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) 2013 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>
39 static int free_range_compar(const void *node1, const void *node2);
41 static kmem_cache_t *dnode_cache;
43 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
44 * turned on when DEBUG is also defined.
51 #define DNODE_STAT_ADD(stat) ((stat)++)
53 #define DNODE_STAT_ADD(stat) /* nothing */
54 #endif /* DNODE_STATS */
56 static dnode_phys_t dnode_phys_zero;
58 int zfs_default_bs = SPA_MINBLOCKSHIFT;
59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
62 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
67 dnode_cons(void *arg, void *unused, int kmflag)
72 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
73 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
74 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
75 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
78 * Every dbuf has a reference, and dropping a tracked reference is
79 * O(number of references), so don't track dn_holds.
81 refcount_create_untracked(&dn->dn_holds);
82 refcount_create(&dn->dn_tx_holds);
83 list_link_init(&dn->dn_link);
85 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
86 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
87 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
88 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
89 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
90 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
91 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
93 for (i = 0; i < TXG_SIZE; i++) {
94 list_link_init(&dn->dn_dirty_link[i]);
95 avl_create(&dn->dn_ranges[i], free_range_compar,
96 sizeof (free_range_t),
97 offsetof(struct free_range, fr_node));
98 list_create(&dn->dn_dirty_records[i],
99 sizeof (dbuf_dirty_record_t),
100 offsetof(dbuf_dirty_record_t, dr_dirty_node));
103 dn->dn_allocated_txg = 0;
105 dn->dn_assigned_txg = 0;
107 dn->dn_dirtyctx_firstset = NULL;
109 dn->dn_have_spill = B_FALSE;
119 dn->dn_dbufs_count = 0;
120 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
121 offsetof(dmu_buf_impl_t, db_link));
124 POINTER_INVALIDATE(&dn->dn_objset);
130 dnode_dest(void *arg, void *unused)
135 rw_destroy(&dn->dn_struct_rwlock);
136 mutex_destroy(&dn->dn_mtx);
137 mutex_destroy(&dn->dn_dbufs_mtx);
138 cv_destroy(&dn->dn_notxholds);
139 refcount_destroy(&dn->dn_holds);
140 refcount_destroy(&dn->dn_tx_holds);
141 ASSERT(!list_link_active(&dn->dn_link));
143 for (i = 0; i < TXG_SIZE; i++) {
144 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
145 avl_destroy(&dn->dn_ranges[i]);
146 list_destroy(&dn->dn_dirty_records[i]);
147 ASSERT0(dn->dn_next_nblkptr[i]);
148 ASSERT0(dn->dn_next_nlevels[i]);
149 ASSERT0(dn->dn_next_indblkshift[i]);
150 ASSERT0(dn->dn_next_bonustype[i]);
151 ASSERT0(dn->dn_rm_spillblk[i]);
152 ASSERT0(dn->dn_next_bonuslen[i]);
153 ASSERT0(dn->dn_next_blksz[i]);
156 ASSERT0(dn->dn_allocated_txg);
157 ASSERT0(dn->dn_free_txg);
158 ASSERT0(dn->dn_assigned_txg);
159 ASSERT0(dn->dn_dirtyctx);
160 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
161 ASSERT3P(dn->dn_bonus, ==, NULL);
162 ASSERT(!dn->dn_have_spill);
163 ASSERT3P(dn->dn_zio, ==, NULL);
164 ASSERT0(dn->dn_oldused);
165 ASSERT0(dn->dn_oldflags);
166 ASSERT0(dn->dn_olduid);
167 ASSERT0(dn->dn_oldgid);
168 ASSERT0(dn->dn_newuid);
169 ASSERT0(dn->dn_newgid);
170 ASSERT0(dn->dn_id_flags);
172 ASSERT0(dn->dn_dbufs_count);
173 list_destroy(&dn->dn_dbufs);
179 ASSERT(dnode_cache == NULL);
180 dnode_cache = kmem_cache_create("dnode_t",
182 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
183 kmem_cache_set_move(dnode_cache, dnode_move);
189 kmem_cache_destroy(dnode_cache);
196 dnode_verify(dnode_t *dn)
198 int drop_struct_lock = FALSE;
201 ASSERT(dn->dn_objset);
202 ASSERT(dn->dn_handle->dnh_dnode == dn);
204 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
206 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
209 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
210 rw_enter(&dn->dn_struct_rwlock, RW_READER);
211 drop_struct_lock = TRUE;
213 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
215 ASSERT3U(dn->dn_indblkshift, >=, 0);
216 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
217 if (dn->dn_datablkshift) {
218 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
219 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
220 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
222 ASSERT3U(dn->dn_nlevels, <=, 30);
223 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
224 ASSERT3U(dn->dn_nblkptr, >=, 1);
225 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
226 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
227 ASSERT3U(dn->dn_datablksz, ==,
228 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
229 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
230 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
231 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
232 for (i = 0; i < TXG_SIZE; i++) {
233 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
236 if (dn->dn_phys->dn_type != DMU_OT_NONE)
237 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
238 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
239 if (dn->dn_dbuf != NULL) {
240 ASSERT3P(dn->dn_phys, ==,
241 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
242 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
244 if (drop_struct_lock)
245 rw_exit(&dn->dn_struct_rwlock);
250 dnode_byteswap(dnode_phys_t *dnp)
252 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
255 if (dnp->dn_type == DMU_OT_NONE) {
256 bzero(dnp, sizeof (dnode_phys_t));
260 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
261 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
262 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
263 dnp->dn_used = BSWAP_64(dnp->dn_used);
266 * dn_nblkptr is only one byte, so it's OK to read it in either
267 * byte order. We can't read dn_bouslen.
269 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
270 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
271 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
272 buf64[i] = BSWAP_64(buf64[i]);
275 * OK to check dn_bonuslen for zero, because it won't matter if
276 * we have the wrong byte order. This is necessary because the
277 * dnode dnode is smaller than a regular dnode.
279 if (dnp->dn_bonuslen != 0) {
281 * Note that the bonus length calculated here may be
282 * longer than the actual bonus buffer. This is because
283 * we always put the bonus buffer after the last block
284 * pointer (instead of packing it against the end of the
287 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
288 size_t len = DN_MAX_BONUSLEN - off;
289 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
290 dmu_object_byteswap_t byteswap =
291 DMU_OT_BYTESWAP(dnp->dn_bonustype);
292 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
295 /* Swap SPILL block if we have one */
296 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
297 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
302 dnode_buf_byteswap(void *vbuf, size_t size)
304 dnode_phys_t *buf = vbuf;
307 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
308 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
310 size >>= DNODE_SHIFT;
311 for (i = 0; i < size; i++) {
318 free_range_compar(const void *node1, const void *node2)
320 const free_range_t *rp1 = node1;
321 const free_range_t *rp2 = node2;
323 if (rp1->fr_blkid < rp2->fr_blkid)
325 else if (rp1->fr_blkid > rp2->fr_blkid)
331 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
333 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
335 dnode_setdirty(dn, tx);
336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
338 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
339 dn->dn_bonuslen = newsize;
341 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
343 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
344 rw_exit(&dn->dn_struct_rwlock);
348 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
350 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
351 dnode_setdirty(dn, tx);
352 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
353 dn->dn_bonustype = newtype;
354 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
355 rw_exit(&dn->dn_struct_rwlock);
359 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
361 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
362 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
363 dnode_setdirty(dn, tx);
364 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
365 dn->dn_have_spill = B_FALSE;
369 dnode_setdblksz(dnode_t *dn, int size)
371 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
372 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
373 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
374 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
375 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
376 dn->dn_datablksz = size;
377 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
378 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
382 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
383 uint64_t object, dnode_handle_t *dnh)
385 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
387 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
391 * Defer setting dn_objset until the dnode is ready to be a candidate
392 * for the dnode_move() callback.
394 dn->dn_object = object;
399 if (dnp->dn_datablkszsec) {
400 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
402 dn->dn_datablksz = 0;
403 dn->dn_datablkszsec = 0;
404 dn->dn_datablkshift = 0;
406 dn->dn_indblkshift = dnp->dn_indblkshift;
407 dn->dn_nlevels = dnp->dn_nlevels;
408 dn->dn_type = dnp->dn_type;
409 dn->dn_nblkptr = dnp->dn_nblkptr;
410 dn->dn_checksum = dnp->dn_checksum;
411 dn->dn_compress = dnp->dn_compress;
412 dn->dn_bonustype = dnp->dn_bonustype;
413 dn->dn_bonuslen = dnp->dn_bonuslen;
414 dn->dn_maxblkid = dnp->dn_maxblkid;
415 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
418 dmu_zfetch_init(&dn->dn_zfetch, dn);
420 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
422 mutex_enter(&os->os_lock);
423 list_insert_head(&os->os_dnodes, dn);
426 * Everything else must be valid before assigning dn_objset makes the
427 * dnode eligible for dnode_move().
430 mutex_exit(&os->os_lock);
432 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
437 * Caller must be holding the dnode handle, which is released upon return.
440 dnode_destroy(dnode_t *dn)
442 objset_t *os = dn->dn_objset;
444 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
446 mutex_enter(&os->os_lock);
447 POINTER_INVALIDATE(&dn->dn_objset);
448 list_remove(&os->os_dnodes, dn);
449 mutex_exit(&os->os_lock);
451 /* the dnode can no longer move, so we can release the handle */
452 zrl_remove(&dn->dn_handle->dnh_zrlock);
454 dn->dn_allocated_txg = 0;
456 dn->dn_assigned_txg = 0;
459 if (dn->dn_dirtyctx_firstset != NULL) {
460 kmem_free(dn->dn_dirtyctx_firstset, 1);
461 dn->dn_dirtyctx_firstset = NULL;
463 if (dn->dn_bonus != NULL) {
464 mutex_enter(&dn->dn_bonus->db_mtx);
465 dbuf_evict(dn->dn_bonus);
470 dn->dn_have_spill = B_FALSE;
479 dmu_zfetch_rele(&dn->dn_zfetch);
480 kmem_cache_free(dnode_cache, dn);
481 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
485 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
486 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
491 blocksize = 1 << zfs_default_bs;
492 else if (blocksize > SPA_MAXBLOCKSIZE)
493 blocksize = SPA_MAXBLOCKSIZE;
495 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
498 ibs = zfs_default_ibs;
500 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
502 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
503 dn->dn_object, tx->tx_txg, blocksize, ibs);
505 ASSERT(dn->dn_type == DMU_OT_NONE);
506 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
507 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
508 ASSERT(ot != DMU_OT_NONE);
509 ASSERT(DMU_OT_IS_VALID(ot));
510 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
511 (bonustype == DMU_OT_SA && bonuslen == 0) ||
512 (bonustype != DMU_OT_NONE && bonuslen != 0));
513 ASSERT(DMU_OT_IS_VALID(bonustype));
514 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
515 ASSERT(dn->dn_type == DMU_OT_NONE);
516 ASSERT0(dn->dn_maxblkid);
517 ASSERT0(dn->dn_allocated_txg);
518 ASSERT0(dn->dn_assigned_txg);
519 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
520 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
521 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
523 for (i = 0; i < TXG_SIZE; i++) {
524 ASSERT0(dn->dn_next_nblkptr[i]);
525 ASSERT0(dn->dn_next_nlevels[i]);
526 ASSERT0(dn->dn_next_indblkshift[i]);
527 ASSERT0(dn->dn_next_bonuslen[i]);
528 ASSERT0(dn->dn_next_bonustype[i]);
529 ASSERT0(dn->dn_rm_spillblk[i]);
530 ASSERT0(dn->dn_next_blksz[i]);
531 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
532 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
533 ASSERT0(avl_numnodes(&dn->dn_ranges[i]));
537 dnode_setdblksz(dn, blocksize);
538 dn->dn_indblkshift = ibs;
540 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
544 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
545 dn->dn_bonustype = bonustype;
546 dn->dn_bonuslen = bonuslen;
547 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
548 dn->dn_compress = ZIO_COMPRESS_INHERIT;
552 if (dn->dn_dirtyctx_firstset) {
553 kmem_free(dn->dn_dirtyctx_firstset, 1);
554 dn->dn_dirtyctx_firstset = NULL;
557 dn->dn_allocated_txg = tx->tx_txg;
560 dnode_setdirty(dn, tx);
561 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
562 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
563 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
564 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
568 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
569 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
573 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
574 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
575 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
576 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
577 ASSERT(tx->tx_txg != 0);
578 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
579 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
580 (bonustype == DMU_OT_SA && bonuslen == 0));
581 ASSERT(DMU_OT_IS_VALID(bonustype));
582 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
584 /* clean up any unreferenced dbufs */
585 dnode_evict_dbufs(dn);
589 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
590 dnode_setdirty(dn, tx);
591 if (dn->dn_datablksz != blocksize) {
592 /* change blocksize */
593 ASSERT(dn->dn_maxblkid == 0 &&
594 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
595 dnode_block_freed(dn, 0)));
596 dnode_setdblksz(dn, blocksize);
597 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
599 if (dn->dn_bonuslen != bonuslen)
600 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
602 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
605 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
606 if (dn->dn_bonustype != bonustype)
607 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
608 if (dn->dn_nblkptr != nblkptr)
609 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
610 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
611 dbuf_rm_spill(dn, tx);
612 dnode_rm_spill(dn, tx);
614 rw_exit(&dn->dn_struct_rwlock);
619 /* change bonus size and type */
620 mutex_enter(&dn->dn_mtx);
621 dn->dn_bonustype = bonustype;
622 dn->dn_bonuslen = bonuslen;
623 dn->dn_nblkptr = nblkptr;
624 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
625 dn->dn_compress = ZIO_COMPRESS_INHERIT;
626 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
628 /* fix up the bonus db_size */
630 dn->dn_bonus->db.db_size =
631 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
632 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
635 dn->dn_allocated_txg = tx->tx_txg;
636 mutex_exit(&dn->dn_mtx);
641 uint64_t dms_dnode_invalid;
642 uint64_t dms_dnode_recheck1;
643 uint64_t dms_dnode_recheck2;
644 uint64_t dms_dnode_special;
645 uint64_t dms_dnode_handle;
646 uint64_t dms_dnode_rwlock;
647 uint64_t dms_dnode_active;
649 #endif /* DNODE_STATS */
652 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
656 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
657 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
658 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
659 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
662 ndn->dn_objset = odn->dn_objset;
663 ndn->dn_object = odn->dn_object;
664 ndn->dn_dbuf = odn->dn_dbuf;
665 ndn->dn_handle = odn->dn_handle;
666 ndn->dn_phys = odn->dn_phys;
667 ndn->dn_type = odn->dn_type;
668 ndn->dn_bonuslen = odn->dn_bonuslen;
669 ndn->dn_bonustype = odn->dn_bonustype;
670 ndn->dn_nblkptr = odn->dn_nblkptr;
671 ndn->dn_checksum = odn->dn_checksum;
672 ndn->dn_compress = odn->dn_compress;
673 ndn->dn_nlevels = odn->dn_nlevels;
674 ndn->dn_indblkshift = odn->dn_indblkshift;
675 ndn->dn_datablkshift = odn->dn_datablkshift;
676 ndn->dn_datablkszsec = odn->dn_datablkszsec;
677 ndn->dn_datablksz = odn->dn_datablksz;
678 ndn->dn_maxblkid = odn->dn_maxblkid;
679 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
680 sizeof (odn->dn_next_nblkptr));
681 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
682 sizeof (odn->dn_next_nlevels));
683 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
684 sizeof (odn->dn_next_indblkshift));
685 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
686 sizeof (odn->dn_next_bonustype));
687 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
688 sizeof (odn->dn_rm_spillblk));
689 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
690 sizeof (odn->dn_next_bonuslen));
691 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
692 sizeof (odn->dn_next_blksz));
693 for (i = 0; i < TXG_SIZE; i++) {
694 list_move_tail(&ndn->dn_dirty_records[i],
695 &odn->dn_dirty_records[i]);
697 bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
698 ndn->dn_allocated_txg = odn->dn_allocated_txg;
699 ndn->dn_free_txg = odn->dn_free_txg;
700 ndn->dn_assigned_txg = odn->dn_assigned_txg;
701 ndn->dn_dirtyctx = odn->dn_dirtyctx;
702 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
703 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
704 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
705 ASSERT(list_is_empty(&ndn->dn_dbufs));
706 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
707 ndn->dn_dbufs_count = odn->dn_dbufs_count;
708 ndn->dn_bonus = odn->dn_bonus;
709 ndn->dn_have_spill = odn->dn_have_spill;
710 ndn->dn_zio = odn->dn_zio;
711 ndn->dn_oldused = odn->dn_oldused;
712 ndn->dn_oldflags = odn->dn_oldflags;
713 ndn->dn_olduid = odn->dn_olduid;
714 ndn->dn_oldgid = odn->dn_oldgid;
715 ndn->dn_newuid = odn->dn_newuid;
716 ndn->dn_newgid = odn->dn_newgid;
717 ndn->dn_id_flags = odn->dn_id_flags;
718 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
719 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
720 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
721 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
722 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
725 * Update back pointers. Updating the handle fixes the back pointer of
726 * every descendant dbuf as well as the bonus dbuf.
728 ASSERT(ndn->dn_handle->dnh_dnode == odn);
729 ndn->dn_handle->dnh_dnode = ndn;
730 if (ndn->dn_zfetch.zf_dnode == odn) {
731 ndn->dn_zfetch.zf_dnode = ndn;
735 * Invalidate the original dnode by clearing all of its back pointers.
738 odn->dn_handle = NULL;
739 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
740 offsetof(dmu_buf_impl_t, db_link));
741 odn->dn_dbufs_count = 0;
742 odn->dn_bonus = NULL;
743 odn->dn_zfetch.zf_dnode = NULL;
746 * Set the low bit of the objset pointer to ensure that dnode_move()
747 * recognizes the dnode as invalid in any subsequent callback.
749 POINTER_INVALIDATE(&odn->dn_objset);
752 * Satisfy the destructor.
754 for (i = 0; i < TXG_SIZE; i++) {
755 list_create(&odn->dn_dirty_records[i],
756 sizeof (dbuf_dirty_record_t),
757 offsetof(dbuf_dirty_record_t, dr_dirty_node));
758 odn->dn_ranges[i].avl_root = NULL;
759 odn->dn_ranges[i].avl_numnodes = 0;
760 odn->dn_next_nlevels[i] = 0;
761 odn->dn_next_indblkshift[i] = 0;
762 odn->dn_next_bonustype[i] = 0;
763 odn->dn_rm_spillblk[i] = 0;
764 odn->dn_next_bonuslen[i] = 0;
765 odn->dn_next_blksz[i] = 0;
767 odn->dn_allocated_txg = 0;
768 odn->dn_free_txg = 0;
769 odn->dn_assigned_txg = 0;
770 odn->dn_dirtyctx = 0;
771 odn->dn_dirtyctx_firstset = NULL;
772 odn->dn_have_spill = B_FALSE;
775 odn->dn_oldflags = 0;
780 odn->dn_id_flags = 0;
786 odn->dn_moved = (uint8_t)-1;
793 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
795 dnode_t *odn = buf, *ndn = newbuf;
801 * The dnode is on the objset's list of known dnodes if the objset
802 * pointer is valid. We set the low bit of the objset pointer when
803 * freeing the dnode to invalidate it, and the memory patterns written
804 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
805 * A newly created dnode sets the objset pointer last of all to indicate
806 * that the dnode is known and in a valid state to be moved by this
810 if (!POINTER_IS_VALID(os)) {
811 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
812 return (KMEM_CBRC_DONT_KNOW);
816 * Ensure that the objset does not go away during the move.
818 rw_enter(&os_lock, RW_WRITER);
819 if (os != odn->dn_objset) {
821 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
822 return (KMEM_CBRC_DONT_KNOW);
826 * If the dnode is still valid, then so is the objset. We know that no
827 * valid objset can be freed while we hold os_lock, so we can safely
828 * ensure that the objset remains in use.
830 mutex_enter(&os->os_lock);
833 * Recheck the objset pointer in case the dnode was removed just before
834 * acquiring the lock.
836 if (os != odn->dn_objset) {
837 mutex_exit(&os->os_lock);
839 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
840 return (KMEM_CBRC_DONT_KNOW);
844 * At this point we know that as long as we hold os->os_lock, the dnode
845 * cannot be freed and fields within the dnode can be safely accessed.
846 * The objset listing this dnode cannot go away as long as this dnode is
850 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
851 mutex_exit(&os->os_lock);
852 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
853 return (KMEM_CBRC_NO);
855 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
858 * Lock the dnode handle to prevent the dnode from obtaining any new
859 * holds. This also prevents the descendant dbufs and the bonus dbuf
860 * from accessing the dnode, so that we can discount their holds. The
861 * handle is safe to access because we know that while the dnode cannot
862 * go away, neither can its handle. Once we hold dnh_zrlock, we can
863 * safely move any dnode referenced only by dbufs.
865 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
866 mutex_exit(&os->os_lock);
867 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
868 return (KMEM_CBRC_LATER);
872 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
873 * We need to guarantee that there is a hold for every dbuf in order to
874 * determine whether the dnode is actively referenced. Falsely matching
875 * a dbuf to an active hold would lead to an unsafe move. It's possible
876 * that a thread already having an active dnode hold is about to add a
877 * dbuf, and we can't compare hold and dbuf counts while the add is in
880 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
881 zrl_exit(&odn->dn_handle->dnh_zrlock);
882 mutex_exit(&os->os_lock);
883 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
884 return (KMEM_CBRC_LATER);
888 * A dbuf may be removed (evicted) without an active dnode hold. In that
889 * case, the dbuf count is decremented under the handle lock before the
890 * dbuf's hold is released. This order ensures that if we count the hold
891 * after the dbuf is removed but before its hold is released, we will
892 * treat the unmatched hold as active and exit safely. If we count the
893 * hold before the dbuf is removed, the hold is discounted, and the
894 * removal is blocked until the move completes.
896 refcount = refcount_count(&odn->dn_holds);
897 ASSERT(refcount >= 0);
898 dbufs = odn->dn_dbufs_count;
900 /* We can't have more dbufs than dnode holds. */
901 ASSERT3U(dbufs, <=, refcount);
902 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
905 if (refcount > dbufs) {
906 rw_exit(&odn->dn_struct_rwlock);
907 zrl_exit(&odn->dn_handle->dnh_zrlock);
908 mutex_exit(&os->os_lock);
909 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
910 return (KMEM_CBRC_LATER);
913 rw_exit(&odn->dn_struct_rwlock);
916 * At this point we know that anyone with a hold on the dnode is not
917 * actively referencing it. The dnode is known and in a valid state to
918 * move. We're holding the locks needed to execute the critical section.
920 dnode_move_impl(odn, ndn);
922 list_link_replace(&odn->dn_link, &ndn->dn_link);
923 /* If the dnode was safe to move, the refcount cannot have changed. */
924 ASSERT(refcount == refcount_count(&ndn->dn_holds));
925 ASSERT(dbufs == ndn->dn_dbufs_count);
926 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
927 mutex_exit(&os->os_lock);
929 return (KMEM_CBRC_YES);
935 dnode_special_close(dnode_handle_t *dnh)
937 dnode_t *dn = dnh->dnh_dnode;
940 * Wait for final references to the dnode to clear. This can
941 * only happen if the arc is asyncronously evicting state that
942 * has a hold on this dnode while we are trying to evict this
945 while (refcount_count(&dn->dn_holds) > 0)
947 zrl_add(&dnh->dnh_zrlock);
948 dnode_destroy(dn); /* implicit zrl_remove() */
949 zrl_destroy(&dnh->dnh_zrlock);
950 dnh->dnh_dnode = NULL;
954 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
957 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
959 zrl_init(&dnh->dnh_zrlock);
965 dnode_buf_pageout(dmu_buf_t *db, void *arg)
967 dnode_children_t *children_dnodes = arg;
969 int epb = db->db_size >> DNODE_SHIFT;
971 ASSERT(epb == children_dnodes->dnc_count);
973 for (i = 0; i < epb; i++) {
974 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
978 * The dnode handle lock guards against the dnode moving to
979 * another valid address, so there is no need here to guard
980 * against changes to or from NULL.
982 if (dnh->dnh_dnode == NULL) {
983 zrl_destroy(&dnh->dnh_zrlock);
987 zrl_add(&dnh->dnh_zrlock);
990 * If there are holds on this dnode, then there should
991 * be holds on the dnode's containing dbuf as well; thus
992 * it wouldn't be eligible for eviction and this function
993 * would not have been called.
995 ASSERT(refcount_is_zero(&dn->dn_holds));
996 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
998 dnode_destroy(dn); /* implicit zrl_remove() */
999 zrl_destroy(&dnh->dnh_zrlock);
1000 dnh->dnh_dnode = NULL;
1002 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1003 (epb - 1) * sizeof (dnode_handle_t));
1008 * EINVAL - invalid object number.
1010 * succeeds even for free dnodes.
1013 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1014 void *tag, dnode_t **dnp)
1017 int drop_struct_lock = FALSE;
1022 dnode_children_t *children_dnodes;
1023 dnode_handle_t *dnh;
1026 * If you are holding the spa config lock as writer, you shouldn't
1027 * be asking the DMU to do *anything* unless it's the root pool
1028 * which may require us to read from the root filesystem while
1029 * holding some (not all) of the locks as writer.
1031 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1032 (spa_is_root(os->os_spa) &&
1033 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1035 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1036 dn = (object == DMU_USERUSED_OBJECT) ?
1037 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1039 return (SET_ERROR(ENOENT));
1041 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1042 return (SET_ERROR(ENOENT));
1043 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1044 return (SET_ERROR(EEXIST));
1046 (void) refcount_add(&dn->dn_holds, tag);
1051 if (object == 0 || object >= DN_MAX_OBJECT)
1052 return (SET_ERROR(EINVAL));
1054 mdn = DMU_META_DNODE(os);
1055 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1059 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1060 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1061 drop_struct_lock = TRUE;
1064 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1066 db = dbuf_hold(mdn, blk, FTAG);
1067 if (drop_struct_lock)
1068 rw_exit(&mdn->dn_struct_rwlock);
1070 return (SET_ERROR(EIO));
1071 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1073 dbuf_rele(db, FTAG);
1077 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1078 epb = db->db.db_size >> DNODE_SHIFT;
1080 idx = object & (epb-1);
1082 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1083 children_dnodes = dmu_buf_get_user(&db->db);
1084 if (children_dnodes == NULL) {
1086 dnode_children_t *winner;
1087 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1088 (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1089 children_dnodes->dnc_count = epb;
1090 dnh = &children_dnodes->dnc_children[0];
1091 for (i = 0; i < epb; i++) {
1092 zrl_init(&dnh[i].dnh_zrlock);
1093 dnh[i].dnh_dnode = NULL;
1095 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1096 dnode_buf_pageout)) {
1097 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1098 (epb - 1) * sizeof (dnode_handle_t));
1099 children_dnodes = winner;
1102 ASSERT(children_dnodes->dnc_count == epb);
1104 dnh = &children_dnodes->dnc_children[idx];
1105 zrl_add(&dnh->dnh_zrlock);
1106 if ((dn = dnh->dnh_dnode) == NULL) {
1107 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1110 dn = dnode_create(os, phys, db, object, dnh);
1111 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1112 if (winner != NULL) {
1113 zrl_add(&dnh->dnh_zrlock);
1114 dnode_destroy(dn); /* implicit zrl_remove() */
1119 mutex_enter(&dn->dn_mtx);
1121 if (dn->dn_free_txg ||
1122 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1123 ((flag & DNODE_MUST_BE_FREE) &&
1124 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1125 mutex_exit(&dn->dn_mtx);
1126 zrl_remove(&dnh->dnh_zrlock);
1127 dbuf_rele(db, FTAG);
1128 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1130 mutex_exit(&dn->dn_mtx);
1132 if (refcount_add(&dn->dn_holds, tag) == 1)
1133 dbuf_add_ref(db, dnh);
1134 /* Now we can rely on the hold to prevent the dnode from moving. */
1135 zrl_remove(&dnh->dnh_zrlock);
1138 ASSERT3P(dn->dn_dbuf, ==, db);
1139 ASSERT3U(dn->dn_object, ==, object);
1140 dbuf_rele(db, FTAG);
1147 * Return held dnode if the object is allocated, NULL if not.
1150 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1152 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1156 * Can only add a reference if there is already at least one
1157 * reference on the dnode. Returns FALSE if unable to add a
1161 dnode_add_ref(dnode_t *dn, void *tag)
1163 mutex_enter(&dn->dn_mtx);
1164 if (refcount_is_zero(&dn->dn_holds)) {
1165 mutex_exit(&dn->dn_mtx);
1168 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1169 mutex_exit(&dn->dn_mtx);
1174 dnode_rele(dnode_t *dn, void *tag)
1177 /* Get while the hold prevents the dnode from moving. */
1178 dmu_buf_impl_t *db = dn->dn_dbuf;
1179 dnode_handle_t *dnh = dn->dn_handle;
1181 mutex_enter(&dn->dn_mtx);
1182 refs = refcount_remove(&dn->dn_holds, tag);
1183 mutex_exit(&dn->dn_mtx);
1186 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1187 * indirectly by dbuf_rele() while relying on the dnode handle to
1188 * prevent the dnode from moving, since releasing the last hold could
1189 * result in the dnode's parent dbuf evicting its dnode handles. For
1190 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1191 * other direct or indirect hold on the dnode must first drop the dnode
1194 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1196 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1197 if (refs == 0 && db != NULL) {
1199 * Another thread could add a hold to the dnode handle in
1200 * dnode_hold_impl() while holding the parent dbuf. Since the
1201 * hold on the parent dbuf prevents the handle from being
1202 * destroyed, the hold on the handle is OK. We can't yet assert
1203 * that the handle has zero references, but that will be
1204 * asserted anyway when the handle gets destroyed.
1211 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1213 objset_t *os = dn->dn_objset;
1214 uint64_t txg = tx->tx_txg;
1216 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1217 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1224 mutex_enter(&dn->dn_mtx);
1225 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1226 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1227 mutex_exit(&dn->dn_mtx);
1231 * Determine old uid/gid when necessary
1233 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1235 mutex_enter(&os->os_lock);
1238 * If we are already marked dirty, we're done.
1240 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1241 mutex_exit(&os->os_lock);
1245 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1246 ASSERT(dn->dn_datablksz != 0);
1247 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1248 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1249 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1251 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1252 dn->dn_object, txg);
1254 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1255 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1257 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1260 mutex_exit(&os->os_lock);
1263 * The dnode maintains a hold on its containing dbuf as
1264 * long as there are holds on it. Each instantiated child
1265 * dbuf maintains a hold on the dnode. When the last child
1266 * drops its hold, the dnode will drop its hold on the
1267 * containing dbuf. We add a "dirty hold" here so that the
1268 * dnode will hang around after we finish processing its
1271 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1273 (void) dbuf_dirty(dn->dn_dbuf, tx);
1275 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1279 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1281 int txgoff = tx->tx_txg & TXG_MASK;
1283 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1285 /* we should be the only holder... hopefully */
1286 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1288 mutex_enter(&dn->dn_mtx);
1289 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1290 mutex_exit(&dn->dn_mtx);
1293 dn->dn_free_txg = tx->tx_txg;
1294 mutex_exit(&dn->dn_mtx);
1297 * If the dnode is already dirty, it needs to be moved from
1298 * the dirty list to the free list.
1300 mutex_enter(&dn->dn_objset->os_lock);
1301 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1302 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1303 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1304 mutex_exit(&dn->dn_objset->os_lock);
1306 mutex_exit(&dn->dn_objset->os_lock);
1307 dnode_setdirty(dn, tx);
1312 * Try to change the block size for the indicated dnode. This can only
1313 * succeed if there are no blocks allocated or dirty beyond first block
1316 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1318 dmu_buf_impl_t *db, *db_next;
1322 size = SPA_MINBLOCKSIZE;
1323 if (size > SPA_MAXBLOCKSIZE)
1324 size = SPA_MAXBLOCKSIZE;
1326 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1328 if (ibs == dn->dn_indblkshift)
1331 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1334 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1336 /* Check for any allocated blocks beyond the first */
1337 if (dn->dn_phys->dn_maxblkid != 0)
1340 mutex_enter(&dn->dn_dbufs_mtx);
1341 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1342 db_next = list_next(&dn->dn_dbufs, db);
1344 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1345 db->db_blkid != DMU_SPILL_BLKID) {
1346 mutex_exit(&dn->dn_dbufs_mtx);
1350 mutex_exit(&dn->dn_dbufs_mtx);
1352 if (ibs && dn->dn_nlevels != 1)
1355 /* resize the old block */
1356 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1358 dbuf_new_size(db, size, tx);
1359 else if (err != ENOENT)
1362 dnode_setdblksz(dn, size);
1363 dnode_setdirty(dn, tx);
1364 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1366 dn->dn_indblkshift = ibs;
1367 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1369 /* rele after we have fixed the blocksize in the dnode */
1371 dbuf_rele(db, FTAG);
1373 rw_exit(&dn->dn_struct_rwlock);
1377 rw_exit(&dn->dn_struct_rwlock);
1378 return (SET_ERROR(ENOTSUP));
1381 /* read-holding callers must not rely on the lock being continuously held */
1383 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1385 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1386 int epbs, new_nlevels;
1389 ASSERT(blkid != DMU_BONUS_BLKID);
1392 RW_READ_HELD(&dn->dn_struct_rwlock) :
1393 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1396 * if we have a read-lock, check to see if we need to do any work
1397 * before upgrading to a write-lock.
1400 if (blkid <= dn->dn_maxblkid)
1403 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1404 rw_exit(&dn->dn_struct_rwlock);
1405 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1409 if (blkid <= dn->dn_maxblkid)
1412 dn->dn_maxblkid = blkid;
1415 * Compute the number of levels necessary to support the new maxblkid.
1418 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1419 for (sz = dn->dn_nblkptr;
1420 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1423 if (new_nlevels > dn->dn_nlevels) {
1424 int old_nlevels = dn->dn_nlevels;
1427 dbuf_dirty_record_t *new, *dr, *dr_next;
1429 dn->dn_nlevels = new_nlevels;
1431 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1432 dn->dn_next_nlevels[txgoff] = new_nlevels;
1434 /* dirty the left indirects */
1435 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1437 new = dbuf_dirty(db, tx);
1438 dbuf_rele(db, FTAG);
1440 /* transfer the dirty records to the new indirect */
1441 mutex_enter(&dn->dn_mtx);
1442 mutex_enter(&new->dt.di.dr_mtx);
1443 list = &dn->dn_dirty_records[txgoff];
1444 for (dr = list_head(list); dr; dr = dr_next) {
1445 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1446 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1447 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1448 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1449 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1450 list_remove(&dn->dn_dirty_records[txgoff], dr);
1451 list_insert_tail(&new->dt.di.dr_children, dr);
1452 dr->dr_parent = new;
1455 mutex_exit(&new->dt.di.dr_mtx);
1456 mutex_exit(&dn->dn_mtx);
1461 rw_downgrade(&dn->dn_struct_rwlock);
1465 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1467 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1470 free_range_t rp_tofind;
1471 uint64_t endblk = blkid + nblks;
1473 ASSERT(MUTEX_HELD(&dn->dn_mtx));
1474 ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1476 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1477 blkid, nblks, tx->tx_txg);
1478 rp_tofind.fr_blkid = blkid;
1479 rp = avl_find(tree, &rp_tofind, &where);
1481 rp = avl_nearest(tree, where, AVL_BEFORE);
1483 rp = avl_nearest(tree, where, AVL_AFTER);
1485 while (rp && (rp->fr_blkid <= blkid + nblks)) {
1486 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1487 free_range_t *nrp = AVL_NEXT(tree, rp);
1489 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1490 /* clear this entire range */
1491 avl_remove(tree, rp);
1492 kmem_free(rp, sizeof (free_range_t));
1493 } else if (blkid <= rp->fr_blkid &&
1494 endblk > rp->fr_blkid && endblk < fr_endblk) {
1495 /* clear the beginning of this range */
1496 rp->fr_blkid = endblk;
1497 rp->fr_nblks = fr_endblk - endblk;
1498 } else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1499 endblk >= fr_endblk) {
1500 /* clear the end of this range */
1501 rp->fr_nblks = blkid - rp->fr_blkid;
1502 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1503 /* clear a chunk out of this range */
1504 free_range_t *new_rp =
1505 kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1507 new_rp->fr_blkid = endblk;
1508 new_rp->fr_nblks = fr_endblk - endblk;
1509 avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1510 rp->fr_nblks = blkid - rp->fr_blkid;
1512 /* there may be no overlap */
1518 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1521 uint64_t blkoff, blkid, nblks;
1522 int blksz, blkshift, head, tail;
1526 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1527 blksz = dn->dn_datablksz;
1528 blkshift = dn->dn_datablkshift;
1529 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1532 len = UINT64_MAX - off;
1537 * First, block align the region to free:
1540 head = P2NPHASE(off, blksz);
1541 blkoff = P2PHASE(off, blksz);
1542 if ((off >> blkshift) > dn->dn_maxblkid)
1545 ASSERT(dn->dn_maxblkid == 0);
1546 if (off == 0 && len >= blksz) {
1547 /* Freeing the whole block; fast-track this request */
1551 } else if (off >= blksz) {
1552 /* Freeing past end-of-data */
1555 /* Freeing part of the block. */
1557 ASSERT3U(head, >, 0);
1561 /* zero out any partial block data at the start of the range */
1563 ASSERT3U(blkoff + head, ==, blksz);
1566 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1570 /* don't dirty if it isn't on disk and isn't dirty */
1571 if (db->db_last_dirty ||
1572 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1573 rw_exit(&dn->dn_struct_rwlock);
1574 dbuf_will_dirty(db, tx);
1575 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1576 data = db->db.db_data;
1577 bzero(data + blkoff, head);
1579 dbuf_rele(db, FTAG);
1585 /* If the range was less than one block, we're done */
1589 /* If the remaining range is past end of file, we're done */
1590 if ((off >> blkshift) > dn->dn_maxblkid)
1593 ASSERT(ISP2(blksz));
1597 tail = P2PHASE(len, blksz);
1599 ASSERT0(P2PHASE(off, blksz));
1600 /* zero out any partial block data at the end of the range */
1604 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1605 TRUE, FTAG, &db) == 0) {
1606 /* don't dirty if not on disk and not dirty */
1607 if (db->db_last_dirty ||
1608 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1609 rw_exit(&dn->dn_struct_rwlock);
1610 dbuf_will_dirty(db, tx);
1611 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1612 bzero(db->db.db_data, tail);
1614 dbuf_rele(db, FTAG);
1619 /* If the range did not include a full block, we are done */
1623 ASSERT(IS_P2ALIGNED(off, blksz));
1624 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1625 blkid = off >> blkshift;
1626 nblks = len >> blkshift;
1631 * Read in and mark all the level-1 indirects dirty,
1632 * so that they will stay in memory until syncing phase.
1633 * Always dirty the first and last indirect to make sure
1634 * we dirty all the partial indirects.
1636 if (dn->dn_nlevels > 1) {
1637 uint64_t i, first, last;
1638 int shift = epbs + dn->dn_datablkshift;
1640 first = blkid >> epbs;
1641 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1642 dbuf_will_dirty(db, tx);
1643 dbuf_rele(db, FTAG);
1646 last = dn->dn_maxblkid >> epbs;
1648 last = (blkid + nblks - 1) >> epbs;
1649 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1650 dbuf_will_dirty(db, tx);
1651 dbuf_rele(db, FTAG);
1653 for (i = first + 1; i < last; i++) {
1654 uint64_t ibyte = i << shift;
1657 err = dnode_next_offset(dn,
1658 DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1660 if (err == ESRCH || i >= last)
1663 db = dbuf_hold_level(dn, 1, i, FTAG);
1665 dbuf_will_dirty(db, tx);
1666 dbuf_rele(db, FTAG);
1672 * Add this range to the dnode range list.
1673 * We will finish up this free operation in the syncing phase.
1675 mutex_enter(&dn->dn_mtx);
1676 dnode_clear_range(dn, blkid, nblks, tx);
1678 free_range_t *rp, *found;
1680 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1682 /* Add new range to dn_ranges */
1683 rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1684 rp->fr_blkid = blkid;
1685 rp->fr_nblks = nblks;
1686 found = avl_find(tree, rp, &where);
1687 ASSERT(found == NULL);
1688 avl_insert(tree, rp, where);
1689 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1690 blkid, nblks, tx->tx_txg);
1692 mutex_exit(&dn->dn_mtx);
1694 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1695 dnode_setdirty(dn, tx);
1697 if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1698 dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1700 rw_exit(&dn->dn_struct_rwlock);
1704 dnode_spill_freed(dnode_t *dn)
1708 mutex_enter(&dn->dn_mtx);
1709 for (i = 0; i < TXG_SIZE; i++) {
1710 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1713 mutex_exit(&dn->dn_mtx);
1714 return (i < TXG_SIZE);
1717 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1719 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1721 free_range_t range_tofind;
1722 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1725 if (blkid == DMU_BONUS_BLKID)
1729 * If we're in the process of opening the pool, dp will not be
1730 * set yet, but there shouldn't be anything dirty.
1735 if (dn->dn_free_txg)
1738 if (blkid == DMU_SPILL_BLKID)
1739 return (dnode_spill_freed(dn));
1741 range_tofind.fr_blkid = blkid;
1742 mutex_enter(&dn->dn_mtx);
1743 for (i = 0; i < TXG_SIZE; i++) {
1744 free_range_t *range_found;
1747 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1749 ASSERT(range_found->fr_nblks > 0);
1752 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1754 range_found->fr_blkid + range_found->fr_nblks > blkid)
1757 mutex_exit(&dn->dn_mtx);
1758 return (i < TXG_SIZE);
1761 /* call from syncing context when we actually write/free space for this dnode */
1763 dnode_diduse_space(dnode_t *dn, int64_t delta)
1766 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1768 (u_longlong_t)dn->dn_phys->dn_used,
1771 mutex_enter(&dn->dn_mtx);
1772 space = DN_USED_BYTES(dn->dn_phys);
1774 ASSERT3U(space + delta, >=, space); /* no overflow */
1776 ASSERT3U(space, >=, -delta); /* no underflow */
1779 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1780 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1781 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1782 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1784 dn->dn_phys->dn_used = space;
1785 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1787 mutex_exit(&dn->dn_mtx);
1791 * Call when we think we're going to write/free space in open context.
1792 * Be conservative (ie. OK to write less than this or free more than
1793 * this, but don't write more or free less).
1796 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1798 objset_t *os = dn->dn_objset;
1799 dsl_dataset_t *ds = os->os_dsl_dataset;
1802 space = spa_get_asize(os->os_spa, space);
1805 dsl_dir_willuse_space(ds->ds_dir, space, tx);
1807 dmu_tx_willuse_space(tx, space);
1811 * Scans a block at the indicated "level" looking for a hole or data,
1812 * depending on 'flags'.
1814 * If level > 0, then we are scanning an indirect block looking at its
1815 * pointers. If level == 0, then we are looking at a block of dnodes.
1817 * If we don't find what we are looking for in the block, we return ESRCH.
1818 * Otherwise, return with *offset pointing to the beginning (if searching
1819 * forwards) or end (if searching backwards) of the range covered by the
1820 * block pointer we matched on (or dnode).
1822 * The basic search algorithm used below by dnode_next_offset() is to
1823 * use this function to search up the block tree (widen the search) until
1824 * we find something (i.e., we don't return ESRCH) and then search back
1825 * down the tree (narrow the search) until we reach our original search
1829 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1830 int lvl, uint64_t blkfill, uint64_t txg)
1832 dmu_buf_impl_t *db = NULL;
1834 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1835 uint64_t epb = 1ULL << epbs;
1836 uint64_t minfill, maxfill;
1838 int i, inc, error, span;
1840 dprintf("probing object %llu offset %llx level %d of %u\n",
1841 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1843 hole = ((flags & DNODE_FIND_HOLE) != 0);
1844 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1845 ASSERT(txg == 0 || !hole);
1847 if (lvl == dn->dn_phys->dn_nlevels) {
1849 epb = dn->dn_phys->dn_nblkptr;
1850 data = dn->dn_phys->dn_blkptr;
1852 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1853 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1855 if (error != ENOENT)
1860 * This can only happen when we are searching up
1861 * the block tree for data. We don't really need to
1862 * adjust the offset, as we will just end up looking
1863 * at the pointer to this block in its parent, and its
1864 * going to be unallocated, so we will skip over it.
1866 return (SET_ERROR(ESRCH));
1868 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1870 dbuf_rele(db, FTAG);
1873 data = db->db.db_data;
1877 (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1879 * This can only happen when we are searching up the tree
1880 * and these conditions mean that we need to keep climbing.
1882 error = SET_ERROR(ESRCH);
1883 } else if (lvl == 0) {
1884 dnode_phys_t *dnp = data;
1886 ASSERT(dn->dn_type == DMU_OT_DNODE);
1888 for (i = (*offset >> span) & (blkfill - 1);
1889 i >= 0 && i < blkfill; i += inc) {
1890 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1892 *offset += (1ULL << span) * inc;
1894 if (i < 0 || i == blkfill)
1895 error = SET_ERROR(ESRCH);
1897 blkptr_t *bp = data;
1898 uint64_t start = *offset;
1899 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1901 maxfill = blkfill << ((lvl - 1) * epbs);
1908 *offset = *offset >> span;
1909 for (i = BF64_GET(*offset, 0, epbs);
1910 i >= 0 && i < epb; i += inc) {
1911 if (bp[i].blk_fill >= minfill &&
1912 bp[i].blk_fill <= maxfill &&
1913 (hole || bp[i].blk_birth > txg))
1915 if (inc > 0 || *offset > 0)
1918 *offset = *offset << span;
1920 /* traversing backwards; position offset at the end */
1921 ASSERT3U(*offset, <=, start);
1922 *offset = MIN(*offset + (1ULL << span) - 1, start);
1923 } else if (*offset < start) {
1926 if (i < 0 || i >= epb)
1927 error = SET_ERROR(ESRCH);
1931 dbuf_rele(db, FTAG);
1937 * Find the next hole, data, or sparse region at or after *offset.
1938 * The value 'blkfill' tells us how many items we expect to find
1939 * in an L0 data block; this value is 1 for normal objects,
1940 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1941 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1945 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1946 * Finds the next/previous hole/data in a file.
1947 * Used in dmu_offset_next().
1949 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1950 * Finds the next free/allocated dnode an objset's meta-dnode.
1951 * Only finds objects that have new contents since txg (ie.
1952 * bonus buffer changes and content removal are ignored).
1953 * Used in dmu_object_next().
1955 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1956 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1957 * Used in dmu_object_alloc().
1960 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1961 int minlvl, uint64_t blkfill, uint64_t txg)
1963 uint64_t initial_offset = *offset;
1967 if (!(flags & DNODE_FIND_HAVELOCK))
1968 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1970 if (dn->dn_phys->dn_nlevels == 0) {
1971 error = SET_ERROR(ESRCH);
1975 if (dn->dn_datablkshift == 0) {
1976 if (*offset < dn->dn_datablksz) {
1977 if (flags & DNODE_FIND_HOLE)
1978 *offset = dn->dn_datablksz;
1980 error = SET_ERROR(ESRCH);
1985 maxlvl = dn->dn_phys->dn_nlevels;
1987 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1988 error = dnode_next_offset_level(dn,
1989 flags, offset, lvl, blkfill, txg);
1994 while (error == 0 && --lvl >= minlvl) {
1995 error = dnode_next_offset_level(dn,
1996 flags, offset, lvl, blkfill, txg);
1999 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2000 initial_offset < *offset : initial_offset > *offset))
2001 error = SET_ERROR(ESRCH);
2003 if (!(flags & DNODE_FIND_HAVELOCK))
2004 rw_exit(&dn->dn_struct_rwlock);