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 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);
77 refcount_create(&dn->dn_holds);
78 refcount_create(&dn->dn_tx_holds);
79 list_link_init(&dn->dn_link);
81 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
82 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
83 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
84 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
85 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
86 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
87 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
89 for (i = 0; i < TXG_SIZE; i++) {
90 list_link_init(&dn->dn_dirty_link[i]);
91 avl_create(&dn->dn_ranges[i], free_range_compar,
92 sizeof (free_range_t),
93 offsetof(struct free_range, fr_node));
94 list_create(&dn->dn_dirty_records[i],
95 sizeof (dbuf_dirty_record_t),
96 offsetof(dbuf_dirty_record_t, dr_dirty_node));
99 dn->dn_allocated_txg = 0;
101 dn->dn_assigned_txg = 0;
103 dn->dn_dirtyctx_firstset = NULL;
105 dn->dn_have_spill = B_FALSE;
115 dn->dn_dbufs_count = 0;
116 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
117 offsetof(dmu_buf_impl_t, db_link));
120 POINTER_INVALIDATE(&dn->dn_objset);
126 dnode_dest(void *arg, void *unused)
131 rw_destroy(&dn->dn_struct_rwlock);
132 mutex_destroy(&dn->dn_mtx);
133 mutex_destroy(&dn->dn_dbufs_mtx);
134 cv_destroy(&dn->dn_notxholds);
135 refcount_destroy(&dn->dn_holds);
136 refcount_destroy(&dn->dn_tx_holds);
137 ASSERT(!list_link_active(&dn->dn_link));
139 for (i = 0; i < TXG_SIZE; i++) {
140 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
141 avl_destroy(&dn->dn_ranges[i]);
142 list_destroy(&dn->dn_dirty_records[i]);
143 ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
144 ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
145 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
146 ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
147 ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
148 ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
149 ASSERT3U(dn->dn_next_blksz[i], ==, 0);
152 ASSERT3U(dn->dn_allocated_txg, ==, 0);
153 ASSERT3U(dn->dn_free_txg, ==, 0);
154 ASSERT3U(dn->dn_assigned_txg, ==, 0);
155 ASSERT3U(dn->dn_dirtyctx, ==, 0);
156 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
157 ASSERT3P(dn->dn_bonus, ==, NULL);
158 ASSERT(!dn->dn_have_spill);
159 ASSERT3P(dn->dn_zio, ==, NULL);
160 ASSERT3U(dn->dn_oldused, ==, 0);
161 ASSERT3U(dn->dn_oldflags, ==, 0);
162 ASSERT3U(dn->dn_olduid, ==, 0);
163 ASSERT3U(dn->dn_oldgid, ==, 0);
164 ASSERT3U(dn->dn_newuid, ==, 0);
165 ASSERT3U(dn->dn_newgid, ==, 0);
166 ASSERT3U(dn->dn_id_flags, ==, 0);
168 ASSERT3U(dn->dn_dbufs_count, ==, 0);
169 list_destroy(&dn->dn_dbufs);
175 ASSERT(dnode_cache == NULL);
176 dnode_cache = kmem_cache_create("dnode_t",
178 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
179 kmem_cache_set_move(dnode_cache, dnode_move);
185 kmem_cache_destroy(dnode_cache);
192 dnode_verify(dnode_t *dn)
194 int drop_struct_lock = FALSE;
197 ASSERT(dn->dn_objset);
198 ASSERT(dn->dn_handle->dnh_dnode == dn);
200 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
202 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
205 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
206 rw_enter(&dn->dn_struct_rwlock, RW_READER);
207 drop_struct_lock = TRUE;
209 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
211 ASSERT3U(dn->dn_indblkshift, >=, 0);
212 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
213 if (dn->dn_datablkshift) {
214 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
215 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
216 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
218 ASSERT3U(dn->dn_nlevels, <=, 30);
219 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
220 ASSERT3U(dn->dn_nblkptr, >=, 1);
221 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
222 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
223 ASSERT3U(dn->dn_datablksz, ==,
224 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
225 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
226 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
227 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
228 for (i = 0; i < TXG_SIZE; i++) {
229 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
232 if (dn->dn_phys->dn_type != DMU_OT_NONE)
233 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
234 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
235 if (dn->dn_dbuf != NULL) {
236 ASSERT3P(dn->dn_phys, ==,
237 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
238 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
240 if (drop_struct_lock)
241 rw_exit(&dn->dn_struct_rwlock);
246 dnode_byteswap(dnode_phys_t *dnp)
248 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
251 if (dnp->dn_type == DMU_OT_NONE) {
252 bzero(dnp, sizeof (dnode_phys_t));
256 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
257 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
258 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
259 dnp->dn_used = BSWAP_64(dnp->dn_used);
262 * dn_nblkptr is only one byte, so it's OK to read it in either
263 * byte order. We can't read dn_bouslen.
265 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
266 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
267 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
268 buf64[i] = BSWAP_64(buf64[i]);
271 * OK to check dn_bonuslen for zero, because it won't matter if
272 * we have the wrong byte order. This is necessary because the
273 * dnode dnode is smaller than a regular dnode.
275 if (dnp->dn_bonuslen != 0) {
277 * Note that the bonus length calculated here may be
278 * longer than the actual bonus buffer. This is because
279 * we always put the bonus buffer after the last block
280 * pointer (instead of packing it against the end of the
283 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
284 size_t len = DN_MAX_BONUSLEN - off;
285 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
286 dmu_object_byteswap_t byteswap =
287 DMU_OT_BYTESWAP(dnp->dn_bonustype);
288 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
291 /* Swap SPILL block if we have one */
292 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
293 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
298 dnode_buf_byteswap(void *vbuf, size_t size)
300 dnode_phys_t *buf = vbuf;
303 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
304 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
306 size >>= DNODE_SHIFT;
307 for (i = 0; i < size; i++) {
314 free_range_compar(const void *node1, const void *node2)
316 const free_range_t *rp1 = node1;
317 const free_range_t *rp2 = node2;
319 if (rp1->fr_blkid < rp2->fr_blkid)
321 else if (rp1->fr_blkid > rp2->fr_blkid)
327 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
329 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
331 dnode_setdirty(dn, tx);
332 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
333 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
334 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
335 dn->dn_bonuslen = newsize;
337 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
339 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
340 rw_exit(&dn->dn_struct_rwlock);
344 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
346 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
347 dnode_setdirty(dn, tx);
348 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
349 dn->dn_bonustype = newtype;
350 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
351 rw_exit(&dn->dn_struct_rwlock);
355 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
357 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
358 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
359 dnode_setdirty(dn, tx);
360 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
361 dn->dn_have_spill = B_FALSE;
365 dnode_setdblksz(dnode_t *dn, int size)
367 ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0);
368 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
369 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
370 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
371 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
372 dn->dn_datablksz = size;
373 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
374 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
378 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
379 uint64_t object, dnode_handle_t *dnh)
381 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
383 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
387 * Defer setting dn_objset until the dnode is ready to be a candidate
388 * for the dnode_move() callback.
390 dn->dn_object = object;
395 if (dnp->dn_datablkszsec) {
396 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
398 dn->dn_datablksz = 0;
399 dn->dn_datablkszsec = 0;
400 dn->dn_datablkshift = 0;
402 dn->dn_indblkshift = dnp->dn_indblkshift;
403 dn->dn_nlevels = dnp->dn_nlevels;
404 dn->dn_type = dnp->dn_type;
405 dn->dn_nblkptr = dnp->dn_nblkptr;
406 dn->dn_checksum = dnp->dn_checksum;
407 dn->dn_compress = dnp->dn_compress;
408 dn->dn_bonustype = dnp->dn_bonustype;
409 dn->dn_bonuslen = dnp->dn_bonuslen;
410 dn->dn_maxblkid = dnp->dn_maxblkid;
411 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
414 dmu_zfetch_init(&dn->dn_zfetch, dn);
416 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
418 mutex_enter(&os->os_lock);
419 list_insert_head(&os->os_dnodes, dn);
422 * Everything else must be valid before assigning dn_objset makes the
423 * dnode eligible for dnode_move().
426 mutex_exit(&os->os_lock);
428 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
433 * Caller must be holding the dnode handle, which is released upon return.
436 dnode_destroy(dnode_t *dn)
438 objset_t *os = dn->dn_objset;
440 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
442 mutex_enter(&os->os_lock);
443 POINTER_INVALIDATE(&dn->dn_objset);
444 list_remove(&os->os_dnodes, dn);
445 mutex_exit(&os->os_lock);
447 /* the dnode can no longer move, so we can release the handle */
448 zrl_remove(&dn->dn_handle->dnh_zrlock);
450 dn->dn_allocated_txg = 0;
452 dn->dn_assigned_txg = 0;
455 if (dn->dn_dirtyctx_firstset != NULL) {
456 kmem_free(dn->dn_dirtyctx_firstset, 1);
457 dn->dn_dirtyctx_firstset = NULL;
459 if (dn->dn_bonus != NULL) {
460 mutex_enter(&dn->dn_bonus->db_mtx);
461 dbuf_evict(dn->dn_bonus);
466 dn->dn_have_spill = B_FALSE;
475 dmu_zfetch_rele(&dn->dn_zfetch);
476 kmem_cache_free(dnode_cache, dn);
477 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
481 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
482 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
487 blocksize = 1 << zfs_default_bs;
488 else if (blocksize > SPA_MAXBLOCKSIZE)
489 blocksize = SPA_MAXBLOCKSIZE;
491 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
494 ibs = zfs_default_ibs;
496 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
498 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
499 dn->dn_object, tx->tx_txg, blocksize, ibs);
501 ASSERT(dn->dn_type == DMU_OT_NONE);
502 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
503 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
504 ASSERT(ot != DMU_OT_NONE);
505 ASSERT(DMU_OT_IS_VALID(ot));
506 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
507 (bonustype == DMU_OT_SA && bonuslen == 0) ||
508 (bonustype != DMU_OT_NONE && bonuslen != 0));
509 ASSERT(DMU_OT_IS_VALID(bonustype));
510 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
511 ASSERT(dn->dn_type == DMU_OT_NONE);
512 ASSERT3U(dn->dn_maxblkid, ==, 0);
513 ASSERT3U(dn->dn_allocated_txg, ==, 0);
514 ASSERT3U(dn->dn_assigned_txg, ==, 0);
515 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
516 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
517 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
519 for (i = 0; i < TXG_SIZE; i++) {
520 ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
521 ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
522 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
523 ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
524 ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
525 ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
526 ASSERT3U(dn->dn_next_blksz[i], ==, 0);
527 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
528 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
529 ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0);
533 dnode_setdblksz(dn, blocksize);
534 dn->dn_indblkshift = ibs;
536 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
540 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
541 dn->dn_bonustype = bonustype;
542 dn->dn_bonuslen = bonuslen;
543 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
544 dn->dn_compress = ZIO_COMPRESS_INHERIT;
548 if (dn->dn_dirtyctx_firstset) {
549 kmem_free(dn->dn_dirtyctx_firstset, 1);
550 dn->dn_dirtyctx_firstset = NULL;
553 dn->dn_allocated_txg = tx->tx_txg;
556 dnode_setdirty(dn, tx);
557 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
558 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
559 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
560 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
564 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
565 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
569 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
570 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
571 ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0);
572 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
573 ASSERT(tx->tx_txg != 0);
574 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
575 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
576 (bonustype == DMU_OT_SA && bonuslen == 0));
577 ASSERT(DMU_OT_IS_VALID(bonustype));
578 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
580 /* clean up any unreferenced dbufs */
581 dnode_evict_dbufs(dn);
585 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
586 dnode_setdirty(dn, tx);
587 if (dn->dn_datablksz != blocksize) {
588 /* change blocksize */
589 ASSERT(dn->dn_maxblkid == 0 &&
590 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
591 dnode_block_freed(dn, 0)));
592 dnode_setdblksz(dn, blocksize);
593 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
595 if (dn->dn_bonuslen != bonuslen)
596 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
598 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
601 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
602 if (dn->dn_bonustype != bonustype)
603 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
604 if (dn->dn_nblkptr != nblkptr)
605 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
606 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
607 dbuf_rm_spill(dn, tx);
608 dnode_rm_spill(dn, tx);
610 rw_exit(&dn->dn_struct_rwlock);
615 /* change bonus size and type */
616 mutex_enter(&dn->dn_mtx);
617 dn->dn_bonustype = bonustype;
618 dn->dn_bonuslen = bonuslen;
619 dn->dn_nblkptr = nblkptr;
620 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
621 dn->dn_compress = ZIO_COMPRESS_INHERIT;
622 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
624 /* fix up the bonus db_size */
626 dn->dn_bonus->db.db_size =
627 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
628 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
631 dn->dn_allocated_txg = tx->tx_txg;
632 mutex_exit(&dn->dn_mtx);
637 uint64_t dms_dnode_invalid;
638 uint64_t dms_dnode_recheck1;
639 uint64_t dms_dnode_recheck2;
640 uint64_t dms_dnode_special;
641 uint64_t dms_dnode_handle;
642 uint64_t dms_dnode_rwlock;
643 uint64_t dms_dnode_active;
645 #endif /* DNODE_STATS */
648 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
652 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
653 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
654 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
655 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
658 ndn->dn_objset = odn->dn_objset;
659 ndn->dn_object = odn->dn_object;
660 ndn->dn_dbuf = odn->dn_dbuf;
661 ndn->dn_handle = odn->dn_handle;
662 ndn->dn_phys = odn->dn_phys;
663 ndn->dn_type = odn->dn_type;
664 ndn->dn_bonuslen = odn->dn_bonuslen;
665 ndn->dn_bonustype = odn->dn_bonustype;
666 ndn->dn_nblkptr = odn->dn_nblkptr;
667 ndn->dn_checksum = odn->dn_checksum;
668 ndn->dn_compress = odn->dn_compress;
669 ndn->dn_nlevels = odn->dn_nlevels;
670 ndn->dn_indblkshift = odn->dn_indblkshift;
671 ndn->dn_datablkshift = odn->dn_datablkshift;
672 ndn->dn_datablkszsec = odn->dn_datablkszsec;
673 ndn->dn_datablksz = odn->dn_datablksz;
674 ndn->dn_maxblkid = odn->dn_maxblkid;
675 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
676 sizeof (odn->dn_next_nblkptr));
677 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
678 sizeof (odn->dn_next_nlevels));
679 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
680 sizeof (odn->dn_next_indblkshift));
681 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
682 sizeof (odn->dn_next_bonustype));
683 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
684 sizeof (odn->dn_rm_spillblk));
685 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
686 sizeof (odn->dn_next_bonuslen));
687 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
688 sizeof (odn->dn_next_blksz));
689 for (i = 0; i < TXG_SIZE; i++) {
690 list_move_tail(&ndn->dn_dirty_records[i],
691 &odn->dn_dirty_records[i]);
693 bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
694 ndn->dn_allocated_txg = odn->dn_allocated_txg;
695 ndn->dn_free_txg = odn->dn_free_txg;
696 ndn->dn_assigned_txg = odn->dn_assigned_txg;
697 ndn->dn_dirtyctx = odn->dn_dirtyctx;
698 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
699 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
700 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
701 ASSERT(list_is_empty(&ndn->dn_dbufs));
702 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
703 ndn->dn_dbufs_count = odn->dn_dbufs_count;
704 ndn->dn_bonus = odn->dn_bonus;
705 ndn->dn_have_spill = odn->dn_have_spill;
706 ndn->dn_zio = odn->dn_zio;
707 ndn->dn_oldused = odn->dn_oldused;
708 ndn->dn_oldflags = odn->dn_oldflags;
709 ndn->dn_olduid = odn->dn_olduid;
710 ndn->dn_oldgid = odn->dn_oldgid;
711 ndn->dn_newuid = odn->dn_newuid;
712 ndn->dn_newgid = odn->dn_newgid;
713 ndn->dn_id_flags = odn->dn_id_flags;
714 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
715 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
716 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
717 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
718 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
721 * Update back pointers. Updating the handle fixes the back pointer of
722 * every descendant dbuf as well as the bonus dbuf.
724 ASSERT(ndn->dn_handle->dnh_dnode == odn);
725 ndn->dn_handle->dnh_dnode = ndn;
726 if (ndn->dn_zfetch.zf_dnode == odn) {
727 ndn->dn_zfetch.zf_dnode = ndn;
731 * Invalidate the original dnode by clearing all of its back pointers.
734 odn->dn_handle = NULL;
735 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
736 offsetof(dmu_buf_impl_t, db_link));
737 odn->dn_dbufs_count = 0;
738 odn->dn_bonus = NULL;
739 odn->dn_zfetch.zf_dnode = NULL;
742 * Set the low bit of the objset pointer to ensure that dnode_move()
743 * recognizes the dnode as invalid in any subsequent callback.
745 POINTER_INVALIDATE(&odn->dn_objset);
748 * Satisfy the destructor.
750 for (i = 0; i < TXG_SIZE; i++) {
751 list_create(&odn->dn_dirty_records[i],
752 sizeof (dbuf_dirty_record_t),
753 offsetof(dbuf_dirty_record_t, dr_dirty_node));
754 odn->dn_ranges[i].avl_root = NULL;
755 odn->dn_ranges[i].avl_numnodes = 0;
756 odn->dn_next_nlevels[i] = 0;
757 odn->dn_next_indblkshift[i] = 0;
758 odn->dn_next_bonustype[i] = 0;
759 odn->dn_rm_spillblk[i] = 0;
760 odn->dn_next_bonuslen[i] = 0;
761 odn->dn_next_blksz[i] = 0;
763 odn->dn_allocated_txg = 0;
764 odn->dn_free_txg = 0;
765 odn->dn_assigned_txg = 0;
766 odn->dn_dirtyctx = 0;
767 odn->dn_dirtyctx_firstset = NULL;
768 odn->dn_have_spill = B_FALSE;
771 odn->dn_oldflags = 0;
776 odn->dn_id_flags = 0;
782 odn->dn_moved = (uint8_t)-1;
789 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
791 dnode_t *odn = buf, *ndn = newbuf;
797 * The dnode is on the objset's list of known dnodes if the objset
798 * pointer is valid. We set the low bit of the objset pointer when
799 * freeing the dnode to invalidate it, and the memory patterns written
800 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
801 * A newly created dnode sets the objset pointer last of all to indicate
802 * that the dnode is known and in a valid state to be moved by this
806 if (!POINTER_IS_VALID(os)) {
807 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
808 return (KMEM_CBRC_DONT_KNOW);
812 * Ensure that the objset does not go away during the move.
814 rw_enter(&os_lock, RW_WRITER);
815 if (os != odn->dn_objset) {
817 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
818 return (KMEM_CBRC_DONT_KNOW);
822 * If the dnode is still valid, then so is the objset. We know that no
823 * valid objset can be freed while we hold os_lock, so we can safely
824 * ensure that the objset remains in use.
826 mutex_enter(&os->os_lock);
829 * Recheck the objset pointer in case the dnode was removed just before
830 * acquiring the lock.
832 if (os != odn->dn_objset) {
833 mutex_exit(&os->os_lock);
835 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
836 return (KMEM_CBRC_DONT_KNOW);
840 * At this point we know that as long as we hold os->os_lock, the dnode
841 * cannot be freed and fields within the dnode can be safely accessed.
842 * The objset listing this dnode cannot go away as long as this dnode is
846 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
847 mutex_exit(&os->os_lock);
848 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
849 return (KMEM_CBRC_NO);
851 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
854 * Lock the dnode handle to prevent the dnode from obtaining any new
855 * holds. This also prevents the descendant dbufs and the bonus dbuf
856 * from accessing the dnode, so that we can discount their holds. The
857 * handle is safe to access because we know that while the dnode cannot
858 * go away, neither can its handle. Once we hold dnh_zrlock, we can
859 * safely move any dnode referenced only by dbufs.
861 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
862 mutex_exit(&os->os_lock);
863 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
864 return (KMEM_CBRC_LATER);
868 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
869 * We need to guarantee that there is a hold for every dbuf in order to
870 * determine whether the dnode is actively referenced. Falsely matching
871 * a dbuf to an active hold would lead to an unsafe move. It's possible
872 * that a thread already having an active dnode hold is about to add a
873 * dbuf, and we can't compare hold and dbuf counts while the add is in
876 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
877 zrl_exit(&odn->dn_handle->dnh_zrlock);
878 mutex_exit(&os->os_lock);
879 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
880 return (KMEM_CBRC_LATER);
884 * A dbuf may be removed (evicted) without an active dnode hold. In that
885 * case, the dbuf count is decremented under the handle lock before the
886 * dbuf's hold is released. This order ensures that if we count the hold
887 * after the dbuf is removed but before its hold is released, we will
888 * treat the unmatched hold as active and exit safely. If we count the
889 * hold before the dbuf is removed, the hold is discounted, and the
890 * removal is blocked until the move completes.
892 refcount = refcount_count(&odn->dn_holds);
893 ASSERT(refcount >= 0);
894 dbufs = odn->dn_dbufs_count;
896 /* We can't have more dbufs than dnode holds. */
897 ASSERT3U(dbufs, <=, refcount);
898 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
901 if (refcount > dbufs) {
902 rw_exit(&odn->dn_struct_rwlock);
903 zrl_exit(&odn->dn_handle->dnh_zrlock);
904 mutex_exit(&os->os_lock);
905 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
906 return (KMEM_CBRC_LATER);
909 rw_exit(&odn->dn_struct_rwlock);
912 * At this point we know that anyone with a hold on the dnode is not
913 * actively referencing it. The dnode is known and in a valid state to
914 * move. We're holding the locks needed to execute the critical section.
916 dnode_move_impl(odn, ndn);
918 list_link_replace(&odn->dn_link, &ndn->dn_link);
919 /* If the dnode was safe to move, the refcount cannot have changed. */
920 ASSERT(refcount == refcount_count(&ndn->dn_holds));
921 ASSERT(dbufs == ndn->dn_dbufs_count);
922 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
923 mutex_exit(&os->os_lock);
925 return (KMEM_CBRC_YES);
931 dnode_special_close(dnode_handle_t *dnh)
933 dnode_t *dn = dnh->dnh_dnode;
936 * Wait for final references to the dnode to clear. This can
937 * only happen if the arc is asyncronously evicting state that
938 * has a hold on this dnode while we are trying to evict this
941 while (refcount_count(&dn->dn_holds) > 0)
943 zrl_add(&dnh->dnh_zrlock);
944 dnode_destroy(dn); /* implicit zrl_remove() */
945 zrl_destroy(&dnh->dnh_zrlock);
946 dnh->dnh_dnode = NULL;
950 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
953 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
955 zrl_init(&dnh->dnh_zrlock);
961 dnode_buf_pageout(dmu_buf_t *db, void *arg)
963 dnode_children_t *children_dnodes = arg;
965 int epb = db->db_size >> DNODE_SHIFT;
967 ASSERT(epb == children_dnodes->dnc_count);
969 for (i = 0; i < epb; i++) {
970 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
974 * The dnode handle lock guards against the dnode moving to
975 * another valid address, so there is no need here to guard
976 * against changes to or from NULL.
978 if (dnh->dnh_dnode == NULL) {
979 zrl_destroy(&dnh->dnh_zrlock);
983 zrl_add(&dnh->dnh_zrlock);
986 * If there are holds on this dnode, then there should
987 * be holds on the dnode's containing dbuf as well; thus
988 * it wouldn't be eligible for eviction and this function
989 * would not have been called.
991 ASSERT(refcount_is_zero(&dn->dn_holds));
992 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
994 dnode_destroy(dn); /* implicit zrl_remove() */
995 zrl_destroy(&dnh->dnh_zrlock);
996 dnh->dnh_dnode = NULL;
998 kmem_free(children_dnodes, sizeof (dnode_children_t) +
999 (epb - 1) * sizeof (dnode_handle_t));
1004 * EINVAL - invalid object number.
1006 * succeeds even for free dnodes.
1009 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1010 void *tag, dnode_t **dnp)
1013 int drop_struct_lock = FALSE;
1018 dnode_children_t *children_dnodes;
1019 dnode_handle_t *dnh;
1022 * If you are holding the spa config lock as writer, you shouldn't
1023 * be asking the DMU to do *anything* unless it's the root pool
1024 * which may require us to read from the root filesystem while
1025 * holding some (not all) of the locks as writer.
1027 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1028 (spa_is_root(os->os_spa) &&
1029 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1031 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1032 dn = (object == DMU_USERUSED_OBJECT) ?
1033 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1037 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1039 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1042 (void) refcount_add(&dn->dn_holds, tag);
1047 if (object == 0 || object >= DN_MAX_OBJECT)
1050 mdn = DMU_META_DNODE(os);
1051 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1055 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1056 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1057 drop_struct_lock = TRUE;
1060 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1062 db = dbuf_hold(mdn, blk, FTAG);
1063 if (drop_struct_lock)
1064 rw_exit(&mdn->dn_struct_rwlock);
1067 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1069 dbuf_rele(db, FTAG);
1073 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1074 epb = db->db.db_size >> DNODE_SHIFT;
1076 idx = object & (epb-1);
1078 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1079 children_dnodes = dmu_buf_get_user(&db->db);
1080 if (children_dnodes == NULL) {
1082 dnode_children_t *winner;
1083 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1084 (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1085 children_dnodes->dnc_count = epb;
1086 dnh = &children_dnodes->dnc_children[0];
1087 for (i = 0; i < epb; i++) {
1088 zrl_init(&dnh[i].dnh_zrlock);
1089 dnh[i].dnh_dnode = NULL;
1091 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1092 dnode_buf_pageout)) {
1093 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1094 (epb - 1) * sizeof (dnode_handle_t));
1095 children_dnodes = winner;
1098 ASSERT(children_dnodes->dnc_count == epb);
1100 dnh = &children_dnodes->dnc_children[idx];
1101 zrl_add(&dnh->dnh_zrlock);
1102 if ((dn = dnh->dnh_dnode) == NULL) {
1103 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1106 dn = dnode_create(os, phys, db, object, dnh);
1107 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1108 if (winner != NULL) {
1109 zrl_add(&dnh->dnh_zrlock);
1110 dnode_destroy(dn); /* implicit zrl_remove() */
1115 mutex_enter(&dn->dn_mtx);
1117 if (dn->dn_free_txg ||
1118 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1119 ((flag & DNODE_MUST_BE_FREE) &&
1120 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1121 mutex_exit(&dn->dn_mtx);
1122 zrl_remove(&dnh->dnh_zrlock);
1123 dbuf_rele(db, FTAG);
1124 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1126 mutex_exit(&dn->dn_mtx);
1128 if (refcount_add(&dn->dn_holds, tag) == 1)
1129 dbuf_add_ref(db, dnh);
1130 /* Now we can rely on the hold to prevent the dnode from moving. */
1131 zrl_remove(&dnh->dnh_zrlock);
1134 ASSERT3P(dn->dn_dbuf, ==, db);
1135 ASSERT3U(dn->dn_object, ==, object);
1136 dbuf_rele(db, FTAG);
1143 * Return held dnode if the object is allocated, NULL if not.
1146 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1148 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1152 * Can only add a reference if there is already at least one
1153 * reference on the dnode. Returns FALSE if unable to add a
1157 dnode_add_ref(dnode_t *dn, void *tag)
1159 mutex_enter(&dn->dn_mtx);
1160 if (refcount_is_zero(&dn->dn_holds)) {
1161 mutex_exit(&dn->dn_mtx);
1164 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1165 mutex_exit(&dn->dn_mtx);
1170 dnode_rele(dnode_t *dn, void *tag)
1173 /* Get while the hold prevents the dnode from moving. */
1174 dmu_buf_impl_t *db = dn->dn_dbuf;
1175 dnode_handle_t *dnh = dn->dn_handle;
1177 mutex_enter(&dn->dn_mtx);
1178 refs = refcount_remove(&dn->dn_holds, tag);
1179 mutex_exit(&dn->dn_mtx);
1182 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1183 * indirectly by dbuf_rele() while relying on the dnode handle to
1184 * prevent the dnode from moving, since releasing the last hold could
1185 * result in the dnode's parent dbuf evicting its dnode handles. For
1186 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1187 * other direct or indirect hold on the dnode must first drop the dnode
1190 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1192 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1193 if (refs == 0 && db != NULL) {
1195 * Another thread could add a hold to the dnode handle in
1196 * dnode_hold_impl() while holding the parent dbuf. Since the
1197 * hold on the parent dbuf prevents the handle from being
1198 * destroyed, the hold on the handle is OK. We can't yet assert
1199 * that the handle has zero references, but that will be
1200 * asserted anyway when the handle gets destroyed.
1207 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1209 objset_t *os = dn->dn_objset;
1210 uint64_t txg = tx->tx_txg;
1212 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1213 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1220 mutex_enter(&dn->dn_mtx);
1221 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1222 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1223 mutex_exit(&dn->dn_mtx);
1227 * Determine old uid/gid when necessary
1229 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1231 mutex_enter(&os->os_lock);
1234 * If we are already marked dirty, we're done.
1236 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1237 mutex_exit(&os->os_lock);
1241 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1242 ASSERT(dn->dn_datablksz != 0);
1243 ASSERT3U(dn->dn_next_bonuslen[txg&TXG_MASK], ==, 0);
1244 ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0);
1245 ASSERT3U(dn->dn_next_bonustype[txg&TXG_MASK], ==, 0);
1247 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1248 dn->dn_object, txg);
1250 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1251 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1253 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1256 mutex_exit(&os->os_lock);
1259 * The dnode maintains a hold on its containing dbuf as
1260 * long as there are holds on it. Each instantiated child
1261 * dbuf maintains a hold on the dnode. When the last child
1262 * drops its hold, the dnode will drop its hold on the
1263 * containing dbuf. We add a "dirty hold" here so that the
1264 * dnode will hang around after we finish processing its
1267 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1269 (void) dbuf_dirty(dn->dn_dbuf, tx);
1271 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1275 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1277 int txgoff = tx->tx_txg & TXG_MASK;
1279 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1281 /* we should be the only holder... hopefully */
1282 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1284 mutex_enter(&dn->dn_mtx);
1285 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1286 mutex_exit(&dn->dn_mtx);
1289 dn->dn_free_txg = tx->tx_txg;
1290 mutex_exit(&dn->dn_mtx);
1293 * If the dnode is already dirty, it needs to be moved from
1294 * the dirty list to the free list.
1296 mutex_enter(&dn->dn_objset->os_lock);
1297 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1298 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1299 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1300 mutex_exit(&dn->dn_objset->os_lock);
1302 mutex_exit(&dn->dn_objset->os_lock);
1303 dnode_setdirty(dn, tx);
1308 * Try to change the block size for the indicated dnode. This can only
1309 * succeed if there are no blocks allocated or dirty beyond first block
1312 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1314 dmu_buf_impl_t *db, *db_next;
1318 size = SPA_MINBLOCKSIZE;
1319 if (size > SPA_MAXBLOCKSIZE)
1320 size = SPA_MAXBLOCKSIZE;
1322 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1324 if (ibs == dn->dn_indblkshift)
1327 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1330 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1332 /* Check for any allocated blocks beyond the first */
1333 if (dn->dn_phys->dn_maxblkid != 0)
1336 mutex_enter(&dn->dn_dbufs_mtx);
1337 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1338 db_next = list_next(&dn->dn_dbufs, db);
1340 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1341 db->db_blkid != DMU_SPILL_BLKID) {
1342 mutex_exit(&dn->dn_dbufs_mtx);
1346 mutex_exit(&dn->dn_dbufs_mtx);
1348 if (ibs && dn->dn_nlevels != 1)
1351 /* resize the old block */
1352 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1354 dbuf_new_size(db, size, tx);
1355 else if (err != ENOENT)
1358 dnode_setdblksz(dn, size);
1359 dnode_setdirty(dn, tx);
1360 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1362 dn->dn_indblkshift = ibs;
1363 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1365 /* rele after we have fixed the blocksize in the dnode */
1367 dbuf_rele(db, FTAG);
1369 rw_exit(&dn->dn_struct_rwlock);
1373 rw_exit(&dn->dn_struct_rwlock);
1377 /* read-holding callers must not rely on the lock being continuously held */
1379 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1381 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1382 int epbs, new_nlevels;
1385 ASSERT(blkid != DMU_BONUS_BLKID);
1388 RW_READ_HELD(&dn->dn_struct_rwlock) :
1389 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1392 * if we have a read-lock, check to see if we need to do any work
1393 * before upgrading to a write-lock.
1396 if (blkid <= dn->dn_maxblkid)
1399 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1400 rw_exit(&dn->dn_struct_rwlock);
1401 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1405 if (blkid <= dn->dn_maxblkid)
1408 dn->dn_maxblkid = blkid;
1411 * Compute the number of levels necessary to support the new maxblkid.
1414 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1415 for (sz = dn->dn_nblkptr;
1416 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1419 if (new_nlevels > dn->dn_nlevels) {
1420 int old_nlevels = dn->dn_nlevels;
1423 dbuf_dirty_record_t *new, *dr, *dr_next;
1425 dn->dn_nlevels = new_nlevels;
1427 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1428 dn->dn_next_nlevels[txgoff] = new_nlevels;
1430 /* dirty the left indirects */
1431 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1433 new = dbuf_dirty(db, tx);
1434 dbuf_rele(db, FTAG);
1436 /* transfer the dirty records to the new indirect */
1437 mutex_enter(&dn->dn_mtx);
1438 mutex_enter(&new->dt.di.dr_mtx);
1439 list = &dn->dn_dirty_records[txgoff];
1440 for (dr = list_head(list); dr; dr = dr_next) {
1441 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1442 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1443 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1444 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1445 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1446 list_remove(&dn->dn_dirty_records[txgoff], dr);
1447 list_insert_tail(&new->dt.di.dr_children, dr);
1448 dr->dr_parent = new;
1451 mutex_exit(&new->dt.di.dr_mtx);
1452 mutex_exit(&dn->dn_mtx);
1457 rw_downgrade(&dn->dn_struct_rwlock);
1461 dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1463 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1466 free_range_t rp_tofind;
1467 uint64_t endblk = blkid + nblks;
1469 ASSERT(MUTEX_HELD(&dn->dn_mtx));
1470 ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1472 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1473 blkid, nblks, tx->tx_txg);
1474 rp_tofind.fr_blkid = blkid;
1475 rp = avl_find(tree, &rp_tofind, &where);
1477 rp = avl_nearest(tree, where, AVL_BEFORE);
1479 rp = avl_nearest(tree, where, AVL_AFTER);
1481 while (rp && (rp->fr_blkid <= blkid + nblks)) {
1482 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1483 free_range_t *nrp = AVL_NEXT(tree, rp);
1485 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1486 /* clear this entire range */
1487 avl_remove(tree, rp);
1488 kmem_free(rp, sizeof (free_range_t));
1489 } else if (blkid <= rp->fr_blkid &&
1490 endblk > rp->fr_blkid && endblk < fr_endblk) {
1491 /* clear the beginning of this range */
1492 rp->fr_blkid = endblk;
1493 rp->fr_nblks = fr_endblk - endblk;
1494 } else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1495 endblk >= fr_endblk) {
1496 /* clear the end of this range */
1497 rp->fr_nblks = blkid - rp->fr_blkid;
1498 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1499 /* clear a chunk out of this range */
1500 free_range_t *new_rp =
1501 kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1503 new_rp->fr_blkid = endblk;
1504 new_rp->fr_nblks = fr_endblk - endblk;
1505 avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1506 rp->fr_nblks = blkid - rp->fr_blkid;
1508 /* there may be no overlap */
1514 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1517 uint64_t blkoff, blkid, nblks;
1518 int blksz, blkshift, head, tail;
1522 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1523 blksz = dn->dn_datablksz;
1524 blkshift = dn->dn_datablkshift;
1525 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1528 len = UINT64_MAX - off;
1533 * First, block align the region to free:
1536 head = P2NPHASE(off, blksz);
1537 blkoff = P2PHASE(off, blksz);
1538 if ((off >> blkshift) > dn->dn_maxblkid)
1541 ASSERT(dn->dn_maxblkid == 0);
1542 if (off == 0 && len >= blksz) {
1543 /* Freeing the whole block; fast-track this request */
1547 } else if (off >= blksz) {
1548 /* Freeing past end-of-data */
1551 /* Freeing part of the block. */
1553 ASSERT3U(head, >, 0);
1557 /* zero out any partial block data at the start of the range */
1559 ASSERT3U(blkoff + head, ==, blksz);
1562 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1566 /* don't dirty if it isn't on disk and isn't dirty */
1567 if (db->db_last_dirty ||
1568 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1569 rw_exit(&dn->dn_struct_rwlock);
1570 dbuf_will_dirty(db, tx);
1571 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1572 data = db->db.db_data;
1573 bzero(data + blkoff, head);
1575 dbuf_rele(db, FTAG);
1581 /* If the range was less than one block, we're done */
1585 /* If the remaining range is past end of file, we're done */
1586 if ((off >> blkshift) > dn->dn_maxblkid)
1589 ASSERT(ISP2(blksz));
1593 tail = P2PHASE(len, blksz);
1595 ASSERT3U(P2PHASE(off, blksz), ==, 0);
1596 /* zero out any partial block data at the end of the range */
1600 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1601 TRUE, FTAG, &db) == 0) {
1602 /* don't dirty if not on disk and not dirty */
1603 if (db->db_last_dirty ||
1604 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1605 rw_exit(&dn->dn_struct_rwlock);
1606 dbuf_will_dirty(db, tx);
1607 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1608 bzero(db->db.db_data, tail);
1610 dbuf_rele(db, FTAG);
1615 /* If the range did not include a full block, we are done */
1619 ASSERT(IS_P2ALIGNED(off, blksz));
1620 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1621 blkid = off >> blkshift;
1622 nblks = len >> blkshift;
1627 * Read in and mark all the level-1 indirects dirty,
1628 * so that they will stay in memory until syncing phase.
1629 * Always dirty the first and last indirect to make sure
1630 * we dirty all the partial indirects.
1632 if (dn->dn_nlevels > 1) {
1633 uint64_t i, first, last;
1634 int shift = epbs + dn->dn_datablkshift;
1636 first = blkid >> epbs;
1637 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1638 dbuf_will_dirty(db, tx);
1639 dbuf_rele(db, FTAG);
1642 last = dn->dn_maxblkid >> epbs;
1644 last = (blkid + nblks - 1) >> epbs;
1645 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1646 dbuf_will_dirty(db, tx);
1647 dbuf_rele(db, FTAG);
1649 for (i = first + 1; i < last; i++) {
1650 uint64_t ibyte = i << shift;
1653 err = dnode_next_offset(dn,
1654 DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1656 if (err == ESRCH || i >= last)
1659 db = dbuf_hold_level(dn, 1, i, FTAG);
1661 dbuf_will_dirty(db, tx);
1662 dbuf_rele(db, FTAG);
1668 * Add this range to the dnode range list.
1669 * We will finish up this free operation in the syncing phase.
1671 mutex_enter(&dn->dn_mtx);
1672 dnode_clear_range(dn, blkid, nblks, tx);
1674 free_range_t *rp, *found;
1676 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1678 /* Add new range to dn_ranges */
1679 rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP);
1680 rp->fr_blkid = blkid;
1681 rp->fr_nblks = nblks;
1682 found = avl_find(tree, rp, &where);
1683 ASSERT(found == NULL);
1684 avl_insert(tree, rp, where);
1685 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1686 blkid, nblks, tx->tx_txg);
1688 mutex_exit(&dn->dn_mtx);
1690 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1691 dnode_setdirty(dn, tx);
1693 if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1694 dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1696 rw_exit(&dn->dn_struct_rwlock);
1700 dnode_spill_freed(dnode_t *dn)
1704 mutex_enter(&dn->dn_mtx);
1705 for (i = 0; i < TXG_SIZE; i++) {
1706 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1709 mutex_exit(&dn->dn_mtx);
1710 return (i < TXG_SIZE);
1713 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1715 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1717 free_range_t range_tofind;
1718 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1721 if (blkid == DMU_BONUS_BLKID)
1725 * If we're in the process of opening the pool, dp will not be
1726 * set yet, but there shouldn't be anything dirty.
1731 if (dn->dn_free_txg)
1734 if (blkid == DMU_SPILL_BLKID)
1735 return (dnode_spill_freed(dn));
1737 range_tofind.fr_blkid = blkid;
1738 mutex_enter(&dn->dn_mtx);
1739 for (i = 0; i < TXG_SIZE; i++) {
1740 free_range_t *range_found;
1743 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1745 ASSERT(range_found->fr_nblks > 0);
1748 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1750 range_found->fr_blkid + range_found->fr_nblks > blkid)
1753 mutex_exit(&dn->dn_mtx);
1754 return (i < TXG_SIZE);
1757 /* call from syncing context when we actually write/free space for this dnode */
1759 dnode_diduse_space(dnode_t *dn, int64_t delta)
1762 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1764 (u_longlong_t)dn->dn_phys->dn_used,
1767 mutex_enter(&dn->dn_mtx);
1768 space = DN_USED_BYTES(dn->dn_phys);
1770 ASSERT3U(space + delta, >=, space); /* no overflow */
1772 ASSERT3U(space, >=, -delta); /* no underflow */
1775 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1776 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1777 ASSERT3U(P2PHASE(space, 1<<DEV_BSHIFT), ==, 0);
1778 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1780 dn->dn_phys->dn_used = space;
1781 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1783 mutex_exit(&dn->dn_mtx);
1787 * Call when we think we're going to write/free space in open context.
1788 * Be conservative (ie. OK to write less than this or free more than
1789 * this, but don't write more or free less).
1792 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1794 objset_t *os = dn->dn_objset;
1795 dsl_dataset_t *ds = os->os_dsl_dataset;
1798 space = spa_get_asize(os->os_spa, space);
1801 dsl_dir_willuse_space(ds->ds_dir, space, tx);
1803 dmu_tx_willuse_space(tx, space);
1807 * This function scans a block at the indicated "level" looking for
1808 * a hole or data (depending on 'flags'). If level > 0, then we are
1809 * scanning an indirect block looking at its pointers. If level == 0,
1810 * then we are looking at a block of dnodes. If we don't find what we
1811 * are looking for in the block, we return ESRCH. Otherwise, return
1812 * with *offset pointing to the beginning (if searching forwards) or
1813 * end (if searching backwards) of the range covered by the block
1814 * pointer we matched on (or dnode).
1816 * The basic search algorithm used below by dnode_next_offset() is to
1817 * use this function to search up the block tree (widen the search) until
1818 * we find something (i.e., we don't return ESRCH) and then search back
1819 * down the tree (narrow the search) until we reach our original search
1823 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1824 int lvl, uint64_t blkfill, uint64_t txg)
1826 dmu_buf_impl_t *db = NULL;
1828 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1829 uint64_t epb = 1ULL << epbs;
1830 uint64_t minfill, maxfill;
1832 int i, inc, error, span;
1834 dprintf("probing object %llu offset %llx level %d of %u\n",
1835 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1837 hole = ((flags & DNODE_FIND_HOLE) != 0);
1838 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1839 ASSERT(txg == 0 || !hole);
1841 if (lvl == dn->dn_phys->dn_nlevels) {
1843 epb = dn->dn_phys->dn_nblkptr;
1844 data = dn->dn_phys->dn_blkptr;
1846 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1847 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1849 if (error != ENOENT)
1854 * This can only happen when we are searching up
1855 * the block tree for data. We don't really need to
1856 * adjust the offset, as we will just end up looking
1857 * at the pointer to this block in its parent, and its
1858 * going to be unallocated, so we will skip over it.
1862 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1864 dbuf_rele(db, FTAG);
1867 data = db->db.db_data;
1871 (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1873 * This can only happen when we are searching up the tree
1874 * and these conditions mean that we need to keep climbing.
1877 } else if (lvl == 0) {
1878 dnode_phys_t *dnp = data;
1880 ASSERT(dn->dn_type == DMU_OT_DNODE);
1882 for (i = (*offset >> span) & (blkfill - 1);
1883 i >= 0 && i < blkfill; i += inc) {
1884 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1886 *offset += (1ULL << span) * inc;
1888 if (i < 0 || i == blkfill)
1891 blkptr_t *bp = data;
1892 uint64_t start = *offset;
1893 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1895 maxfill = blkfill << ((lvl - 1) * epbs);
1902 *offset = *offset >> span;
1903 for (i = BF64_GET(*offset, 0, epbs);
1904 i >= 0 && i < epb; i += inc) {
1905 if (bp[i].blk_fill >= minfill &&
1906 bp[i].blk_fill <= maxfill &&
1907 (hole || bp[i].blk_birth > txg))
1909 if (inc > 0 || *offset > 0)
1912 *offset = *offset << span;
1914 /* traversing backwards; position offset at the end */
1915 ASSERT3U(*offset, <=, start);
1916 *offset = MIN(*offset + (1ULL << span) - 1, start);
1917 } else if (*offset < start) {
1920 if (i < 0 || i >= epb)
1925 dbuf_rele(db, FTAG);
1931 * Find the next hole, data, or sparse region at or after *offset.
1932 * The value 'blkfill' tells us how many items we expect to find
1933 * in an L0 data block; this value is 1 for normal objects,
1934 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1935 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1939 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1940 * Finds the next/previous hole/data in a file.
1941 * Used in dmu_offset_next().
1943 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1944 * Finds the next free/allocated dnode an objset's meta-dnode.
1945 * Only finds objects that have new contents since txg (ie.
1946 * bonus buffer changes and content removal are ignored).
1947 * Used in dmu_object_next().
1949 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1950 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1951 * Used in dmu_object_alloc().
1954 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1955 int minlvl, uint64_t blkfill, uint64_t txg)
1957 uint64_t initial_offset = *offset;
1961 if (!(flags & DNODE_FIND_HAVELOCK))
1962 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1964 if (dn->dn_phys->dn_nlevels == 0) {
1969 if (dn->dn_datablkshift == 0) {
1970 if (*offset < dn->dn_datablksz) {
1971 if (flags & DNODE_FIND_HOLE)
1972 *offset = dn->dn_datablksz;
1979 maxlvl = dn->dn_phys->dn_nlevels;
1981 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1982 error = dnode_next_offset_level(dn,
1983 flags, offset, lvl, blkfill, txg);
1988 while (error == 0 && --lvl >= minlvl) {
1989 error = dnode_next_offset_level(dn,
1990 flags, offset, lvl, blkfill, txg);
1993 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1994 initial_offset < *offset : initial_offset > *offset))
1997 if (!(flags & DNODE_FIND_HAVELOCK))
1998 rw_exit(&dn->dn_struct_rwlock);