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.
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_dnode.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_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 ASSERTV(static dnode_phys_t dnode_phys_zero);
80 int zfs_default_bs = SPA_MINBLOCKSHIFT;
81 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
84 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
88 dbuf_compare(const void *x1, const void *x2)
90 const dmu_buf_impl_t *d1 = x1;
91 const dmu_buf_impl_t *d2 = x2;
93 int cmp = AVL_CMP(d1->db_level, d2->db_level);
97 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
101 if (d1->db_state == DB_SEARCH) {
102 ASSERT3S(d2->db_state, !=, DB_SEARCH);
104 } else if (d2->db_state == DB_SEARCH) {
105 ASSERT3S(d1->db_state, !=, DB_SEARCH);
109 return (AVL_PCMP(d1, d2));
114 dnode_cons(void *arg, void *unused, int kmflag)
119 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
120 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
121 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
122 cv_init(&dn->dn_notxholds, 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 refcount_create_untracked(&dn->dn_holds);
129 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 refcount_destroy(&dn->dn_holds);
188 refcount_destroy(&dn->dn_tx_holds);
189 ASSERT(!list_link_active(&dn->dn_link));
191 for (i = 0; i < TXG_SIZE; i++) {
192 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
193 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
194 list_destroy(&dn->dn_dirty_records[i]);
195 ASSERT0(dn->dn_next_nblkptr[i]);
196 ASSERT0(dn->dn_next_nlevels[i]);
197 ASSERT0(dn->dn_next_indblkshift[i]);
198 ASSERT0(dn->dn_next_bonustype[i]);
199 ASSERT0(dn->dn_rm_spillblk[i]);
200 ASSERT0(dn->dn_next_bonuslen[i]);
201 ASSERT0(dn->dn_next_blksz[i]);
202 ASSERT0(dn->dn_next_maxblkid[i]);
205 ASSERT0(dn->dn_allocated_txg);
206 ASSERT0(dn->dn_free_txg);
207 ASSERT0(dn->dn_assigned_txg);
208 ASSERT0(dn->dn_dirty_txg);
209 ASSERT0(dn->dn_dirtyctx);
210 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
211 ASSERT3P(dn->dn_bonus, ==, NULL);
212 ASSERT(!dn->dn_have_spill);
213 ASSERT3P(dn->dn_zio, ==, NULL);
214 ASSERT0(dn->dn_oldused);
215 ASSERT0(dn->dn_oldflags);
216 ASSERT0(dn->dn_olduid);
217 ASSERT0(dn->dn_oldgid);
218 ASSERT0(dn->dn_oldprojid);
219 ASSERT0(dn->dn_newuid);
220 ASSERT0(dn->dn_newgid);
221 ASSERT0(dn->dn_newprojid);
222 ASSERT0(dn->dn_id_flags);
224 ASSERT0(dn->dn_dbufs_count);
225 avl_destroy(&dn->dn_dbufs);
231 ASSERT(dnode_cache == NULL);
232 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
233 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
234 kmem_cache_set_move(dnode_cache, dnode_move);
236 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
237 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
239 if (dnode_ksp != NULL) {
240 dnode_ksp->ks_data = &dnode_stats;
241 kstat_install(dnode_ksp);
248 if (dnode_ksp != NULL) {
249 kstat_delete(dnode_ksp);
253 kmem_cache_destroy(dnode_cache);
260 dnode_verify(dnode_t *dn)
262 int drop_struct_lock = FALSE;
265 ASSERT(dn->dn_objset);
266 ASSERT(dn->dn_handle->dnh_dnode == dn);
268 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
270 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
273 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
274 rw_enter(&dn->dn_struct_rwlock, RW_READER);
275 drop_struct_lock = TRUE;
277 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
279 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
280 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
281 if (dn->dn_datablkshift) {
282 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
283 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
284 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
286 ASSERT3U(dn->dn_nlevels, <=, 30);
287 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
288 ASSERT3U(dn->dn_nblkptr, >=, 1);
289 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
290 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
291 ASSERT3U(dn->dn_datablksz, ==,
292 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
293 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
294 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
295 dn->dn_bonuslen, <=, max_bonuslen);
296 for (i = 0; i < TXG_SIZE; i++) {
297 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
300 if (dn->dn_phys->dn_type != DMU_OT_NONE)
301 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
302 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
303 if (dn->dn_dbuf != NULL) {
304 ASSERT3P(dn->dn_phys, ==,
305 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
306 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
308 if (drop_struct_lock)
309 rw_exit(&dn->dn_struct_rwlock);
314 dnode_byteswap(dnode_phys_t *dnp)
316 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
319 if (dnp->dn_type == DMU_OT_NONE) {
320 bzero(dnp, sizeof (dnode_phys_t));
324 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
325 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
326 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
327 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
328 dnp->dn_used = BSWAP_64(dnp->dn_used);
331 * dn_nblkptr is only one byte, so it's OK to read it in either
332 * byte order. We can't read dn_bouslen.
334 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
335 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
336 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
337 buf64[i] = BSWAP_64(buf64[i]);
340 * OK to check dn_bonuslen for zero, because it won't matter if
341 * we have the wrong byte order. This is necessary because the
342 * dnode dnode is smaller than a regular dnode.
344 if (dnp->dn_bonuslen != 0) {
346 * Note that the bonus length calculated here may be
347 * longer than the actual bonus buffer. This is because
348 * we always put the bonus buffer after the last block
349 * pointer (instead of packing it against the end of the
352 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
353 int slots = dnp->dn_extra_slots + 1;
354 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
355 dmu_object_byteswap_t byteswap;
356 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
357 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
358 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
361 /* Swap SPILL block if we have one */
362 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
363 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
367 dnode_buf_byteswap(void *vbuf, size_t size)
371 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
372 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
375 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
379 if (dnp->dn_type != DMU_OT_NONE)
380 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
385 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
387 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
389 dnode_setdirty(dn, tx);
390 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
391 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
392 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
393 dn->dn_bonuslen = newsize;
395 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
397 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
398 rw_exit(&dn->dn_struct_rwlock);
402 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
404 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
405 dnode_setdirty(dn, tx);
406 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
407 dn->dn_bonustype = newtype;
408 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
409 rw_exit(&dn->dn_struct_rwlock);
413 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
415 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
416 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
417 dnode_setdirty(dn, tx);
418 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
419 dn->dn_have_spill = B_FALSE;
423 dnode_setdblksz(dnode_t *dn, int size)
425 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
426 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
427 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
428 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
429 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
430 dn->dn_datablksz = size;
431 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
432 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
436 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
437 uint64_t object, dnode_handle_t *dnh)
441 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
442 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
446 * Defer setting dn_objset until the dnode is ready to be a candidate
447 * for the dnode_move() callback.
449 dn->dn_object = object;
454 if (dnp->dn_datablkszsec) {
455 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
457 dn->dn_datablksz = 0;
458 dn->dn_datablkszsec = 0;
459 dn->dn_datablkshift = 0;
461 dn->dn_indblkshift = dnp->dn_indblkshift;
462 dn->dn_nlevels = dnp->dn_nlevels;
463 dn->dn_type = dnp->dn_type;
464 dn->dn_nblkptr = dnp->dn_nblkptr;
465 dn->dn_checksum = dnp->dn_checksum;
466 dn->dn_compress = dnp->dn_compress;
467 dn->dn_bonustype = dnp->dn_bonustype;
468 dn->dn_bonuslen = dnp->dn_bonuslen;
469 dn->dn_num_slots = dnp->dn_extra_slots + 1;
470 dn->dn_maxblkid = dnp->dn_maxblkid;
471 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
474 dmu_zfetch_init(&dn->dn_zfetch, dn);
476 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
477 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
478 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
480 mutex_enter(&os->os_lock);
483 * Exclude special dnodes from os_dnodes so an empty os_dnodes
484 * signifies that the special dnodes have no references from
485 * their children (the entries in os_dnodes). This allows
486 * dnode_destroy() to easily determine if the last child has
487 * been removed and then complete eviction of the objset.
489 if (!DMU_OBJECT_IS_SPECIAL(object))
490 list_insert_head(&os->os_dnodes, dn);
494 * Everything else must be valid before assigning dn_objset
495 * makes the dnode eligible for dnode_move().
500 mutex_exit(&os->os_lock);
502 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
508 * Caller must be holding the dnode handle, which is released upon return.
511 dnode_destroy(dnode_t *dn)
513 objset_t *os = dn->dn_objset;
514 boolean_t complete_os_eviction = B_FALSE;
516 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
518 mutex_enter(&os->os_lock);
519 POINTER_INVALIDATE(&dn->dn_objset);
520 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
521 list_remove(&os->os_dnodes, dn);
522 complete_os_eviction =
523 list_is_empty(&os->os_dnodes) &&
524 list_link_active(&os->os_evicting_node);
526 mutex_exit(&os->os_lock);
528 /* the dnode can no longer move, so we can release the handle */
529 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
530 zrl_remove(&dn->dn_handle->dnh_zrlock);
532 dn->dn_allocated_txg = 0;
534 dn->dn_assigned_txg = 0;
535 dn->dn_dirty_txg = 0;
538 if (dn->dn_dirtyctx_firstset != NULL) {
539 kmem_free(dn->dn_dirtyctx_firstset, 1);
540 dn->dn_dirtyctx_firstset = NULL;
542 if (dn->dn_bonus != NULL) {
543 mutex_enter(&dn->dn_bonus->db_mtx);
544 dbuf_destroy(dn->dn_bonus);
549 dn->dn_have_spill = B_FALSE;
554 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
557 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
560 dmu_zfetch_fini(&dn->dn_zfetch);
561 kmem_cache_free(dnode_cache, dn);
562 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
564 if (complete_os_eviction)
565 dmu_objset_evict_done(os);
569 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
570 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
574 ASSERT3U(dn_slots, >, 0);
575 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
576 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
577 ASSERT3U(blocksize, <=,
578 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
580 blocksize = 1 << zfs_default_bs;
582 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
585 ibs = zfs_default_ibs;
587 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
589 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
590 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
591 DNODE_STAT_BUMP(dnode_allocate);
593 ASSERT(dn->dn_type == DMU_OT_NONE);
594 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
595 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
596 ASSERT(ot != DMU_OT_NONE);
597 ASSERT(DMU_OT_IS_VALID(ot));
598 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
599 (bonustype == DMU_OT_SA && bonuslen == 0) ||
600 (bonustype != DMU_OT_NONE && bonuslen != 0));
601 ASSERT(DMU_OT_IS_VALID(bonustype));
602 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
603 ASSERT(dn->dn_type == DMU_OT_NONE);
604 ASSERT0(dn->dn_maxblkid);
605 ASSERT0(dn->dn_allocated_txg);
606 ASSERT0(dn->dn_assigned_txg);
607 ASSERT0(dn->dn_dirty_txg);
608 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
609 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
610 ASSERT(avl_is_empty(&dn->dn_dbufs));
612 for (i = 0; i < TXG_SIZE; i++) {
613 ASSERT0(dn->dn_next_nblkptr[i]);
614 ASSERT0(dn->dn_next_nlevels[i]);
615 ASSERT0(dn->dn_next_indblkshift[i]);
616 ASSERT0(dn->dn_next_bonuslen[i]);
617 ASSERT0(dn->dn_next_bonustype[i]);
618 ASSERT0(dn->dn_rm_spillblk[i]);
619 ASSERT0(dn->dn_next_blksz[i]);
620 ASSERT0(dn->dn_next_maxblkid[i]);
621 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
622 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
623 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
627 dnode_setdblksz(dn, blocksize);
628 dn->dn_indblkshift = ibs;
630 dn->dn_num_slots = dn_slots;
631 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
634 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
635 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
639 dn->dn_bonustype = bonustype;
640 dn->dn_bonuslen = bonuslen;
641 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
642 dn->dn_compress = ZIO_COMPRESS_INHERIT;
646 if (dn->dn_dirtyctx_firstset) {
647 kmem_free(dn->dn_dirtyctx_firstset, 1);
648 dn->dn_dirtyctx_firstset = NULL;
651 dn->dn_allocated_txg = tx->tx_txg;
654 dnode_setdirty(dn, tx);
655 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
656 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
657 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
658 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
662 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
663 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
667 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
668 ASSERT3U(blocksize, <=,
669 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
670 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
671 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
672 ASSERT(tx->tx_txg != 0);
673 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
674 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
675 (bonustype == DMU_OT_SA && bonuslen == 0));
676 ASSERT(DMU_OT_IS_VALID(bonustype));
677 ASSERT3U(bonuslen, <=,
678 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
679 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
681 dnode_free_interior_slots(dn);
682 DNODE_STAT_BUMP(dnode_reallocate);
684 /* clean up any unreferenced dbufs */
685 dnode_evict_dbufs(dn);
689 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
690 dnode_setdirty(dn, tx);
691 if (dn->dn_datablksz != blocksize) {
692 /* change blocksize */
693 ASSERT(dn->dn_maxblkid == 0 &&
694 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
695 dnode_block_freed(dn, 0)));
696 dnode_setdblksz(dn, blocksize);
697 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
699 if (dn->dn_bonuslen != bonuslen)
700 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
702 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
705 nblkptr = MIN(DN_MAX_NBLKPTR,
706 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
708 if (dn->dn_bonustype != bonustype)
709 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
710 if (dn->dn_nblkptr != nblkptr)
711 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
712 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
713 dbuf_rm_spill(dn, tx);
714 dnode_rm_spill(dn, tx);
716 rw_exit(&dn->dn_struct_rwlock);
721 /* change bonus size and type */
722 mutex_enter(&dn->dn_mtx);
723 dn->dn_bonustype = bonustype;
724 dn->dn_bonuslen = bonuslen;
725 dn->dn_num_slots = dn_slots;
726 dn->dn_nblkptr = nblkptr;
727 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
728 dn->dn_compress = ZIO_COMPRESS_INHERIT;
729 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
731 /* fix up the bonus db_size */
733 dn->dn_bonus->db.db_size =
734 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
735 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
736 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
739 dn->dn_allocated_txg = tx->tx_txg;
740 mutex_exit(&dn->dn_mtx);
745 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
749 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
750 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
751 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
752 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
755 ndn->dn_objset = odn->dn_objset;
756 ndn->dn_object = odn->dn_object;
757 ndn->dn_dbuf = odn->dn_dbuf;
758 ndn->dn_handle = odn->dn_handle;
759 ndn->dn_phys = odn->dn_phys;
760 ndn->dn_type = odn->dn_type;
761 ndn->dn_bonuslen = odn->dn_bonuslen;
762 ndn->dn_bonustype = odn->dn_bonustype;
763 ndn->dn_nblkptr = odn->dn_nblkptr;
764 ndn->dn_checksum = odn->dn_checksum;
765 ndn->dn_compress = odn->dn_compress;
766 ndn->dn_nlevels = odn->dn_nlevels;
767 ndn->dn_indblkshift = odn->dn_indblkshift;
768 ndn->dn_datablkshift = odn->dn_datablkshift;
769 ndn->dn_datablkszsec = odn->dn_datablkszsec;
770 ndn->dn_datablksz = odn->dn_datablksz;
771 ndn->dn_maxblkid = odn->dn_maxblkid;
772 ndn->dn_num_slots = odn->dn_num_slots;
773 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
774 sizeof (odn->dn_next_nblkptr));
775 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
776 sizeof (odn->dn_next_nlevels));
777 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
778 sizeof (odn->dn_next_indblkshift));
779 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
780 sizeof (odn->dn_next_bonustype));
781 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
782 sizeof (odn->dn_rm_spillblk));
783 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
784 sizeof (odn->dn_next_bonuslen));
785 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
786 sizeof (odn->dn_next_blksz));
787 bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
788 sizeof (odn->dn_next_maxblkid));
789 for (i = 0; i < TXG_SIZE; i++) {
790 list_move_tail(&ndn->dn_dirty_records[i],
791 &odn->dn_dirty_records[i]);
793 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
794 sizeof (odn->dn_free_ranges));
795 ndn->dn_allocated_txg = odn->dn_allocated_txg;
796 ndn->dn_free_txg = odn->dn_free_txg;
797 ndn->dn_assigned_txg = odn->dn_assigned_txg;
798 ndn->dn_dirty_txg = odn->dn_dirty_txg;
799 ndn->dn_dirtyctx = odn->dn_dirtyctx;
800 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
801 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
802 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
803 ASSERT(avl_is_empty(&ndn->dn_dbufs));
804 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
805 ndn->dn_dbufs_count = odn->dn_dbufs_count;
806 ndn->dn_bonus = odn->dn_bonus;
807 ndn->dn_have_spill = odn->dn_have_spill;
808 ndn->dn_zio = odn->dn_zio;
809 ndn->dn_oldused = odn->dn_oldused;
810 ndn->dn_oldflags = odn->dn_oldflags;
811 ndn->dn_olduid = odn->dn_olduid;
812 ndn->dn_oldgid = odn->dn_oldgid;
813 ndn->dn_oldprojid = odn->dn_oldprojid;
814 ndn->dn_newuid = odn->dn_newuid;
815 ndn->dn_newgid = odn->dn_newgid;
816 ndn->dn_newprojid = odn->dn_newprojid;
817 ndn->dn_id_flags = odn->dn_id_flags;
818 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
819 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
820 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
823 * Update back pointers. Updating the handle fixes the back pointer of
824 * every descendant dbuf as well as the bonus dbuf.
826 ASSERT(ndn->dn_handle->dnh_dnode == odn);
827 ndn->dn_handle->dnh_dnode = ndn;
828 if (ndn->dn_zfetch.zf_dnode == odn) {
829 ndn->dn_zfetch.zf_dnode = ndn;
833 * Invalidate the original dnode by clearing all of its back pointers.
836 odn->dn_handle = NULL;
837 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
838 offsetof(dmu_buf_impl_t, db_link));
839 odn->dn_dbufs_count = 0;
840 odn->dn_bonus = NULL;
841 odn->dn_zfetch.zf_dnode = NULL;
844 * Set the low bit of the objset pointer to ensure that dnode_move()
845 * recognizes the dnode as invalid in any subsequent callback.
847 POINTER_INVALIDATE(&odn->dn_objset);
850 * Satisfy the destructor.
852 for (i = 0; i < TXG_SIZE; i++) {
853 list_create(&odn->dn_dirty_records[i],
854 sizeof (dbuf_dirty_record_t),
855 offsetof(dbuf_dirty_record_t, dr_dirty_node));
856 odn->dn_free_ranges[i] = NULL;
857 odn->dn_next_nlevels[i] = 0;
858 odn->dn_next_indblkshift[i] = 0;
859 odn->dn_next_bonustype[i] = 0;
860 odn->dn_rm_spillblk[i] = 0;
861 odn->dn_next_bonuslen[i] = 0;
862 odn->dn_next_blksz[i] = 0;
864 odn->dn_allocated_txg = 0;
865 odn->dn_free_txg = 0;
866 odn->dn_assigned_txg = 0;
867 odn->dn_dirty_txg = 0;
868 odn->dn_dirtyctx = 0;
869 odn->dn_dirtyctx_firstset = NULL;
870 odn->dn_have_spill = B_FALSE;
873 odn->dn_oldflags = 0;
876 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
879 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
880 odn->dn_id_flags = 0;
886 odn->dn_moved = (uint8_t)-1;
891 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
893 dnode_t *odn = buf, *ndn = newbuf;
899 * The dnode is on the objset's list of known dnodes if the objset
900 * pointer is valid. We set the low bit of the objset pointer when
901 * freeing the dnode to invalidate it, and the memory patterns written
902 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
903 * A newly created dnode sets the objset pointer last of all to indicate
904 * that the dnode is known and in a valid state to be moved by this
908 if (!POINTER_IS_VALID(os)) {
909 DNODE_STAT_BUMP(dnode_move_invalid);
910 return (KMEM_CBRC_DONT_KNOW);
914 * Ensure that the objset does not go away during the move.
916 rw_enter(&os_lock, RW_WRITER);
917 if (os != odn->dn_objset) {
919 DNODE_STAT_BUMP(dnode_move_recheck1);
920 return (KMEM_CBRC_DONT_KNOW);
924 * If the dnode is still valid, then so is the objset. We know that no
925 * valid objset can be freed while we hold os_lock, so we can safely
926 * ensure that the objset remains in use.
928 mutex_enter(&os->os_lock);
931 * Recheck the objset pointer in case the dnode was removed just before
932 * acquiring the lock.
934 if (os != odn->dn_objset) {
935 mutex_exit(&os->os_lock);
937 DNODE_STAT_BUMP(dnode_move_recheck2);
938 return (KMEM_CBRC_DONT_KNOW);
942 * At this point we know that as long as we hold os->os_lock, the dnode
943 * cannot be freed and fields within the dnode can be safely accessed.
944 * The objset listing this dnode cannot go away as long as this dnode is
948 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
949 mutex_exit(&os->os_lock);
950 DNODE_STAT_BUMP(dnode_move_special);
951 return (KMEM_CBRC_NO);
953 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
956 * Lock the dnode handle to prevent the dnode from obtaining any new
957 * holds. This also prevents the descendant dbufs and the bonus dbuf
958 * from accessing the dnode, so that we can discount their holds. The
959 * handle is safe to access because we know that while the dnode cannot
960 * go away, neither can its handle. Once we hold dnh_zrlock, we can
961 * safely move any dnode referenced only by dbufs.
963 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
964 mutex_exit(&os->os_lock);
965 DNODE_STAT_BUMP(dnode_move_handle);
966 return (KMEM_CBRC_LATER);
970 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
971 * We need to guarantee that there is a hold for every dbuf in order to
972 * determine whether the dnode is actively referenced. Falsely matching
973 * a dbuf to an active hold would lead to an unsafe move. It's possible
974 * that a thread already having an active dnode hold is about to add a
975 * dbuf, and we can't compare hold and dbuf counts while the add is in
978 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
979 zrl_exit(&odn->dn_handle->dnh_zrlock);
980 mutex_exit(&os->os_lock);
981 DNODE_STAT_BUMP(dnode_move_rwlock);
982 return (KMEM_CBRC_LATER);
986 * A dbuf may be removed (evicted) without an active dnode hold. In that
987 * case, the dbuf count is decremented under the handle lock before the
988 * dbuf's hold is released. This order ensures that if we count the hold
989 * after the dbuf is removed but before its hold is released, we will
990 * treat the unmatched hold as active and exit safely. If we count the
991 * hold before the dbuf is removed, the hold is discounted, and the
992 * removal is blocked until the move completes.
994 refcount = refcount_count(&odn->dn_holds);
995 ASSERT(refcount >= 0);
996 dbufs = odn->dn_dbufs_count;
998 /* We can't have more dbufs than dnode holds. */
999 ASSERT3U(dbufs, <=, refcount);
1000 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1003 if (refcount > dbufs) {
1004 rw_exit(&odn->dn_struct_rwlock);
1005 zrl_exit(&odn->dn_handle->dnh_zrlock);
1006 mutex_exit(&os->os_lock);
1007 DNODE_STAT_BUMP(dnode_move_active);
1008 return (KMEM_CBRC_LATER);
1011 rw_exit(&odn->dn_struct_rwlock);
1014 * At this point we know that anyone with a hold on the dnode is not
1015 * actively referencing it. The dnode is known and in a valid state to
1016 * move. We're holding the locks needed to execute the critical section.
1018 dnode_move_impl(odn, ndn);
1020 list_link_replace(&odn->dn_link, &ndn->dn_link);
1021 /* If the dnode was safe to move, the refcount cannot have changed. */
1022 ASSERT(refcount == refcount_count(&ndn->dn_holds));
1023 ASSERT(dbufs == ndn->dn_dbufs_count);
1024 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1025 mutex_exit(&os->os_lock);
1027 return (KMEM_CBRC_YES);
1029 #endif /* _KERNEL */
1032 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1034 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1036 for (int i = idx; i < idx + slots; i++) {
1037 dnode_handle_t *dnh = &children->dnc_children[i];
1038 zrl_add(&dnh->dnh_zrlock);
1043 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1045 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1047 for (int i = idx; i < idx + slots; i++) {
1048 dnode_handle_t *dnh = &children->dnc_children[i];
1050 if (zrl_is_locked(&dnh->dnh_zrlock))
1051 zrl_exit(&dnh->dnh_zrlock);
1053 zrl_remove(&dnh->dnh_zrlock);
1058 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1060 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1062 for (int i = idx; i < idx + slots; i++) {
1063 dnode_handle_t *dnh = &children->dnc_children[i];
1065 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1066 for (int j = idx; j < i; j++) {
1067 dnh = &children->dnc_children[j];
1068 zrl_exit(&dnh->dnh_zrlock);
1079 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1081 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1083 for (int i = idx; i < idx + slots; i++) {
1084 dnode_handle_t *dnh = &children->dnc_children[i];
1085 dnh->dnh_dnode = ptr;
1090 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1092 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1095 * If all dnode slots are either already free or
1096 * evictable return B_TRUE.
1098 for (int i = idx; i < idx + slots; i++) {
1099 dnode_handle_t *dnh = &children->dnc_children[i];
1100 dnode_t *dn = dnh->dnh_dnode;
1102 if (dn == DN_SLOT_FREE) {
1104 } else if (DN_SLOT_IS_PTR(dn)) {
1105 mutex_enter(&dn->dn_mtx);
1106 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1107 !DNODE_IS_DIRTY(dn));
1108 mutex_exit(&dn->dn_mtx);
1123 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1125 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1127 for (int i = idx; i < idx + slots; i++) {
1128 dnode_handle_t *dnh = &children->dnc_children[i];
1130 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1132 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1133 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1134 dnode_destroy(dnh->dnh_dnode);
1135 dnh->dnh_dnode = DN_SLOT_FREE;
1141 dnode_free_interior_slots(dnode_t *dn)
1143 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1144 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1145 int idx = (dn->dn_object & (epb - 1)) + 1;
1146 int slots = dn->dn_num_slots - 1;
1151 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1153 while (!dnode_slots_tryenter(children, idx, slots))
1154 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1156 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1157 dnode_slots_rele(children, idx, slots);
1161 dnode_special_close(dnode_handle_t *dnh)
1163 dnode_t *dn = dnh->dnh_dnode;
1166 * Wait for final references to the dnode to clear. This can
1167 * only happen if the arc is asynchronously evicting state that
1168 * has a hold on this dnode while we are trying to evict this
1171 while (refcount_count(&dn->dn_holds) > 0)
1173 ASSERT(dn->dn_dbuf == NULL ||
1174 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1175 zrl_add(&dnh->dnh_zrlock);
1176 dnode_destroy(dn); /* implicit zrl_remove() */
1177 zrl_destroy(&dnh->dnh_zrlock);
1178 dnh->dnh_dnode = NULL;
1182 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1183 dnode_handle_t *dnh)
1187 zrl_init(&dnh->dnh_zrlock);
1188 zrl_tryenter(&dnh->dnh_zrlock);
1190 dn = dnode_create(os, dnp, NULL, object, dnh);
1193 zrl_exit(&dnh->dnh_zrlock);
1197 dnode_buf_evict_async(void *dbu)
1199 dnode_children_t *dnc = dbu;
1201 DNODE_STAT_BUMP(dnode_buf_evict);
1203 for (int i = 0; i < dnc->dnc_count; i++) {
1204 dnode_handle_t *dnh = &dnc->dnc_children[i];
1208 * The dnode handle lock guards against the dnode moving to
1209 * another valid address, so there is no need here to guard
1210 * against changes to or from NULL.
1212 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1213 zrl_destroy(&dnh->dnh_zrlock);
1214 dnh->dnh_dnode = DN_SLOT_UNINIT;
1218 zrl_add(&dnh->dnh_zrlock);
1219 dn = dnh->dnh_dnode;
1221 * If there are holds on this dnode, then there should
1222 * be holds on the dnode's containing dbuf as well; thus
1223 * it wouldn't be eligible for eviction and this function
1224 * would not have been called.
1226 ASSERT(refcount_is_zero(&dn->dn_holds));
1227 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1229 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1230 zrl_destroy(&dnh->dnh_zrlock);
1231 dnh->dnh_dnode = DN_SLOT_UNINIT;
1233 kmem_free(dnc, sizeof (dnode_children_t) +
1234 dnc->dnc_count * sizeof (dnode_handle_t));
1238 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1239 * to ensure the hole at the specified object offset is large enough to
1240 * hold the dnode being created. The slots parameter is also used to ensure
1241 * a dnode does not span multiple dnode blocks. In both of these cases, if
1242 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1243 * are only possible when using DNODE_MUST_BE_FREE.
1245 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1246 * dnode_hold_impl() will check if the requested dnode is already consumed
1247 * as an extra dnode slot by an large dnode, in which case it returns
1251 * EINVAL - Invalid object number or flags.
1252 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1253 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1254 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1255 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1256 * EIO - I/O error when reading the meta dnode dbuf.
1258 * succeeds even for free dnodes.
1261 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1262 void *tag, dnode_t **dnp)
1265 int drop_struct_lock = FALSE;
1270 dnode_children_t *dnc;
1271 dnode_phys_t *dn_block;
1272 dnode_handle_t *dnh;
1274 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1275 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1278 * If you are holding the spa config lock as writer, you shouldn't
1279 * be asking the DMU to do *anything* unless it's the root pool
1280 * which may require us to read from the root filesystem while
1281 * holding some (not all) of the locks as writer.
1283 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1284 (spa_is_root(os->os_spa) &&
1285 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1287 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1288 object == DMU_PROJECTUSED_OBJECT) {
1289 if (object == DMU_USERUSED_OBJECT)
1290 dn = DMU_USERUSED_DNODE(os);
1291 else if (object == DMU_GROUPUSED_OBJECT)
1292 dn = DMU_GROUPUSED_DNODE(os);
1294 dn = DMU_PROJECTUSED_DNODE(os);
1296 return (SET_ERROR(ENOENT));
1298 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1299 return (SET_ERROR(ENOENT));
1300 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1301 return (SET_ERROR(EEXIST));
1303 (void) refcount_add(&dn->dn_holds, tag);
1308 if (object == 0 || object >= DN_MAX_OBJECT)
1309 return (SET_ERROR(EINVAL));
1311 mdn = DMU_META_DNODE(os);
1312 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1316 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1317 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1318 drop_struct_lock = TRUE;
1321 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1323 db = dbuf_hold(mdn, blk, FTAG);
1324 if (drop_struct_lock)
1325 rw_exit(&mdn->dn_struct_rwlock);
1327 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1328 return (SET_ERROR(EIO));
1332 * We do not need to decrypt to read the dnode so it doesn't matter
1333 * if we get the encrypted or decrypted version.
1335 err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1337 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1338 dbuf_rele(db, FTAG);
1342 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1343 epb = db->db.db_size >> DNODE_SHIFT;
1345 idx = object & (epb - 1);
1346 dn_block = (dnode_phys_t *)db->db.db_data;
1348 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1349 dnc = dmu_buf_get_user(&db->db);
1352 dnode_children_t *winner;
1355 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1356 epb * sizeof (dnode_handle_t), KM_SLEEP);
1357 dnc->dnc_count = epb;
1358 dnh = &dnc->dnc_children[0];
1360 /* Initialize dnode slot status from dnode_phys_t */
1361 for (int i = 0; i < epb; i++) {
1362 zrl_init(&dnh[i].dnh_zrlock);
1369 if (dn_block[i].dn_type != DMU_OT_NONE) {
1370 int interior = dn_block[i].dn_extra_slots;
1372 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1373 dnode_set_slots(dnc, i + 1, interior,
1377 dnh[i].dnh_dnode = DN_SLOT_FREE;
1382 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1383 dnode_buf_evict_async, NULL);
1384 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1385 if (winner != NULL) {
1387 for (int i = 0; i < epb; i++)
1388 zrl_destroy(&dnh[i].dnh_zrlock);
1390 kmem_free(dnc, sizeof (dnode_children_t) +
1391 epb * sizeof (dnode_handle_t));
1396 ASSERT(dnc->dnc_count == epb);
1397 dn = DN_SLOT_UNINIT;
1399 if (flag & DNODE_MUST_BE_ALLOCATED) {
1402 while (dn == DN_SLOT_UNINIT) {
1403 dnode_slots_hold(dnc, idx, slots);
1404 dnh = &dnc->dnc_children[idx];
1406 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1407 dn = dnh->dnh_dnode;
1409 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1410 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1411 dnode_slots_rele(dnc, idx, slots);
1412 dbuf_rele(db, FTAG);
1413 return (SET_ERROR(EEXIST));
1414 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1415 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1416 dnode_slots_rele(dnc, idx, slots);
1417 dbuf_rele(db, FTAG);
1418 return (SET_ERROR(ENOENT));
1421 dnode_slots_rele(dnc, idx, slots);
1422 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1423 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1428 * Someone else won the race and called dnode_create()
1429 * after we checked DN_SLOT_IS_PTR() above but before
1430 * we acquired the lock.
1432 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1433 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1434 dn = dnh->dnh_dnode;
1436 dn = dnode_create(os, dn_block + idx, db,
1441 mutex_enter(&dn->dn_mtx);
1442 if (dn->dn_type == DMU_OT_NONE) {
1443 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1444 mutex_exit(&dn->dn_mtx);
1445 dnode_slots_rele(dnc, idx, slots);
1446 dbuf_rele(db, FTAG);
1447 return (SET_ERROR(ENOENT));
1450 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1451 } else if (flag & DNODE_MUST_BE_FREE) {
1453 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1454 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1455 dbuf_rele(db, FTAG);
1456 return (SET_ERROR(ENOSPC));
1459 while (dn == DN_SLOT_UNINIT) {
1460 dnode_slots_hold(dnc, idx, slots);
1462 if (!dnode_check_slots_free(dnc, idx, slots)) {
1463 DNODE_STAT_BUMP(dnode_hold_free_misses);
1464 dnode_slots_rele(dnc, idx, slots);
1465 dbuf_rele(db, FTAG);
1466 return (SET_ERROR(ENOSPC));
1469 dnode_slots_rele(dnc, idx, slots);
1470 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1471 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1475 if (!dnode_check_slots_free(dnc, idx, slots)) {
1476 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1477 dnode_slots_rele(dnc, idx, slots);
1478 dbuf_rele(db, FTAG);
1479 return (SET_ERROR(ENOSPC));
1483 * Allocated but otherwise free dnodes which would
1484 * be in the interior of a multi-slot dnodes need
1485 * to be freed. Single slot dnodes can be safely
1486 * re-purposed as a performance optimization.
1489 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1491 dnh = &dnc->dnc_children[idx];
1492 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1493 dn = dnh->dnh_dnode;
1495 dn = dnode_create(os, dn_block + idx, db,
1500 mutex_enter(&dn->dn_mtx);
1501 if (!refcount_is_zero(&dn->dn_holds)) {
1502 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1503 mutex_exit(&dn->dn_mtx);
1504 dnode_slots_rele(dnc, idx, slots);
1505 dbuf_rele(db, FTAG);
1506 return (SET_ERROR(EEXIST));
1509 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1510 DNODE_STAT_BUMP(dnode_hold_free_hits);
1512 dbuf_rele(db, FTAG);
1513 return (SET_ERROR(EINVAL));
1516 if (dn->dn_free_txg) {
1517 DNODE_STAT_BUMP(dnode_hold_free_txg);
1519 mutex_exit(&dn->dn_mtx);
1520 dnode_slots_rele(dnc, idx, slots);
1521 dbuf_rele(db, FTAG);
1522 return (SET_ERROR(type == DMU_OT_NONE ? ENOENT : EEXIST));
1525 if (refcount_add(&dn->dn_holds, tag) == 1)
1526 dbuf_add_ref(db, dnh);
1528 mutex_exit(&dn->dn_mtx);
1530 /* Now we can rely on the hold to prevent the dnode from moving. */
1531 dnode_slots_rele(dnc, idx, slots);
1534 ASSERT3P(dn->dn_dbuf, ==, db);
1535 ASSERT3U(dn->dn_object, ==, object);
1536 dbuf_rele(db, FTAG);
1543 * Return held dnode if the object is allocated, NULL if not.
1546 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1548 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1553 * Can only add a reference if there is already at least one
1554 * reference on the dnode. Returns FALSE if unable to add a
1558 dnode_add_ref(dnode_t *dn, void *tag)
1560 mutex_enter(&dn->dn_mtx);
1561 if (refcount_is_zero(&dn->dn_holds)) {
1562 mutex_exit(&dn->dn_mtx);
1565 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1566 mutex_exit(&dn->dn_mtx);
1571 dnode_rele(dnode_t *dn, void *tag)
1573 mutex_enter(&dn->dn_mtx);
1574 dnode_rele_and_unlock(dn, tag, B_FALSE);
1578 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1581 /* Get while the hold prevents the dnode from moving. */
1582 dmu_buf_impl_t *db = dn->dn_dbuf;
1583 dnode_handle_t *dnh = dn->dn_handle;
1585 refs = refcount_remove(&dn->dn_holds, tag);
1586 mutex_exit(&dn->dn_mtx);
1589 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1590 * indirectly by dbuf_rele() while relying on the dnode handle to
1591 * prevent the dnode from moving, since releasing the last hold could
1592 * result in the dnode's parent dbuf evicting its dnode handles. For
1593 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1594 * other direct or indirect hold on the dnode must first drop the dnode
1597 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1599 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1600 if (refs == 0 && db != NULL) {
1602 * Another thread could add a hold to the dnode handle in
1603 * dnode_hold_impl() while holding the parent dbuf. Since the
1604 * hold on the parent dbuf prevents the handle from being
1605 * destroyed, the hold on the handle is OK. We can't yet assert
1606 * that the handle has zero references, but that will be
1607 * asserted anyway when the handle gets destroyed.
1609 mutex_enter(&db->db_mtx);
1610 dbuf_rele_and_unlock(db, dnh, evicting);
1615 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1617 objset_t *os = dn->dn_objset;
1618 uint64_t txg = tx->tx_txg;
1620 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1621 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1628 mutex_enter(&dn->dn_mtx);
1629 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1630 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1631 mutex_exit(&dn->dn_mtx);
1635 * Determine old uid/gid when necessary
1637 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1639 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1640 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1643 * If we are already marked dirty, we're done.
1645 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1646 multilist_sublist_unlock(mls);
1650 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1651 !avl_is_empty(&dn->dn_dbufs));
1652 ASSERT(dn->dn_datablksz != 0);
1653 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1654 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1655 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1657 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1658 dn->dn_object, txg);
1660 multilist_sublist_insert_head(mls, dn);
1662 multilist_sublist_unlock(mls);
1665 * The dnode maintains a hold on its containing dbuf as
1666 * long as there are holds on it. Each instantiated child
1667 * dbuf maintains a hold on the dnode. When the last child
1668 * drops its hold, the dnode will drop its hold on the
1669 * containing dbuf. We add a "dirty hold" here so that the
1670 * dnode will hang around after we finish processing its
1673 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1675 (void) dbuf_dirty(dn->dn_dbuf, tx);
1677 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1681 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1683 mutex_enter(&dn->dn_mtx);
1684 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1685 mutex_exit(&dn->dn_mtx);
1688 dn->dn_free_txg = tx->tx_txg;
1689 mutex_exit(&dn->dn_mtx);
1691 dnode_setdirty(dn, tx);
1695 * Try to change the block size for the indicated dnode. This can only
1696 * succeed if there are no blocks allocated or dirty beyond first block
1699 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1704 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1706 size = SPA_MINBLOCKSIZE;
1708 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1710 if (ibs == dn->dn_indblkshift)
1713 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1716 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1718 /* Check for any allocated blocks beyond the first */
1719 if (dn->dn_maxblkid != 0)
1722 mutex_enter(&dn->dn_dbufs_mtx);
1723 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1724 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1725 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1726 db->db_blkid != DMU_SPILL_BLKID) {
1727 mutex_exit(&dn->dn_dbufs_mtx);
1731 mutex_exit(&dn->dn_dbufs_mtx);
1733 if (ibs && dn->dn_nlevels != 1)
1736 /* resize the old block */
1737 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1739 dbuf_new_size(db, size, tx);
1740 else if (err != ENOENT)
1743 dnode_setdblksz(dn, size);
1744 dnode_setdirty(dn, tx);
1745 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1747 dn->dn_indblkshift = ibs;
1748 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1750 /* rele after we have fixed the blocksize in the dnode */
1752 dbuf_rele(db, FTAG);
1754 rw_exit(&dn->dn_struct_rwlock);
1758 rw_exit(&dn->dn_struct_rwlock);
1759 return (SET_ERROR(ENOTSUP));
1763 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1765 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1766 int old_nlevels = dn->dn_nlevels;
1769 dbuf_dirty_record_t *new, *dr, *dr_next;
1771 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1773 dn->dn_nlevels = new_nlevels;
1775 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1776 dn->dn_next_nlevels[txgoff] = new_nlevels;
1778 /* dirty the left indirects */
1779 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1781 new = dbuf_dirty(db, tx);
1782 dbuf_rele(db, FTAG);
1784 /* transfer the dirty records to the new indirect */
1785 mutex_enter(&dn->dn_mtx);
1786 mutex_enter(&new->dt.di.dr_mtx);
1787 list = &dn->dn_dirty_records[txgoff];
1788 for (dr = list_head(list); dr; dr = dr_next) {
1789 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1790 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1791 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1792 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1793 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1794 list_remove(&dn->dn_dirty_records[txgoff], dr);
1795 list_insert_tail(&new->dt.di.dr_children, dr);
1796 dr->dr_parent = new;
1799 mutex_exit(&new->dt.di.dr_mtx);
1800 mutex_exit(&dn->dn_mtx);
1804 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1808 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1810 if (dn->dn_nlevels == nlevels) {
1813 } else if (nlevels < dn->dn_nlevels) {
1814 ret = SET_ERROR(EINVAL);
1818 dnode_set_nlevels_impl(dn, nlevels, tx);
1821 rw_exit(&dn->dn_struct_rwlock);
1825 /* read-holding callers must not rely on the lock being continuously held */
1827 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1829 int epbs, new_nlevels;
1832 ASSERT(blkid != DMU_BONUS_BLKID);
1835 RW_READ_HELD(&dn->dn_struct_rwlock) :
1836 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1839 * if we have a read-lock, check to see if we need to do any work
1840 * before upgrading to a write-lock.
1843 if (blkid <= dn->dn_maxblkid)
1846 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1847 rw_exit(&dn->dn_struct_rwlock);
1848 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1852 if (blkid <= dn->dn_maxblkid)
1855 dn->dn_maxblkid = blkid;
1856 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] = blkid;
1859 * Compute the number of levels necessary to support the new maxblkid.
1862 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1863 for (sz = dn->dn_nblkptr;
1864 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1867 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1869 if (new_nlevels > dn->dn_nlevels)
1870 dnode_set_nlevels_impl(dn, new_nlevels, tx);
1874 rw_downgrade(&dn->dn_struct_rwlock);
1878 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1880 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1882 dmu_buf_will_dirty(&db->db, tx);
1883 dbuf_rele(db, FTAG);
1888 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1891 uint64_t blkoff, blkid, nblks;
1892 int blksz, blkshift, head, tail;
1896 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1897 blksz = dn->dn_datablksz;
1898 blkshift = dn->dn_datablkshift;
1899 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1901 if (len == DMU_OBJECT_END) {
1902 len = UINT64_MAX - off;
1907 * First, block align the region to free:
1910 head = P2NPHASE(off, blksz);
1911 blkoff = P2PHASE(off, blksz);
1912 if ((off >> blkshift) > dn->dn_maxblkid)
1915 ASSERT(dn->dn_maxblkid == 0);
1916 if (off == 0 && len >= blksz) {
1918 * Freeing the whole block; fast-track this request.
1919 * Note that we won't dirty any indirect blocks,
1920 * which is fine because we will be freeing the entire
1921 * file and thus all indirect blocks will be freed
1922 * by free_children().
1927 } else if (off >= blksz) {
1928 /* Freeing past end-of-data */
1931 /* Freeing part of the block. */
1933 ASSERT3U(head, >, 0);
1937 /* zero out any partial block data at the start of the range */
1939 ASSERT3U(blkoff + head, ==, blksz);
1942 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1943 TRUE, FALSE, FTAG, &db) == 0) {
1946 /* don't dirty if it isn't on disk and isn't dirty */
1947 if (db->db_last_dirty ||
1948 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1949 rw_exit(&dn->dn_struct_rwlock);
1950 dmu_buf_will_dirty(&db->db, tx);
1951 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1952 data = db->db.db_data;
1953 bzero(data + blkoff, head);
1955 dbuf_rele(db, FTAG);
1961 /* If the range was less than one block, we're done */
1965 /* If the remaining range is past end of file, we're done */
1966 if ((off >> blkshift) > dn->dn_maxblkid)
1969 ASSERT(ISP2(blksz));
1973 tail = P2PHASE(len, blksz);
1975 ASSERT0(P2PHASE(off, blksz));
1976 /* zero out any partial block data at the end of the range */
1980 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1981 TRUE, FALSE, FTAG, &db) == 0) {
1982 /* don't dirty if not on disk and not dirty */
1983 if (db->db_last_dirty ||
1984 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1985 rw_exit(&dn->dn_struct_rwlock);
1986 dmu_buf_will_dirty(&db->db, tx);
1987 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1988 bzero(db->db.db_data, tail);
1990 dbuf_rele(db, FTAG);
1995 /* If the range did not include a full block, we are done */
1999 ASSERT(IS_P2ALIGNED(off, blksz));
2000 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2001 blkid = off >> blkshift;
2002 nblks = len >> blkshift;
2007 * Dirty all the indirect blocks in this range. Note that only
2008 * the first and last indirect blocks can actually be written
2009 * (if they were partially freed) -- they must be dirtied, even if
2010 * they do not exist on disk yet. The interior blocks will
2011 * be freed by free_children(), so they will not actually be written.
2012 * Even though these interior blocks will not be written, we
2013 * dirty them for two reasons:
2015 * - It ensures that the indirect blocks remain in memory until
2016 * syncing context. (They have already been prefetched by
2017 * dmu_tx_hold_free(), so we don't have to worry about reading
2018 * them serially here.)
2020 * - The dirty space accounting will put pressure on the txg sync
2021 * mechanism to begin syncing, and to delay transactions if there
2022 * is a large amount of freeing. Even though these indirect
2023 * blocks will not be written, we could need to write the same
2024 * amount of space if we copy the freed BPs into deadlists.
2026 if (dn->dn_nlevels > 1) {
2027 uint64_t first, last;
2029 first = blkid >> epbs;
2030 dnode_dirty_l1(dn, first, tx);
2032 last = dn->dn_maxblkid >> epbs;
2034 last = (blkid + nblks - 1) >> epbs;
2036 dnode_dirty_l1(dn, last, tx);
2038 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2040 for (uint64_t i = first + 1; i < last; i++) {
2042 * Set i to the blockid of the next non-hole
2043 * level-1 indirect block at or after i. Note
2044 * that dnode_next_offset() operates in terms of
2045 * level-0-equivalent bytes.
2047 uint64_t ibyte = i << shift;
2048 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2055 * Normally we should not see an error, either
2056 * from dnode_next_offset() or dbuf_hold_level()
2057 * (except for ESRCH from dnode_next_offset).
2058 * If there is an i/o error, then when we read
2059 * this block in syncing context, it will use
2060 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2061 * to the "failmode" property. dnode_next_offset()
2062 * doesn't have a flag to indicate MUSTSUCCEED.
2067 dnode_dirty_l1(dn, i, tx);
2073 * Add this range to the dnode range list.
2074 * We will finish up this free operation in the syncing phase.
2076 mutex_enter(&dn->dn_mtx);
2078 int txgoff = tx->tx_txg & TXG_MASK;
2079 if (dn->dn_free_ranges[txgoff] == NULL) {
2080 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2082 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2083 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2085 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2086 blkid, nblks, tx->tx_txg);
2087 mutex_exit(&dn->dn_mtx);
2089 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2090 dnode_setdirty(dn, tx);
2093 rw_exit(&dn->dn_struct_rwlock);
2097 dnode_spill_freed(dnode_t *dn)
2101 mutex_enter(&dn->dn_mtx);
2102 for (i = 0; i < TXG_SIZE; i++) {
2103 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2106 mutex_exit(&dn->dn_mtx);
2107 return (i < TXG_SIZE);
2110 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2112 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2114 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2117 if (blkid == DMU_BONUS_BLKID)
2121 * If we're in the process of opening the pool, dp will not be
2122 * set yet, but there shouldn't be anything dirty.
2127 if (dn->dn_free_txg)
2130 if (blkid == DMU_SPILL_BLKID)
2131 return (dnode_spill_freed(dn));
2133 mutex_enter(&dn->dn_mtx);
2134 for (i = 0; i < TXG_SIZE; i++) {
2135 if (dn->dn_free_ranges[i] != NULL &&
2136 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2139 mutex_exit(&dn->dn_mtx);
2140 return (i < TXG_SIZE);
2143 /* call from syncing context when we actually write/free space for this dnode */
2145 dnode_diduse_space(dnode_t *dn, int64_t delta)
2148 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2150 (u_longlong_t)dn->dn_phys->dn_used,
2153 mutex_enter(&dn->dn_mtx);
2154 space = DN_USED_BYTES(dn->dn_phys);
2156 ASSERT3U(space + delta, >=, space); /* no overflow */
2158 ASSERT3U(space, >=, -delta); /* no underflow */
2161 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2162 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2163 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2164 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2166 dn->dn_phys->dn_used = space;
2167 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2169 mutex_exit(&dn->dn_mtx);
2173 * Scans a block at the indicated "level" looking for a hole or data,
2174 * depending on 'flags'.
2176 * If level > 0, then we are scanning an indirect block looking at its
2177 * pointers. If level == 0, then we are looking at a block of dnodes.
2179 * If we don't find what we are looking for in the block, we return ESRCH.
2180 * Otherwise, return with *offset pointing to the beginning (if searching
2181 * forwards) or end (if searching backwards) of the range covered by the
2182 * block pointer we matched on (or dnode).
2184 * The basic search algorithm used below by dnode_next_offset() is to
2185 * use this function to search up the block tree (widen the search) until
2186 * we find something (i.e., we don't return ESRCH) and then search back
2187 * down the tree (narrow the search) until we reach our original search
2191 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2192 int lvl, uint64_t blkfill, uint64_t txg)
2194 dmu_buf_impl_t *db = NULL;
2196 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2197 uint64_t epb = 1ULL << epbs;
2198 uint64_t minfill, maxfill;
2200 int i, inc, error, span;
2202 hole = ((flags & DNODE_FIND_HOLE) != 0);
2203 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2204 ASSERT(txg == 0 || !hole);
2206 if (lvl == dn->dn_phys->dn_nlevels) {
2208 epb = dn->dn_phys->dn_nblkptr;
2209 data = dn->dn_phys->dn_blkptr;
2211 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2212 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2214 if (error != ENOENT)
2219 * This can only happen when we are searching up
2220 * the block tree for data. We don't really need to
2221 * adjust the offset, as we will just end up looking
2222 * at the pointer to this block in its parent, and its
2223 * going to be unallocated, so we will skip over it.
2225 return (SET_ERROR(ESRCH));
2227 error = dbuf_read(db, NULL,
2228 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2230 dbuf_rele(db, FTAG);
2233 data = db->db.db_data;
2237 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2238 db->db_blkptr->blk_birth <= txg ||
2239 BP_IS_HOLE(db->db_blkptr))) {
2241 * This can only happen when we are searching up the tree
2242 * and these conditions mean that we need to keep climbing.
2244 error = SET_ERROR(ESRCH);
2245 } else if (lvl == 0) {
2246 dnode_phys_t *dnp = data;
2248 ASSERT(dn->dn_type == DMU_OT_DNODE);
2249 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2251 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2252 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2253 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2258 error = SET_ERROR(ESRCH);
2260 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2263 blkptr_t *bp = data;
2264 uint64_t start = *offset;
2265 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2267 maxfill = blkfill << ((lvl - 1) * epbs);
2274 if (span >= 8 * sizeof (*offset)) {
2275 /* This only happens on the highest indirection level */
2276 ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2279 *offset = *offset >> span;
2282 for (i = BF64_GET(*offset, 0, epbs);
2283 i >= 0 && i < epb; i += inc) {
2284 if (BP_GET_FILL(&bp[i]) >= minfill &&
2285 BP_GET_FILL(&bp[i]) <= maxfill &&
2286 (hole || bp[i].blk_birth > txg))
2288 if (inc > 0 || *offset > 0)
2292 if (span >= 8 * sizeof (*offset)) {
2295 *offset = *offset << span;
2299 /* traversing backwards; position offset at the end */
2300 ASSERT3U(*offset, <=, start);
2301 *offset = MIN(*offset + (1ULL << span) - 1, start);
2302 } else if (*offset < start) {
2305 if (i < 0 || i >= epb)
2306 error = SET_ERROR(ESRCH);
2310 dbuf_rele(db, FTAG);
2316 * Find the next hole, data, or sparse region at or after *offset.
2317 * The value 'blkfill' tells us how many items we expect to find
2318 * in an L0 data block; this value is 1 for normal objects,
2319 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2320 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2324 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2325 * Finds the next/previous hole/data in a file.
2326 * Used in dmu_offset_next().
2328 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2329 * Finds the next free/allocated dnode an objset's meta-dnode.
2330 * Only finds objects that have new contents since txg (ie.
2331 * bonus buffer changes and content removal are ignored).
2332 * Used in dmu_object_next().
2334 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2335 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2336 * Used in dmu_object_alloc().
2339 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2340 int minlvl, uint64_t blkfill, uint64_t txg)
2342 uint64_t initial_offset = *offset;
2346 if (!(flags & DNODE_FIND_HAVELOCK))
2347 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2349 if (dn->dn_phys->dn_nlevels == 0) {
2350 error = SET_ERROR(ESRCH);
2354 if (dn->dn_datablkshift == 0) {
2355 if (*offset < dn->dn_datablksz) {
2356 if (flags & DNODE_FIND_HOLE)
2357 *offset = dn->dn_datablksz;
2359 error = SET_ERROR(ESRCH);
2364 maxlvl = dn->dn_phys->dn_nlevels;
2366 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2367 error = dnode_next_offset_level(dn,
2368 flags, offset, lvl, blkfill, txg);
2373 while (error == 0 && --lvl >= minlvl) {
2374 error = dnode_next_offset_level(dn,
2375 flags, offset, lvl, blkfill, txg);
2379 * There's always a "virtual hole" at the end of the object, even
2380 * if all BP's which physically exist are non-holes.
2382 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2383 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2387 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2388 initial_offset < *offset : initial_offset > *offset))
2389 error = SET_ERROR(ESRCH);
2391 if (!(flags & DNODE_FIND_HAVELOCK))
2392 rw_exit(&dn->dn_struct_rwlock);