4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
28 #include <sys/dnode.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/range_tree.h>
39 #include <sys/trace_dnode.h>
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 ASSERTV(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 dn->dn_free_ranges[i] = NULL;
96 list_create(&dn->dn_dirty_records[i],
97 sizeof (dbuf_dirty_record_t),
98 offsetof(dbuf_dirty_record_t, dr_dirty_node));
101 dn->dn_allocated_txg = 0;
103 dn->dn_assigned_txg = 0;
105 dn->dn_dirtyctx_firstset = NULL;
107 dn->dn_have_spill = B_FALSE;
117 dn->dn_dbufs_count = 0;
118 dn->dn_unlisted_l0_blkid = 0;
119 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
120 offsetof(dmu_buf_impl_t, db_link));
128 dnode_dest(void *arg, void *unused)
133 rw_destroy(&dn->dn_struct_rwlock);
134 mutex_destroy(&dn->dn_mtx);
135 mutex_destroy(&dn->dn_dbufs_mtx);
136 cv_destroy(&dn->dn_notxholds);
137 refcount_destroy(&dn->dn_holds);
138 refcount_destroy(&dn->dn_tx_holds);
139 ASSERT(!list_link_active(&dn->dn_link));
141 for (i = 0; i < TXG_SIZE; i++) {
142 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
143 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
144 list_destroy(&dn->dn_dirty_records[i]);
145 ASSERT0(dn->dn_next_nblkptr[i]);
146 ASSERT0(dn->dn_next_nlevels[i]);
147 ASSERT0(dn->dn_next_indblkshift[i]);
148 ASSERT0(dn->dn_next_bonustype[i]);
149 ASSERT0(dn->dn_rm_spillblk[i]);
150 ASSERT0(dn->dn_next_bonuslen[i]);
151 ASSERT0(dn->dn_next_blksz[i]);
154 ASSERT0(dn->dn_allocated_txg);
155 ASSERT0(dn->dn_free_txg);
156 ASSERT0(dn->dn_assigned_txg);
157 ASSERT0(dn->dn_dirtyctx);
158 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
159 ASSERT3P(dn->dn_bonus, ==, NULL);
160 ASSERT(!dn->dn_have_spill);
161 ASSERT3P(dn->dn_zio, ==, NULL);
162 ASSERT0(dn->dn_oldused);
163 ASSERT0(dn->dn_oldflags);
164 ASSERT0(dn->dn_olduid);
165 ASSERT0(dn->dn_oldgid);
166 ASSERT0(dn->dn_newuid);
167 ASSERT0(dn->dn_newgid);
168 ASSERT0(dn->dn_id_flags);
170 ASSERT0(dn->dn_dbufs_count);
171 ASSERT0(dn->dn_unlisted_l0_blkid);
172 list_destroy(&dn->dn_dbufs);
178 ASSERT(dnode_cache == NULL);
179 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
180 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
181 kmem_cache_set_move(dnode_cache, dnode_move);
187 kmem_cache_destroy(dnode_cache);
194 dnode_verify(dnode_t *dn)
196 int drop_struct_lock = FALSE;
199 ASSERT(dn->dn_objset);
200 ASSERT(dn->dn_handle->dnh_dnode == dn);
202 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
204 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
207 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
208 rw_enter(&dn->dn_struct_rwlock, RW_READER);
209 drop_struct_lock = TRUE;
211 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
213 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
214 if (dn->dn_datablkshift) {
215 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
216 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
217 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
219 ASSERT3U(dn->dn_nlevels, <=, 30);
220 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
221 ASSERT3U(dn->dn_nblkptr, >=, 1);
222 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
223 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
224 ASSERT3U(dn->dn_datablksz, ==,
225 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
226 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
227 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
228 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
229 for (i = 0; i < TXG_SIZE; i++) {
230 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
233 if (dn->dn_phys->dn_type != DMU_OT_NONE)
234 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
235 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
236 if (dn->dn_dbuf != NULL) {
237 ASSERT3P(dn->dn_phys, ==,
238 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
239 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
241 if (drop_struct_lock)
242 rw_exit(&dn->dn_struct_rwlock);
247 dnode_byteswap(dnode_phys_t *dnp)
249 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
252 if (dnp->dn_type == DMU_OT_NONE) {
253 bzero(dnp, sizeof (dnode_phys_t));
257 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
258 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
259 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
260 dnp->dn_used = BSWAP_64(dnp->dn_used);
263 * dn_nblkptr is only one byte, so it's OK to read it in either
264 * byte order. We can't read dn_bouslen.
266 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
267 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
268 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
269 buf64[i] = BSWAP_64(buf64[i]);
272 * OK to check dn_bonuslen for zero, because it won't matter if
273 * we have the wrong byte order. This is necessary because the
274 * dnode dnode is smaller than a regular dnode.
276 if (dnp->dn_bonuslen != 0) {
278 * Note that the bonus length calculated here may be
279 * longer than the actual bonus buffer. This is because
280 * we always put the bonus buffer after the last block
281 * pointer (instead of packing it against the end of the
284 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
285 size_t len = DN_MAX_BONUSLEN - off;
286 dmu_object_byteswap_t byteswap;
287 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
288 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
289 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
292 /* Swap SPILL block if we have one */
293 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
294 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
299 dnode_buf_byteswap(void *vbuf, size_t size)
301 dnode_phys_t *buf = vbuf;
304 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
305 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
307 size >>= DNODE_SHIFT;
308 for (i = 0; i < size; i++) {
315 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
317 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
319 dnode_setdirty(dn, tx);
320 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
321 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
322 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
323 dn->dn_bonuslen = newsize;
325 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
327 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
328 rw_exit(&dn->dn_struct_rwlock);
332 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
334 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
335 dnode_setdirty(dn, tx);
336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 dn->dn_bonustype = newtype;
338 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
339 rw_exit(&dn->dn_struct_rwlock);
343 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
345 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
346 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
347 dnode_setdirty(dn, tx);
348 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
349 dn->dn_have_spill = B_FALSE;
353 dnode_setdblksz(dnode_t *dn, int size)
355 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
356 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
357 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
358 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
359 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
360 dn->dn_datablksz = size;
361 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
362 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
366 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
367 uint64_t object, dnode_handle_t *dnh)
369 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_PUSHPAGE);
371 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
375 * Defer setting dn_objset until the dnode is ready to be a candidate
376 * for the dnode_move() callback.
378 dn->dn_object = object;
383 if (dnp->dn_datablkszsec) {
384 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
386 dn->dn_datablksz = 0;
387 dn->dn_datablkszsec = 0;
388 dn->dn_datablkshift = 0;
390 dn->dn_indblkshift = dnp->dn_indblkshift;
391 dn->dn_nlevels = dnp->dn_nlevels;
392 dn->dn_type = dnp->dn_type;
393 dn->dn_nblkptr = dnp->dn_nblkptr;
394 dn->dn_checksum = dnp->dn_checksum;
395 dn->dn_compress = dnp->dn_compress;
396 dn->dn_bonustype = dnp->dn_bonustype;
397 dn->dn_bonuslen = dnp->dn_bonuslen;
398 dn->dn_maxblkid = dnp->dn_maxblkid;
399 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
402 dmu_zfetch_init(&dn->dn_zfetch, dn);
404 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
406 mutex_enter(&os->os_lock);
407 list_insert_head(&os->os_dnodes, dn);
410 * Everything else must be valid before assigning dn_objset makes the
411 * dnode eligible for dnode_move().
414 mutex_exit(&os->os_lock);
416 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
421 * Caller must be holding the dnode handle, which is released upon return.
424 dnode_destroy(dnode_t *dn)
426 objset_t *os = dn->dn_objset;
428 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
430 mutex_enter(&os->os_lock);
431 POINTER_INVALIDATE(&dn->dn_objset);
432 list_remove(&os->os_dnodes, dn);
433 mutex_exit(&os->os_lock);
435 /* the dnode can no longer move, so we can release the handle */
436 zrl_remove(&dn->dn_handle->dnh_zrlock);
438 dn->dn_allocated_txg = 0;
440 dn->dn_assigned_txg = 0;
443 if (dn->dn_dirtyctx_firstset != NULL) {
444 kmem_free(dn->dn_dirtyctx_firstset, 1);
445 dn->dn_dirtyctx_firstset = NULL;
447 if (dn->dn_bonus != NULL) {
448 mutex_enter(&dn->dn_bonus->db_mtx);
449 dbuf_evict(dn->dn_bonus);
454 dn->dn_have_spill = B_FALSE;
462 dn->dn_unlisted_l0_blkid = 0;
464 dmu_zfetch_rele(&dn->dn_zfetch);
465 kmem_cache_free(dnode_cache, dn);
466 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
470 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
471 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
476 blocksize = 1 << zfs_default_bs;
477 else if (blocksize > SPA_MAXBLOCKSIZE)
478 blocksize = SPA_MAXBLOCKSIZE;
480 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
483 ibs = zfs_default_ibs;
485 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
487 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
488 dn->dn_object, tx->tx_txg, blocksize, ibs);
490 ASSERT(dn->dn_type == DMU_OT_NONE);
491 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
492 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
493 ASSERT(ot != DMU_OT_NONE);
494 ASSERT(DMU_OT_IS_VALID(ot));
495 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
496 (bonustype == DMU_OT_SA && bonuslen == 0) ||
497 (bonustype != DMU_OT_NONE && bonuslen != 0));
498 ASSERT(DMU_OT_IS_VALID(bonustype));
499 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
500 ASSERT(dn->dn_type == DMU_OT_NONE);
501 ASSERT0(dn->dn_maxblkid);
502 ASSERT0(dn->dn_allocated_txg);
503 ASSERT0(dn->dn_assigned_txg);
504 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
505 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
506 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
508 for (i = 0; i < TXG_SIZE; i++) {
509 ASSERT0(dn->dn_next_nblkptr[i]);
510 ASSERT0(dn->dn_next_nlevels[i]);
511 ASSERT0(dn->dn_next_indblkshift[i]);
512 ASSERT0(dn->dn_next_bonuslen[i]);
513 ASSERT0(dn->dn_next_bonustype[i]);
514 ASSERT0(dn->dn_rm_spillblk[i]);
515 ASSERT0(dn->dn_next_blksz[i]);
516 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
517 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
518 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
522 dnode_setdblksz(dn, blocksize);
523 dn->dn_indblkshift = ibs;
525 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
529 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
530 dn->dn_bonustype = bonustype;
531 dn->dn_bonuslen = bonuslen;
532 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
533 dn->dn_compress = ZIO_COMPRESS_INHERIT;
537 if (dn->dn_dirtyctx_firstset) {
538 kmem_free(dn->dn_dirtyctx_firstset, 1);
539 dn->dn_dirtyctx_firstset = NULL;
542 dn->dn_allocated_txg = tx->tx_txg;
545 dnode_setdirty(dn, tx);
546 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
547 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
548 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
549 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
553 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
554 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
558 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
559 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
560 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
561 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
562 ASSERT(tx->tx_txg != 0);
563 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
564 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
565 (bonustype == DMU_OT_SA && bonuslen == 0));
566 ASSERT(DMU_OT_IS_VALID(bonustype));
567 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
569 /* clean up any unreferenced dbufs */
570 dnode_evict_dbufs(dn);
574 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
575 dnode_setdirty(dn, tx);
576 if (dn->dn_datablksz != blocksize) {
577 /* change blocksize */
578 ASSERT(dn->dn_maxblkid == 0 &&
579 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
580 dnode_block_freed(dn, 0)));
581 dnode_setdblksz(dn, blocksize);
582 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
584 if (dn->dn_bonuslen != bonuslen)
585 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
587 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
590 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
591 if (dn->dn_bonustype != bonustype)
592 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
593 if (dn->dn_nblkptr != nblkptr)
594 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
595 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
596 dbuf_rm_spill(dn, tx);
597 dnode_rm_spill(dn, tx);
599 rw_exit(&dn->dn_struct_rwlock);
604 /* change bonus size and type */
605 mutex_enter(&dn->dn_mtx);
606 dn->dn_bonustype = bonustype;
607 dn->dn_bonuslen = bonuslen;
608 dn->dn_nblkptr = nblkptr;
609 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
610 dn->dn_compress = ZIO_COMPRESS_INHERIT;
611 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
613 /* fix up the bonus db_size */
615 dn->dn_bonus->db.db_size =
616 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
617 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
620 dn->dn_allocated_txg = tx->tx_txg;
621 mutex_exit(&dn->dn_mtx);
627 uint64_t dms_dnode_invalid;
628 uint64_t dms_dnode_recheck1;
629 uint64_t dms_dnode_recheck2;
630 uint64_t dms_dnode_special;
631 uint64_t dms_dnode_handle;
632 uint64_t dms_dnode_rwlock;
633 uint64_t dms_dnode_active;
635 #endif /* DNODE_STATS */
638 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
642 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
643 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
644 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
645 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
648 ndn->dn_objset = odn->dn_objset;
649 ndn->dn_object = odn->dn_object;
650 ndn->dn_dbuf = odn->dn_dbuf;
651 ndn->dn_handle = odn->dn_handle;
652 ndn->dn_phys = odn->dn_phys;
653 ndn->dn_type = odn->dn_type;
654 ndn->dn_bonuslen = odn->dn_bonuslen;
655 ndn->dn_bonustype = odn->dn_bonustype;
656 ndn->dn_nblkptr = odn->dn_nblkptr;
657 ndn->dn_checksum = odn->dn_checksum;
658 ndn->dn_compress = odn->dn_compress;
659 ndn->dn_nlevels = odn->dn_nlevels;
660 ndn->dn_indblkshift = odn->dn_indblkshift;
661 ndn->dn_datablkshift = odn->dn_datablkshift;
662 ndn->dn_datablkszsec = odn->dn_datablkszsec;
663 ndn->dn_datablksz = odn->dn_datablksz;
664 ndn->dn_maxblkid = odn->dn_maxblkid;
665 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
666 sizeof (odn->dn_next_nblkptr));
667 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
668 sizeof (odn->dn_next_nlevels));
669 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
670 sizeof (odn->dn_next_indblkshift));
671 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
672 sizeof (odn->dn_next_bonustype));
673 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
674 sizeof (odn->dn_rm_spillblk));
675 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
676 sizeof (odn->dn_next_bonuslen));
677 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
678 sizeof (odn->dn_next_blksz));
679 for (i = 0; i < TXG_SIZE; i++) {
680 list_move_tail(&ndn->dn_dirty_records[i],
681 &odn->dn_dirty_records[i]);
683 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
684 sizeof (odn->dn_free_ranges));
685 ndn->dn_allocated_txg = odn->dn_allocated_txg;
686 ndn->dn_free_txg = odn->dn_free_txg;
687 ndn->dn_assigned_txg = odn->dn_assigned_txg;
688 ndn->dn_dirtyctx = odn->dn_dirtyctx;
689 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
690 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
691 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
692 ASSERT(list_is_empty(&ndn->dn_dbufs));
693 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
694 ndn->dn_dbufs_count = odn->dn_dbufs_count;
695 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
696 ndn->dn_bonus = odn->dn_bonus;
697 ndn->dn_have_spill = odn->dn_have_spill;
698 ndn->dn_zio = odn->dn_zio;
699 ndn->dn_oldused = odn->dn_oldused;
700 ndn->dn_oldflags = odn->dn_oldflags;
701 ndn->dn_olduid = odn->dn_olduid;
702 ndn->dn_oldgid = odn->dn_oldgid;
703 ndn->dn_newuid = odn->dn_newuid;
704 ndn->dn_newgid = odn->dn_newgid;
705 ndn->dn_id_flags = odn->dn_id_flags;
706 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
707 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
708 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
709 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
710 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
713 * Update back pointers. Updating the handle fixes the back pointer of
714 * every descendant dbuf as well as the bonus dbuf.
716 ASSERT(ndn->dn_handle->dnh_dnode == odn);
717 ndn->dn_handle->dnh_dnode = ndn;
718 if (ndn->dn_zfetch.zf_dnode == odn) {
719 ndn->dn_zfetch.zf_dnode = ndn;
723 * Invalidate the original dnode by clearing all of its back pointers.
726 odn->dn_handle = NULL;
727 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
728 offsetof(dmu_buf_impl_t, db_link));
729 odn->dn_dbufs_count = 0;
730 odn->dn_unlisted_l0_blkid = 0;
731 odn->dn_bonus = NULL;
732 odn->dn_zfetch.zf_dnode = NULL;
735 * Set the low bit of the objset pointer to ensure that dnode_move()
736 * recognizes the dnode as invalid in any subsequent callback.
738 POINTER_INVALIDATE(&odn->dn_objset);
741 * Satisfy the destructor.
743 for (i = 0; i < TXG_SIZE; i++) {
744 list_create(&odn->dn_dirty_records[i],
745 sizeof (dbuf_dirty_record_t),
746 offsetof(dbuf_dirty_record_t, dr_dirty_node));
747 odn->dn_free_ranges[i] = NULL;
748 odn->dn_next_nlevels[i] = 0;
749 odn->dn_next_indblkshift[i] = 0;
750 odn->dn_next_bonustype[i] = 0;
751 odn->dn_rm_spillblk[i] = 0;
752 odn->dn_next_bonuslen[i] = 0;
753 odn->dn_next_blksz[i] = 0;
755 odn->dn_allocated_txg = 0;
756 odn->dn_free_txg = 0;
757 odn->dn_assigned_txg = 0;
758 odn->dn_dirtyctx = 0;
759 odn->dn_dirtyctx_firstset = NULL;
760 odn->dn_have_spill = B_FALSE;
763 odn->dn_oldflags = 0;
768 odn->dn_id_flags = 0;
774 odn->dn_moved = (uint8_t)-1;
779 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
781 dnode_t *odn = buf, *ndn = newbuf;
787 * The dnode is on the objset's list of known dnodes if the objset
788 * pointer is valid. We set the low bit of the objset pointer when
789 * freeing the dnode to invalidate it, and the memory patterns written
790 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
791 * A newly created dnode sets the objset pointer last of all to indicate
792 * that the dnode is known and in a valid state to be moved by this
796 if (!POINTER_IS_VALID(os)) {
797 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
798 return (KMEM_CBRC_DONT_KNOW);
802 * Ensure that the objset does not go away during the move.
804 rw_enter(&os_lock, RW_WRITER);
805 if (os != odn->dn_objset) {
807 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
808 return (KMEM_CBRC_DONT_KNOW);
812 * If the dnode is still valid, then so is the objset. We know that no
813 * valid objset can be freed while we hold os_lock, so we can safely
814 * ensure that the objset remains in use.
816 mutex_enter(&os->os_lock);
819 * Recheck the objset pointer in case the dnode was removed just before
820 * acquiring the lock.
822 if (os != odn->dn_objset) {
823 mutex_exit(&os->os_lock);
825 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
826 return (KMEM_CBRC_DONT_KNOW);
830 * At this point we know that as long as we hold os->os_lock, the dnode
831 * cannot be freed and fields within the dnode can be safely accessed.
832 * The objset listing this dnode cannot go away as long as this dnode is
836 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
837 mutex_exit(&os->os_lock);
838 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
839 return (KMEM_CBRC_NO);
841 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
844 * Lock the dnode handle to prevent the dnode from obtaining any new
845 * holds. This also prevents the descendant dbufs and the bonus dbuf
846 * from accessing the dnode, so that we can discount their holds. The
847 * handle is safe to access because we know that while the dnode cannot
848 * go away, neither can its handle. Once we hold dnh_zrlock, we can
849 * safely move any dnode referenced only by dbufs.
851 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
852 mutex_exit(&os->os_lock);
853 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
854 return (KMEM_CBRC_LATER);
858 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
859 * We need to guarantee that there is a hold for every dbuf in order to
860 * determine whether the dnode is actively referenced. Falsely matching
861 * a dbuf to an active hold would lead to an unsafe move. It's possible
862 * that a thread already having an active dnode hold is about to add a
863 * dbuf, and we can't compare hold and dbuf counts while the add is in
866 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
867 zrl_exit(&odn->dn_handle->dnh_zrlock);
868 mutex_exit(&os->os_lock);
869 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
870 return (KMEM_CBRC_LATER);
874 * A dbuf may be removed (evicted) without an active dnode hold. In that
875 * case, the dbuf count is decremented under the handle lock before the
876 * dbuf's hold is released. This order ensures that if we count the hold
877 * after the dbuf is removed but before its hold is released, we will
878 * treat the unmatched hold as active and exit safely. If we count the
879 * hold before the dbuf is removed, the hold is discounted, and the
880 * removal is blocked until the move completes.
882 refcount = refcount_count(&odn->dn_holds);
883 ASSERT(refcount >= 0);
884 dbufs = odn->dn_dbufs_count;
886 /* We can't have more dbufs than dnode holds. */
887 ASSERT3U(dbufs, <=, refcount);
888 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
891 if (refcount > dbufs) {
892 rw_exit(&odn->dn_struct_rwlock);
893 zrl_exit(&odn->dn_handle->dnh_zrlock);
894 mutex_exit(&os->os_lock);
895 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
896 return (KMEM_CBRC_LATER);
899 rw_exit(&odn->dn_struct_rwlock);
902 * At this point we know that anyone with a hold on the dnode is not
903 * actively referencing it. The dnode is known and in a valid state to
904 * move. We're holding the locks needed to execute the critical section.
906 dnode_move_impl(odn, ndn);
908 list_link_replace(&odn->dn_link, &ndn->dn_link);
909 /* If the dnode was safe to move, the refcount cannot have changed. */
910 ASSERT(refcount == refcount_count(&ndn->dn_holds));
911 ASSERT(dbufs == ndn->dn_dbufs_count);
912 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
913 mutex_exit(&os->os_lock);
915 return (KMEM_CBRC_YES);
920 dnode_special_close(dnode_handle_t *dnh)
922 dnode_t *dn = dnh->dnh_dnode;
925 * Wait for final references to the dnode to clear. This can
926 * only happen if the arc is asyncronously evicting state that
927 * has a hold on this dnode while we are trying to evict this
930 while (refcount_count(&dn->dn_holds) > 0)
932 zrl_add(&dnh->dnh_zrlock);
933 dnode_destroy(dn); /* implicit zrl_remove() */
934 zrl_destroy(&dnh->dnh_zrlock);
935 dnh->dnh_dnode = NULL;
939 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
942 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
944 zrl_init(&dnh->dnh_zrlock);
950 dnode_buf_pageout(dmu_buf_t *db, void *arg)
952 dnode_children_t *children_dnodes = arg;
954 int epb = db->db_size >> DNODE_SHIFT;
956 ASSERT(epb == children_dnodes->dnc_count);
958 for (i = 0; i < epb; i++) {
959 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
963 * The dnode handle lock guards against the dnode moving to
964 * another valid address, so there is no need here to guard
965 * against changes to or from NULL.
967 if (dnh->dnh_dnode == NULL) {
968 zrl_destroy(&dnh->dnh_zrlock);
972 zrl_add(&dnh->dnh_zrlock);
975 * If there are holds on this dnode, then there should
976 * be holds on the dnode's containing dbuf as well; thus
977 * it wouldn't be eligible for eviction and this function
978 * would not have been called.
980 ASSERT(refcount_is_zero(&dn->dn_holds));
981 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
983 dnode_destroy(dn); /* implicit zrl_remove() */
984 zrl_destroy(&dnh->dnh_zrlock);
985 dnh->dnh_dnode = NULL;
987 kmem_free(children_dnodes, sizeof (dnode_children_t) +
988 (epb - 1) * sizeof (dnode_handle_t));
993 * EINVAL - invalid object number.
995 * succeeds even for free dnodes.
998 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
999 void *tag, dnode_t **dnp)
1002 int drop_struct_lock = FALSE;
1007 dnode_children_t *children_dnodes;
1008 dnode_handle_t *dnh;
1011 * If you are holding the spa config lock as writer, you shouldn't
1012 * be asking the DMU to do *anything* unless it's the root pool
1013 * which may require us to read from the root filesystem while
1014 * holding some (not all) of the locks as writer.
1016 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1017 (spa_is_root(os->os_spa) &&
1018 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1020 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1021 dn = (object == DMU_USERUSED_OBJECT) ?
1022 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1024 return (SET_ERROR(ENOENT));
1026 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1027 return (SET_ERROR(ENOENT));
1028 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1029 return (SET_ERROR(EEXIST));
1031 (void) refcount_add(&dn->dn_holds, tag);
1036 if (object == 0 || object >= DN_MAX_OBJECT)
1037 return (SET_ERROR(EINVAL));
1039 mdn = DMU_META_DNODE(os);
1040 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1044 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1045 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1046 drop_struct_lock = TRUE;
1049 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1051 db = dbuf_hold(mdn, blk, FTAG);
1052 if (drop_struct_lock)
1053 rw_exit(&mdn->dn_struct_rwlock);
1055 return (SET_ERROR(EIO));
1056 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1058 dbuf_rele(db, FTAG);
1062 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1063 epb = db->db.db_size >> DNODE_SHIFT;
1065 idx = object & (epb-1);
1067 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1068 children_dnodes = dmu_buf_get_user(&db->db);
1069 if (children_dnodes == NULL) {
1071 dnode_children_t *winner;
1072 children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1073 (epb - 1) * sizeof (dnode_handle_t), KM_PUSHPAGE);
1074 children_dnodes->dnc_count = epb;
1075 dnh = &children_dnodes->dnc_children[0];
1076 for (i = 0; i < epb; i++) {
1077 zrl_init(&dnh[i].dnh_zrlock);
1078 dnh[i].dnh_dnode = NULL;
1080 if ((winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1081 dnode_buf_pageout))) {
1082 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1083 (epb - 1) * sizeof (dnode_handle_t));
1084 children_dnodes = winner;
1087 ASSERT(children_dnodes->dnc_count == epb);
1089 dnh = &children_dnodes->dnc_children[idx];
1090 zrl_add(&dnh->dnh_zrlock);
1091 if ((dn = dnh->dnh_dnode) == NULL) {
1092 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1095 dn = dnode_create(os, phys, db, object, dnh);
1096 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1097 if (winner != NULL) {
1098 zrl_add(&dnh->dnh_zrlock);
1099 dnode_destroy(dn); /* implicit zrl_remove() */
1104 mutex_enter(&dn->dn_mtx);
1106 if (dn->dn_free_txg ||
1107 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1108 ((flag & DNODE_MUST_BE_FREE) &&
1109 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1110 mutex_exit(&dn->dn_mtx);
1111 zrl_remove(&dnh->dnh_zrlock);
1112 dbuf_rele(db, FTAG);
1113 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1115 mutex_exit(&dn->dn_mtx);
1117 if (refcount_add(&dn->dn_holds, tag) == 1)
1118 dbuf_add_ref(db, dnh);
1119 /* Now we can rely on the hold to prevent the dnode from moving. */
1120 zrl_remove(&dnh->dnh_zrlock);
1123 ASSERT3P(dn->dn_dbuf, ==, db);
1124 ASSERT3U(dn->dn_object, ==, object);
1125 dbuf_rele(db, FTAG);
1132 * Return held dnode if the object is allocated, NULL if not.
1135 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1137 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1141 * Can only add a reference if there is already at least one
1142 * reference on the dnode. Returns FALSE if unable to add a
1146 dnode_add_ref(dnode_t *dn, void *tag)
1148 mutex_enter(&dn->dn_mtx);
1149 if (refcount_is_zero(&dn->dn_holds)) {
1150 mutex_exit(&dn->dn_mtx);
1153 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1154 mutex_exit(&dn->dn_mtx);
1159 dnode_rele(dnode_t *dn, void *tag)
1162 /* Get while the hold prevents the dnode from moving. */
1163 dmu_buf_impl_t *db = dn->dn_dbuf;
1164 dnode_handle_t *dnh = dn->dn_handle;
1166 mutex_enter(&dn->dn_mtx);
1167 refs = refcount_remove(&dn->dn_holds, tag);
1168 mutex_exit(&dn->dn_mtx);
1171 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1172 * indirectly by dbuf_rele() while relying on the dnode handle to
1173 * prevent the dnode from moving, since releasing the last hold could
1174 * result in the dnode's parent dbuf evicting its dnode handles. For
1175 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1176 * other direct or indirect hold on the dnode must first drop the dnode
1179 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1181 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1182 if (refs == 0 && db != NULL) {
1184 * Another thread could add a hold to the dnode handle in
1185 * dnode_hold_impl() while holding the parent dbuf. Since the
1186 * hold on the parent dbuf prevents the handle from being
1187 * destroyed, the hold on the handle is OK. We can't yet assert
1188 * that the handle has zero references, but that will be
1189 * asserted anyway when the handle gets destroyed.
1196 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1198 objset_t *os = dn->dn_objset;
1199 uint64_t txg = tx->tx_txg;
1201 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1202 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1209 mutex_enter(&dn->dn_mtx);
1210 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1211 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1212 mutex_exit(&dn->dn_mtx);
1216 * Determine old uid/gid when necessary
1218 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1220 mutex_enter(&os->os_lock);
1223 * If we are already marked dirty, we're done.
1225 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1226 mutex_exit(&os->os_lock);
1230 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1231 ASSERT(dn->dn_datablksz != 0);
1232 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1233 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1234 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1236 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1237 dn->dn_object, txg);
1239 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1240 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1242 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1245 mutex_exit(&os->os_lock);
1248 * The dnode maintains a hold on its containing dbuf as
1249 * long as there are holds on it. Each instantiated child
1250 * dbuf maintains a hold on the dnode. When the last child
1251 * drops its hold, the dnode will drop its hold on the
1252 * containing dbuf. We add a "dirty hold" here so that the
1253 * dnode will hang around after we finish processing its
1256 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1258 (void) dbuf_dirty(dn->dn_dbuf, tx);
1260 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1264 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1266 int txgoff = tx->tx_txg & TXG_MASK;
1268 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1270 /* we should be the only holder... hopefully */
1271 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1273 mutex_enter(&dn->dn_mtx);
1274 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1275 mutex_exit(&dn->dn_mtx);
1278 dn->dn_free_txg = tx->tx_txg;
1279 mutex_exit(&dn->dn_mtx);
1282 * If the dnode is already dirty, it needs to be moved from
1283 * the dirty list to the free list.
1285 mutex_enter(&dn->dn_objset->os_lock);
1286 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1287 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1288 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1289 mutex_exit(&dn->dn_objset->os_lock);
1291 mutex_exit(&dn->dn_objset->os_lock);
1292 dnode_setdirty(dn, tx);
1297 * Try to change the block size for the indicated dnode. This can only
1298 * succeed if there are no blocks allocated or dirty beyond first block
1301 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1303 dmu_buf_impl_t *db, *db_next;
1307 size = SPA_MINBLOCKSIZE;
1308 if (size > SPA_MAXBLOCKSIZE)
1309 size = SPA_MAXBLOCKSIZE;
1311 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1313 if (ibs == dn->dn_indblkshift)
1316 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1319 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1321 /* Check for any allocated blocks beyond the first */
1322 if (dn->dn_maxblkid != 0)
1325 mutex_enter(&dn->dn_dbufs_mtx);
1326 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1327 db_next = list_next(&dn->dn_dbufs, db);
1329 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1330 db->db_blkid != DMU_SPILL_BLKID) {
1331 mutex_exit(&dn->dn_dbufs_mtx);
1335 mutex_exit(&dn->dn_dbufs_mtx);
1337 if (ibs && dn->dn_nlevels != 1)
1340 /* resize the old block */
1341 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1343 dbuf_new_size(db, size, tx);
1344 else if (err != ENOENT)
1347 dnode_setdblksz(dn, size);
1348 dnode_setdirty(dn, tx);
1349 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1351 dn->dn_indblkshift = ibs;
1352 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1354 /* rele after we have fixed the blocksize in the dnode */
1356 dbuf_rele(db, FTAG);
1358 rw_exit(&dn->dn_struct_rwlock);
1362 rw_exit(&dn->dn_struct_rwlock);
1363 return (SET_ERROR(ENOTSUP));
1366 /* read-holding callers must not rely on the lock being continuously held */
1368 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1370 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1371 int epbs, new_nlevels;
1374 ASSERT(blkid != DMU_BONUS_BLKID);
1377 RW_READ_HELD(&dn->dn_struct_rwlock) :
1378 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1381 * if we have a read-lock, check to see if we need to do any work
1382 * before upgrading to a write-lock.
1385 if (blkid <= dn->dn_maxblkid)
1388 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1389 rw_exit(&dn->dn_struct_rwlock);
1390 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1394 if (blkid <= dn->dn_maxblkid)
1397 dn->dn_maxblkid = blkid;
1400 * Compute the number of levels necessary to support the new maxblkid.
1403 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1404 for (sz = dn->dn_nblkptr;
1405 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1408 if (new_nlevels > dn->dn_nlevels) {
1409 int old_nlevels = dn->dn_nlevels;
1412 dbuf_dirty_record_t *new, *dr, *dr_next;
1414 dn->dn_nlevels = new_nlevels;
1416 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1417 dn->dn_next_nlevels[txgoff] = new_nlevels;
1419 /* dirty the left indirects */
1420 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1422 new = dbuf_dirty(db, tx);
1423 dbuf_rele(db, FTAG);
1425 /* transfer the dirty records to the new indirect */
1426 mutex_enter(&dn->dn_mtx);
1427 mutex_enter(&new->dt.di.dr_mtx);
1428 list = &dn->dn_dirty_records[txgoff];
1429 for (dr = list_head(list); dr; dr = dr_next) {
1430 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1431 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1432 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1433 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1434 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1435 list_remove(&dn->dn_dirty_records[txgoff], dr);
1436 list_insert_tail(&new->dt.di.dr_children, dr);
1437 dr->dr_parent = new;
1440 mutex_exit(&new->dt.di.dr_mtx);
1441 mutex_exit(&dn->dn_mtx);
1446 rw_downgrade(&dn->dn_struct_rwlock);
1450 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1453 uint64_t blkoff, blkid, nblks;
1454 int blksz, blkshift, head, tail;
1458 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1459 blksz = dn->dn_datablksz;
1460 blkshift = dn->dn_datablkshift;
1461 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1463 if (len == DMU_OBJECT_END) {
1464 len = UINT64_MAX - off;
1469 * First, block align the region to free:
1472 head = P2NPHASE(off, blksz);
1473 blkoff = P2PHASE(off, blksz);
1474 if ((off >> blkshift) > dn->dn_maxblkid)
1477 ASSERT(dn->dn_maxblkid == 0);
1478 if (off == 0 && len >= blksz) {
1480 * Freeing the whole block; fast-track this request.
1481 * Note that we won't dirty any indirect blocks,
1482 * which is fine because we will be freeing the entire
1483 * file and thus all indirect blocks will be freed
1484 * by free_children().
1489 } else if (off >= blksz) {
1490 /* Freeing past end-of-data */
1493 /* Freeing part of the block. */
1495 ASSERT3U(head, >, 0);
1499 /* zero out any partial block data at the start of the range */
1501 ASSERT3U(blkoff + head, ==, blksz);
1504 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1508 /* don't dirty if it isn't on disk and isn't dirty */
1509 if (db->db_last_dirty ||
1510 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1511 rw_exit(&dn->dn_struct_rwlock);
1512 dmu_buf_will_dirty(&db->db, tx);
1513 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1514 data = db->db.db_data;
1515 bzero(data + blkoff, head);
1517 dbuf_rele(db, FTAG);
1523 /* If the range was less than one block, we're done */
1527 /* If the remaining range is past end of file, we're done */
1528 if ((off >> blkshift) > dn->dn_maxblkid)
1531 ASSERT(ISP2(blksz));
1535 tail = P2PHASE(len, blksz);
1537 ASSERT0(P2PHASE(off, blksz));
1538 /* zero out any partial block data at the end of the range */
1542 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1543 TRUE, FTAG, &db) == 0) {
1544 /* don't dirty if not on disk and not dirty */
1545 if (db->db_last_dirty ||
1546 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1547 rw_exit(&dn->dn_struct_rwlock);
1548 dmu_buf_will_dirty(&db->db, tx);
1549 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1550 bzero(db->db.db_data, tail);
1552 dbuf_rele(db, FTAG);
1557 /* If the range did not include a full block, we are done */
1561 ASSERT(IS_P2ALIGNED(off, blksz));
1562 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1563 blkid = off >> blkshift;
1564 nblks = len >> blkshift;
1569 * Dirty the first and last indirect blocks, as they (and/or their
1570 * parents) will need to be written out if they were only
1571 * partially freed. Interior indirect blocks will be themselves freed,
1572 * by free_children(), so they need not be dirtied. Note that these
1573 * interior blocks have already been prefetched by dmu_tx_hold_free().
1575 if (dn->dn_nlevels > 1) {
1576 uint64_t first, last;
1578 first = blkid >> epbs;
1579 if ((db = dbuf_hold_level(dn, 1, first, FTAG))) {
1580 dmu_buf_will_dirty(&db->db, tx);
1581 dbuf_rele(db, FTAG);
1584 last = dn->dn_maxblkid >> epbs;
1586 last = (blkid + nblks - 1) >> epbs;
1587 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1588 dmu_buf_will_dirty(&db->db, tx);
1589 dbuf_rele(db, FTAG);
1595 * Add this range to the dnode range list.
1596 * We will finish up this free operation in the syncing phase.
1598 mutex_enter(&dn->dn_mtx);
1600 int txgoff = tx->tx_txg & TXG_MASK;
1601 if (dn->dn_free_ranges[txgoff] == NULL) {
1602 dn->dn_free_ranges[txgoff] =
1603 range_tree_create(NULL, NULL, &dn->dn_mtx);
1605 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1606 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1608 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1609 blkid, nblks, tx->tx_txg);
1610 mutex_exit(&dn->dn_mtx);
1612 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1613 dnode_setdirty(dn, tx);
1616 rw_exit(&dn->dn_struct_rwlock);
1620 dnode_spill_freed(dnode_t *dn)
1624 mutex_enter(&dn->dn_mtx);
1625 for (i = 0; i < TXG_SIZE; i++) {
1626 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1629 mutex_exit(&dn->dn_mtx);
1630 return (i < TXG_SIZE);
1633 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1635 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1637 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1640 if (blkid == DMU_BONUS_BLKID)
1644 * If we're in the process of opening the pool, dp will not be
1645 * set yet, but there shouldn't be anything dirty.
1650 if (dn->dn_free_txg)
1653 if (blkid == DMU_SPILL_BLKID)
1654 return (dnode_spill_freed(dn));
1656 mutex_enter(&dn->dn_mtx);
1657 for (i = 0; i < TXG_SIZE; i++) {
1658 if (dn->dn_free_ranges[i] != NULL &&
1659 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1662 mutex_exit(&dn->dn_mtx);
1663 return (i < TXG_SIZE);
1666 /* call from syncing context when we actually write/free space for this dnode */
1668 dnode_diduse_space(dnode_t *dn, int64_t delta)
1671 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1673 (u_longlong_t)dn->dn_phys->dn_used,
1676 mutex_enter(&dn->dn_mtx);
1677 space = DN_USED_BYTES(dn->dn_phys);
1679 ASSERT3U(space + delta, >=, space); /* no overflow */
1681 ASSERT3U(space, >=, -delta); /* no underflow */
1684 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1685 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1686 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1687 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1689 dn->dn_phys->dn_used = space;
1690 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1692 mutex_exit(&dn->dn_mtx);
1696 * Call when we think we're going to write/free space in open context to track
1697 * the amount of memory in use by the currently open txg.
1700 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1702 objset_t *os = dn->dn_objset;
1703 dsl_dataset_t *ds = os->os_dsl_dataset;
1704 int64_t aspace = spa_get_asize(os->os_spa, space);
1707 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1708 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1711 dmu_tx_willuse_space(tx, aspace);
1715 * Scans a block at the indicated "level" looking for a hole or data,
1716 * depending on 'flags'.
1718 * If level > 0, then we are scanning an indirect block looking at its
1719 * pointers. If level == 0, then we are looking at a block of dnodes.
1721 * If we don't find what we are looking for in the block, we return ESRCH.
1722 * Otherwise, return with *offset pointing to the beginning (if searching
1723 * forwards) or end (if searching backwards) of the range covered by the
1724 * block pointer we matched on (or dnode).
1726 * The basic search algorithm used below by dnode_next_offset() is to
1727 * use this function to search up the block tree (widen the search) until
1728 * we find something (i.e., we don't return ESRCH) and then search back
1729 * down the tree (narrow the search) until we reach our original search
1733 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1734 int lvl, uint64_t blkfill, uint64_t txg)
1736 dmu_buf_impl_t *db = NULL;
1738 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1739 uint64_t epb = 1ULL << epbs;
1740 uint64_t minfill, maxfill;
1742 int i, inc, error, span;
1744 dprintf("probing object %llu offset %llx level %d of %u\n",
1745 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1747 hole = ((flags & DNODE_FIND_HOLE) != 0);
1748 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1749 ASSERT(txg == 0 || !hole);
1751 if (lvl == dn->dn_phys->dn_nlevels) {
1753 epb = dn->dn_phys->dn_nblkptr;
1754 data = dn->dn_phys->dn_blkptr;
1756 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1757 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1759 if (error != ENOENT)
1764 * This can only happen when we are searching up
1765 * the block tree for data. We don't really need to
1766 * adjust the offset, as we will just end up looking
1767 * at the pointer to this block in its parent, and its
1768 * going to be unallocated, so we will skip over it.
1770 return (SET_ERROR(ESRCH));
1772 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1774 dbuf_rele(db, FTAG);
1777 data = db->db.db_data;
1781 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1782 db->db_blkptr->blk_birth <= txg ||
1783 BP_IS_HOLE(db->db_blkptr))) {
1785 * This can only happen when we are searching up the tree
1786 * and these conditions mean that we need to keep climbing.
1788 error = SET_ERROR(ESRCH);
1789 } else if (lvl == 0) {
1790 dnode_phys_t *dnp = data;
1792 ASSERT(dn->dn_type == DMU_OT_DNODE);
1794 for (i = (*offset >> span) & (blkfill - 1);
1795 i >= 0 && i < blkfill; i += inc) {
1796 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1798 *offset += (1ULL << span) * inc;
1800 if (i < 0 || i == blkfill)
1801 error = SET_ERROR(ESRCH);
1803 blkptr_t *bp = data;
1804 uint64_t start = *offset;
1805 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1807 maxfill = blkfill << ((lvl - 1) * epbs);
1814 *offset = *offset >> span;
1815 for (i = BF64_GET(*offset, 0, epbs);
1816 i >= 0 && i < epb; i += inc) {
1817 if (BP_GET_FILL(&bp[i]) >= minfill &&
1818 BP_GET_FILL(&bp[i]) <= maxfill &&
1819 (hole || bp[i].blk_birth > txg))
1821 if (inc > 0 || *offset > 0)
1824 *offset = *offset << span;
1826 /* traversing backwards; position offset at the end */
1827 ASSERT3U(*offset, <=, start);
1828 *offset = MIN(*offset + (1ULL << span) - 1, start);
1829 } else if (*offset < start) {
1832 if (i < 0 || i >= epb)
1833 error = SET_ERROR(ESRCH);
1837 dbuf_rele(db, FTAG);
1843 * Find the next hole, data, or sparse region at or after *offset.
1844 * The value 'blkfill' tells us how many items we expect to find
1845 * in an L0 data block; this value is 1 for normal objects,
1846 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1847 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1851 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1852 * Finds the next/previous hole/data in a file.
1853 * Used in dmu_offset_next().
1855 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1856 * Finds the next free/allocated dnode an objset's meta-dnode.
1857 * Only finds objects that have new contents since txg (ie.
1858 * bonus buffer changes and content removal are ignored).
1859 * Used in dmu_object_next().
1861 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1862 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1863 * Used in dmu_object_alloc().
1866 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1867 int minlvl, uint64_t blkfill, uint64_t txg)
1869 uint64_t initial_offset = *offset;
1873 if (!(flags & DNODE_FIND_HAVELOCK))
1874 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1876 if (dn->dn_phys->dn_nlevels == 0) {
1877 error = SET_ERROR(ESRCH);
1881 if (dn->dn_datablkshift == 0) {
1882 if (*offset < dn->dn_datablksz) {
1883 if (flags & DNODE_FIND_HOLE)
1884 *offset = dn->dn_datablksz;
1886 error = SET_ERROR(ESRCH);
1891 maxlvl = dn->dn_phys->dn_nlevels;
1893 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1894 error = dnode_next_offset_level(dn,
1895 flags, offset, lvl, blkfill, txg);
1900 while (error == 0 && --lvl >= minlvl) {
1901 error = dnode_next_offset_level(dn,
1902 flags, offset, lvl, blkfill, txg);
1906 * There's always a "virtual hole" at the end of the object, even
1907 * if all BP's which physically exist are non-holes.
1909 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1910 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1914 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1915 initial_offset < *offset : initial_offset > *offset))
1916 error = SET_ERROR(ESRCH);
1918 if (!(flags & DNODE_FIND_HAVELOCK))
1919 rw_exit(&dn->dn_struct_rwlock);