4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
26 #include <sys/zfs_context.h>
28 #include <sys/dnode.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/range_tree.h>
40 static kmem_cache_t *dnode_cache;
42 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
43 * turned on when DEBUG is also defined.
50 #define DNODE_STAT_ADD(stat) ((stat)++)
52 #define DNODE_STAT_ADD(stat) /* nothing */
53 #endif /* DNODE_STATS */
55 static dnode_phys_t dnode_phys_zero;
57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
66 dnode_cons(void *arg, void *unused, int kmflag)
71 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
72 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
73 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
74 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
77 * Every dbuf has a reference, and dropping a tracked reference is
78 * O(number of references), so don't track dn_holds.
80 refcount_create_untracked(&dn->dn_holds);
81 refcount_create(&dn->dn_tx_holds);
82 list_link_init(&dn->dn_link);
84 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
85 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
86 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
87 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
88 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
89 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
90 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
92 for (i = 0; i < TXG_SIZE; i++) {
93 list_link_init(&dn->dn_dirty_link[i]);
94 dn->dn_free_ranges[i] = NULL;
95 list_create(&dn->dn_dirty_records[i],
96 sizeof (dbuf_dirty_record_t),
97 offsetof(dbuf_dirty_record_t, dr_dirty_node));
100 dn->dn_allocated_txg = 0;
102 dn->dn_assigned_txg = 0;
104 dn->dn_dirtyctx_firstset = NULL;
106 dn->dn_have_spill = B_FALSE;
116 dn->dn_dbufs_count = 0;
117 dn->dn_unlisted_l0_blkid = 0;
118 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
119 offsetof(dmu_buf_impl_t, db_link));
122 POINTER_INVALIDATE(&dn->dn_objset);
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",
181 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
182 kmem_cache_set_move(dnode_cache, dnode_move);
188 kmem_cache_destroy(dnode_cache);
195 dnode_verify(dnode_t *dn)
197 int drop_struct_lock = FALSE;
200 ASSERT(dn->dn_objset);
201 ASSERT(dn->dn_handle->dnh_dnode == dn);
203 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
205 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
208 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
209 rw_enter(&dn->dn_struct_rwlock, RW_READER);
210 drop_struct_lock = TRUE;
212 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
214 ASSERT3U(dn->dn_indblkshift, >=, 0);
215 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
216 if (dn->dn_datablkshift) {
217 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
218 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
219 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
221 ASSERT3U(dn->dn_nlevels, <=, 30);
222 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
223 ASSERT3U(dn->dn_nblkptr, >=, 1);
224 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
225 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
226 ASSERT3U(dn->dn_datablksz, ==,
227 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
228 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
229 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
230 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
231 for (i = 0; i < TXG_SIZE; i++) {
232 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
235 if (dn->dn_phys->dn_type != DMU_OT_NONE)
236 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
237 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
238 if (dn->dn_dbuf != NULL) {
239 ASSERT3P(dn->dn_phys, ==,
240 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
241 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
243 if (drop_struct_lock)
244 rw_exit(&dn->dn_struct_rwlock);
249 dnode_byteswap(dnode_phys_t *dnp)
251 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
254 if (dnp->dn_type == DMU_OT_NONE) {
255 bzero(dnp, sizeof (dnode_phys_t));
259 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
260 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
261 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
262 dnp->dn_used = BSWAP_64(dnp->dn_used);
265 * dn_nblkptr is only one byte, so it's OK to read it in either
266 * byte order. We can't read dn_bouslen.
268 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
269 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
270 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
271 buf64[i] = BSWAP_64(buf64[i]);
274 * OK to check dn_bonuslen for zero, because it won't matter if
275 * we have the wrong byte order. This is necessary because the
276 * dnode dnode is smaller than a regular dnode.
278 if (dnp->dn_bonuslen != 0) {
280 * Note that the bonus length calculated here may be
281 * longer than the actual bonus buffer. This is because
282 * we always put the bonus buffer after the last block
283 * pointer (instead of packing it against the end of the
286 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
287 size_t len = DN_MAX_BONUSLEN - off;
288 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
289 dmu_object_byteswap_t byteswap =
290 DMU_OT_BYTESWAP(dnp->dn_bonustype);
291 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
294 /* Swap SPILL block if we have one */
295 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
296 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
301 dnode_buf_byteswap(void *vbuf, size_t size)
303 dnode_phys_t *buf = vbuf;
306 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
307 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
309 size >>= DNODE_SHIFT;
310 for (i = 0; i < size; i++) {
317 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
319 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
321 dnode_setdirty(dn, tx);
322 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
323 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
324 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
325 dn->dn_bonuslen = newsize;
327 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
329 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
330 rw_exit(&dn->dn_struct_rwlock);
334 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
336 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
337 dnode_setdirty(dn, tx);
338 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
339 dn->dn_bonustype = newtype;
340 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
341 rw_exit(&dn->dn_struct_rwlock);
345 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
347 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
348 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
349 dnode_setdirty(dn, tx);
350 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
351 dn->dn_have_spill = B_FALSE;
355 dnode_setdblksz(dnode_t *dn, int size)
357 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
358 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
359 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
360 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
361 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
362 dn->dn_datablksz = size;
363 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
364 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
368 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
369 uint64_t object, dnode_handle_t *dnh)
371 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
373 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
377 * Defer setting dn_objset until the dnode is ready to be a candidate
378 * for the dnode_move() callback.
380 dn->dn_object = object;
385 if (dnp->dn_datablkszsec) {
386 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
388 dn->dn_datablksz = 0;
389 dn->dn_datablkszsec = 0;
390 dn->dn_datablkshift = 0;
392 dn->dn_indblkshift = dnp->dn_indblkshift;
393 dn->dn_nlevels = dnp->dn_nlevels;
394 dn->dn_type = dnp->dn_type;
395 dn->dn_nblkptr = dnp->dn_nblkptr;
396 dn->dn_checksum = dnp->dn_checksum;
397 dn->dn_compress = dnp->dn_compress;
398 dn->dn_bonustype = dnp->dn_bonustype;
399 dn->dn_bonuslen = dnp->dn_bonuslen;
400 dn->dn_maxblkid = dnp->dn_maxblkid;
401 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
404 dmu_zfetch_init(&dn->dn_zfetch, dn);
406 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
408 mutex_enter(&os->os_lock);
409 list_insert_head(&os->os_dnodes, dn);
412 * Everything else must be valid before assigning dn_objset makes the
413 * dnode eligible for dnode_move().
416 mutex_exit(&os->os_lock);
418 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
423 * Caller must be holding the dnode handle, which is released upon return.
426 dnode_destroy(dnode_t *dn)
428 objset_t *os = dn->dn_objset;
430 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
432 mutex_enter(&os->os_lock);
433 POINTER_INVALIDATE(&dn->dn_objset);
434 list_remove(&os->os_dnodes, dn);
435 mutex_exit(&os->os_lock);
437 /* the dnode can no longer move, so we can release the handle */
438 zrl_remove(&dn->dn_handle->dnh_zrlock);
440 dn->dn_allocated_txg = 0;
442 dn->dn_assigned_txg = 0;
445 if (dn->dn_dirtyctx_firstset != NULL) {
446 kmem_free(dn->dn_dirtyctx_firstset, 1);
447 dn->dn_dirtyctx_firstset = NULL;
449 if (dn->dn_bonus != NULL) {
450 mutex_enter(&dn->dn_bonus->db_mtx);
451 dbuf_evict(dn->dn_bonus);
456 dn->dn_have_spill = B_FALSE;
464 dn->dn_unlisted_l0_blkid = 0;
466 dmu_zfetch_rele(&dn->dn_zfetch);
467 kmem_cache_free(dnode_cache, dn);
468 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
472 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
473 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
478 blocksize = 1 << zfs_default_bs;
479 else if (blocksize > SPA_MAXBLOCKSIZE)
480 blocksize = SPA_MAXBLOCKSIZE;
482 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
485 ibs = zfs_default_ibs;
487 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
489 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
490 dn->dn_object, tx->tx_txg, blocksize, ibs);
492 ASSERT(dn->dn_type == DMU_OT_NONE);
493 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
494 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
495 ASSERT(ot != DMU_OT_NONE);
496 ASSERT(DMU_OT_IS_VALID(ot));
497 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
498 (bonustype == DMU_OT_SA && bonuslen == 0) ||
499 (bonustype != DMU_OT_NONE && bonuslen != 0));
500 ASSERT(DMU_OT_IS_VALID(bonustype));
501 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
502 ASSERT(dn->dn_type == DMU_OT_NONE);
503 ASSERT0(dn->dn_maxblkid);
504 ASSERT0(dn->dn_allocated_txg);
505 ASSERT0(dn->dn_assigned_txg);
506 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
507 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
508 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
510 for (i = 0; i < TXG_SIZE; i++) {
511 ASSERT0(dn->dn_next_nblkptr[i]);
512 ASSERT0(dn->dn_next_nlevels[i]);
513 ASSERT0(dn->dn_next_indblkshift[i]);
514 ASSERT0(dn->dn_next_bonuslen[i]);
515 ASSERT0(dn->dn_next_bonustype[i]);
516 ASSERT0(dn->dn_rm_spillblk[i]);
517 ASSERT0(dn->dn_next_blksz[i]);
518 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
519 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
520 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
524 dnode_setdblksz(dn, blocksize);
525 dn->dn_indblkshift = ibs;
527 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
531 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
532 dn->dn_bonustype = bonustype;
533 dn->dn_bonuslen = bonuslen;
534 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
535 dn->dn_compress = ZIO_COMPRESS_INHERIT;
539 if (dn->dn_dirtyctx_firstset) {
540 kmem_free(dn->dn_dirtyctx_firstset, 1);
541 dn->dn_dirtyctx_firstset = NULL;
544 dn->dn_allocated_txg = tx->tx_txg;
547 dnode_setdirty(dn, tx);
548 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
549 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
550 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
551 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
555 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
556 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
560 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
561 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
562 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
563 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
564 ASSERT(tx->tx_txg != 0);
565 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
566 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
567 (bonustype == DMU_OT_SA && bonuslen == 0));
568 ASSERT(DMU_OT_IS_VALID(bonustype));
569 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
571 /* clean up any unreferenced dbufs */
572 dnode_evict_dbufs(dn);
576 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
577 dnode_setdirty(dn, tx);
578 if (dn->dn_datablksz != blocksize) {
579 /* change blocksize */
580 ASSERT(dn->dn_maxblkid == 0 &&
581 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
582 dnode_block_freed(dn, 0)));
583 dnode_setdblksz(dn, blocksize);
584 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
586 if (dn->dn_bonuslen != bonuslen)
587 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
589 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
592 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
593 if (dn->dn_bonustype != bonustype)
594 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
595 if (dn->dn_nblkptr != nblkptr)
596 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
597 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
598 dbuf_rm_spill(dn, tx);
599 dnode_rm_spill(dn, tx);
601 rw_exit(&dn->dn_struct_rwlock);
606 /* change bonus size and type */
607 mutex_enter(&dn->dn_mtx);
608 dn->dn_bonustype = bonustype;
609 dn->dn_bonuslen = bonuslen;
610 dn->dn_nblkptr = nblkptr;
611 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
612 dn->dn_compress = ZIO_COMPRESS_INHERIT;
613 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
615 /* fix up the bonus db_size */
617 dn->dn_bonus->db.db_size =
618 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
619 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
622 dn->dn_allocated_txg = tx->tx_txg;
623 mutex_exit(&dn->dn_mtx);
628 uint64_t dms_dnode_invalid;
629 uint64_t dms_dnode_recheck1;
630 uint64_t dms_dnode_recheck2;
631 uint64_t dms_dnode_special;
632 uint64_t dms_dnode_handle;
633 uint64_t dms_dnode_rwlock;
634 uint64_t dms_dnode_active;
636 #endif /* DNODE_STATS */
639 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
643 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
644 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
645 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
646 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
649 ndn->dn_objset = odn->dn_objset;
650 ndn->dn_object = odn->dn_object;
651 ndn->dn_dbuf = odn->dn_dbuf;
652 ndn->dn_handle = odn->dn_handle;
653 ndn->dn_phys = odn->dn_phys;
654 ndn->dn_type = odn->dn_type;
655 ndn->dn_bonuslen = odn->dn_bonuslen;
656 ndn->dn_bonustype = odn->dn_bonustype;
657 ndn->dn_nblkptr = odn->dn_nblkptr;
658 ndn->dn_checksum = odn->dn_checksum;
659 ndn->dn_compress = odn->dn_compress;
660 ndn->dn_nlevels = odn->dn_nlevels;
661 ndn->dn_indblkshift = odn->dn_indblkshift;
662 ndn->dn_datablkshift = odn->dn_datablkshift;
663 ndn->dn_datablkszsec = odn->dn_datablkszsec;
664 ndn->dn_datablksz = odn->dn_datablksz;
665 ndn->dn_maxblkid = odn->dn_maxblkid;
666 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
667 sizeof (odn->dn_next_nblkptr));
668 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
669 sizeof (odn->dn_next_nlevels));
670 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
671 sizeof (odn->dn_next_indblkshift));
672 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
673 sizeof (odn->dn_next_bonustype));
674 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
675 sizeof (odn->dn_rm_spillblk));
676 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
677 sizeof (odn->dn_next_bonuslen));
678 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
679 sizeof (odn->dn_next_blksz));
680 for (i = 0; i < TXG_SIZE; i++) {
681 list_move_tail(&ndn->dn_dirty_records[i],
682 &odn->dn_dirty_records[i]);
684 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
685 sizeof (odn->dn_free_ranges));
686 ndn->dn_allocated_txg = odn->dn_allocated_txg;
687 ndn->dn_free_txg = odn->dn_free_txg;
688 ndn->dn_assigned_txg = odn->dn_assigned_txg;
689 ndn->dn_dirtyctx = odn->dn_dirtyctx;
690 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
691 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
692 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
693 ASSERT(list_is_empty(&ndn->dn_dbufs));
694 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
695 ndn->dn_dbufs_count = odn->dn_dbufs_count;
696 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
697 ndn->dn_bonus = odn->dn_bonus;
698 ndn->dn_have_spill = odn->dn_have_spill;
699 ndn->dn_zio = odn->dn_zio;
700 ndn->dn_oldused = odn->dn_oldused;
701 ndn->dn_oldflags = odn->dn_oldflags;
702 ndn->dn_olduid = odn->dn_olduid;
703 ndn->dn_oldgid = odn->dn_oldgid;
704 ndn->dn_newuid = odn->dn_newuid;
705 ndn->dn_newgid = odn->dn_newgid;
706 ndn->dn_id_flags = odn->dn_id_flags;
707 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
708 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
709 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
710 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
711 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
714 * Update back pointers. Updating the handle fixes the back pointer of
715 * every descendant dbuf as well as the bonus dbuf.
717 ASSERT(ndn->dn_handle->dnh_dnode == odn);
718 ndn->dn_handle->dnh_dnode = ndn;
719 if (ndn->dn_zfetch.zf_dnode == odn) {
720 ndn->dn_zfetch.zf_dnode = ndn;
724 * Invalidate the original dnode by clearing all of its back pointers.
727 odn->dn_handle = NULL;
728 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
729 offsetof(dmu_buf_impl_t, db_link));
730 odn->dn_dbufs_count = 0;
731 odn->dn_unlisted_l0_blkid = 0;
732 odn->dn_bonus = NULL;
733 odn->dn_zfetch.zf_dnode = NULL;
736 * Set the low bit of the objset pointer to ensure that dnode_move()
737 * recognizes the dnode as invalid in any subsequent callback.
739 POINTER_INVALIDATE(&odn->dn_objset);
742 * Satisfy the destructor.
744 for (i = 0; i < TXG_SIZE; i++) {
745 list_create(&odn->dn_dirty_records[i],
746 sizeof (dbuf_dirty_record_t),
747 offsetof(dbuf_dirty_record_t, dr_dirty_node));
748 odn->dn_free_ranges[i] = NULL;
749 odn->dn_next_nlevels[i] = 0;
750 odn->dn_next_indblkshift[i] = 0;
751 odn->dn_next_bonustype[i] = 0;
752 odn->dn_rm_spillblk[i] = 0;
753 odn->dn_next_bonuslen[i] = 0;
754 odn->dn_next_blksz[i] = 0;
756 odn->dn_allocated_txg = 0;
757 odn->dn_free_txg = 0;
758 odn->dn_assigned_txg = 0;
759 odn->dn_dirtyctx = 0;
760 odn->dn_dirtyctx_firstset = NULL;
761 odn->dn_have_spill = B_FALSE;
764 odn->dn_oldflags = 0;
769 odn->dn_id_flags = 0;
775 odn->dn_moved = (uint8_t)-1;
782 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
784 dnode_t *odn = buf, *ndn = newbuf;
790 * The dnode is on the objset's list of known dnodes if the objset
791 * pointer is valid. We set the low bit of the objset pointer when
792 * freeing the dnode to invalidate it, and the memory patterns written
793 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
794 * A newly created dnode sets the objset pointer last of all to indicate
795 * that the dnode is known and in a valid state to be moved by this
799 if (!POINTER_IS_VALID(os)) {
800 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
801 return (KMEM_CBRC_DONT_KNOW);
805 * Ensure that the objset does not go away during the move.
807 rw_enter(&os_lock, RW_WRITER);
808 if (os != odn->dn_objset) {
810 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
811 return (KMEM_CBRC_DONT_KNOW);
815 * If the dnode is still valid, then so is the objset. We know that no
816 * valid objset can be freed while we hold os_lock, so we can safely
817 * ensure that the objset remains in use.
819 mutex_enter(&os->os_lock);
822 * Recheck the objset pointer in case the dnode was removed just before
823 * acquiring the lock.
825 if (os != odn->dn_objset) {
826 mutex_exit(&os->os_lock);
828 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
829 return (KMEM_CBRC_DONT_KNOW);
833 * At this point we know that as long as we hold os->os_lock, the dnode
834 * cannot be freed and fields within the dnode can be safely accessed.
835 * The objset listing this dnode cannot go away as long as this dnode is
839 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
840 mutex_exit(&os->os_lock);
841 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
842 return (KMEM_CBRC_NO);
844 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
847 * Lock the dnode handle to prevent the dnode from obtaining any new
848 * holds. This also prevents the descendant dbufs and the bonus dbuf
849 * from accessing the dnode, so that we can discount their holds. The
850 * handle is safe to access because we know that while the dnode cannot
851 * go away, neither can its handle. Once we hold dnh_zrlock, we can
852 * safely move any dnode referenced only by dbufs.
854 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
855 mutex_exit(&os->os_lock);
856 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
857 return (KMEM_CBRC_LATER);
861 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
862 * We need to guarantee that there is a hold for every dbuf in order to
863 * determine whether the dnode is actively referenced. Falsely matching
864 * a dbuf to an active hold would lead to an unsafe move. It's possible
865 * that a thread already having an active dnode hold is about to add a
866 * dbuf, and we can't compare hold and dbuf counts while the add is in
869 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
870 zrl_exit(&odn->dn_handle->dnh_zrlock);
871 mutex_exit(&os->os_lock);
872 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
873 return (KMEM_CBRC_LATER);
877 * A dbuf may be removed (evicted) without an active dnode hold. In that
878 * case, the dbuf count is decremented under the handle lock before the
879 * dbuf's hold is released. This order ensures that if we count the hold
880 * after the dbuf is removed but before its hold is released, we will
881 * treat the unmatched hold as active and exit safely. If we count the
882 * hold before the dbuf is removed, the hold is discounted, and the
883 * removal is blocked until the move completes.
885 refcount = refcount_count(&odn->dn_holds);
886 ASSERT(refcount >= 0);
887 dbufs = odn->dn_dbufs_count;
889 /* We can't have more dbufs than dnode holds. */
890 ASSERT3U(dbufs, <=, refcount);
891 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
894 if (refcount > dbufs) {
895 rw_exit(&odn->dn_struct_rwlock);
896 zrl_exit(&odn->dn_handle->dnh_zrlock);
897 mutex_exit(&os->os_lock);
898 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
899 return (KMEM_CBRC_LATER);
902 rw_exit(&odn->dn_struct_rwlock);
905 * At this point we know that anyone with a hold on the dnode is not
906 * actively referencing it. The dnode is known and in a valid state to
907 * move. We're holding the locks needed to execute the critical section.
909 dnode_move_impl(odn, ndn);
911 list_link_replace(&odn->dn_link, &ndn->dn_link);
912 /* If the dnode was safe to move, the refcount cannot have changed. */
913 ASSERT(refcount == refcount_count(&ndn->dn_holds));
914 ASSERT(dbufs == ndn->dn_dbufs_count);
915 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
916 mutex_exit(&os->os_lock);
918 return (KMEM_CBRC_YES);
924 dnode_special_close(dnode_handle_t *dnh)
926 dnode_t *dn = dnh->dnh_dnode;
929 * Wait for final references to the dnode to clear. This can
930 * only happen if the arc is asyncronously evicting state that
931 * has a hold on this dnode while we are trying to evict this
934 while (refcount_count(&dn->dn_holds) > 0)
936 zrl_add(&dnh->dnh_zrlock);
937 dnode_destroy(dn); /* implicit zrl_remove() */
938 zrl_destroy(&dnh->dnh_zrlock);
939 dnh->dnh_dnode = NULL;
943 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
946 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
948 zrl_init(&dnh->dnh_zrlock);
954 dnode_buf_pageout(dmu_buf_t *db, void *arg)
956 dnode_children_t *children_dnodes = arg;
958 int epb = db->db_size >> DNODE_SHIFT;
960 ASSERT(epb == children_dnodes->dnc_count);
962 for (i = 0; i < epb; i++) {
963 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
967 * The dnode handle lock guards against the dnode moving to
968 * another valid address, so there is no need here to guard
969 * against changes to or from NULL.
971 if (dnh->dnh_dnode == NULL) {
972 zrl_destroy(&dnh->dnh_zrlock);
976 zrl_add(&dnh->dnh_zrlock);
979 * If there are holds on this dnode, then there should
980 * be holds on the dnode's containing dbuf as well; thus
981 * it wouldn't be eligible for eviction and this function
982 * would not have been called.
984 ASSERT(refcount_is_zero(&dn->dn_holds));
985 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
987 dnode_destroy(dn); /* implicit zrl_remove() */
988 zrl_destroy(&dnh->dnh_zrlock);
989 dnh->dnh_dnode = NULL;
991 kmem_free(children_dnodes, sizeof (dnode_children_t) +
992 (epb - 1) * sizeof (dnode_handle_t));
997 * EINVAL - invalid object number.
999 * succeeds even for free dnodes.
1002 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1003 void *tag, dnode_t **dnp)
1006 int drop_struct_lock = FALSE;
1011 dnode_children_t *children_dnodes;
1012 dnode_handle_t *dnh;
1015 * If you are holding the spa config lock as writer, you shouldn't
1016 * be asking the DMU to do *anything* unless it's the root pool
1017 * which may require us to read from the root filesystem while
1018 * holding some (not all) of the locks as writer.
1020 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1021 (spa_is_root(os->os_spa) &&
1022 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1024 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1025 dn = (object == DMU_USERUSED_OBJECT) ?
1026 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1028 return (SET_ERROR(ENOENT));
1030 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1031 return (SET_ERROR(ENOENT));
1032 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1033 return (SET_ERROR(EEXIST));
1035 (void) refcount_add(&dn->dn_holds, tag);
1040 if (object == 0 || object >= DN_MAX_OBJECT)
1041 return (SET_ERROR(EINVAL));
1043 mdn = DMU_META_DNODE(os);
1044 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1048 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1049 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1050 drop_struct_lock = TRUE;
1053 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1055 db = dbuf_hold(mdn, blk, FTAG);
1056 if (drop_struct_lock)
1057 rw_exit(&mdn->dn_struct_rwlock);
1059 return (SET_ERROR(EIO));
1060 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1062 dbuf_rele(db, FTAG);
1066 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1067 epb = db->db.db_size >> DNODE_SHIFT;
1069 idx = object & (epb-1);
1071 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1072 children_dnodes = dmu_buf_get_user(&db->db);
1073 if (children_dnodes == NULL) {
1075 dnode_children_t *winner;
1076 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1077 (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1078 children_dnodes->dnc_count = epb;
1079 dnh = &children_dnodes->dnc_children[0];
1080 for (i = 0; i < epb; i++) {
1081 zrl_init(&dnh[i].dnh_zrlock);
1082 dnh[i].dnh_dnode = NULL;
1084 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1085 dnode_buf_pageout)) {
1086 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1087 (epb - 1) * sizeof (dnode_handle_t));
1088 children_dnodes = winner;
1091 ASSERT(children_dnodes->dnc_count == epb);
1093 dnh = &children_dnodes->dnc_children[idx];
1094 zrl_add(&dnh->dnh_zrlock);
1095 if ((dn = dnh->dnh_dnode) == NULL) {
1096 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1099 dn = dnode_create(os, phys, db, object, dnh);
1100 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1101 if (winner != NULL) {
1102 zrl_add(&dnh->dnh_zrlock);
1103 dnode_destroy(dn); /* implicit zrl_remove() */
1108 mutex_enter(&dn->dn_mtx);
1110 if (dn->dn_free_txg ||
1111 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1112 ((flag & DNODE_MUST_BE_FREE) &&
1113 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1114 mutex_exit(&dn->dn_mtx);
1115 zrl_remove(&dnh->dnh_zrlock);
1116 dbuf_rele(db, FTAG);
1117 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1119 mutex_exit(&dn->dn_mtx);
1121 if (refcount_add(&dn->dn_holds, tag) == 1)
1122 dbuf_add_ref(db, dnh);
1123 /* Now we can rely on the hold to prevent the dnode from moving. */
1124 zrl_remove(&dnh->dnh_zrlock);
1127 ASSERT3P(dn->dn_dbuf, ==, db);
1128 ASSERT3U(dn->dn_object, ==, object);
1129 dbuf_rele(db, FTAG);
1136 * Return held dnode if the object is allocated, NULL if not.
1139 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1141 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1145 * Can only add a reference if there is already at least one
1146 * reference on the dnode. Returns FALSE if unable to add a
1150 dnode_add_ref(dnode_t *dn, void *tag)
1152 mutex_enter(&dn->dn_mtx);
1153 if (refcount_is_zero(&dn->dn_holds)) {
1154 mutex_exit(&dn->dn_mtx);
1157 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1158 mutex_exit(&dn->dn_mtx);
1163 dnode_rele(dnode_t *dn, void *tag)
1166 /* Get while the hold prevents the dnode from moving. */
1167 dmu_buf_impl_t *db = dn->dn_dbuf;
1168 dnode_handle_t *dnh = dn->dn_handle;
1170 mutex_enter(&dn->dn_mtx);
1171 refs = refcount_remove(&dn->dn_holds, tag);
1172 mutex_exit(&dn->dn_mtx);
1175 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1176 * indirectly by dbuf_rele() while relying on the dnode handle to
1177 * prevent the dnode from moving, since releasing the last hold could
1178 * result in the dnode's parent dbuf evicting its dnode handles. For
1179 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1180 * other direct or indirect hold on the dnode must first drop the dnode
1183 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1185 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1186 if (refs == 0 && db != NULL) {
1188 * Another thread could add a hold to the dnode handle in
1189 * dnode_hold_impl() while holding the parent dbuf. Since the
1190 * hold on the parent dbuf prevents the handle from being
1191 * destroyed, the hold on the handle is OK. We can't yet assert
1192 * that the handle has zero references, but that will be
1193 * asserted anyway when the handle gets destroyed.
1200 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1202 objset_t *os = dn->dn_objset;
1203 uint64_t txg = tx->tx_txg;
1205 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1206 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1213 mutex_enter(&dn->dn_mtx);
1214 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1215 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1216 mutex_exit(&dn->dn_mtx);
1220 * Determine old uid/gid when necessary
1222 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1224 mutex_enter(&os->os_lock);
1227 * If we are already marked dirty, we're done.
1229 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1230 mutex_exit(&os->os_lock);
1234 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1235 ASSERT(dn->dn_datablksz != 0);
1236 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1237 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1238 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1240 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1241 dn->dn_object, txg);
1243 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1244 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1246 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1249 mutex_exit(&os->os_lock);
1252 * The dnode maintains a hold on its containing dbuf as
1253 * long as there are holds on it. Each instantiated child
1254 * dbuf maintains a hold on the dnode. When the last child
1255 * drops its hold, the dnode will drop its hold on the
1256 * containing dbuf. We add a "dirty hold" here so that the
1257 * dnode will hang around after we finish processing its
1260 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1262 (void) dbuf_dirty(dn->dn_dbuf, tx);
1264 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1268 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1270 int txgoff = tx->tx_txg & TXG_MASK;
1272 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1274 /* we should be the only holder... hopefully */
1275 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1277 mutex_enter(&dn->dn_mtx);
1278 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1279 mutex_exit(&dn->dn_mtx);
1282 dn->dn_free_txg = tx->tx_txg;
1283 mutex_exit(&dn->dn_mtx);
1286 * If the dnode is already dirty, it needs to be moved from
1287 * the dirty list to the free list.
1289 mutex_enter(&dn->dn_objset->os_lock);
1290 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1291 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1292 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1293 mutex_exit(&dn->dn_objset->os_lock);
1295 mutex_exit(&dn->dn_objset->os_lock);
1296 dnode_setdirty(dn, tx);
1301 * Try to change the block size for the indicated dnode. This can only
1302 * succeed if there are no blocks allocated or dirty beyond first block
1305 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1307 dmu_buf_impl_t *db, *db_next;
1311 size = SPA_MINBLOCKSIZE;
1312 if (size > SPA_MAXBLOCKSIZE)
1313 size = SPA_MAXBLOCKSIZE;
1315 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1317 if (ibs == dn->dn_indblkshift)
1320 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1323 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1325 /* Check for any allocated blocks beyond the first */
1326 if (dn->dn_maxblkid != 0)
1329 mutex_enter(&dn->dn_dbufs_mtx);
1330 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1331 db_next = list_next(&dn->dn_dbufs, db);
1333 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1334 db->db_blkid != DMU_SPILL_BLKID) {
1335 mutex_exit(&dn->dn_dbufs_mtx);
1339 mutex_exit(&dn->dn_dbufs_mtx);
1341 if (ibs && dn->dn_nlevels != 1)
1344 /* resize the old block */
1345 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1347 dbuf_new_size(db, size, tx);
1348 else if (err != ENOENT)
1351 dnode_setdblksz(dn, size);
1352 dnode_setdirty(dn, tx);
1353 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1355 dn->dn_indblkshift = ibs;
1356 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1358 /* rele after we have fixed the blocksize in the dnode */
1360 dbuf_rele(db, FTAG);
1362 rw_exit(&dn->dn_struct_rwlock);
1366 rw_exit(&dn->dn_struct_rwlock);
1367 return (SET_ERROR(ENOTSUP));
1370 /* read-holding callers must not rely on the lock being continuously held */
1372 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1374 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1375 int epbs, new_nlevels;
1378 ASSERT(blkid != DMU_BONUS_BLKID);
1381 RW_READ_HELD(&dn->dn_struct_rwlock) :
1382 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1385 * if we have a read-lock, check to see if we need to do any work
1386 * before upgrading to a write-lock.
1389 if (blkid <= dn->dn_maxblkid)
1392 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1393 rw_exit(&dn->dn_struct_rwlock);
1394 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1398 if (blkid <= dn->dn_maxblkid)
1401 dn->dn_maxblkid = blkid;
1404 * Compute the number of levels necessary to support the new maxblkid.
1407 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1408 for (sz = dn->dn_nblkptr;
1409 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1412 if (new_nlevels > dn->dn_nlevels) {
1413 int old_nlevels = dn->dn_nlevels;
1416 dbuf_dirty_record_t *new, *dr, *dr_next;
1418 dn->dn_nlevels = new_nlevels;
1420 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1421 dn->dn_next_nlevels[txgoff] = new_nlevels;
1423 /* dirty the left indirects */
1424 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1426 new = dbuf_dirty(db, tx);
1427 dbuf_rele(db, FTAG);
1429 /* transfer the dirty records to the new indirect */
1430 mutex_enter(&dn->dn_mtx);
1431 mutex_enter(&new->dt.di.dr_mtx);
1432 list = &dn->dn_dirty_records[txgoff];
1433 for (dr = list_head(list); dr; dr = dr_next) {
1434 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1435 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1436 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1437 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1438 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1439 list_remove(&dn->dn_dirty_records[txgoff], dr);
1440 list_insert_tail(&new->dt.di.dr_children, dr);
1441 dr->dr_parent = new;
1444 mutex_exit(&new->dt.di.dr_mtx);
1445 mutex_exit(&dn->dn_mtx);
1450 rw_downgrade(&dn->dn_struct_rwlock);
1454 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1457 uint64_t blkoff, blkid, nblks;
1458 int blksz, blkshift, head, tail;
1462 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1463 blksz = dn->dn_datablksz;
1464 blkshift = dn->dn_datablkshift;
1465 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1467 if (len == DMU_OBJECT_END) {
1468 len = UINT64_MAX - off;
1473 * First, block align the region to free:
1476 head = P2NPHASE(off, blksz);
1477 blkoff = P2PHASE(off, blksz);
1478 if ((off >> blkshift) > dn->dn_maxblkid)
1481 ASSERT(dn->dn_maxblkid == 0);
1482 if (off == 0 && len >= blksz) {
1484 * Freeing the whole block; fast-track this request.
1485 * Note that we won't dirty any indirect blocks,
1486 * which is fine because we will be freeing the entire
1487 * file and thus all indirect blocks will be freed
1488 * by free_children().
1493 } else if (off >= blksz) {
1494 /* Freeing past end-of-data */
1497 /* Freeing part of the block. */
1499 ASSERT3U(head, >, 0);
1503 /* zero out any partial block data at the start of the range */
1505 ASSERT3U(blkoff + head, ==, blksz);
1508 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1512 /* don't dirty if it isn't on disk and isn't dirty */
1513 if (db->db_last_dirty ||
1514 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1515 rw_exit(&dn->dn_struct_rwlock);
1516 dmu_buf_will_dirty(&db->db, tx);
1517 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1518 data = db->db.db_data;
1519 bzero(data + blkoff, head);
1521 dbuf_rele(db, FTAG);
1527 /* If the range was less than one block, we're done */
1531 /* If the remaining range is past end of file, we're done */
1532 if ((off >> blkshift) > dn->dn_maxblkid)
1535 ASSERT(ISP2(blksz));
1539 tail = P2PHASE(len, blksz);
1541 ASSERT0(P2PHASE(off, blksz));
1542 /* zero out any partial block data at the end of the range */
1546 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1547 TRUE, FTAG, &db) == 0) {
1548 /* don't dirty if not on disk and not dirty */
1549 if (db->db_last_dirty ||
1550 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1551 rw_exit(&dn->dn_struct_rwlock);
1552 dmu_buf_will_dirty(&db->db, tx);
1553 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1554 bzero(db->db.db_data, tail);
1556 dbuf_rele(db, FTAG);
1561 /* If the range did not include a full block, we are done */
1565 ASSERT(IS_P2ALIGNED(off, blksz));
1566 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1567 blkid = off >> blkshift;
1568 nblks = len >> blkshift;
1573 * Dirty the first and last indirect blocks, as they (and/or their
1574 * parents) will need to be written out if they were only
1575 * partially freed. Interior indirect blocks will be themselves freed,
1576 * by free_children(), so they need not be dirtied. Note that these
1577 * interior blocks have already been prefetched by dmu_tx_hold_free().
1579 if (dn->dn_nlevels > 1) {
1580 uint64_t first, last;
1582 first = blkid >> epbs;
1583 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1584 dmu_buf_will_dirty(&db->db, tx);
1585 dbuf_rele(db, FTAG);
1588 last = dn->dn_maxblkid >> epbs;
1590 last = (blkid + nblks - 1) >> epbs;
1591 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1592 dmu_buf_will_dirty(&db->db, tx);
1593 dbuf_rele(db, FTAG);
1599 * Add this range to the dnode range list.
1600 * We will finish up this free operation in the syncing phase.
1602 mutex_enter(&dn->dn_mtx);
1603 int txgoff = tx->tx_txg & TXG_MASK;
1604 if (dn->dn_free_ranges[txgoff] == NULL) {
1605 dn->dn_free_ranges[txgoff] =
1606 range_tree_create(NULL, NULL, &dn->dn_mtx);
1608 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1609 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1610 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1611 blkid, nblks, tx->tx_txg);
1612 mutex_exit(&dn->dn_mtx);
1614 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1615 dnode_setdirty(dn, tx);
1618 rw_exit(&dn->dn_struct_rwlock);
1622 dnode_spill_freed(dnode_t *dn)
1626 mutex_enter(&dn->dn_mtx);
1627 for (i = 0; i < TXG_SIZE; i++) {
1628 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1631 mutex_exit(&dn->dn_mtx);
1632 return (i < TXG_SIZE);
1635 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1637 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1639 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1642 if (blkid == DMU_BONUS_BLKID)
1646 * If we're in the process of opening the pool, dp will not be
1647 * set yet, but there shouldn't be anything dirty.
1652 if (dn->dn_free_txg)
1655 if (blkid == DMU_SPILL_BLKID)
1656 return (dnode_spill_freed(dn));
1658 mutex_enter(&dn->dn_mtx);
1659 for (i = 0; i < TXG_SIZE; i++) {
1660 if (dn->dn_free_ranges[i] != NULL &&
1661 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1664 mutex_exit(&dn->dn_mtx);
1665 return (i < TXG_SIZE);
1668 /* call from syncing context when we actually write/free space for this dnode */
1670 dnode_diduse_space(dnode_t *dn, int64_t delta)
1673 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1675 (u_longlong_t)dn->dn_phys->dn_used,
1678 mutex_enter(&dn->dn_mtx);
1679 space = DN_USED_BYTES(dn->dn_phys);
1681 ASSERT3U(space + delta, >=, space); /* no overflow */
1683 ASSERT3U(space, >=, -delta); /* no underflow */
1686 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1687 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1688 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1689 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1691 dn->dn_phys->dn_used = space;
1692 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1694 mutex_exit(&dn->dn_mtx);
1698 * Call when we think we're going to write/free space in open context to track
1699 * the amount of memory in use by the currently open txg.
1702 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1704 objset_t *os = dn->dn_objset;
1705 dsl_dataset_t *ds = os->os_dsl_dataset;
1706 int64_t aspace = spa_get_asize(os->os_spa, space);
1709 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1710 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1713 dmu_tx_willuse_space(tx, aspace);
1717 * Scans a block at the indicated "level" looking for a hole or data,
1718 * depending on 'flags'.
1720 * If level > 0, then we are scanning an indirect block looking at its
1721 * pointers. If level == 0, then we are looking at a block of dnodes.
1723 * If we don't find what we are looking for in the block, we return ESRCH.
1724 * Otherwise, return with *offset pointing to the beginning (if searching
1725 * forwards) or end (if searching backwards) of the range covered by the
1726 * block pointer we matched on (or dnode).
1728 * The basic search algorithm used below by dnode_next_offset() is to
1729 * use this function to search up the block tree (widen the search) until
1730 * we find something (i.e., we don't return ESRCH) and then search back
1731 * down the tree (narrow the search) until we reach our original search
1735 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1736 int lvl, uint64_t blkfill, uint64_t txg)
1738 dmu_buf_impl_t *db = NULL;
1740 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1741 uint64_t epb = 1ULL << epbs;
1742 uint64_t minfill, maxfill;
1744 int i, inc, error, span;
1746 dprintf("probing object %llu offset %llx level %d of %u\n",
1747 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1749 hole = ((flags & DNODE_FIND_HOLE) != 0);
1750 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1751 ASSERT(txg == 0 || !hole);
1753 if (lvl == dn->dn_phys->dn_nlevels) {
1755 epb = dn->dn_phys->dn_nblkptr;
1756 data = dn->dn_phys->dn_blkptr;
1758 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1759 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1761 if (error != ENOENT)
1766 * This can only happen when we are searching up
1767 * the block tree for data. We don't really need to
1768 * adjust the offset, as we will just end up looking
1769 * at the pointer to this block in its parent, and its
1770 * going to be unallocated, so we will skip over it.
1772 return (SET_ERROR(ESRCH));
1774 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1776 dbuf_rele(db, FTAG);
1779 data = db->db.db_data;
1783 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1784 db->db_blkptr->blk_birth <= txg ||
1785 BP_IS_HOLE(db->db_blkptr))) {
1787 * This can only happen when we are searching up the tree
1788 * and these conditions mean that we need to keep climbing.
1790 error = SET_ERROR(ESRCH);
1791 } else if (lvl == 0) {
1792 dnode_phys_t *dnp = data;
1794 ASSERT(dn->dn_type == DMU_OT_DNODE);
1796 for (i = (*offset >> span) & (blkfill - 1);
1797 i >= 0 && i < blkfill; i += inc) {
1798 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1800 *offset += (1ULL << span) * inc;
1802 if (i < 0 || i == blkfill)
1803 error = SET_ERROR(ESRCH);
1805 blkptr_t *bp = data;
1806 uint64_t start = *offset;
1807 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1809 maxfill = blkfill << ((lvl - 1) * epbs);
1816 *offset = *offset >> span;
1817 for (i = BF64_GET(*offset, 0, epbs);
1818 i >= 0 && i < epb; i += inc) {
1819 if (bp[i].blk_fill >= minfill &&
1820 bp[i].blk_fill <= maxfill &&
1821 (hole || bp[i].blk_birth > txg))
1823 if (inc > 0 || *offset > 0)
1826 *offset = *offset << span;
1828 /* traversing backwards; position offset at the end */
1829 ASSERT3U(*offset, <=, start);
1830 *offset = MIN(*offset + (1ULL << span) - 1, start);
1831 } else if (*offset < start) {
1834 if (i < 0 || i >= epb)
1835 error = SET_ERROR(ESRCH);
1839 dbuf_rele(db, FTAG);
1845 * Find the next hole, data, or sparse region at or after *offset.
1846 * The value 'blkfill' tells us how many items we expect to find
1847 * in an L0 data block; this value is 1 for normal objects,
1848 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1849 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1853 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1854 * Finds the next/previous hole/data in a file.
1855 * Used in dmu_offset_next().
1857 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1858 * Finds the next free/allocated dnode an objset's meta-dnode.
1859 * Only finds objects that have new contents since txg (ie.
1860 * bonus buffer changes and content removal are ignored).
1861 * Used in dmu_object_next().
1863 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1864 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1865 * Used in dmu_object_alloc().
1868 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1869 int minlvl, uint64_t blkfill, uint64_t txg)
1871 uint64_t initial_offset = *offset;
1875 if (!(flags & DNODE_FIND_HAVELOCK))
1876 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1878 if (dn->dn_phys->dn_nlevels == 0) {
1879 error = SET_ERROR(ESRCH);
1883 if (dn->dn_datablkshift == 0) {
1884 if (*offset < dn->dn_datablksz) {
1885 if (flags & DNODE_FIND_HOLE)
1886 *offset = dn->dn_datablksz;
1888 error = SET_ERROR(ESRCH);
1893 maxlvl = dn->dn_phys->dn_nlevels;
1895 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1896 error = dnode_next_offset_level(dn,
1897 flags, offset, lvl, blkfill, txg);
1902 while (error == 0 && --lvl >= minlvl) {
1903 error = dnode_next_offset_level(dn,
1904 flags, offset, lvl, blkfill, txg);
1907 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1908 initial_offset < *offset : initial_offset > *offset))
1909 error = SET_ERROR(ESRCH);
1911 if (!(flags & DNODE_FIND_HAVELOCK))
1912 rw_exit(&dn->dn_struct_rwlock);