 rw_exit(&dn->dn_struct_rwlock);
315 dnode_byteswap(dnode_phys_t *dnp)
317 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
320 if (dnp->dn_type == DMU_OT_NONE) {
321 bzero(dnp, sizeof (dnode_phys_t));
325 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
326 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
327 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
328 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
329 dnp->dn_used = BSWAP_64(dnp->dn_used);
332 * dn_nblkptr is only one byte, so it's OK to read it in either
333 * byte order. We can't read dn_bouslen.
335 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
336 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
337 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
338 buf64[i] = BSWAP_64(buf64[i]);
341 * OK to check dn_bonuslen for zero, because it won't matter if
342 * we have the wrong byte order. This is necessary because the
343 * dnode dnode is smaller than a regular dnode.
345 if (dnp->dn_bonuslen != 0) {
347 * Note that the bonus length calculated here may be
348 * longer than the actual bonus buffer. This is because
349 * we always put the bonus buffer after the last block
350 * pointer (instead of packing it against the end of the
353 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
354 int slots = dnp->dn_extra_slots + 1;
355 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
356 dmu_object_byteswap_t byteswap;
357 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
358 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
359 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
362 /* Swap SPILL block if we have one */
363 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
364 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
368 dnode_buf_byteswap(void *vbuf, size_t size)
372 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
373 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
376 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
380 if (dnp->dn_type != DMU_OT_NONE)
381 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
386 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
388 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
390 dnode_setdirty(dn, tx);
391 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
392 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
393 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
395 if (newsize < dn->dn_bonuslen) {
396 /* clear any data after the end of the new size */
397 size_t diff = dn->dn_bonuslen - newsize;
398 char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
399 bzero(data_end, diff);
402 dn->dn_bonuslen = newsize;
404 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
406 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
407 rw_exit(&dn->dn_struct_rwlock);
411 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
413 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
414 dnode_setdirty(dn, tx);
415 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
416 dn->dn_bonustype = newtype;
417 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
418 rw_exit(&dn->dn_struct_rwlock);
422 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
424 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
425 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
426 dnode_setdirty(dn, tx);
427 dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
428 dn->dn_have_spill = B_FALSE;
432 dnode_setdblksz(dnode_t *dn, int size)
434 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
435 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
436 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
437 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
438 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
439 dn->dn_datablksz = size;
440 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
441 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
445 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
446 uint64_t object, dnode_handle_t *dnh)
450 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
454 * Defer setting dn_objset until the dnode is ready to be a candidate
455 * for the dnode_move() callback.
457 dn->dn_object = object;
462 if (dnp->dn_datablkszsec) {
463 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
465 dn->dn_datablksz = 0;
466 dn->dn_datablkszsec = 0;
467 dn->dn_datablkshift = 0;
469 dn->dn_indblkshift = dnp->dn_indblkshift;
470 dn->dn_nlevels = dnp->dn_nlevels;
471 dn->dn_type = dnp->dn_type;
472 dn->dn_nblkptr = dnp->dn_nblkptr;
473 dn->dn_checksum = dnp->dn_checksum;
474 dn->dn_compress = dnp->dn_compress;
475 dn->dn_bonustype = dnp->dn_bonustype;
476 dn->dn_bonuslen = dnp->dn_bonuslen;
477 dn->dn_num_slots = dnp->dn_extra_slots + 1;
478 dn->dn_maxblkid = dnp->dn_maxblkid;
479 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
482 dmu_zfetch_init(&dn->dn_zfetch, dn);
484 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
485 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
486 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
488 mutex_enter(&os->os_lock);
491 * Exclude special dnodes from os_dnodes so an empty os_dnodes
492 * signifies that the special dnodes have no references from
493 * their children (the entries in os_dnodes). This allows
494 * dnode_destroy() to easily determine if the last child has
495 * been removed and then complete eviction of the objset.
497 if (!DMU_OBJECT_IS_SPECIAL(object))
498 list_insert_head(&os->os_dnodes, dn);
502 * Everything else must be valid before assigning dn_objset
503 * makes the dnode eligible for dnode_move().
508 mutex_exit(&os->os_lock);
510 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
516 * Caller must be holding the dnode handle, which is released upon return.
519 dnode_destroy(dnode_t *dn)
521 objset_t *os = dn->dn_objset;
522 boolean_t complete_os_eviction = B_FALSE;
524 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
526 mutex_enter(&os->os_lock);
527 POINTER_INVALIDATE(&dn->dn_objset);
528 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
529 list_remove(&os->os_dnodes, dn);
530 complete_os_eviction =
531 list_is_empty(&os->os_dnodes) &&
532 list_link_active(&os->os_evicting_node);
534 mutex_exit(&os->os_lock);
536 /* the dnode can no longer move, so we can release the handle */
537 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
538 zrl_remove(&dn->dn_handle->dnh_zrlock);
540 dn->dn_allocated_txg = 0;
542 dn->dn_assigned_txg = 0;
543 dn->dn_dirty_txg = 0;
546 dn->dn_dirtyctx_firstset = NULL;
547 if (dn->dn_bonus != NULL) {
548 mutex_enter(&dn->dn_bonus->db_mtx);
549 dbuf_destroy(dn->dn_bonus);
554 dn->dn_have_spill = B_FALSE;
559 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
562 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
565 dmu_zfetch_fini(&dn->dn_zfetch);
566 kmem_cache_free(dnode_cache, dn);
567 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
569 if (complete_os_eviction)
570 dmu_objset_evict_done(os);
574 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
575 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
579 ASSERT3U(dn_slots, >, 0);
580 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
581 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
582 ASSERT3U(blocksize, <=,
583 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
585 blocksize = 1 << zfs_default_bs;
587 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
590 ibs = zfs_default_ibs;
592 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
594 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
595 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
596 DNODE_STAT_BUMP(dnode_allocate);
598 ASSERT(dn->dn_type == DMU_OT_NONE);
599 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
600 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
601 ASSERT(ot != DMU_OT_NONE);
602 ASSERT(DMU_OT_IS_VALID(ot));
603 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
604 (bonustype == DMU_OT_SA && bonuslen == 0) ||
605 (bonustype != DMU_OT_NONE && bonuslen != 0));
606 ASSERT(DMU_OT_IS_VALID(bonustype));
607 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
608 ASSERT(dn->dn_type == DMU_OT_NONE);
609 ASSERT0(dn->dn_maxblkid);
610 ASSERT0(dn->dn_allocated_txg);
611 ASSERT0(dn->dn_assigned_txg);
612 ASSERT0(dn->dn_dirty_txg);
613 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
614 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
615 ASSERT(avl_is_empty(&dn->dn_dbufs));
617 for (i = 0; i < TXG_SIZE; i++) {
618 ASSERT0(dn->dn_next_nblkptr[i]);
619 ASSERT0(dn->dn_next_nlevels[i]);
620 ASSERT0(dn->dn_next_indblkshift[i]);
621 ASSERT0(dn->dn_next_bonuslen[i]);
622 ASSERT0(dn->dn_next_bonustype[i]);
623 ASSERT0(dn->dn_rm_spillblk[i]);
624 ASSERT0(dn->dn_next_blksz[i]);
625 ASSERT0(dn->dn_next_maxblkid[i]);
626 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
627 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
628 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
632 dnode_setdblksz(dn, blocksize);
633 dn->dn_indblkshift = ibs;
635 dn->dn_num_slots = dn_slots;
636 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
639 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
640 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
644 dn->dn_bonustype = bonustype;
645 dn->dn_bonuslen = bonuslen;
646 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
647 dn->dn_compress = ZIO_COMPRESS_INHERIT;
651 dn->dn_dirtyctx_firstset = NULL;
653 dn->dn_allocated_txg = tx->tx_txg;
656 dnode_setdirty(dn, tx);
657 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
658 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
659 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
660 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
664 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
665 dmu_object_type_t bonustype, int bonuslen, int dn_slots,
666 boolean_t keep_spill, dmu_tx_t *tx)
670 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
671 ASSERT3U(blocksize, <=,
672 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
673 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
674 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
675 ASSERT(tx->tx_txg != 0);
676 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
677 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
678 (bonustype == DMU_OT_SA && bonuslen == 0));
679 ASSERT(DMU_OT_IS_VALID(bonustype));
680 ASSERT3U(bonuslen, <=,
681 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
682 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
684 dnode_free_interior_slots(dn);
685 DNODE_STAT_BUMP(dnode_reallocate);
687 /* clean up any unreferenced dbufs */
688 dnode_evict_dbufs(dn);
692 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
693 dnode_setdirty(dn, tx);
694 if (dn->dn_datablksz != blocksize) {
695 /* change blocksize */
696 ASSERT0(dn->dn_maxblkid);
697 ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
698 dnode_block_freed(dn, 0));
700 dnode_setdblksz(dn, blocksize);
701 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
703 if (dn->dn_bonuslen != bonuslen)
704 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
706 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
709 nblkptr = MIN(DN_MAX_NBLKPTR,
710 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
712 if (dn->dn_bonustype != bonustype)
713 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
714 if (dn->dn_nblkptr != nblkptr)
715 dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
716 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
717 dbuf_rm_spill(dn, tx);
718 dnode_rm_spill(dn, tx);
721 rw_exit(&dn->dn_struct_rwlock);
726 /* change bonus size and type */
727 mutex_enter(&dn->dn_mtx);
728 dn->dn_bonustype = bonustype;
729 dn->dn_bonuslen = bonuslen;
730 dn->dn_num_slots = dn_slots;
731 dn->dn_nblkptr = nblkptr;
732 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
733 dn->dn_compress = ZIO_COMPRESS_INHERIT;
734 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
736 /* fix up the bonus db_size */
738 dn->dn_bonus->db.db_size =
739 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
740 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
741 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
744 dn->dn_allocated_txg = tx->tx_txg;
745 mutex_exit(&dn->dn_mtx);
750 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
754 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
755 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
756 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
757 ASSERT(!MUTEX_HELD(&odn->dn_zfetch.zf_lock));
760 ndn->dn_objset = odn->dn_objset;
761 ndn->dn_object = odn->dn_object;
762 ndn->dn_dbuf = odn->dn_dbuf;
763 ndn->dn_handle = odn->dn_handle;
764 ndn->dn_phys = odn->dn_phys;
765 ndn->dn_type = odn->dn_type;
766 ndn->dn_bonuslen = odn->dn_bonuslen;
767 ndn->dn_bonustype = odn->dn_bonustype;
768 ndn->dn_nblkptr = odn->dn_nblkptr;
769 ndn->dn_checksum = odn->dn_checksum;
770 ndn->dn_compress = odn->dn_compress;
771 ndn->dn_nlevels = odn->dn_nlevels;
772 ndn->dn_indblkshift = odn->dn_indblkshift;
773 ndn->dn_datablkshift = odn->dn_datablkshift;
774 ndn->dn_datablkszsec = odn->dn_datablkszsec;
775 ndn->dn_datablksz = odn->dn_datablksz;
776 ndn->dn_maxblkid = odn->dn_maxblkid;
777 ndn->dn_num_slots = odn->dn_num_slots;
778 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
779 sizeof (odn->dn_next_type));
780 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
781 sizeof (odn->dn_next_nblkptr));
782 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
783 sizeof (odn->dn_next_nlevels));
784 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
785 sizeof (odn->dn_next_indblkshift));
786 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
787 sizeof (odn->dn_next_bonustype));
788 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
789 sizeof (odn->dn_rm_spillblk));
790 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
791 sizeof (odn->dn_next_bonuslen));
792 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
793 sizeof (odn->dn_next_blksz));
794 bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
795 sizeof (odn->dn_next_maxblkid));
796 for (i = 0; i < TXG_SIZE; i++) {
797 list_move_tail(&ndn->dn_dirty_records[i],
798 &odn->dn_dirty_records[i]);
800 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
801 sizeof (odn->dn_free_ranges));
802 ndn->dn_allocated_txg = odn->dn_allocated_txg;
803 ndn->dn_free_txg = odn->dn_free_txg;
804 ndn->dn_assigned_txg = odn->dn_assigned_txg;
805 ndn->dn_dirty_txg = odn->dn_dirty_txg;
806 ndn->dn_dirtyctx = odn->dn_dirtyctx;
807 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
808 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
809 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
810 ASSERT(avl_is_empty(&ndn->dn_dbufs));
811 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
812 ndn->dn_dbufs_count = odn->dn_dbufs_count;
813 ndn->dn_bonus = odn->dn_bonus;
814 ndn->dn_have_spill = odn->dn_have_spill;
815 ndn->dn_zio = odn->dn_zio;
816 ndn->dn_oldused = odn->dn_oldused;
817 ndn->dn_oldflags = odn->dn_oldflags;
818 ndn->dn_olduid = odn->dn_olduid;
819 ndn->dn_oldgid = odn->dn_oldgid;
820 ndn->dn_oldprojid = odn->dn_oldprojid;
821 ndn->dn_newuid = odn->dn_newuid;
822 ndn->dn_newgid = odn->dn_newgid;
823 ndn->dn_newprojid = odn->dn_newprojid;
824 ndn->dn_id_flags = odn->dn_id_flags;
825 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
826 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
827 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
830 * Update back pointers. Updating the handle fixes the back pointer of
831 * every descendant dbuf as well as the bonus dbuf.
833 ASSERT(ndn->dn_handle->dnh_dnode == odn);
834 ndn->dn_handle->dnh_dnode = ndn;
835 if (ndn->dn_zfetch.zf_dnode == odn) {
836 ndn->dn_zfetch.zf_dnode = ndn;
840 * Invalidate the original dnode by clearing all of its back pointers.
843 odn->dn_handle = NULL;
844 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
845 offsetof(dmu_buf_impl_t, db_link));
846 odn->dn_dbufs_count = 0;
847 odn->dn_bonus = NULL;
848 dmu_zfetch_fini(&odn->dn_zfetch);
851 * Set the low bit of the objset pointer to ensure that dnode_move()
852 * recognizes the dnode as invalid in any subsequent callback.
854 POINTER_INVALIDATE(&odn->dn_objset);
857 * Satisfy the destructor.
859 for (i = 0; i < TXG_SIZE; i++) {
860 list_create(&odn->dn_dirty_records[i],
861 sizeof (dbuf_dirty_record_t),
862 offsetof(dbuf_dirty_record_t, dr_dirty_node));
863 odn->dn_free_ranges[i] = NULL;
864 odn->dn_next_nlevels[i] = 0;
865 odn->dn_next_indblkshift[i] = 0;
866 odn->dn_next_bonustype[i] = 0;
867 odn->dn_rm_spillblk[i] = 0;
868 odn->dn_next_bonuslen[i] = 0;
869 odn->dn_next_blksz[i] = 0;
871 odn->dn_allocated_txg = 0;
872 odn->dn_free_txg = 0;
873 odn->dn_assigned_txg = 0;
874 odn->dn_dirty_txg = 0;
875 odn->dn_dirtyctx = 0;
876 odn->dn_dirtyctx_firstset = NULL;
877 odn->dn_have_spill = B_FALSE;
880 odn->dn_oldflags = 0;
883 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
886 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
887 odn->dn_id_flags = 0;
893 odn->dn_moved = (uint8_t)-1;
898 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
900 dnode_t *odn = buf, *ndn = newbuf;
906 * The dnode is on the objset's list of known dnodes if the objset
907 * pointer is valid. We set the low bit of the objset pointer when
908 * freeing the dnode to invalidate it, and the memory patterns written
909 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
910 * A newly created dnode sets the objset pointer last of all to indicate
911 * that the dnode is known and in a valid state to be moved by this
915 if (!POINTER_IS_VALID(os)) {
916 DNODE_STAT_BUMP(dnode_move_invalid);
917 return (KMEM_CBRC_DONT_KNOW);
921 * Ensure that the objset does not go away during the move.
923 rw_enter(&os_lock, RW_WRITER);
924 if (os != odn->dn_objset) {
926 DNODE_STAT_BUMP(dnode_move_recheck1);
927 return (KMEM_CBRC_DONT_KNOW);
931 * If the dnode is still valid, then so is the objset. We know that no
932 * valid objset can be freed while we hold os_lock, so we can safely
933 * ensure that the objset remains in use.
935 mutex_enter(&os->os_lock);
938 * Recheck the objset pointer in case the dnode was removed just before
939 * acquiring the lock.
941 if (os != odn->dn_objset) {
942 mutex_exit(&os->os_lock);
944 DNODE_STAT_BUMP(dnode_move_recheck2);
945 return (KMEM_CBRC_DONT_KNOW);
949 * At this point we know that as long as we hold os->os_lock, the dnode
950 * cannot be freed and fields within the dnode can be safely accessed.
951 * The objset listing this dnode cannot go away as long as this dnode is
955 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
956 mutex_exit(&os->os_lock);
957 DNODE_STAT_BUMP(dnode_move_special);
958 return (KMEM_CBRC_NO);
960 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
963 * Lock the dnode handle to prevent the dnode from obtaining any new
964 * holds. This also prevents the descendant dbufs and the bonus dbuf
965 * from accessing the dnode, so that we can discount their holds. The
966 * handle is safe to access because we know that while the dnode cannot
967 * go away, neither can its handle. Once we hold dnh_zrlock, we can
968 * safely move any dnode referenced only by dbufs.
970 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
971 mutex_exit(&os->os_lock);
972 DNODE_STAT_BUMP(dnode_move_handle);
973 return (KMEM_CBRC_LATER);
977 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
978 * We need to guarantee that there is a hold for every dbuf in order to
979 * determine whether the dnode is actively referenced. Falsely matching
980 * a dbuf to an active hold would lead to an unsafe move. It's possible
981 * that a thread already having an active dnode hold is about to add a
982 * dbuf, and we can't compare hold and dbuf counts while the add is in
985 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
986 zrl_exit(&odn->dn_handle->dnh_zrlock);
987 mutex_exit(&os->os_lock);
988 DNODE_STAT_BUMP(dnode_move_rwlock);
989 return (KMEM_CBRC_LATER);
993 * A dbuf may be removed (evicted) without an active dnode hold. In that
994 * case, the dbuf count is decremented under the handle lock before the
995 * dbuf's hold is released. This order ensures that if we count the hold
996 * after the dbuf is removed but before its hold is released, we will
997 * treat the unmatched hold as active and exit safely. If we count the
998 * hold before the dbuf is removed, the hold is discounted, and the
999 * removal is blocked until the move completes.
1001 refcount = zfs_refcount_count(&odn->dn_holds);
1002 ASSERT(refcount >= 0);
1003 dbufs = DN_DBUFS_COUNT(odn);
1005 /* We can't have more dbufs than dnode holds. */
1006 ASSERT3U(dbufs, <=, refcount);
1007 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1010 if (refcount > dbufs) {
1011 rw_exit(&odn->dn_struct_rwlock);
1012 zrl_exit(&odn->dn_handle->dnh_zrlock);
1013 mutex_exit(&os->os_lock);
1014 DNODE_STAT_BUMP(dnode_move_active);
1015 return (KMEM_CBRC_LATER);
1018 rw_exit(&odn->dn_struct_rwlock);
1021 * At this point we know that anyone with a hold on the dnode is not
1022 * actively referencing it. The dnode is known and in a valid state to
1023 * move. We're holding the locks needed to execute the critical section.
1025 dnode_move_impl(odn, ndn);
1027 list_link_replace(&odn->dn_link, &ndn->dn_link);
1028 /* If the dnode was safe to move, the refcount cannot have changed. */
1029 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1030 ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1031 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1032 mutex_exit(&os->os_lock);
1034 return (KMEM_CBRC_YES);
1036 #endif /* _KERNEL */
1039 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1041 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1043 for (int i = idx; i < idx + slots; i++) {
1044 dnode_handle_t *dnh = &children->dnc_children[i];
1045 zrl_add(&dnh->dnh_zrlock);
1050 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1052 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1054 for (int i = idx; i < idx + slots; i++) {
1055 dnode_handle_t *dnh = &children->dnc_children[i];
1057 if (zrl_is_locked(&dnh->dnh_zrlock))
1058 zrl_exit(&dnh->dnh_zrlock);
1060 zrl_remove(&dnh->dnh_zrlock);
1065 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1067 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1069 for (int i = idx; i < idx + slots; i++) {
1070 dnode_handle_t *dnh = &children->dnc_children[i];
1072 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1073 for (int j = idx; j < i; j++) {
1074 dnh = &children->dnc_children[j];
1075 zrl_exit(&dnh->dnh_zrlock);
1086 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1088 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1090 for (int i = idx; i < idx + slots; i++) {
1091 dnode_handle_t *dnh = &children->dnc_children[i];
1092 dnh->dnh_dnode = ptr;
1097 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1099 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1102 * If all dnode slots are either already free or
1103 * evictable return B_TRUE.
1105 for (int i = idx; i < idx + slots; i++) {
1106 dnode_handle_t *dnh = &children->dnc_children[i];
1107 dnode_t *dn = dnh->dnh_dnode;
1109 if (dn == DN_SLOT_FREE) {
1111 } else if (DN_SLOT_IS_PTR(dn)) {
1112 mutex_enter(&dn->dn_mtx);
1113 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1114 zfs_refcount_is_zero(&dn->dn_holds) &&
1115 !DNODE_IS_DIRTY(dn));
1116 mutex_exit(&dn->dn_mtx);
1131 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1133 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1135 for (int i = idx; i < idx + slots; i++) {
1136 dnode_handle_t *dnh = &children->dnc_children[i];
1138 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1140 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1141 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1142 dnode_destroy(dnh->dnh_dnode);
1143 dnh->dnh_dnode = DN_SLOT_FREE;
1149 dnode_free_interior_slots(dnode_t *dn)
1151 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1152 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1153 int idx = (dn->dn_object & (epb - 1)) + 1;
1154 int slots = dn->dn_num_slots - 1;
1159 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1161 while (!dnode_slots_tryenter(children, idx, slots)) {
1162 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1166 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1167 dnode_slots_rele(children, idx, slots);
1171 dnode_special_close(dnode_handle_t *dnh)
1173 dnode_t *dn = dnh->dnh_dnode;
1176 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1177 * zfs_refcount_remove()
1179 mutex_enter(&dn->dn_mtx);
1180 if (zfs_refcount_count(&dn->dn_holds) > 0)
1181 cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1182 mutex_exit(&dn->dn_mtx);
1183 ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1185 ASSERT(dn->dn_dbuf == NULL ||
1186 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1187 zrl_add(&dnh->dnh_zrlock);
1188 dnode_destroy(dn); /* implicit zrl_remove() */
1189 zrl_destroy(&dnh->dnh_zrlock);
1190 dnh->dnh_dnode = NULL;
1194 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1195 dnode_handle_t *dnh)
1199 zrl_init(&dnh->dnh_zrlock);
1200 VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1202 dn = dnode_create(os, dnp, NULL, object, dnh);
1205 zrl_exit(&dnh->dnh_zrlock);
1209 dnode_buf_evict_async(void *dbu)
1211 dnode_children_t *dnc = dbu;
1213 DNODE_STAT_BUMP(dnode_buf_evict);
1215 for (int i = 0; i < dnc->dnc_count; i++) {
1216 dnode_handle_t *dnh = &dnc->dnc_children[i];
1220 * The dnode handle lock guards against the dnode moving to
1221 * another valid address, so there is no need here to guard
1222 * against changes to or from NULL.
1224 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1225 zrl_destroy(&dnh->dnh_zrlock);
1226 dnh->dnh_dnode = DN_SLOT_UNINIT;
1230 zrl_add(&dnh->dnh_zrlock);
1231 dn = dnh->dnh_dnode;
1233 * If there are holds on this dnode, then there should
1234 * be holds on the dnode's containing dbuf as well; thus
1235 * it wouldn't be eligible for eviction and this function
1236 * would not have been called.
1238 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1239 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1241 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1242 zrl_destroy(&dnh->dnh_zrlock);
1243 dnh->dnh_dnode = DN_SLOT_UNINIT;
1245 kmem_free(dnc, sizeof (dnode_children_t) +
1246 dnc->dnc_count * sizeof (dnode_handle_t));
1250 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1251 * to ensure the hole at the specified object offset is large enough to
1252 * hold the dnode being created. The slots parameter is also used to ensure
1253 * a dnode does not span multiple dnode blocks. In both of these cases, if
1254 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1255 * are only possible when using DNODE_MUST_BE_FREE.
1257 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1258 * dnode_hold_impl() will check if the requested dnode is already consumed
1259 * as an extra dnode slot by an large dnode, in which case it returns
1262 * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1263 * return whether the hold would succeed or not. tag and dnp should set to
1264 * NULL in this case.
1267 * EINVAL - Invalid object number or flags.
1268 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1269 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1270 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1271 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1272 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1273 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1274 * EIO - I/O error when reading the meta dnode dbuf.
1276 * succeeds even for free dnodes.
1279 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1280 void *tag, dnode_t **dnp)
1283 int drop_struct_lock = FALSE;
1288 dnode_children_t *dnc;
1289 dnode_phys_t *dn_block;
1290 dnode_handle_t *dnh;
1292 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1293 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1294 IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1297 * If you are holding the spa config lock as writer, you shouldn't
1298 * be asking the DMU to do *anything* unless it's the root pool
1299 * which may require us to read from the root filesystem while
1300 * holding some (not all) of the locks as writer.
1302 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1303 (spa_is_root(os->os_spa) &&
1304 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1306 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1308 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1309 object == DMU_PROJECTUSED_OBJECT) {
1310 if (object == DMU_USERUSED_OBJECT)
1311 dn = DMU_USERUSED_DNODE(os);
1312 else if (object == DMU_GROUPUSED_OBJECT)
1313 dn = DMU_GROUPUSED_DNODE(os);
1315 dn = DMU_PROJECTUSED_DNODE(os);
1317 return (SET_ERROR(ENOENT));
1319 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1320 return (SET_ERROR(ENOENT));
1321 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1322 return (SET_ERROR(EEXIST));
1324 /* Don't actually hold if dry run, just return 0 */
1325 if (!(flag & DNODE_DRY_RUN)) {
1326 (void) zfs_refcount_add(&dn->dn_holds, tag);
1332 if (object == 0 || object >= DN_MAX_OBJECT)
1333 return (SET_ERROR(EINVAL));
1335 mdn = DMU_META_DNODE(os);
1336 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1340 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1341 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1342 drop_struct_lock = TRUE;
1345 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1346 db = dbuf_hold(mdn, blk, FTAG);
1347 if (drop_struct_lock)
1348 rw_exit(&mdn->dn_struct_rwlock);
1350 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1351 return (SET_ERROR(EIO));
1355 * We do not need to decrypt to read the dnode so it doesn't matter
1356 * if we get the encrypted or decrypted version.
1358 err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1360 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1361 dbuf_rele(db, FTAG);
1365 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1366 epb = db->db.db_size >> DNODE_SHIFT;
1368 idx = object & (epb - 1);
1369 dn_block = (dnode_phys_t *)db->db.db_data;
1371 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1372 dnc = dmu_buf_get_user(&db->db);
1375 dnode_children_t *winner;
1378 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1379 epb * sizeof (dnode_handle_t), KM_SLEEP);
1380 dnc->dnc_count = epb;
1381 dnh = &dnc->dnc_children[0];
1383 /* Initialize dnode slot status from dnode_phys_t */
1384 for (int i = 0; i < epb; i++) {
1385 zrl_init(&dnh[i].dnh_zrlock);
1392 if (dn_block[i].dn_type != DMU_OT_NONE) {
1393 int interior = dn_block[i].dn_extra_slots;
1395 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1396 dnode_set_slots(dnc, i + 1, interior,
1400 dnh[i].dnh_dnode = DN_SLOT_FREE;
1405 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1406 dnode_buf_evict_async, NULL);
1407 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1408 if (winner != NULL) {
1410 for (int i = 0; i < epb; i++)
1411 zrl_destroy(&dnh[i].dnh_zrlock);
1413 kmem_free(dnc, sizeof (dnode_children_t) +
1414 epb * sizeof (dnode_handle_t));
1419 ASSERT(dnc->dnc_count == epb);
1421 if (flag & DNODE_MUST_BE_ALLOCATED) {
1424 dnode_slots_hold(dnc, idx, slots);
1425 dnh = &dnc->dnc_children[idx];
1427 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1428 dn = dnh->dnh_dnode;
1429 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1430 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1431 dnode_slots_rele(dnc, idx, slots);
1432 dbuf_rele(db, FTAG);
1433 return (SET_ERROR(EEXIST));
1434 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1435 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1436 dnode_slots_rele(dnc, idx, slots);
1437 dbuf_rele(db, FTAG);
1438 return (SET_ERROR(ENOENT));
1440 dnode_slots_rele(dnc, idx, slots);
1441 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1442 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1447 * Someone else won the race and called dnode_create()
1448 * after we checked DN_SLOT_IS_PTR() above but before
1449 * we acquired the lock.
1451 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1452 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1453 dn = dnh->dnh_dnode;
1455 dn = dnode_create(os, dn_block + idx, db,
1460 mutex_enter(&dn->dn_mtx);
1461 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1462 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1463 mutex_exit(&dn->dn_mtx);
1464 dnode_slots_rele(dnc, idx, slots);
1465 dbuf_rele(db, FTAG);
1466 return (SET_ERROR(ENOENT));
1469 /* Don't actually hold if dry run, just return 0 */
1470 if (flag & DNODE_DRY_RUN) {
1471 mutex_exit(&dn->dn_mtx);
1472 dnode_slots_rele(dnc, idx, slots);
1473 dbuf_rele(db, FTAG);
1477 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1478 } else if (flag & DNODE_MUST_BE_FREE) {
1480 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1481 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1482 dbuf_rele(db, FTAG);
1483 return (SET_ERROR(ENOSPC));
1486 dnode_slots_hold(dnc, idx, slots);
1488 if (!dnode_check_slots_free(dnc, idx, slots)) {
1489 DNODE_STAT_BUMP(dnode_hold_free_misses);
1490 dnode_slots_rele(dnc, idx, slots);
1491 dbuf_rele(db, FTAG);
1492 return (SET_ERROR(ENOSPC));
1495 dnode_slots_rele(dnc, idx, slots);
1496 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1497 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1501 if (!dnode_check_slots_free(dnc, idx, slots)) {
1502 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1503 dnode_slots_rele(dnc, idx, slots);
1504 dbuf_rele(db, FTAG);
1505 return (SET_ERROR(ENOSPC));
1509 * Allocated but otherwise free dnodes which would
1510 * be in the interior of a multi-slot dnodes need
1511 * to be freed. Single slot dnodes can be safely
1512 * re-purposed as a performance optimization.
1515 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1517 dnh = &dnc->dnc_children[idx];
1518 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1519 dn = dnh->dnh_dnode;
1521 dn = dnode_create(os, dn_block + idx, db,
1525 mutex_enter(&dn->dn_mtx);
1526 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1527 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1528 mutex_exit(&dn->dn_mtx);
1529 dnode_slots_rele(dnc, idx, slots);
1530 dbuf_rele(db, FTAG);
1531 return (SET_ERROR(EEXIST));
1534 /* Don't actually hold if dry run, just return 0 */
1535 if (flag & DNODE_DRY_RUN) {
1536 mutex_exit(&dn->dn_mtx);
1537 dnode_slots_rele(dnc, idx, slots);
1538 dbuf_rele(db, FTAG);
1542 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1543 DNODE_STAT_BUMP(dnode_hold_free_hits);
1545 dbuf_rele(db, FTAG);
1546 return (SET_ERROR(EINVAL));
1549 ASSERT0(dn->dn_free_txg);
1551 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1552 dbuf_add_ref(db, dnh);
1554 mutex_exit(&dn->dn_mtx);
1556 /* Now we can rely on the hold to prevent the dnode from moving. */
1557 dnode_slots_rele(dnc, idx, slots);
1560 ASSERT3P(dnp, !=, NULL);
1561 ASSERT3P(dn->dn_dbuf, ==, db);
1562 ASSERT3U(dn->dn_object, ==, object);
1563 dbuf_rele(db, FTAG);
1570 * Return held dnode if the object is allocated, NULL if not.
1573 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1575 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1580 * Can only add a reference if there is already at least one
1581 * reference on the dnode. Returns FALSE if unable to add a
1585 dnode_add_ref(dnode_t *dn, void *tag)
1587 mutex_enter(&dn->dn_mtx);
1588 if (zfs_refcount_is_zero(&dn->dn_holds)) {
1589 mutex_exit(&dn->dn_mtx);
1592 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1593 mutex_exit(&dn->dn_mtx);
1598 dnode_rele(dnode_t *dn, void *tag)
1600 mutex_enter(&dn->dn_mtx);
1601 dnode_rele_and_unlock(dn, tag, B_FALSE);
1605 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1608 /* Get while the hold prevents the dnode from moving. */
1609 dmu_buf_impl_t *db = dn->dn_dbuf;
1610 dnode_handle_t *dnh = dn->dn_handle;
1612 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1614 cv_broadcast(&dn->dn_nodnholds);
1615 mutex_exit(&dn->dn_mtx);
1616 /* dnode could get destroyed at this point, so don't use it anymore */
1619 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1620 * indirectly by dbuf_rele() while relying on the dnode handle to
1621 * prevent the dnode from moving, since releasing the last hold could
1622 * result in the dnode's parent dbuf evicting its dnode handles. For
1623 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1624 * other direct or indirect hold on the dnode must first drop the dnode
1627 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1629 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1630 if (refs == 0 && db != NULL) {
1632 * Another thread could add a hold to the dnode handle in
1633 * dnode_hold_impl() while holding the parent dbuf. Since the
1634 * hold on the parent dbuf prevents the handle from being
1635 * destroyed, the hold on the handle is OK. We can't yet assert
1636 * that the handle has zero references, but that will be
1637 * asserted anyway when the handle gets destroyed.
1639 mutex_enter(&db->db_mtx);
1640 dbuf_rele_and_unlock(db, dnh, evicting);
1645 * Test whether we can create a dnode at the specified location.
1648 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1650 return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1651 slots, NULL, NULL));
1655 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1657 objset_t *os = dn->dn_objset;
1658 uint64_t txg = tx->tx_txg;
1660 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1661 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1668 mutex_enter(&dn->dn_mtx);
1669 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1670 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1671 mutex_exit(&dn->dn_mtx);
1675 * Determine old uid/gid when necessary
1677 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1679 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1680 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1683 * If we are already marked dirty, we're done.
1685 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1686 multilist_sublist_unlock(mls);
1690 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1691 !avl_is_empty(&dn->dn_dbufs));
1692 ASSERT(dn->dn_datablksz != 0);
1693 ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1694 ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1695 ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1697 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1698 dn->dn_object, txg);
1700 multilist_sublist_insert_head(mls, dn);
1702 multilist_sublist_unlock(mls);
1705 * The dnode maintains a hold on its containing dbuf as
1706 * long as there are holds on it. Each instantiated child
1707 * dbuf maintains a hold on the dnode. When the last child
1708 * drops its hold, the dnode will drop its hold on the
1709 * containing dbuf. We add a "dirty hold" here so that the
1710 * dnode will hang around after we finish processing its
1713 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1715 (void) dbuf_dirty(dn->dn_dbuf, tx);
1717 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1721 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1723 mutex_enter(&dn->dn_mtx);
1724 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1725 mutex_exit(&dn->dn_mtx);
1728 dn->dn_free_txg = tx->tx_txg;
1729 mutex_exit(&dn->dn_mtx);
1731 dnode_setdirty(dn, tx);
1735 * Try to change the block size for the indicated dnode. This can only
1736 * succeed if there are no blocks allocated or dirty beyond first block
1739 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1744 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1746 size = SPA_MINBLOCKSIZE;
1748 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1750 if (ibs == dn->dn_indblkshift)
1753 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1756 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1758 /* Check for any allocated blocks beyond the first */
1759 if (dn->dn_maxblkid != 0)
1762 mutex_enter(&dn->dn_dbufs_mtx);
1763 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1764 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1765 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1766 db->db_blkid != DMU_SPILL_BLKID) {
1767 mutex_exit(&dn->dn_dbufs_mtx);
1771 mutex_exit(&dn->dn_dbufs_mtx);
1773 if (ibs && dn->dn_nlevels != 1)
1776 /* resize the old block */
1777 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1779 dbuf_new_size(db, size, tx);
1780 } else if (err != ENOENT) {
1784 dnode_setdblksz(dn, size);
1785 dnode_setdirty(dn, tx);
1786 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1788 dn->dn_indblkshift = ibs;
1789 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1791 /* release after we have fixed the blocksize in the dnode */
1793 dbuf_rele(db, FTAG);
1795 rw_exit(&dn->dn_struct_rwlock);
1799 rw_exit(&dn->dn_struct_rwlock);
1800 return (SET_ERROR(ENOTSUP));
1804 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1806 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1807 int old_nlevels = dn->dn_nlevels;
1810 dbuf_dirty_record_t *new, *dr, *dr_next;
1812 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1814 dn->dn_nlevels = new_nlevels;
1816 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1817 dn->dn_next_nlevels[txgoff] = new_nlevels;
1819 /* dirty the left indirects */
1820 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1822 new = dbuf_dirty(db, tx);
1823 dbuf_rele(db, FTAG);
1825 /* transfer the dirty records to the new indirect */
1826 mutex_enter(&dn->dn_mtx);
1827 mutex_enter(&new->dt.di.dr_mtx);
1828 list = &dn->dn_dirty_records[txgoff];
1829 for (dr = list_head(list); dr; dr = dr_next) {
1830 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1831 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1832 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1833 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1834 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1835 list_remove(&dn->dn_dirty_records[txgoff], dr);
1836 list_insert_tail(&new->dt.di.dr_children, dr);
1837 dr->dr_parent = new;
1840 mutex_exit(&new->dt.di.dr_mtx);
1841 mutex_exit(&dn->dn_mtx);
1845 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1849 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1851 if (dn->dn_nlevels == nlevels) {
1854 } else if (nlevels < dn->dn_nlevels) {
1855 ret = SET_ERROR(EINVAL);
1859 dnode_set_nlevels_impl(dn, nlevels, tx);
1862 rw_exit(&dn->dn_struct_rwlock);
1866 /* read-holding callers must not rely on the lock being continuously held */
1868 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1871 int epbs, new_nlevels;
1874 ASSERT(blkid != DMU_BONUS_BLKID);
1877 RW_READ_HELD(&dn->dn_struct_rwlock) :
1878 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1881 * if we have a read-lock, check to see if we need to do any work
1882 * before upgrading to a write-lock.
1885 if (blkid <= dn->dn_maxblkid)
1888 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1889 rw_exit(&dn->dn_struct_rwlock);
1890 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1895 * Raw sends (indicated by the force flag) require that we take the
1896 * given blkid even if the value is lower than the current value.
1898 if (!force && blkid <= dn->dn_maxblkid)
1902 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1903 * to indicate that this field is set. This allows us to set the
1904 * maxblkid to 0 on an existing object in dnode_sync().
1906 dn->dn_maxblkid = blkid;
1907 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1908 blkid | DMU_NEXT_MAXBLKID_SET;
1911 * Compute the number of levels necessary to support the new maxblkid.
1912 * Raw sends will ensure nlevels is set correctly for us.
1915 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1916 for (sz = dn->dn_nblkptr;
1917 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1920 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1923 if (new_nlevels > dn->dn_nlevels)
1924 dnode_set_nlevels_impl(dn, new_nlevels, tx);
1926 ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1931 rw_downgrade(&dn->dn_struct_rwlock);
1935 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1937 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1939 dmu_buf_will_dirty(&db->db, tx);
1940 dbuf_rele(db, FTAG);
1945 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1949 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1952 dmu_buf_impl_t *db_search;
1956 db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1958 mutex_enter(&dn->dn_dbufs_mtx);
1960 db_search->db_level = 1;
1961 db_search->db_blkid = start_blkid + 1;
1962 db_search->db_state = DB_SEARCH;
1965 db = avl_find(&dn->dn_dbufs, db_search, &where);
1967 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1969 if (db == NULL || db->db_level != 1 ||
1970 db->db_blkid >= end_blkid) {
1975 * Setup the next blkid we want to search for.
1977 db_search->db_blkid = db->db_blkid + 1;
1978 ASSERT3U(db->db_blkid, >=, start_blkid);
1981 * If the dbuf transitions to DB_EVICTING while we're trying
1982 * to dirty it, then we will be unable to discover it in
1983 * the dbuf hash table. This will result in a call to
1984 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1985 * lock. To avoid a deadlock, we drop the lock before
1986 * dirtying the level-1 dbuf.
1988 mutex_exit(&dn->dn_dbufs_mtx);
1989 dnode_dirty_l1(dn, db->db_blkid, tx);
1990 mutex_enter(&dn->dn_dbufs_mtx);
1995 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1997 db_search->db_level = 1;
1998 db_search->db_blkid = start_blkid + 1;
1999 db_search->db_state = DB_SEARCH;
2000 db = avl_find(&dn->dn_dbufs, db_search, &where);
2002 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2003 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2004 if (db->db_level != 1 || db->db_blkid >= end_blkid)
2006 if (db->db_state != DB_EVICTING)
2007 ASSERT(db->db_dirtycnt > 0);
2010 kmem_free(db_search, sizeof (dmu_buf_impl_t));
2011 mutex_exit(&dn->dn_dbufs_mtx);
2015 dnode_set_dirtyctx(dnode_t *dn, dmu_tx_t *tx, void *tag)
2018 * Don't set dirtyctx to SYNC if we're just modifying this as we
2019 * initialize the objset.
2021 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
2022 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2025 rrw_enter(&ds->ds_bp_rwlock, RW_READER, tag);
2027 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
2028 if (dmu_tx_is_syncing(tx))
2029 dn->dn_dirtyctx = DN_DIRTY_SYNC;
2031 dn->dn_dirtyctx = DN_DIRTY_OPEN;
2032 dn->dn_dirtyctx_firstset = tag;
2035 rrw_exit(&ds->ds_bp_rwlock, tag);
2041 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2044 uint64_t blkoff, blkid, nblks;
2045 int blksz, blkshift, head, tail;
2049 blksz = dn->dn_datablksz;
2050 blkshift = dn->dn_datablkshift;
2051 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2053 if (len == DMU_OBJECT_END) {
2054 len = UINT64_MAX - off;
2059 * First, block align the region to free:
2062 head = P2NPHASE(off, blksz);
2063 blkoff = P2PHASE(off, blksz);
2064 if ((off >> blkshift) > dn->dn_maxblkid)
2067 ASSERT(dn->dn_maxblkid == 0);
2068 if (off == 0 && len >= blksz) {
2070 * Freeing the whole block; fast-track this request.
2074 if (dn->dn_nlevels > 1) {
2075 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2076 dnode_dirty_l1(dn, 0, tx);
2077 rw_exit(&dn->dn_struct_rwlock);
2080 } else if (off >= blksz) {
2081 /* Freeing past end-of-data */
2084 /* Freeing part of the block. */
2086 ASSERT3U(head, >, 0);
2090 /* zero out any partial block data at the start of the range */
2093 ASSERT3U(blkoff + head, ==, blksz);
2096 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2097 res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2098 TRUE, FALSE, FTAG, &db);
2099 rw_exit(&dn->dn_struct_rwlock);
2104 db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER,
2106 /* don't dirty if it isn't on disk and isn't dirty */
2107 dirty = !list_is_empty(&db->db_dirty_records) ||
2108 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2109 dmu_buf_unlock_parent(db, dblt, FTAG);
2111 dmu_buf_will_dirty(&db->db, tx);
2112 data = db->db.db_data;
2113 bzero(data + blkoff, head);
2115 dbuf_rele(db, FTAG);
2121 /* If the range was less than one block, we're done */
2125 /* If the remaining range is past end of file, we're done */
2126 if ((off >> blkshift) > dn->dn_maxblkid)
2129 ASSERT(ISP2(blksz));
2133 tail = P2PHASE(len, blksz);
2135 ASSERT0(P2PHASE(off, blksz));
2136 /* zero out any partial block data at the end of the range */
2141 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2142 res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2143 TRUE, FALSE, FTAG, &db);
2144 rw_exit(&dn->dn_struct_rwlock);
2147 /* don't dirty if not on disk and not dirty */
2148 db_lock_type_t type = dmu_buf_lock_parent(db, RW_READER,
2150 dirty = !list_is_empty(&db->db_dirty_records) ||
2151 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2152 dmu_buf_unlock_parent(db, type, FTAG);
2154 dmu_buf_will_dirty(&db->db, tx);
2155 bzero(db->db.db_data, tail);
2157 dbuf_rele(db, FTAG);
2162 /* If the range did not include a full block, we are done */
2166 ASSERT(IS_P2ALIGNED(off, blksz));
2167 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2168 blkid = off >> blkshift;
2169 nblks = len >> blkshift;
2174 * Dirty all the indirect blocks in this range. Note that only
2175 * the first and last indirect blocks can actually be written
2176 * (if they were partially freed) -- they must be dirtied, even if
2177 * they do not exist on disk yet. The interior blocks will
2178 * be freed by free_children(), so they will not actually be written.
2179 * Even though these interior blocks will not be written, we
2180 * dirty them for two reasons:
2182 * - It ensures that the indirect blocks remain in memory until
2183 * syncing context. (They have already been prefetched by
2184 * dmu_tx_hold_free(), so we don't have to worry about reading
2185 * them serially here.)
2187 * - The dirty space accounting will put pressure on the txg sync
2188 * mechanism to begin syncing, and to delay transactions if there
2189 * is a large amount of freeing. Even though these indirect
2190 * blocks will not be written, we could need to write the same
2191 * amount of space if we copy the freed BPs into deadlists.
2193 if (dn->dn_nlevels > 1) {
2194 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2195 uint64_t first, last;
2197 first = blkid >> epbs;
2198 dnode_dirty_l1(dn, first, tx);
2200 last = dn->dn_maxblkid >> epbs;
2202 last = (blkid + nblks - 1) >> epbs;
2204 dnode_dirty_l1(dn, last, tx);
2206 dnode_dirty_l1range(dn, first, last, tx);
2208 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2210 for (uint64_t i = first + 1; i < last; i++) {
2212 * Set i to the blockid of the next non-hole
2213 * level-1 indirect block at or after i. Note
2214 * that dnode_next_offset() operates in terms of
2215 * level-0-equivalent bytes.
2217 uint64_t ibyte = i << shift;
2218 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2225 * Normally we should not see an error, either
2226 * from dnode_next_offset() or dbuf_hold_level()
2227 * (except for ESRCH from dnode_next_offset).
2228 * If there is an i/o error, then when we read
2229 * this block in syncing context, it will use
2230 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2231 * to the "failmode" property. dnode_next_offset()
2232 * doesn't have a flag to indicate MUSTSUCCEED.
2237 dnode_dirty_l1(dn, i, tx);
2239 rw_exit(&dn->dn_struct_rwlock);
2244 * Add this range to the dnode range list.
2245 * We will finish up this free operation in the syncing phase.
2247 mutex_enter(&dn->dn_mtx);
2249 int txgoff = tx->tx_txg & TXG_MASK;
2250 if (dn->dn_free_ranges[txgoff] == NULL) {
2251 dn->dn_free_ranges[txgoff] = range_tree_create(NULL,
2252 RANGE_SEG64, NULL, 0, 0);
2254 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2255 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2257 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2258 blkid, nblks, tx->tx_txg);
2259 mutex_exit(&dn->dn_mtx);
2261 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2262 dnode_setdirty(dn, tx);
2266 dnode_spill_freed(dnode_t *dn)
2270 mutex_enter(&dn->dn_mtx);
2271 for (i = 0; i < TXG_SIZE; i++) {
2272 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2275 mutex_exit(&dn->dn_mtx);
2276 return (i < TXG_SIZE);
2279 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2281 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2283 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2286 if (blkid == DMU_BONUS_BLKID)
2290 * If we're in the process of opening the pool, dp will not be
2291 * set yet, but there shouldn't be anything dirty.
2296 if (dn->dn_free_txg)
2299 if (blkid == DMU_SPILL_BLKID)
2300 return (dnode_spill_freed(dn));
2302 mutex_enter(&dn->dn_mtx);
2303 for (i = 0; i < TXG_SIZE; i++) {
2304 if (dn->dn_free_ranges[i] != NULL &&
2305 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2308 mutex_exit(&dn->dn_mtx);
2309 return (i < TXG_SIZE);
2312 /* call from syncing context when we actually write/free space for this dnode */
2314 dnode_diduse_space(dnode_t *dn, int64_t delta)
2317 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2319 (u_longlong_t)dn->dn_phys->dn_used,
2322 mutex_enter(&dn->dn_mtx);
2323 space = DN_USED_BYTES(dn->dn_phys);
2325 ASSERT3U(space + delta, >=, space); /* no overflow */
2327 ASSERT3U(space, >=, -delta); /* no underflow */
2330 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2331 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2332 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2333 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2335 dn->dn_phys->dn_used = space;
2336 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2338 mutex_exit(&dn->dn_mtx);
2342 * Scans a block at the indicated "level" looking for a hole or data,
2343 * depending on 'flags'.
2345 * If level > 0, then we are scanning an indirect block looking at its
2346 * pointers. If level == 0, then we are looking at a block of dnodes.
2348 * If we don't find what we are looking for in the block, we return ESRCH.
2349 * Otherwise, return with *offset pointing to the beginning (if searching
2350 * forwards) or end (if searching backwards) of the range covered by the
2351 * block pointer we matched on (or dnode).
2353 * The basic search algorithm used below by dnode_next_offset() is to
2354 * use this function to search up the block tree (widen the search) until
2355 * we find something (i.e., we don't return ESRCH) and then search back
2356 * down the tree (narrow the search) until we reach our original search
2360 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2361 int lvl, uint64_t blkfill, uint64_t txg)
2363 dmu_buf_impl_t *db = NULL;
2365 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2366 uint64_t epb = 1ULL << epbs;
2367 uint64_t minfill, maxfill;
2369 int i, inc, error, span;
2371 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2373 hole = ((flags & DNODE_FIND_HOLE) != 0);
2374 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2375 ASSERT(txg == 0 || !hole);
2377 if (lvl == dn->dn_phys->dn_nlevels) {
2379 epb = dn->dn_phys->dn_nblkptr;
2380 data = dn->dn_phys->dn_blkptr;
2382 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2383 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2385 if (error != ENOENT)
2390 * This can only happen when we are searching up
2391 * the block tree for data. We don't really need to
2392 * adjust the offset, as we will just end up looking
2393 * at the pointer to this block in its parent, and its
2394 * going to be unallocated, so we will skip over it.
2396 return (SET_ERROR(ESRCH));
2398 error = dbuf_read(db, NULL,
2399 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2401 dbuf_rele(db, FTAG);
2404 data = db->db.db_data;
2405 rw_enter(&db->db_rwlock, RW_READER);
2408 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2409 db->db_blkptr->blk_birth <= txg ||
2410 BP_IS_HOLE(db->db_blkptr))) {
2412 * This can only happen when we are searching up the tree
2413 * and these conditions mean that we need to keep climbing.
2415 error = SET_ERROR(ESRCH);
2416 } else if (lvl == 0) {
2417 dnode_phys_t *dnp = data;
2419 ASSERT(dn->dn_type == DMU_OT_DNODE);
2420 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2422 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2423 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2424 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2429 error = SET_ERROR(ESRCH);
2431 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2434 blkptr_t *bp = data;
2435 uint64_t start = *offset;
2436 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2438 maxfill = blkfill << ((lvl - 1) * epbs);
2445 if (span >= 8 * sizeof (*offset)) {
2446 /* This only happens on the highest indirection level */
2447 ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2450 *offset = *offset >> span;
2453 for (i = BF64_GET(*offset, 0, epbs);
2454 i >= 0 && i < epb; i += inc) {
2455 if (BP_GET_FILL(&bp[i]) >= minfill &&
2456 BP_GET_FILL(&bp[i]) <= maxfill &&
2457 (hole || bp[i].blk_birth > txg))
2459 if (inc > 0 || *offset > 0)
2463 if (span >= 8 * sizeof (*offset)) {
2466 *offset = *offset << span;
2470 /* traversing backwards; position offset at the end */
2471 ASSERT3U(*offset, <=, start);
2472 *offset = MIN(*offset + (1ULL << span) - 1, start);
2473 } else if (*offset < start) {
2476 if (i < 0 || i >= epb)
2477 error = SET_ERROR(ESRCH);
2481 rw_exit(&db->db_rwlock);
2482 dbuf_rele(db, FTAG);
2489 * Find the next hole, data, or sparse region at or after *offset.
2490 * The value 'blkfill' tells us how many items we expect to find
2491 * in an L0 data block; this value is 1 for normal objects,
2492 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2493 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2497 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2498 * Finds the next/previous hole/data in a file.
2499 * Used in dmu_offset_next().
2501 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2502 * Finds the next free/allocated dnode an objset's meta-dnode.
2503 * Only finds objects that have new contents since txg (ie.
2504 * bonus buffer changes and content removal are ignored).
2505 * Used in dmu_object_next().
2507 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2508 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2509 * Used in dmu_object_alloc().
2512 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2513 int minlvl, uint64_t blkfill, uint64_t txg)
2515 uint64_t initial_offset = *offset;
2519 if (!(flags & DNODE_FIND_HAVELOCK))
2520 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2522 if (dn->dn_phys->dn_nlevels == 0) {
2523 error = SET_ERROR(ESRCH);
2527 if (dn->dn_datablkshift == 0) {
2528 if (*offset < dn->dn_datablksz) {
2529 if (flags & DNODE_FIND_HOLE)
2530 *offset = dn->dn_datablksz;
2532 error = SET_ERROR(ESRCH);
2537 maxlvl = dn->dn_phys->dn_nlevels;
2539 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2540 error = dnode_next_offset_level(dn,
2541 flags, offset, lvl, blkfill, txg);
2546 while (error == 0 && --lvl >= minlvl) {
2547 error = dnode_next_offset_level(dn,
2548 flags, offset, lvl, blkfill, txg);
2552 * There's always a "virtual hole" at the end of the object, even
2553 * if all BP's which physically exist are non-holes.
2555 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2556 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2560 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2561 initial_offset < *offset : initial_offset > *offset))
2562 error = SET_ERROR(ESRCH);
2564 if (!(flags & DNODE_FIND_HAVELOCK))
2565 rw_exit(&dn->dn_struct_rwlock);
2570 #if defined(_KERNEL)
2571 EXPORT_SYMBOL(dnode_hold);
2572 EXPORT_SYMBOL(dnode_rele);
2573 EXPORT_SYMBOL(dnode_set_nlevels);
2574 EXPORT_SYMBOL(dnode_set_blksz);
2575 EXPORT_SYMBOL(dnode_free_range);
2576 EXPORT_SYMBOL(dnode_evict_dbufs);
2577 EXPORT_SYMBOL(dnode_evict_bonus);