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, 2014 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
28 #include <sys/dnode.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/range_tree.h>
40 static kmem_cache_t *dnode_cache;
42 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
43 * turned on when DEBUG is also defined.
50 #define DNODE_STAT_ADD(stat) ((stat)++)
52 #define DNODE_STAT_ADD(stat) /* nothing */
53 #endif /* DNODE_STATS */
55 static dnode_phys_t dnode_phys_zero;
57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
65 dbuf_compare(const void *x1, const void *x2)
67 const dmu_buf_impl_t *d1 = x1;
68 const dmu_buf_impl_t *d2 = x2;
70 if (d1->db_level < d2->db_level) {
72 } else if (d1->db_level > d2->db_level) {
76 if (d1->db_blkid < d2->db_blkid) {
78 } else if (d1->db_blkid > d2->db_blkid) {
83 * If a dbuf is being evicted while dn_dbufs_mutex is not held, we set
84 * the db_state to DB_EVICTING but do not remove it from dn_dbufs. If
85 * another thread creates a dbuf of the same blkid before the dbuf is
86 * removed from dn_dbufs, we can reach a state where there are two
87 * dbufs of the same blkid and level in db_dbufs. To maintain the avl
88 * invariant that there cannot be duplicate items, we distinguish
89 * between these two dbufs based on the time they were created.
91 if (d1->db_creation < d2->db_creation) {
93 } else if (d1->db_creation > d2->db_creation) {
103 dnode_cons(void *arg, void *unused, int kmflag)
108 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
109 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
110 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
111 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
114 * Every dbuf has a reference, and dropping a tracked reference is
115 * O(number of references), so don't track dn_holds.
117 refcount_create_untracked(&dn->dn_holds);
118 refcount_create(&dn->dn_tx_holds);
119 list_link_init(&dn->dn_link);
121 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
122 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
123 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
124 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
125 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
126 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
127 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
129 for (i = 0; i < TXG_SIZE; i++) {
130 list_link_init(&dn->dn_dirty_link[i]);
131 dn->dn_free_ranges[i] = NULL;
132 list_create(&dn->dn_dirty_records[i],
133 sizeof (dbuf_dirty_record_t),
134 offsetof(dbuf_dirty_record_t, dr_dirty_node));
137 dn->dn_allocated_txg = 0;
139 dn->dn_assigned_txg = 0;
141 dn->dn_dirtyctx_firstset = NULL;
143 dn->dn_have_spill = B_FALSE;
153 dn->dn_dbufs_count = 0;
154 dn->dn_unlisted_l0_blkid = 0;
155 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
156 offsetof(dmu_buf_impl_t, db_link));
159 POINTER_INVALIDATE(&dn->dn_objset);
165 dnode_dest(void *arg, void *unused)
170 rw_destroy(&dn->dn_struct_rwlock);
171 mutex_destroy(&dn->dn_mtx);
172 mutex_destroy(&dn->dn_dbufs_mtx);
173 cv_destroy(&dn->dn_notxholds);
174 refcount_destroy(&dn->dn_holds);
175 refcount_destroy(&dn->dn_tx_holds);
176 ASSERT(!list_link_active(&dn->dn_link));
178 for (i = 0; i < TXG_SIZE; i++) {
179 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
180 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
181 list_destroy(&dn->dn_dirty_records[i]);
182 ASSERT0(dn->dn_next_nblkptr[i]);
183 ASSERT0(dn->dn_next_nlevels[i]);
184 ASSERT0(dn->dn_next_indblkshift[i]);
185 ASSERT0(dn->dn_next_bonustype[i]);
186 ASSERT0(dn->dn_rm_spillblk[i]);
187 ASSERT0(dn->dn_next_bonuslen[i]);
188 ASSERT0(dn->dn_next_blksz[i]);
191 ASSERT0(dn->dn_allocated_txg);
192 ASSERT0(dn->dn_free_txg);
193 ASSERT0(dn->dn_assigned_txg);
194 ASSERT0(dn->dn_dirtyctx);
195 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
196 ASSERT3P(dn->dn_bonus, ==, NULL);
197 ASSERT(!dn->dn_have_spill);
198 ASSERT3P(dn->dn_zio, ==, NULL);
199 ASSERT0(dn->dn_oldused);
200 ASSERT0(dn->dn_oldflags);
201 ASSERT0(dn->dn_olduid);
202 ASSERT0(dn->dn_oldgid);
203 ASSERT0(dn->dn_newuid);
204 ASSERT0(dn->dn_newgid);
205 ASSERT0(dn->dn_id_flags);
207 ASSERT0(dn->dn_dbufs_count);
208 ASSERT0(dn->dn_unlisted_l0_blkid);
209 avl_destroy(&dn->dn_dbufs);
215 ASSERT(dnode_cache == NULL);
216 dnode_cache = kmem_cache_create("dnode_t",
218 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
219 kmem_cache_set_move(dnode_cache, dnode_move);
225 kmem_cache_destroy(dnode_cache);
232 dnode_verify(dnode_t *dn)
234 int drop_struct_lock = FALSE;
237 ASSERT(dn->dn_objset);
238 ASSERT(dn->dn_handle->dnh_dnode == dn);
240 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
242 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
245 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
246 rw_enter(&dn->dn_struct_rwlock, RW_READER);
247 drop_struct_lock = TRUE;
249 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
251 ASSERT3U(dn->dn_indblkshift, >=, 0);
252 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
253 if (dn->dn_datablkshift) {
254 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
255 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
256 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
258 ASSERT3U(dn->dn_nlevels, <=, 30);
259 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
260 ASSERT3U(dn->dn_nblkptr, >=, 1);
261 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
262 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
263 ASSERT3U(dn->dn_datablksz, ==,
264 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
265 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
266 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
267 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
268 for (i = 0; i < TXG_SIZE; i++) {
269 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
272 if (dn->dn_phys->dn_type != DMU_OT_NONE)
273 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
274 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
275 if (dn->dn_dbuf != NULL) {
276 ASSERT3P(dn->dn_phys, ==,
277 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
278 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
280 if (drop_struct_lock)
281 rw_exit(&dn->dn_struct_rwlock);
286 dnode_byteswap(dnode_phys_t *dnp)
288 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
291 if (dnp->dn_type == DMU_OT_NONE) {
292 bzero(dnp, sizeof (dnode_phys_t));
296 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
297 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
298 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
299 dnp->dn_used = BSWAP_64(dnp->dn_used);
302 * dn_nblkptr is only one byte, so it's OK to read it in either
303 * byte order. We can't read dn_bouslen.
305 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
306 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
307 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
308 buf64[i] = BSWAP_64(buf64[i]);
311 * OK to check dn_bonuslen for zero, because it won't matter if
312 * we have the wrong byte order. This is necessary because the
313 * dnode dnode is smaller than a regular dnode.
315 if (dnp->dn_bonuslen != 0) {
317 * Note that the bonus length calculated here may be
318 * longer than the actual bonus buffer. This is because
319 * we always put the bonus buffer after the last block
320 * pointer (instead of packing it against the end of the
323 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
324 size_t len = DN_MAX_BONUSLEN - off;
325 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
326 dmu_object_byteswap_t byteswap =
327 DMU_OT_BYTESWAP(dnp->dn_bonustype);
328 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
331 /* Swap SPILL block if we have one */
332 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
333 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
338 dnode_buf_byteswap(void *vbuf, size_t size)
340 dnode_phys_t *buf = vbuf;
343 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
344 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
346 size >>= DNODE_SHIFT;
347 for (i = 0; i < size; i++) {
354 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
356 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
358 dnode_setdirty(dn, tx);
359 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
360 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
361 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
362 dn->dn_bonuslen = newsize;
364 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
366 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
367 rw_exit(&dn->dn_struct_rwlock);
371 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
373 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
374 dnode_setdirty(dn, tx);
375 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
376 dn->dn_bonustype = newtype;
377 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
378 rw_exit(&dn->dn_struct_rwlock);
382 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
384 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
385 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
386 dnode_setdirty(dn, tx);
387 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
388 dn->dn_have_spill = B_FALSE;
392 dnode_setdblksz(dnode_t *dn, int size)
394 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
395 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
396 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
397 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
398 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
399 dn->dn_datablksz = size;
400 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
401 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
405 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
406 uint64_t object, dnode_handle_t *dnh)
408 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
410 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
414 * Defer setting dn_objset until the dnode is ready to be a candidate
415 * for the dnode_move() callback.
417 dn->dn_object = object;
422 if (dnp->dn_datablkszsec) {
423 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
425 dn->dn_datablksz = 0;
426 dn->dn_datablkszsec = 0;
427 dn->dn_datablkshift = 0;
429 dn->dn_indblkshift = dnp->dn_indblkshift;
430 dn->dn_nlevels = dnp->dn_nlevels;
431 dn->dn_type = dnp->dn_type;
432 dn->dn_nblkptr = dnp->dn_nblkptr;
433 dn->dn_checksum = dnp->dn_checksum;
434 dn->dn_compress = dnp->dn_compress;
435 dn->dn_bonustype = dnp->dn_bonustype;
436 dn->dn_bonuslen = dnp->dn_bonuslen;
437 dn->dn_maxblkid = dnp->dn_maxblkid;
438 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
441 dmu_zfetch_init(&dn->dn_zfetch, dn);
443 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
445 mutex_enter(&os->os_lock);
446 list_insert_head(&os->os_dnodes, dn);
449 * Everything else must be valid before assigning dn_objset makes the
450 * dnode eligible for dnode_move().
453 mutex_exit(&os->os_lock);
455 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
460 * Caller must be holding the dnode handle, which is released upon return.
463 dnode_destroy(dnode_t *dn)
465 objset_t *os = dn->dn_objset;
467 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
469 mutex_enter(&os->os_lock);
470 POINTER_INVALIDATE(&dn->dn_objset);
471 list_remove(&os->os_dnodes, dn);
472 mutex_exit(&os->os_lock);
474 /* the dnode can no longer move, so we can release the handle */
475 zrl_remove(&dn->dn_handle->dnh_zrlock);
477 dn->dn_allocated_txg = 0;
479 dn->dn_assigned_txg = 0;
482 if (dn->dn_dirtyctx_firstset != NULL) {
483 kmem_free(dn->dn_dirtyctx_firstset, 1);
484 dn->dn_dirtyctx_firstset = NULL;
486 if (dn->dn_bonus != NULL) {
487 mutex_enter(&dn->dn_bonus->db_mtx);
488 dbuf_evict(dn->dn_bonus);
493 dn->dn_have_spill = B_FALSE;
501 dn->dn_unlisted_l0_blkid = 0;
503 dmu_zfetch_rele(&dn->dn_zfetch);
504 kmem_cache_free(dnode_cache, dn);
505 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
509 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
510 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
515 blocksize = 1 << zfs_default_bs;
516 else if (blocksize > SPA_MAXBLOCKSIZE)
517 blocksize = SPA_MAXBLOCKSIZE;
519 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
522 ibs = zfs_default_ibs;
524 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
526 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
527 dn->dn_object, tx->tx_txg, blocksize, ibs);
529 ASSERT(dn->dn_type == DMU_OT_NONE);
530 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
531 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
532 ASSERT(ot != DMU_OT_NONE);
533 ASSERT(DMU_OT_IS_VALID(ot));
534 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
535 (bonustype == DMU_OT_SA && bonuslen == 0) ||
536 (bonustype != DMU_OT_NONE && bonuslen != 0));
537 ASSERT(DMU_OT_IS_VALID(bonustype));
538 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
539 ASSERT(dn->dn_type == DMU_OT_NONE);
540 ASSERT0(dn->dn_maxblkid);
541 ASSERT0(dn->dn_allocated_txg);
542 ASSERT0(dn->dn_assigned_txg);
543 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
544 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
545 ASSERT(avl_is_empty(&dn->dn_dbufs));
547 for (i = 0; i < TXG_SIZE; i++) {
548 ASSERT0(dn->dn_next_nblkptr[i]);
549 ASSERT0(dn->dn_next_nlevels[i]);
550 ASSERT0(dn->dn_next_indblkshift[i]);
551 ASSERT0(dn->dn_next_bonuslen[i]);
552 ASSERT0(dn->dn_next_bonustype[i]);
553 ASSERT0(dn->dn_rm_spillblk[i]);
554 ASSERT0(dn->dn_next_blksz[i]);
555 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
556 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
557 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
561 dnode_setdblksz(dn, blocksize);
562 dn->dn_indblkshift = ibs;
564 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
568 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
569 dn->dn_bonustype = bonustype;
570 dn->dn_bonuslen = bonuslen;
571 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
572 dn->dn_compress = ZIO_COMPRESS_INHERIT;
576 if (dn->dn_dirtyctx_firstset) {
577 kmem_free(dn->dn_dirtyctx_firstset, 1);
578 dn->dn_dirtyctx_firstset = NULL;
581 dn->dn_allocated_txg = tx->tx_txg;
584 dnode_setdirty(dn, tx);
585 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
586 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
587 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
588 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
592 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
593 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
597 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
598 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
599 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
600 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
601 ASSERT(tx->tx_txg != 0);
602 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
603 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
604 (bonustype == DMU_OT_SA && bonuslen == 0));
605 ASSERT(DMU_OT_IS_VALID(bonustype));
606 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
608 /* clean up any unreferenced dbufs */
609 dnode_evict_dbufs(dn);
613 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
614 dnode_setdirty(dn, tx);
615 if (dn->dn_datablksz != blocksize) {
616 /* change blocksize */
617 ASSERT(dn->dn_maxblkid == 0 &&
618 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
619 dnode_block_freed(dn, 0)));
620 dnode_setdblksz(dn, blocksize);
621 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
623 if (dn->dn_bonuslen != bonuslen)
624 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
626 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
629 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
630 if (dn->dn_bonustype != bonustype)
631 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
632 if (dn->dn_nblkptr != nblkptr)
633 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
634 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
635 dbuf_rm_spill(dn, tx);
636 dnode_rm_spill(dn, tx);
638 rw_exit(&dn->dn_struct_rwlock);
643 /* change bonus size and type */
644 mutex_enter(&dn->dn_mtx);
645 dn->dn_bonustype = bonustype;
646 dn->dn_bonuslen = bonuslen;
647 dn->dn_nblkptr = nblkptr;
648 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
649 dn->dn_compress = ZIO_COMPRESS_INHERIT;
650 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
652 /* fix up the bonus db_size */
654 dn->dn_bonus->db.db_size =
655 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
656 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
659 dn->dn_allocated_txg = tx->tx_txg;
660 mutex_exit(&dn->dn_mtx);
665 uint64_t dms_dnode_invalid;
666 uint64_t dms_dnode_recheck1;
667 uint64_t dms_dnode_recheck2;
668 uint64_t dms_dnode_special;
669 uint64_t dms_dnode_handle;
670 uint64_t dms_dnode_rwlock;
671 uint64_t dms_dnode_active;
673 #endif /* DNODE_STATS */
676 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
680 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
681 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
682 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
683 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
686 ndn->dn_objset = odn->dn_objset;
687 ndn->dn_object = odn->dn_object;
688 ndn->dn_dbuf = odn->dn_dbuf;
689 ndn->dn_handle = odn->dn_handle;
690 ndn->dn_phys = odn->dn_phys;
691 ndn->dn_type = odn->dn_type;
692 ndn->dn_bonuslen = odn->dn_bonuslen;
693 ndn->dn_bonustype = odn->dn_bonustype;
694 ndn->dn_nblkptr = odn->dn_nblkptr;
695 ndn->dn_checksum = odn->dn_checksum;
696 ndn->dn_compress = odn->dn_compress;
697 ndn->dn_nlevels = odn->dn_nlevels;
698 ndn->dn_indblkshift = odn->dn_indblkshift;
699 ndn->dn_datablkshift = odn->dn_datablkshift;
700 ndn->dn_datablkszsec = odn->dn_datablkszsec;
701 ndn->dn_datablksz = odn->dn_datablksz;
702 ndn->dn_maxblkid = odn->dn_maxblkid;
703 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
704 sizeof (odn->dn_next_nblkptr));
705 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
706 sizeof (odn->dn_next_nlevels));
707 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
708 sizeof (odn->dn_next_indblkshift));
709 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
710 sizeof (odn->dn_next_bonustype));
711 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
712 sizeof (odn->dn_rm_spillblk));
713 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
714 sizeof (odn->dn_next_bonuslen));
715 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
716 sizeof (odn->dn_next_blksz));
717 for (i = 0; i < TXG_SIZE; i++) {
718 list_move_tail(&ndn->dn_dirty_records[i],
719 &odn->dn_dirty_records[i]);
721 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
722 sizeof (odn->dn_free_ranges));
723 ndn->dn_allocated_txg = odn->dn_allocated_txg;
724 ndn->dn_free_txg = odn->dn_free_txg;
725 ndn->dn_assigned_txg = odn->dn_assigned_txg;
726 ndn->dn_dirtyctx = odn->dn_dirtyctx;
727 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
728 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
729 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
730 ASSERT(avl_is_empty(&ndn->dn_dbufs));
731 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
732 ndn->dn_dbufs_count = odn->dn_dbufs_count;
733 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
734 ndn->dn_bonus = odn->dn_bonus;
735 ndn->dn_have_spill = odn->dn_have_spill;
736 ndn->dn_zio = odn->dn_zio;
737 ndn->dn_oldused = odn->dn_oldused;
738 ndn->dn_oldflags = odn->dn_oldflags;
739 ndn->dn_olduid = odn->dn_olduid;
740 ndn->dn_oldgid = odn->dn_oldgid;
741 ndn->dn_newuid = odn->dn_newuid;
742 ndn->dn_newgid = odn->dn_newgid;
743 ndn->dn_id_flags = odn->dn_id_flags;
744 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
745 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
746 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
747 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
748 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
751 * Update back pointers. Updating the handle fixes the back pointer of
752 * every descendant dbuf as well as the bonus dbuf.
754 ASSERT(ndn->dn_handle->dnh_dnode == odn);
755 ndn->dn_handle->dnh_dnode = ndn;
756 if (ndn->dn_zfetch.zf_dnode == odn) {
757 ndn->dn_zfetch.zf_dnode = ndn;
761 * Invalidate the original dnode by clearing all of its back pointers.
764 odn->dn_handle = NULL;
765 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
766 offsetof(dmu_buf_impl_t, db_link));
767 odn->dn_dbufs_count = 0;
768 odn->dn_unlisted_l0_blkid = 0;
769 odn->dn_bonus = NULL;
770 odn->dn_zfetch.zf_dnode = NULL;
773 * Set the low bit of the objset pointer to ensure that dnode_move()
774 * recognizes the dnode as invalid in any subsequent callback.
776 POINTER_INVALIDATE(&odn->dn_objset);
779 * Satisfy the destructor.
781 for (i = 0; i < TXG_SIZE; i++) {
782 list_create(&odn->dn_dirty_records[i],
783 sizeof (dbuf_dirty_record_t),
784 offsetof(dbuf_dirty_record_t, dr_dirty_node));
785 odn->dn_free_ranges[i] = NULL;
786 odn->dn_next_nlevels[i] = 0;
787 odn->dn_next_indblkshift[i] = 0;
788 odn->dn_next_bonustype[i] = 0;
789 odn->dn_rm_spillblk[i] = 0;
790 odn->dn_next_bonuslen[i] = 0;
791 odn->dn_next_blksz[i] = 0;
793 odn->dn_allocated_txg = 0;
794 odn->dn_free_txg = 0;
795 odn->dn_assigned_txg = 0;
796 odn->dn_dirtyctx = 0;
797 odn->dn_dirtyctx_firstset = NULL;
798 odn->dn_have_spill = B_FALSE;
801 odn->dn_oldflags = 0;
806 odn->dn_id_flags = 0;
812 odn->dn_moved = (uint8_t)-1;
819 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
821 dnode_t *odn = buf, *ndn = newbuf;
827 * The dnode is on the objset's list of known dnodes if the objset
828 * pointer is valid. We set the low bit of the objset pointer when
829 * freeing the dnode to invalidate it, and the memory patterns written
830 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
831 * A newly created dnode sets the objset pointer last of all to indicate
832 * that the dnode is known and in a valid state to be moved by this
836 if (!POINTER_IS_VALID(os)) {
837 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
838 return (KMEM_CBRC_DONT_KNOW);
842 * Ensure that the objset does not go away during the move.
844 rw_enter(&os_lock, RW_WRITER);
845 if (os != odn->dn_objset) {
847 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
848 return (KMEM_CBRC_DONT_KNOW);
852 * If the dnode is still valid, then so is the objset. We know that no
853 * valid objset can be freed while we hold os_lock, so we can safely
854 * ensure that the objset remains in use.
856 mutex_enter(&os->os_lock);
859 * Recheck the objset pointer in case the dnode was removed just before
860 * acquiring the lock.
862 if (os != odn->dn_objset) {
863 mutex_exit(&os->os_lock);
865 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
866 return (KMEM_CBRC_DONT_KNOW);
870 * At this point we know that as long as we hold os->os_lock, the dnode
871 * cannot be freed and fields within the dnode can be safely accessed.
872 * The objset listing this dnode cannot go away as long as this dnode is
876 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
877 mutex_exit(&os->os_lock);
878 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
879 return (KMEM_CBRC_NO);
881 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
884 * Lock the dnode handle to prevent the dnode from obtaining any new
885 * holds. This also prevents the descendant dbufs and the bonus dbuf
886 * from accessing the dnode, so that we can discount their holds. The
887 * handle is safe to access because we know that while the dnode cannot
888 * go away, neither can its handle. Once we hold dnh_zrlock, we can
889 * safely move any dnode referenced only by dbufs.
891 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
892 mutex_exit(&os->os_lock);
893 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
894 return (KMEM_CBRC_LATER);
898 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
899 * We need to guarantee that there is a hold for every dbuf in order to
900 * determine whether the dnode is actively referenced. Falsely matching
901 * a dbuf to an active hold would lead to an unsafe move. It's possible
902 * that a thread already having an active dnode hold is about to add a
903 * dbuf, and we can't compare hold and dbuf counts while the add is in
906 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
907 zrl_exit(&odn->dn_handle->dnh_zrlock);
908 mutex_exit(&os->os_lock);
909 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
910 return (KMEM_CBRC_LATER);
914 * A dbuf may be removed (evicted) without an active dnode hold. In that
915 * case, the dbuf count is decremented under the handle lock before the
916 * dbuf's hold is released. This order ensures that if we count the hold
917 * after the dbuf is removed but before its hold is released, we will
918 * treat the unmatched hold as active and exit safely. If we count the
919 * hold before the dbuf is removed, the hold is discounted, and the
920 * removal is blocked until the move completes.
922 refcount = refcount_count(&odn->dn_holds);
923 ASSERT(refcount >= 0);
924 dbufs = odn->dn_dbufs_count;
926 /* We can't have more dbufs than dnode holds. */
927 ASSERT3U(dbufs, <=, refcount);
928 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
931 if (refcount > dbufs) {
932 rw_exit(&odn->dn_struct_rwlock);
933 zrl_exit(&odn->dn_handle->dnh_zrlock);
934 mutex_exit(&os->os_lock);
935 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
936 return (KMEM_CBRC_LATER);
939 rw_exit(&odn->dn_struct_rwlock);
942 * At this point we know that anyone with a hold on the dnode is not
943 * actively referencing it. The dnode is known and in a valid state to
944 * move. We're holding the locks needed to execute the critical section.
946 dnode_move_impl(odn, ndn);
948 list_link_replace(&odn->dn_link, &ndn->dn_link);
949 /* If the dnode was safe to move, the refcount cannot have changed. */
950 ASSERT(refcount == refcount_count(&ndn->dn_holds));
951 ASSERT(dbufs == ndn->dn_dbufs_count);
952 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
953 mutex_exit(&os->os_lock);
955 return (KMEM_CBRC_YES);
961 dnode_special_close(dnode_handle_t *dnh)
963 dnode_t *dn = dnh->dnh_dnode;
966 * Wait for final references to the dnode to clear. This can
967 * only happen if the arc is asyncronously evicting state that
968 * has a hold on this dnode while we are trying to evict this
971 while (refcount_count(&dn->dn_holds) > 0)
973 zrl_add(&dnh->dnh_zrlock);
974 dnode_destroy(dn); /* implicit zrl_remove() */
975 zrl_destroy(&dnh->dnh_zrlock);
976 dnh->dnh_dnode = NULL;
980 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
983 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
985 zrl_init(&dnh->dnh_zrlock);
991 dnode_buf_pageout(dmu_buf_t *db, void *arg)
993 dnode_children_t *children_dnodes = arg;
995 int epb = db->db_size >> DNODE_SHIFT;
997 ASSERT(epb == children_dnodes->dnc_count);
999 for (i = 0; i < epb; i++) {
1000 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1004 * The dnode handle lock guards against the dnode moving to
1005 * another valid address, so there is no need here to guard
1006 * against changes to or from NULL.
1008 if (dnh->dnh_dnode == NULL) {
1009 zrl_destroy(&dnh->dnh_zrlock);
1013 zrl_add(&dnh->dnh_zrlock);
1014 dn = dnh->dnh_dnode;
1016 * If there are holds on this dnode, then there should
1017 * be holds on the dnode's containing dbuf as well; thus
1018 * it wouldn't be eligible for eviction and this function
1019 * would not have been called.
1021 ASSERT(refcount_is_zero(&dn->dn_holds));
1022 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1024 dnode_destroy(dn); /* implicit zrl_remove() */
1025 zrl_destroy(&dnh->dnh_zrlock);
1026 dnh->dnh_dnode = NULL;
1028 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1029 epb * sizeof (dnode_handle_t));
1034 * EINVAL - invalid object number.
1036 * succeeds even for free dnodes.
1039 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1040 void *tag, dnode_t **dnp)
1043 int drop_struct_lock = FALSE;
1048 dnode_children_t *children_dnodes;
1049 dnode_handle_t *dnh;
1052 * If you are holding the spa config lock as writer, you shouldn't
1053 * be asking the DMU to do *anything* unless it's the root pool
1054 * which may require us to read from the root filesystem while
1055 * holding some (not all) of the locks as writer.
1057 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1058 (spa_is_root(os->os_spa) &&
1059 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1061 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1062 dn = (object == DMU_USERUSED_OBJECT) ?
1063 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1065 return (SET_ERROR(ENOENT));
1067 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1068 return (SET_ERROR(ENOENT));
1069 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1070 return (SET_ERROR(EEXIST));
1072 (void) refcount_add(&dn->dn_holds, tag);
1077 if (object == 0 || object >= DN_MAX_OBJECT)
1078 return (SET_ERROR(EINVAL));
1080 mdn = DMU_META_DNODE(os);
1081 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1085 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1086 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1087 drop_struct_lock = TRUE;
1090 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1092 db = dbuf_hold(mdn, blk, FTAG);
1093 if (drop_struct_lock)
1094 rw_exit(&mdn->dn_struct_rwlock);
1096 return (SET_ERROR(EIO));
1097 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1099 dbuf_rele(db, FTAG);
1103 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1104 epb = db->db.db_size >> DNODE_SHIFT;
1106 idx = object & (epb-1);
1108 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1109 children_dnodes = dmu_buf_get_user(&db->db);
1110 if (children_dnodes == NULL) {
1112 dnode_children_t *winner;
1113 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1114 epb * sizeof (dnode_handle_t), KM_SLEEP);
1115 children_dnodes->dnc_count = epb;
1116 dnh = &children_dnodes->dnc_children[0];
1117 for (i = 0; i < epb; i++) {
1118 zrl_init(&dnh[i].dnh_zrlock);
1119 dnh[i].dnh_dnode = NULL;
1121 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1122 dnode_buf_pageout)) {
1124 for (i = 0; i < epb; i++) {
1125 zrl_destroy(&dnh[i].dnh_zrlock);
1128 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1129 epb * sizeof (dnode_handle_t));
1130 children_dnodes = winner;
1133 ASSERT(children_dnodes->dnc_count == epb);
1135 dnh = &children_dnodes->dnc_children[idx];
1136 zrl_add(&dnh->dnh_zrlock);
1137 if ((dn = dnh->dnh_dnode) == NULL) {
1138 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1141 dn = dnode_create(os, phys, db, object, dnh);
1142 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1143 if (winner != NULL) {
1144 zrl_add(&dnh->dnh_zrlock);
1145 dnode_destroy(dn); /* implicit zrl_remove() */
1150 mutex_enter(&dn->dn_mtx);
1152 if (dn->dn_free_txg ||
1153 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1154 ((flag & DNODE_MUST_BE_FREE) &&
1155 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1156 mutex_exit(&dn->dn_mtx);
1157 zrl_remove(&dnh->dnh_zrlock);
1158 dbuf_rele(db, FTAG);
1159 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1161 mutex_exit(&dn->dn_mtx);
1163 if (refcount_add(&dn->dn_holds, tag) == 1)
1164 dbuf_add_ref(db, dnh);
1165 /* Now we can rely on the hold to prevent the dnode from moving. */
1166 zrl_remove(&dnh->dnh_zrlock);
1169 ASSERT3P(dn->dn_dbuf, ==, db);
1170 ASSERT3U(dn->dn_object, ==, object);
1171 dbuf_rele(db, FTAG);
1178 * Return held dnode if the object is allocated, NULL if not.
1181 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1183 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1187 * Can only add a reference if there is already at least one
1188 * reference on the dnode. Returns FALSE if unable to add a
1192 dnode_add_ref(dnode_t *dn, void *tag)
1194 mutex_enter(&dn->dn_mtx);
1195 if (refcount_is_zero(&dn->dn_holds)) {
1196 mutex_exit(&dn->dn_mtx);
1199 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1200 mutex_exit(&dn->dn_mtx);
1205 dnode_rele(dnode_t *dn, void *tag)
1208 /* Get while the hold prevents the dnode from moving. */
1209 dmu_buf_impl_t *db = dn->dn_dbuf;
1210 dnode_handle_t *dnh = dn->dn_handle;
1212 mutex_enter(&dn->dn_mtx);
1213 refs = refcount_remove(&dn->dn_holds, tag);
1214 mutex_exit(&dn->dn_mtx);
1217 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1218 * indirectly by dbuf_rele() while relying on the dnode handle to
1219 * prevent the dnode from moving, since releasing the last hold could
1220 * result in the dnode's parent dbuf evicting its dnode handles. For
1221 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1222 * other direct or indirect hold on the dnode must first drop the dnode
1225 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1227 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1228 if (refs == 0 && db != NULL) {
1230 * Another thread could add a hold to the dnode handle in
1231 * dnode_hold_impl() while holding the parent dbuf. Since the
1232 * hold on the parent dbuf prevents the handle from being
1233 * destroyed, the hold on the handle is OK. We can't yet assert
1234 * that the handle has zero references, but that will be
1235 * asserted anyway when the handle gets destroyed.
1242 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1244 objset_t *os = dn->dn_objset;
1245 uint64_t txg = tx->tx_txg;
1247 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1248 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1255 mutex_enter(&dn->dn_mtx);
1256 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1257 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1258 mutex_exit(&dn->dn_mtx);
1262 * Determine old uid/gid when necessary
1264 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1266 mutex_enter(&os->os_lock);
1269 * If we are already marked dirty, we're done.
1271 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1272 mutex_exit(&os->os_lock);
1276 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1277 !avl_is_empty(&dn->dn_dbufs));
1278 ASSERT(dn->dn_datablksz != 0);
1279 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1280 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1281 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1283 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1284 dn->dn_object, txg);
1286 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1287 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1289 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1292 mutex_exit(&os->os_lock);
1295 * The dnode maintains a hold on its containing dbuf as
1296 * long as there are holds on it. Each instantiated child
1297 * dbuf maintains a hold on the dnode. When the last child
1298 * drops its hold, the dnode will drop its hold on the
1299 * containing dbuf. We add a "dirty hold" here so that the
1300 * dnode will hang around after we finish processing its
1303 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1305 (void) dbuf_dirty(dn->dn_dbuf, tx);
1307 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1311 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1313 int txgoff = tx->tx_txg & TXG_MASK;
1315 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1317 /* we should be the only holder... hopefully */
1318 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1320 mutex_enter(&dn->dn_mtx);
1321 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1322 mutex_exit(&dn->dn_mtx);
1325 dn->dn_free_txg = tx->tx_txg;
1326 mutex_exit(&dn->dn_mtx);
1329 * If the dnode is already dirty, it needs to be moved from
1330 * the dirty list to the free list.
1332 mutex_enter(&dn->dn_objset->os_lock);
1333 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1334 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1335 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1336 mutex_exit(&dn->dn_objset->os_lock);
1338 mutex_exit(&dn->dn_objset->os_lock);
1339 dnode_setdirty(dn, tx);
1344 * Try to change the block size for the indicated dnode. This can only
1345 * succeed if there are no blocks allocated or dirty beyond first block
1348 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1354 size = SPA_MINBLOCKSIZE;
1355 if (size > SPA_MAXBLOCKSIZE)
1356 size = SPA_MAXBLOCKSIZE;
1358 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1360 if (ibs == dn->dn_indblkshift)
1363 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1366 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1368 /* Check for any allocated blocks beyond the first */
1369 if (dn->dn_maxblkid != 0)
1372 mutex_enter(&dn->dn_dbufs_mtx);
1373 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1374 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1375 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1376 db->db_blkid != DMU_SPILL_BLKID) {
1377 mutex_exit(&dn->dn_dbufs_mtx);
1381 mutex_exit(&dn->dn_dbufs_mtx);
1383 if (ibs && dn->dn_nlevels != 1)
1386 /* resize the old block */
1387 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1389 dbuf_new_size(db, size, tx);
1390 else if (err != ENOENT)
1393 dnode_setdblksz(dn, size);
1394 dnode_setdirty(dn, tx);
1395 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1397 dn->dn_indblkshift = ibs;
1398 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1400 /* rele after we have fixed the blocksize in the dnode */
1402 dbuf_rele(db, FTAG);
1404 rw_exit(&dn->dn_struct_rwlock);
1408 rw_exit(&dn->dn_struct_rwlock);
1409 return (SET_ERROR(ENOTSUP));
1412 /* read-holding callers must not rely on the lock being continuously held */
1414 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1416 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1417 int epbs, new_nlevels;
1420 ASSERT(blkid != DMU_BONUS_BLKID);
1423 RW_READ_HELD(&dn->dn_struct_rwlock) :
1424 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1427 * if we have a read-lock, check to see if we need to do any work
1428 * before upgrading to a write-lock.
1431 if (blkid <= dn->dn_maxblkid)
1434 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1435 rw_exit(&dn->dn_struct_rwlock);
1436 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1440 if (blkid <= dn->dn_maxblkid)
1443 dn->dn_maxblkid = blkid;
1446 * Compute the number of levels necessary to support the new maxblkid.
1449 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1450 for (sz = dn->dn_nblkptr;
1451 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1454 if (new_nlevels > dn->dn_nlevels) {
1455 int old_nlevels = dn->dn_nlevels;
1458 dbuf_dirty_record_t *new, *dr, *dr_next;
1460 dn->dn_nlevels = new_nlevels;
1462 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1463 dn->dn_next_nlevels[txgoff] = new_nlevels;
1465 /* dirty the left indirects */
1466 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1468 new = dbuf_dirty(db, tx);
1469 dbuf_rele(db, FTAG);
1471 /* transfer the dirty records to the new indirect */
1472 mutex_enter(&dn->dn_mtx);
1473 mutex_enter(&new->dt.di.dr_mtx);
1474 list = &dn->dn_dirty_records[txgoff];
1475 for (dr = list_head(list); dr; dr = dr_next) {
1476 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1477 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1478 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1479 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1480 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1481 list_remove(&dn->dn_dirty_records[txgoff], dr);
1482 list_insert_tail(&new->dt.di.dr_children, dr);
1483 dr->dr_parent = new;
1486 mutex_exit(&new->dt.di.dr_mtx);
1487 mutex_exit(&dn->dn_mtx);
1492 rw_downgrade(&dn->dn_struct_rwlock);
1496 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1499 uint64_t blkoff, blkid, nblks;
1500 int blksz, blkshift, head, tail;
1504 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1505 blksz = dn->dn_datablksz;
1506 blkshift = dn->dn_datablkshift;
1507 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1509 if (len == DMU_OBJECT_END) {
1510 len = UINT64_MAX - off;
1515 * First, block align the region to free:
1518 head = P2NPHASE(off, blksz);
1519 blkoff = P2PHASE(off, blksz);
1520 if ((off >> blkshift) > dn->dn_maxblkid)
1523 ASSERT(dn->dn_maxblkid == 0);
1524 if (off == 0 && len >= blksz) {
1526 * Freeing the whole block; fast-track this request.
1527 * Note that we won't dirty any indirect blocks,
1528 * which is fine because we will be freeing the entire
1529 * file and thus all indirect blocks will be freed
1530 * by free_children().
1535 } else if (off >= blksz) {
1536 /* Freeing past end-of-data */
1539 /* Freeing part of the block. */
1541 ASSERT3U(head, >, 0);
1545 /* zero out any partial block data at the start of the range */
1547 ASSERT3U(blkoff + head, ==, blksz);
1550 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1554 /* don't dirty if it isn't on disk and isn't dirty */
1555 if (db->db_last_dirty ||
1556 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1557 rw_exit(&dn->dn_struct_rwlock);
1558 dmu_buf_will_dirty(&db->db, tx);
1559 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1560 data = db->db.db_data;
1561 bzero(data + blkoff, head);
1563 dbuf_rele(db, FTAG);
1569 /* If the range was less than one block, we're done */
1573 /* If the remaining range is past end of file, we're done */
1574 if ((off >> blkshift) > dn->dn_maxblkid)
1577 ASSERT(ISP2(blksz));
1581 tail = P2PHASE(len, blksz);
1583 ASSERT0(P2PHASE(off, blksz));
1584 /* zero out any partial block data at the end of the range */
1588 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1589 TRUE, FTAG, &db) == 0) {
1590 /* don't dirty if not on disk and not dirty */
1591 if (db->db_last_dirty ||
1592 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1593 rw_exit(&dn->dn_struct_rwlock);
1594 dmu_buf_will_dirty(&db->db, tx);
1595 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1596 bzero(db->db.db_data, tail);
1598 dbuf_rele(db, FTAG);
1603 /* If the range did not include a full block, we are done */
1607 ASSERT(IS_P2ALIGNED(off, blksz));
1608 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1609 blkid = off >> blkshift;
1610 nblks = len >> blkshift;
1615 * Dirty the first and last indirect blocks, as they (and/or their
1616 * parents) will need to be written out if they were only
1617 * partially freed. Interior indirect blocks will be themselves freed,
1618 * by free_children(), so they need not be dirtied. Note that these
1619 * interior blocks have already been prefetched by dmu_tx_hold_free().
1621 if (dn->dn_nlevels > 1) {
1622 uint64_t first, last;
1624 first = blkid >> epbs;
1625 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1626 dmu_buf_will_dirty(&db->db, tx);
1627 dbuf_rele(db, FTAG);
1630 last = dn->dn_maxblkid >> epbs;
1632 last = (blkid + nblks - 1) >> epbs;
1633 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1634 dmu_buf_will_dirty(&db->db, tx);
1635 dbuf_rele(db, FTAG);
1641 * Add this range to the dnode range list.
1642 * We will finish up this free operation in the syncing phase.
1644 mutex_enter(&dn->dn_mtx);
1645 int txgoff = tx->tx_txg & TXG_MASK;
1646 if (dn->dn_free_ranges[txgoff] == NULL) {
1647 dn->dn_free_ranges[txgoff] =
1648 range_tree_create(NULL, NULL, &dn->dn_mtx);
1650 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1651 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1652 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1653 blkid, nblks, tx->tx_txg);
1654 mutex_exit(&dn->dn_mtx);
1656 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1657 dnode_setdirty(dn, tx);
1660 rw_exit(&dn->dn_struct_rwlock);
1664 dnode_spill_freed(dnode_t *dn)
1668 mutex_enter(&dn->dn_mtx);
1669 for (i = 0; i < TXG_SIZE; i++) {
1670 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1673 mutex_exit(&dn->dn_mtx);
1674 return (i < TXG_SIZE);
1677 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1679 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1681 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1684 if (blkid == DMU_BONUS_BLKID)
1688 * If we're in the process of opening the pool, dp will not be
1689 * set yet, but there shouldn't be anything dirty.
1694 if (dn->dn_free_txg)
1697 if (blkid == DMU_SPILL_BLKID)
1698 return (dnode_spill_freed(dn));
1700 mutex_enter(&dn->dn_mtx);
1701 for (i = 0; i < TXG_SIZE; i++) {
1702 if (dn->dn_free_ranges[i] != NULL &&
1703 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1706 mutex_exit(&dn->dn_mtx);
1707 return (i < TXG_SIZE);
1710 /* call from syncing context when we actually write/free space for this dnode */
1712 dnode_diduse_space(dnode_t *dn, int64_t delta)
1715 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1717 (u_longlong_t)dn->dn_phys->dn_used,
1720 mutex_enter(&dn->dn_mtx);
1721 space = DN_USED_BYTES(dn->dn_phys);
1723 ASSERT3U(space + delta, >=, space); /* no overflow */
1725 ASSERT3U(space, >=, -delta); /* no underflow */
1728 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1729 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1730 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1731 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1733 dn->dn_phys->dn_used = space;
1734 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1736 mutex_exit(&dn->dn_mtx);
1740 * Call when we think we're going to write/free space in open context to track
1741 * the amount of memory in use by the currently open txg.
1744 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1746 objset_t *os = dn->dn_objset;
1747 dsl_dataset_t *ds = os->os_dsl_dataset;
1748 int64_t aspace = spa_get_asize(os->os_spa, space);
1751 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1752 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1755 dmu_tx_willuse_space(tx, aspace);
1759 * Scans a block at the indicated "level" looking for a hole or data,
1760 * depending on 'flags'.
1762 * If level > 0, then we are scanning an indirect block looking at its
1763 * pointers. If level == 0, then we are looking at a block of dnodes.
1765 * If we don't find what we are looking for in the block, we return ESRCH.
1766 * Otherwise, return with *offset pointing to the beginning (if searching
1767 * forwards) or end (if searching backwards) of the range covered by the
1768 * block pointer we matched on (or dnode).
1770 * The basic search algorithm used below by dnode_next_offset() is to
1771 * use this function to search up the block tree (widen the search) until
1772 * we find something (i.e., we don't return ESRCH) and then search back
1773 * down the tree (narrow the search) until we reach our original search
1777 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1778 int lvl, uint64_t blkfill, uint64_t txg)
1780 dmu_buf_impl_t *db = NULL;
1782 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1783 uint64_t epb = 1ULL << epbs;
1784 uint64_t minfill, maxfill;
1786 int i, inc, error, span;
1788 dprintf("probing object %llu offset %llx level %d of %u\n",
1789 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1791 hole = ((flags & DNODE_FIND_HOLE) != 0);
1792 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1793 ASSERT(txg == 0 || !hole);
1795 if (lvl == dn->dn_phys->dn_nlevels) {
1797 epb = dn->dn_phys->dn_nblkptr;
1798 data = dn->dn_phys->dn_blkptr;
1800 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1801 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1803 if (error != ENOENT)
1808 * This can only happen when we are searching up
1809 * the block tree for data. We don't really need to
1810 * adjust the offset, as we will just end up looking
1811 * at the pointer to this block in its parent, and its
1812 * going to be unallocated, so we will skip over it.
1814 return (SET_ERROR(ESRCH));
1816 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1818 dbuf_rele(db, FTAG);
1821 data = db->db.db_data;
1825 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1826 db->db_blkptr->blk_birth <= txg ||
1827 BP_IS_HOLE(db->db_blkptr))) {
1829 * This can only happen when we are searching up the tree
1830 * and these conditions mean that we need to keep climbing.
1832 error = SET_ERROR(ESRCH);
1833 } else if (lvl == 0) {
1834 dnode_phys_t *dnp = data;
1836 ASSERT(dn->dn_type == DMU_OT_DNODE);
1838 for (i = (*offset >> span) & (blkfill - 1);
1839 i >= 0 && i < blkfill; i += inc) {
1840 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1842 *offset += (1ULL << span) * inc;
1844 if (i < 0 || i == blkfill)
1845 error = SET_ERROR(ESRCH);
1847 blkptr_t *bp = data;
1848 uint64_t start = *offset;
1849 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1851 maxfill = blkfill << ((lvl - 1) * epbs);
1858 *offset = *offset >> span;
1859 for (i = BF64_GET(*offset, 0, epbs);
1860 i >= 0 && i < epb; i += inc) {
1861 if (BP_GET_FILL(&bp[i]) >= minfill &&
1862 BP_GET_FILL(&bp[i]) <= maxfill &&
1863 (hole || bp[i].blk_birth > txg))
1865 if (inc > 0 || *offset > 0)
1868 *offset = *offset << span;
1870 /* traversing backwards; position offset at the end */
1871 ASSERT3U(*offset, <=, start);
1872 *offset = MIN(*offset + (1ULL << span) - 1, start);
1873 } else if (*offset < start) {
1876 if (i < 0 || i >= epb)
1877 error = SET_ERROR(ESRCH);
1881 dbuf_rele(db, FTAG);
1887 * Find the next hole, data, or sparse region at or after *offset.
1888 * The value 'blkfill' tells us how many items we expect to find
1889 * in an L0 data block; this value is 1 for normal objects,
1890 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1891 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1895 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1896 * Finds the next/previous hole/data in a file.
1897 * Used in dmu_offset_next().
1899 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1900 * Finds the next free/allocated dnode an objset's meta-dnode.
1901 * Only finds objects that have new contents since txg (ie.
1902 * bonus buffer changes and content removal are ignored).
1903 * Used in dmu_object_next().
1905 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1906 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1907 * Used in dmu_object_alloc().
1910 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1911 int minlvl, uint64_t blkfill, uint64_t txg)
1913 uint64_t initial_offset = *offset;
1917 if (!(flags & DNODE_FIND_HAVELOCK))
1918 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1920 if (dn->dn_phys->dn_nlevels == 0) {
1921 error = SET_ERROR(ESRCH);
1925 if (dn->dn_datablkshift == 0) {
1926 if (*offset < dn->dn_datablksz) {
1927 if (flags & DNODE_FIND_HOLE)
1928 *offset = dn->dn_datablksz;
1930 error = SET_ERROR(ESRCH);
1935 maxlvl = dn->dn_phys->dn_nlevels;
1937 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1938 error = dnode_next_offset_level(dn,
1939 flags, offset, lvl, blkfill, txg);
1944 while (error == 0 && --lvl >= minlvl) {
1945 error = dnode_next_offset_level(dn,
1946 flags, offset, lvl, blkfill, txg);
1949 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1950 initial_offset < *offset : initial_offset > *offset))
1951 error = SET_ERROR(ESRCH);
1953 if (!(flags & DNODE_FIND_HAVELOCK))
1954 rw_exit(&dn->dn_struct_rwlock);