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.
25 #include <sys/zfs_context.h>
27 #include <sys/dnode.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dsl_dataset.h>
36 #include <sys/dmu_zfetch.h>
38 static int free_range_compar(const void *node1, const void *node2);
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 ASSERTV(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);
76 refcount_create(&dn->dn_holds);
77 refcount_create(&dn->dn_tx_holds);
78 list_link_init(&dn->dn_link);
80 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
81 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
82 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
83 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
84 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
85 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
86 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
88 for (i = 0; i < TXG_SIZE; i++) {
89 list_link_init(&dn->dn_dirty_link[i]);
90 avl_create(&dn->dn_ranges[i], free_range_compar,
91 sizeof (free_range_t),
92 offsetof(struct free_range, fr_node));
93 list_create(&dn->dn_dirty_records[i],
94 sizeof (dbuf_dirty_record_t),
95 offsetof(dbuf_dirty_record_t, dr_dirty_node));
98 dn->dn_allocated_txg = 0;
100 dn->dn_assigned_txg = 0;
102 dn->dn_dirtyctx_firstset = NULL;
104 dn->dn_have_spill = B_FALSE;
114 dn->dn_dbufs_count = 0;
115 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
116 offsetof(dmu_buf_impl_t, db_link));
124 dnode_dest(void *arg, void *unused)
129 rw_destroy(&dn->dn_struct_rwlock);
130 mutex_destroy(&dn->dn_mtx);
131 mutex_destroy(&dn->dn_dbufs_mtx);
132 cv_destroy(&dn->dn_notxholds);
133 refcount_destroy(&dn->dn_holds);
134 refcount_destroy(&dn->dn_tx_holds);
135 ASSERT(!list_link_active(&dn->dn_link));
137 for (i = 0; i < TXG_SIZE; i++) {
138 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
139 avl_destroy(&dn->dn_ranges[i]);
140 list_destroy(&dn->dn_dirty_records[i]);
141 ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
142 ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
143 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
144 ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
145 ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
146 ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
147 ASSERT3U(dn->dn_next_blksz[i], ==, 0);
150 ASSERT3U(dn->dn_allocated_txg, ==, 0);
151 ASSERT3U(dn->dn_free_txg, ==, 0);
152 ASSERT3U(dn->dn_assigned_txg, ==, 0);
153 ASSERT3U(dn->dn_dirtyctx, ==, 0);
154 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
155 ASSERT3P(dn->dn_bonus, ==, NULL);
156 ASSERT(!dn->dn_have_spill);
157 ASSERT3P(dn->dn_zio, ==, NULL);
158 ASSERT3U(dn->dn_oldused, ==, 0);
159 ASSERT3U(dn->dn_oldflags, ==, 0);
160 ASSERT3U(dn->dn_olduid, ==, 0);
161 ASSERT3U(dn->dn_oldgid, ==, 0);
162 ASSERT3U(dn->dn_newuid, ==, 0);
163 ASSERT3U(dn->dn_newgid, ==, 0);
164 ASSERT3U(dn->dn_id_flags, ==, 0);
166 ASSERT3U(dn->dn_dbufs_count, ==, 0);
167 list_destroy(&dn->dn_dbufs);
173 ASSERT(dnode_cache == NULL);
174 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
175 0, dnode_cons, dnode_dest, NULL, NULL, NULL, KMC_KMEM);
176 kmem_cache_set_move(dnode_cache, dnode_move);
182 kmem_cache_destroy(dnode_cache);
189 dnode_verify(dnode_t *dn)
191 int drop_struct_lock = FALSE;
194 ASSERT(dn->dn_objset);
195 ASSERT(dn->dn_handle->dnh_dnode == dn);
197 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
199 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
202 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
203 rw_enter(&dn->dn_struct_rwlock, RW_READER);
204 drop_struct_lock = TRUE;
206 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
208 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
209 if (dn->dn_datablkshift) {
210 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
211 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
212 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
214 ASSERT3U(dn->dn_nlevels, <=, 30);
215 ASSERT3U(dn->dn_type, <=, DMU_OT_NUMTYPES);
216 ASSERT3U(dn->dn_nblkptr, >=, 1);
217 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
218 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
219 ASSERT3U(dn->dn_datablksz, ==,
220 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
221 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
222 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
223 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
224 for (i = 0; i < TXG_SIZE; i++) {
225 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
228 if (dn->dn_phys->dn_type != DMU_OT_NONE)
229 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
230 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
231 if (dn->dn_dbuf != NULL) {
232 ASSERT3P(dn->dn_phys, ==,
233 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
234 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
236 if (drop_struct_lock)
237 rw_exit(&dn->dn_struct_rwlock);
242 dnode_byteswap(dnode_phys_t *dnp)
244 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
247 if (dnp->dn_type == DMU_OT_NONE) {
248 bzero(dnp, sizeof (dnode_phys_t));
252 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
253 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
254 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
255 dnp->dn_used = BSWAP_64(dnp->dn_used);
258 * dn_nblkptr is only one byte, so it's OK to read it in either
259 * byte order. We can't read dn_bouslen.
261 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
262 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
263 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
264 buf64[i] = BSWAP_64(buf64[i]);
267 * OK to check dn_bonuslen for zero, because it won't matter if
268 * we have the wrong byte order. This is necessary because the
269 * dnode dnode is smaller than a regular dnode.
271 if (dnp->dn_bonuslen != 0) {
273 * Note that the bonus length calculated here may be
274 * longer than the actual bonus buffer. This is because
275 * we always put the bonus buffer after the last block
276 * pointer (instead of packing it against the end of the
279 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
280 size_t len = DN_MAX_BONUSLEN - off;
281 ASSERT3U(dnp->dn_bonustype, <, DMU_OT_NUMTYPES);
282 dmu_ot[dnp->dn_bonustype].ot_byteswap(dnp->dn_bonus + off, len);
285 /* Swap SPILL block if we have one */
286 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
287 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
292 dnode_buf_byteswap(void *vbuf, size_t size)
294 dnode_phys_t *buf = vbuf;
297 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
298 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
300 size >>= DNODE_SHIFT;
301 for (i = 0; i < size; i++) {
308 free_range_compar(const void *node1, const void *node2)
310 const free_range_t *rp1 = node1;
311 const free_range_t *rp2 = node2;
313 if (rp1->fr_blkid < rp2->fr_blkid)
315 else if (rp1->fr_blkid > rp2->fr_blkid)
321 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
323 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
325 dnode_setdirty(dn, tx);
326 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
327 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
328 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
329 dn->dn_bonuslen = newsize;
331 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
333 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
334 rw_exit(&dn->dn_struct_rwlock);
338 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
340 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
341 dnode_setdirty(dn, tx);
342 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
343 dn->dn_bonustype = newtype;
344 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
345 rw_exit(&dn->dn_struct_rwlock);
349 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
351 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
352 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
353 dnode_setdirty(dn, tx);
354 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
355 dn->dn_have_spill = B_FALSE;
359 dnode_setdblksz(dnode_t *dn, int size)
361 ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0);
362 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
363 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
364 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
365 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
366 dn->dn_datablksz = size;
367 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
368 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0;
372 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
373 uint64_t object, dnode_handle_t *dnh)
375 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_PUSHPAGE);
377 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
381 * Defer setting dn_objset until the dnode is ready to be a candidate
382 * for the dnode_move() callback.
384 dn->dn_object = object;
389 if (dnp->dn_datablkszsec) {
390 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
392 dn->dn_datablksz = 0;
393 dn->dn_datablkszsec = 0;
394 dn->dn_datablkshift = 0;
396 dn->dn_indblkshift = dnp->dn_indblkshift;
397 dn->dn_nlevels = dnp->dn_nlevels;
398 dn->dn_type = dnp->dn_type;
399 dn->dn_nblkptr = dnp->dn_nblkptr;
400 dn->dn_checksum = dnp->dn_checksum;
401 dn->dn_compress = dnp->dn_compress;
402 dn->dn_bonustype = dnp->dn_bonustype;
403 dn->dn_bonuslen = dnp->dn_bonuslen;
404 dn->dn_maxblkid = dnp->dn_maxblkid;
405 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
408 dmu_zfetch_init(&dn->dn_zfetch, dn);
410 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES);
412 mutex_enter(&os->os_lock);
413 list_insert_head(&os->os_dnodes, dn);
416 * Everything else must be valid before assigning dn_objset makes the
417 * dnode eligible for dnode_move().
420 mutex_exit(&os->os_lock);
422 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
427 * Caller must be holding the dnode handle, which is released upon return.
430 dnode_destroy(dnode_t *dn)
432 objset_t *os = dn->dn_objset;
434 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
436 mutex_enter(&os->os_lock);
437 POINTER_INVALIDATE(&dn->dn_objset);
438 list_remove(&os->os_dnodes, dn);
439 mutex_exit(&os->os_lock);
441 /* the dnode can no longer move, so we can release the handle */
442 zrl_remove(&dn->dn_handle->dnh_zrlock);
444 dn->dn_allocated_txg = 0;
446 dn->dn_assigned_txg = 0;
449 if (dn->dn_dirtyctx_firstset != NULL) {
450 kmem_free(dn->dn_dirtyctx_firstset, 1);
451 dn->dn_dirtyctx_firstset = NULL;
453 if (dn->dn_bonus != NULL) {
454 mutex_enter(&dn->dn_bonus->db_mtx);
455 dbuf_evict(dn->dn_bonus);
460 dn->dn_have_spill = B_FALSE;
469 dmu_zfetch_rele(&dn->dn_zfetch);
470 kmem_cache_free(dnode_cache, dn);
471 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
475 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
476 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
481 blocksize = 1 << zfs_default_bs;
482 else if (blocksize > SPA_MAXBLOCKSIZE)
483 blocksize = SPA_MAXBLOCKSIZE;
485 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
488 ibs = zfs_default_ibs;
490 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
492 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
493 dn->dn_object, tx->tx_txg, blocksize, ibs);
495 ASSERT(dn->dn_type == DMU_OT_NONE);
496 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
497 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
498 ASSERT(ot != DMU_OT_NONE);
499 ASSERT3U(ot, <, DMU_OT_NUMTYPES);
500 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
501 (bonustype == DMU_OT_SA && bonuslen == 0) ||
502 (bonustype != DMU_OT_NONE && bonuslen != 0));
503 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
504 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
505 ASSERT(dn->dn_type == DMU_OT_NONE);
506 ASSERT3U(dn->dn_maxblkid, ==, 0);
507 ASSERT3U(dn->dn_allocated_txg, ==, 0);
508 ASSERT3U(dn->dn_assigned_txg, ==, 0);
509 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
510 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
511 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
513 for (i = 0; i < TXG_SIZE; i++) {
514 ASSERT3U(dn->dn_next_nblkptr[i], ==, 0);
515 ASSERT3U(dn->dn_next_nlevels[i], ==, 0);
516 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0);
517 ASSERT3U(dn->dn_next_bonuslen[i], ==, 0);
518 ASSERT3U(dn->dn_next_bonustype[i], ==, 0);
519 ASSERT3U(dn->dn_rm_spillblk[i], ==, 0);
520 ASSERT3U(dn->dn_next_blksz[i], ==, 0);
521 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
522 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
523 ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0);
527 dnode_setdblksz(dn, blocksize);
528 dn->dn_indblkshift = ibs;
530 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
534 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
535 dn->dn_bonustype = bonustype;
536 dn->dn_bonuslen = bonuslen;
537 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
538 dn->dn_compress = ZIO_COMPRESS_INHERIT;
542 if (dn->dn_dirtyctx_firstset) {
543 kmem_free(dn->dn_dirtyctx_firstset, 1);
544 dn->dn_dirtyctx_firstset = NULL;
547 dn->dn_allocated_txg = tx->tx_txg;
550 dnode_setdirty(dn, tx);
551 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
552 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
553 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
554 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
558 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
559 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
563 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
564 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
565 ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0);
566 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
567 ASSERT(tx->tx_txg != 0);
568 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
569 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
570 (bonustype == DMU_OT_SA && bonuslen == 0));
571 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES);
572 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
574 /* clean up any unreferenced dbufs */
575 dnode_evict_dbufs(dn);
579 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
580 dnode_setdirty(dn, tx);
581 if (dn->dn_datablksz != blocksize) {
582 /* change blocksize */
583 ASSERT(dn->dn_maxblkid == 0 &&
584 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
585 dnode_block_freed(dn, 0)));
586 dnode_setdblksz(dn, blocksize);
587 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
589 if (dn->dn_bonuslen != bonuslen)
590 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
592 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
595 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
596 if (dn->dn_bonustype != bonustype)
597 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
598 if (dn->dn_nblkptr != nblkptr)
599 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
600 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
601 dbuf_rm_spill(dn, tx);
602 dnode_rm_spill(dn, tx);
604 rw_exit(&dn->dn_struct_rwlock);
609 /* change bonus size and type */
610 mutex_enter(&dn->dn_mtx);
611 dn->dn_bonustype = bonustype;
612 dn->dn_bonuslen = bonuslen;
613 dn->dn_nblkptr = nblkptr;
614 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
615 dn->dn_compress = ZIO_COMPRESS_INHERIT;
616 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
618 /* fix up the bonus db_size */
620 dn->dn_bonus->db.db_size =
621 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
622 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
625 dn->dn_allocated_txg = tx->tx_txg;
626 mutex_exit(&dn->dn_mtx);
632 uint64_t dms_dnode_invalid;
633 uint64_t dms_dnode_recheck1;
634 uint64_t dms_dnode_recheck2;
635 uint64_t dms_dnode_special;
636 uint64_t dms_dnode_handle;
637 uint64_t dms_dnode_rwlock;
638 uint64_t dms_dnode_active;
640 #endif /* DNODE_STATS */
643 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
647 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
648 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
649 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
650 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
653 ndn->dn_objset = odn->dn_objset;
654 ndn->dn_object = odn->dn_object;
655 ndn->dn_dbuf = odn->dn_dbuf;
656 ndn->dn_handle = odn->dn_handle;
657 ndn->dn_phys = odn->dn_phys;
658 ndn->dn_type = odn->dn_type;
659 ndn->dn_bonuslen = odn->dn_bonuslen;
660 ndn->dn_bonustype = odn->dn_bonustype;
661 ndn->dn_nblkptr = odn->dn_nblkptr;
662 ndn->dn_checksum = odn->dn_checksum;
663 ndn->dn_compress = odn->dn_compress;
664 ndn->dn_nlevels = odn->dn_nlevels;
665 ndn->dn_indblkshift = odn->dn_indblkshift;
666 ndn->dn_datablkshift = odn->dn_datablkshift;
667 ndn->dn_datablkszsec = odn->dn_datablkszsec;
668 ndn->dn_datablksz = odn->dn_datablksz;
669 ndn->dn_maxblkid = odn->dn_maxblkid;
670 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
671 sizeof (odn->dn_next_nblkptr));
672 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
673 sizeof (odn->dn_next_nlevels));
674 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
675 sizeof (odn->dn_next_indblkshift));
676 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
677 sizeof (odn->dn_next_bonustype));
678 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
679 sizeof (odn->dn_rm_spillblk));
680 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
681 sizeof (odn->dn_next_bonuslen));
682 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
683 sizeof (odn->dn_next_blksz));
684 for (i = 0; i < TXG_SIZE; i++) {
685 list_move_tail(&ndn->dn_dirty_records[i],
686 &odn->dn_dirty_records[i]);
688 bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges));
689 ndn->dn_allocated_txg = odn->dn_allocated_txg;
690 ndn->dn_free_txg = odn->dn_free_txg;
691 ndn->dn_assigned_txg = odn->dn_assigned_txg;
692 ndn->dn_dirtyctx = odn->dn_dirtyctx;
693 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
694 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
695 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
696 ASSERT(list_is_empty(&ndn->dn_dbufs));
697 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
698 ndn->dn_dbufs_count = odn->dn_dbufs_count;
699 ndn->dn_bonus = odn->dn_bonus;
700 ndn->dn_have_spill = odn->dn_have_spill;
701 ndn->dn_zio = odn->dn_zio;
702 ndn->dn_oldused = odn->dn_oldused;
703 ndn->dn_oldflags = odn->dn_oldflags;
704 ndn->dn_olduid = odn->dn_olduid;
705 ndn->dn_oldgid = odn->dn_oldgid;
706 ndn->dn_newuid = odn->dn_newuid;
707 ndn->dn_newgid = odn->dn_newgid;
708 ndn->dn_id_flags = odn->dn_id_flags;
709 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
710 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
711 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
712 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
713 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
716 * Update back pointers. Updating the handle fixes the back pointer of
717 * every descendant dbuf as well as the bonus dbuf.
719 ASSERT(ndn->dn_handle->dnh_dnode == odn);
720 ndn->dn_handle->dnh_dnode = ndn;
721 if (ndn->dn_zfetch.zf_dnode == odn) {
722 ndn->dn_zfetch.zf_dnode = ndn;
726 * Invalidate the original dnode by clearing all of its back pointers.
729 odn->dn_handle = NULL;
730 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
731 offsetof(dmu_buf_impl_t, db_link));
732 odn->dn_dbufs_count = 0;
733 odn->dn_bonus = NULL;
734 odn->dn_zfetch.zf_dnode = NULL;
737 * Set the low bit of the objset pointer to ensure that dnode_move()
738 * recognizes the dnode as invalid in any subsequent callback.
740 POINTER_INVALIDATE(&odn->dn_objset);
743 * Satisfy the destructor.
745 for (i = 0; i < TXG_SIZE; i++) {
746 list_create(&odn->dn_dirty_records[i],
747 sizeof (dbuf_dirty_record_t),
748 offsetof(dbuf_dirty_record_t, dr_dirty_node));
749 odn->dn_ranges[i].avl_root = NULL;
750 odn->dn_ranges[i].avl_numnodes = 0;
751 odn->dn_next_nlevels[i] = 0;
752 odn->dn_next_indblkshift[i] = 0;
753 odn->dn_next_bonustype[i] = 0;
754 odn->dn_rm_spillblk[i] = 0;
755 odn->dn_next_bonuslen[i] = 0;
756 odn->dn_next_blksz[i] = 0;
758 odn->dn_allocated_txg = 0;
759 odn->dn_free_txg = 0;
760 odn->dn_assigned_txg = 0;
761 odn->dn_dirtyctx = 0;
762 odn->dn_dirtyctx_firstset = NULL;
763 odn->dn_have_spill = B_FALSE;
766 odn->dn_oldflags = 0;
771 odn->dn_id_flags = 0;
777 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);
923 dnode_special_close(dnode_handle_t *dnh)
925 dnode_t *dn = dnh->dnh_dnode;
928 * Wait for final references to the dnode to clear. This can
929 * only happen if the arc is asyncronously evicting state that
930 * has a hold on this dnode while we are trying to evict this
933 while (refcount_count(&dn->dn_holds) > 0)
935 zrl_add(&dnh->dnh_zrlock);
936 dnode_destroy(dn); /* implicit zrl_remove() */
937 zrl_destroy(&dnh->dnh_zrlock);
938 dnh->dnh_dnode = NULL;
942 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
945 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
947 zrl_init(&dnh->dnh_zrlock);
953 dnode_buf_pageout(dmu_buf_t *db, void *arg)
955 dnode_children_t *children_dnodes = arg;
957 int epb = db->db_size >> DNODE_SHIFT;
959 ASSERT(epb == children_dnodes->dnc_count);
961 for (i = 0; i < epb; i++) {
962 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
966 * The dnode handle lock guards against the dnode moving to
967 * another valid address, so there is no need here to guard
968 * against changes to or from NULL.
970 if (dnh->dnh_dnode == NULL) {
971 zrl_destroy(&dnh->dnh_zrlock);
975 zrl_add(&dnh->dnh_zrlock);
978 * If there are holds on this dnode, then there should
979 * be holds on the dnode's containing dbuf as well; thus
980 * it wouldn't be eligible for eviction and this function
981 * would not have been called.
983 ASSERT(refcount_is_zero(&dn->dn_holds));
984 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
986 dnode_destroy(dn); /* implicit zrl_remove() */
987 zrl_destroy(&dnh->dnh_zrlock);
988 dnh->dnh_dnode = NULL;
990 kmem_free(children_dnodes, sizeof (dnode_children_t) +
991 (epb - 1) * sizeof (dnode_handle_t));
996 * EINVAL - invalid object number.
998 * succeeds even for free dnodes.
1001 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1002 void *tag, dnode_t **dnp)
1005 int drop_struct_lock = FALSE;
1010 dnode_children_t *children_dnodes;
1011 dnode_handle_t *dnh;
1014 * If you are holding the spa config lock as writer, you shouldn't
1015 * be asking the DMU to do *anything* unless it's the root pool
1016 * which may require us to read from the root filesystem while
1017 * holding some (not all) of the locks as writer.
1019 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1020 (spa_is_root(os->os_spa) &&
1021 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1023 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1024 dn = (object == DMU_USERUSED_OBJECT) ?
1025 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1029 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1031 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1034 (void) refcount_add(&dn->dn_holds, tag);
1039 if (object == 0 || object >= DN_MAX_OBJECT)
1042 mdn = DMU_META_DNODE(os);
1043 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1047 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1048 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1049 drop_struct_lock = TRUE;
1052 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1054 db = dbuf_hold(mdn, blk, FTAG);
1055 if (drop_struct_lock)
1056 rw_exit(&mdn->dn_struct_rwlock);
1059 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1061 dbuf_rele(db, FTAG);
1065 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1066 epb = db->db.db_size >> DNODE_SHIFT;
1068 idx = object & (epb-1);
1070 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1071 children_dnodes = dmu_buf_get_user(&db->db);
1072 if (children_dnodes == NULL) {
1074 dnode_children_t *winner;
1075 children_dnodes = kmem_alloc(sizeof (dnode_children_t) +
1076 (epb - 1) * sizeof (dnode_handle_t),
1077 KM_PUSHPAGE | KM_NODEBUG);
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 ASSERT3U(dn->dn_next_bonuslen[txg&TXG_MASK], ==, 0);
1237 ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0);
1238 ASSERT3U(dn->dn_next_bonustype[txg&TXG_MASK], ==, 0);
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_phys->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);
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_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx)
1456 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1459 free_range_t rp_tofind;
1460 uint64_t endblk = blkid + nblks;
1462 ASSERT(MUTEX_HELD(&dn->dn_mtx));
1463 ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */
1465 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1466 blkid, nblks, tx->tx_txg);
1467 rp_tofind.fr_blkid = blkid;
1468 rp = avl_find(tree, &rp_tofind, &where);
1470 rp = avl_nearest(tree, where, AVL_BEFORE);
1472 rp = avl_nearest(tree, where, AVL_AFTER);
1474 while (rp && (rp->fr_blkid <= blkid + nblks)) {
1475 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks;
1476 free_range_t *nrp = AVL_NEXT(tree, rp);
1478 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) {
1479 /* clear this entire range */
1480 avl_remove(tree, rp);
1481 kmem_free(rp, sizeof (free_range_t));
1482 } else if (blkid <= rp->fr_blkid &&
1483 endblk > rp->fr_blkid && endblk < fr_endblk) {
1484 /* clear the beginning of this range */
1485 rp->fr_blkid = endblk;
1486 rp->fr_nblks = fr_endblk - endblk;
1487 } else if (blkid > rp->fr_blkid && blkid < fr_endblk &&
1488 endblk >= fr_endblk) {
1489 /* clear the end of this range */
1490 rp->fr_nblks = blkid - rp->fr_blkid;
1491 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) {
1492 /* clear a chunk out of this range */
1493 free_range_t *new_rp =
1494 kmem_alloc(sizeof (free_range_t), KM_PUSHPAGE);
1496 new_rp->fr_blkid = endblk;
1497 new_rp->fr_nblks = fr_endblk - endblk;
1498 avl_insert_here(tree, new_rp, rp, AVL_AFTER);
1499 rp->fr_nblks = blkid - rp->fr_blkid;
1501 /* there may be no overlap */
1507 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1510 uint64_t blkoff, blkid, nblks;
1511 int blksz, blkshift, head, tail;
1515 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1516 blksz = dn->dn_datablksz;
1517 blkshift = dn->dn_datablkshift;
1518 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1521 len = UINT64_MAX - off;
1526 * First, block align the region to free:
1529 head = P2NPHASE(off, blksz);
1530 blkoff = P2PHASE(off, blksz);
1531 if ((off >> blkshift) > dn->dn_maxblkid)
1534 ASSERT(dn->dn_maxblkid == 0);
1535 if (off == 0 && len >= blksz) {
1536 /* Freeing the whole block; fast-track this request */
1540 } else if (off >= blksz) {
1541 /* Freeing past end-of-data */
1544 /* Freeing part of the block. */
1546 ASSERT3U(head, >, 0);
1550 /* zero out any partial block data at the start of the range */
1552 ASSERT3U(blkoff + head, ==, blksz);
1555 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1559 /* don't dirty if it isn't on disk and isn't dirty */
1560 if (db->db_last_dirty ||
1561 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1562 rw_exit(&dn->dn_struct_rwlock);
1563 dbuf_will_dirty(db, tx);
1564 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1565 data = db->db.db_data;
1566 bzero(data + blkoff, head);
1568 dbuf_rele(db, FTAG);
1574 /* If the range was less than one block, we're done */
1578 /* If the remaining range is past end of file, we're done */
1579 if ((off >> blkshift) > dn->dn_maxblkid)
1582 ASSERT(ISP2(blksz));
1586 tail = P2PHASE(len, blksz);
1588 ASSERT3U(P2PHASE(off, blksz), ==, 0);
1589 /* zero out any partial block data at the end of the range */
1593 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1594 TRUE, FTAG, &db) == 0) {
1595 /* don't dirty if not on disk and not dirty */
1596 if (db->db_last_dirty ||
1597 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1598 rw_exit(&dn->dn_struct_rwlock);
1599 dbuf_will_dirty(db, tx);
1600 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1601 bzero(db->db.db_data, tail);
1603 dbuf_rele(db, FTAG);
1608 /* If the range did not include a full block, we are done */
1612 ASSERT(IS_P2ALIGNED(off, blksz));
1613 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1614 blkid = off >> blkshift;
1615 nblks = len >> blkshift;
1620 * Read in and mark all the level-1 indirects dirty,
1621 * so that they will stay in memory until syncing phase.
1622 * Always dirty the first and last indirect to make sure
1623 * we dirty all the partial indirects.
1625 if (dn->dn_nlevels > 1) {
1626 uint64_t i, first, last;
1627 int shift = epbs + dn->dn_datablkshift;
1629 first = blkid >> epbs;
1630 if ((db = dbuf_hold_level(dn, 1, first, FTAG))) {
1631 dbuf_will_dirty(db, tx);
1632 dbuf_rele(db, FTAG);
1635 last = dn->dn_maxblkid >> epbs;
1637 last = (blkid + nblks - 1) >> epbs;
1638 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1639 dbuf_will_dirty(db, tx);
1640 dbuf_rele(db, FTAG);
1642 for (i = first + 1; i < last; i++) {
1643 uint64_t ibyte = i << shift;
1646 err = dnode_next_offset(dn,
1647 DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0);
1649 if (err == ESRCH || i >= last)
1652 db = dbuf_hold_level(dn, 1, i, FTAG);
1654 dbuf_will_dirty(db, tx);
1655 dbuf_rele(db, FTAG);
1661 * Add this range to the dnode range list.
1662 * We will finish up this free operation in the syncing phase.
1664 mutex_enter(&dn->dn_mtx);
1665 dnode_clear_range(dn, blkid, nblks, tx);
1667 free_range_t *rp, *found;
1669 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK];
1671 /* Add new range to dn_ranges */
1672 rp = kmem_alloc(sizeof (free_range_t), KM_PUSHPAGE);
1673 rp->fr_blkid = blkid;
1674 rp->fr_nblks = nblks;
1675 found = avl_find(tree, rp, &where);
1676 ASSERT(found == NULL);
1677 avl_insert(tree, rp, where);
1678 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1679 blkid, nblks, tx->tx_txg);
1681 mutex_exit(&dn->dn_mtx);
1683 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1684 dnode_setdirty(dn, tx);
1686 if (trunc && dn->dn_maxblkid >= (off >> blkshift))
1687 dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0);
1689 rw_exit(&dn->dn_struct_rwlock);
1693 dnode_spill_freed(dnode_t *dn)
1697 mutex_enter(&dn->dn_mtx);
1698 for (i = 0; i < TXG_SIZE; i++) {
1699 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1702 mutex_exit(&dn->dn_mtx);
1703 return (i < TXG_SIZE);
1706 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1708 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1710 free_range_t range_tofind;
1711 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1714 if (blkid == DMU_BONUS_BLKID)
1718 * If we're in the process of opening the pool, dp will not be
1719 * set yet, but there shouldn't be anything dirty.
1724 if (dn->dn_free_txg)
1727 if (blkid == DMU_SPILL_BLKID)
1728 return (dnode_spill_freed(dn));
1730 range_tofind.fr_blkid = blkid;
1731 mutex_enter(&dn->dn_mtx);
1732 for (i = 0; i < TXG_SIZE; i++) {
1733 free_range_t *range_found;
1736 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx);
1738 ASSERT(range_found->fr_nblks > 0);
1741 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE);
1743 range_found->fr_blkid + range_found->fr_nblks > blkid)
1746 mutex_exit(&dn->dn_mtx);
1747 return (i < TXG_SIZE);
1750 /* call from syncing context when we actually write/free space for this dnode */
1752 dnode_diduse_space(dnode_t *dn, int64_t delta)
1755 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1757 (u_longlong_t)dn->dn_phys->dn_used,
1760 mutex_enter(&dn->dn_mtx);
1761 space = DN_USED_BYTES(dn->dn_phys);
1763 ASSERT3U(space + delta, >=, space); /* no overflow */
1765 ASSERT3U(space, >=, -delta); /* no underflow */
1768 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1769 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1770 ASSERT3U(P2PHASE(space, 1<<DEV_BSHIFT), ==, 0);
1771 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1773 dn->dn_phys->dn_used = space;
1774 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1776 mutex_exit(&dn->dn_mtx);
1780 * Call when we think we're going to write/free space in open context.
1781 * Be conservative (ie. OK to write less than this or free more than
1782 * this, but don't write more or free less).
1785 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1787 objset_t *os = dn->dn_objset;
1788 dsl_dataset_t *ds = os->os_dsl_dataset;
1791 space = spa_get_asize(os->os_spa, space);
1794 dsl_dir_willuse_space(ds->ds_dir, space, tx);
1796 dmu_tx_willuse_space(tx, space);
1800 * This function scans a block at the indicated "level" looking for
1801 * a hole or data (depending on 'flags'). If level > 0, then we are
1802 * scanning an indirect block looking at its pointers. If level == 0,
1803 * then we are looking at a block of dnodes. If we don't find what we
1804 * are looking for in the block, we return ESRCH. Otherwise, return
1805 * with *offset pointing to the beginning (if searching forwards) or
1806 * end (if searching backwards) of the range covered by the block
1807 * pointer we matched on (or dnode).
1809 * The basic search algorithm used below by dnode_next_offset() is to
1810 * use this function to search up the block tree (widen the search) until
1811 * we find something (i.e., we don't return ESRCH) and then search back
1812 * down the tree (narrow the search) until we reach our original search
1816 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1817 int lvl, uint64_t blkfill, uint64_t txg)
1819 dmu_buf_impl_t *db = NULL;
1821 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1822 uint64_t epb = 1ULL << epbs;
1823 uint64_t minfill, maxfill;
1825 int i, inc, error, span;
1827 dprintf("probing object %llu offset %llx level %d of %u\n",
1828 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1830 hole = ((flags & DNODE_FIND_HOLE) != 0);
1831 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1832 ASSERT(txg == 0 || !hole);
1834 if (lvl == dn->dn_phys->dn_nlevels) {
1836 epb = dn->dn_phys->dn_nblkptr;
1837 data = dn->dn_phys->dn_blkptr;
1839 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1840 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1842 if (error != ENOENT)
1847 * This can only happen when we are searching up
1848 * the block tree for data. We don't really need to
1849 * adjust the offset, as we will just end up looking
1850 * at the pointer to this block in its parent, and its
1851 * going to be unallocated, so we will skip over it.
1855 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1857 dbuf_rele(db, FTAG);
1860 data = db->db.db_data;
1864 (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) {
1866 * This can only happen when we are searching up the tree
1867 * and these conditions mean that we need to keep climbing.
1870 } else if (lvl == 0) {
1871 dnode_phys_t *dnp = data;
1873 ASSERT(dn->dn_type == DMU_OT_DNODE);
1875 for (i = (*offset >> span) & (blkfill - 1);
1876 i >= 0 && i < blkfill; i += inc) {
1877 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1879 *offset += (1ULL << span) * inc;
1881 if (i < 0 || i == blkfill)
1884 blkptr_t *bp = data;
1885 uint64_t start = *offset;
1886 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1888 maxfill = blkfill << ((lvl - 1) * epbs);
1895 *offset = *offset >> span;
1896 for (i = BF64_GET(*offset, 0, epbs);
1897 i >= 0 && i < epb; i += inc) {
1898 if (bp[i].blk_fill >= minfill &&
1899 bp[i].blk_fill <= maxfill &&
1900 (hole || bp[i].blk_birth > txg))
1902 if (inc > 0 || *offset > 0)
1905 *offset = *offset << span;
1907 /* traversing backwards; position offset at the end */
1908 ASSERT3U(*offset, <=, start);
1909 *offset = MIN(*offset + (1ULL << span) - 1, start);
1910 } else if (*offset < start) {
1913 if (i < 0 || i >= epb)
1918 dbuf_rele(db, FTAG);
1924 * Find the next hole, data, or sparse region at or after *offset.
1925 * The value 'blkfill' tells us how many items we expect to find
1926 * in an L0 data block; this value is 1 for normal objects,
1927 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1928 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1932 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1933 * Finds the next/previous hole/data in a file.
1934 * Used in dmu_offset_next().
1936 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1937 * Finds the next free/allocated dnode an objset's meta-dnode.
1938 * Only finds objects that have new contents since txg (ie.
1939 * bonus buffer changes and content removal are ignored).
1940 * Used in dmu_object_next().
1942 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1943 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1944 * Used in dmu_object_alloc().
1947 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1948 int minlvl, uint64_t blkfill, uint64_t txg)
1950 uint64_t initial_offset = *offset;
1954 if (!(flags & DNODE_FIND_HAVELOCK))
1955 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1957 if (dn->dn_phys->dn_nlevels == 0) {
1962 if (dn->dn_datablkshift == 0) {
1963 if (*offset < dn->dn_datablksz) {
1964 if (flags & DNODE_FIND_HOLE)
1965 *offset = dn->dn_datablksz;
1972 maxlvl = dn->dn_phys->dn_nlevels;
1974 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1975 error = dnode_next_offset_level(dn,
1976 flags, offset, lvl, blkfill, txg);
1981 while (error == 0 && --lvl >= minlvl) {
1982 error = dnode_next_offset_level(dn,
1983 flags, offset, lvl, blkfill, txg);
1986 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1987 initial_offset < *offset : initial_offset > *offset))
1990 if (!(flags & DNODE_FIND_HAVELOCK))
1991 rw_exit(&dn->dn_struct_rwlock);