]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dnode.c
projects / mirror_zfs.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Don't allow dnode allocation if dn_holds != 0
[mirror_zfs.git] / module / zfs / dnode.c
1 /*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 #include <sys/trace_dnode.h>
41 #include <sys/zfs_project.h>
42
43 dnode_stats_t dnode_stats = {
44 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64 },
45 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64 },
46 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64 },
47 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64 },
48 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64 },
49 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64 },
50 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64 },
51 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64 },
52 { "dnode_hold_free_hits", KSTAT_DATA_UINT64 },
53 { "dnode_hold_free_misses", KSTAT_DATA_UINT64 },
54 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64 },
55 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64 },
56 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64 },
57 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64 },
58 { "dnode_hold_free_txg", KSTAT_DATA_UINT64 },
59 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 },
60 { "dnode_allocate", KSTAT_DATA_UINT64 },
61 { "dnode_reallocate", KSTAT_DATA_UINT64 },
62 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
63 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
64 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
65 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
66 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
67 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
68 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
69 { "dnode_move_special", KSTAT_DATA_UINT64 },
70 { "dnode_move_handle", KSTAT_DATA_UINT64 },
71 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
72 { "dnode_move_active", KSTAT_DATA_UINT64 },
73 };
74
75 static kstat_t *dnode_ksp;
76 static kmem_cache_t *dnode_cache;
77
78 ASSERTV(static dnode_phys_t dnode_phys_zero);
79
80 int zfs_default_bs = SPA_MINBLOCKSHIFT;
81 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
82
83 #ifdef _KERNEL
84 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
85 #endif /* _KERNEL */
86
87 static int
88 dbuf_compare(const void *x1, const void *x2)
89 {
90 const dmu_buf_impl_t *d1 = x1;
91 const dmu_buf_impl_t *d2 = x2;
92
93 int cmp = AVL_CMP(d1->db_level, d2->db_level);
94 if (likely(cmp))
95 return (cmp);
96
97 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
98 if (likely(cmp))
99 return (cmp);
100
101 if (d1->db_state == DB_SEARCH) {
102 ASSERT3S(d2->db_state, !=, DB_SEARCH);
103 return (-1);
104 } else if (d2->db_state == DB_SEARCH) {
105 ASSERT3S(d1->db_state, !=, DB_SEARCH);
106 return (1);
107 }
108
109 return (AVL_PCMP(d1, d2));
110 }
111
112 /* ARGSUSED */
113 static int
114 dnode_cons(void *arg, void *unused, int kmflag)
115 {
116 dnode_t *dn = arg;
117 int i;
118
119 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
120 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
121 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
122 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
123
124 /*
125 * Every dbuf has a reference, and dropping a tracked reference is
126 * O(number of references), so don't track dn_holds.
127 */
128 zfs_refcount_create_untracked(&dn->dn_holds);
129 zfs_refcount_create(&dn->dn_tx_holds);
130 list_link_init(&dn->dn_link);
131
132 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
133 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
134 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
135 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
136 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
137 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
138 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
139 bzero(&dn->dn_next_maxblkid[0], sizeof (dn->dn_next_maxblkid));
140
141 for (i = 0; i < TXG_SIZE; i++) {
142 multilist_link_init(&dn->dn_dirty_link[i]);
143 dn->dn_free_ranges[i] = NULL;
144 list_create(&dn->dn_dirty_records[i],
145 sizeof (dbuf_dirty_record_t),
146 offsetof(dbuf_dirty_record_t, dr_dirty_node));
147 }
148
149 dn->dn_allocated_txg = 0;
150 dn->dn_free_txg = 0;
151 dn->dn_assigned_txg = 0;
152 dn->dn_dirty_txg = 0;
153 dn->dn_dirtyctx = 0;
154 dn->dn_dirtyctx_firstset = NULL;
155 dn->dn_bonus = NULL;
156 dn->dn_have_spill = B_FALSE;
157 dn->dn_zio = NULL;
158 dn->dn_oldused = 0;
159 dn->dn_oldflags = 0;
160 dn->dn_olduid = 0;
161 dn->dn_oldgid = 0;
162 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
163 dn->dn_newuid = 0;
164 dn->dn_newgid = 0;
165 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
166 dn->dn_id_flags = 0;
167
168 dn->dn_dbufs_count = 0;
169 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
170 offsetof(dmu_buf_impl_t, db_link));
171
172 dn->dn_moved = 0;
173 return (0);
174 }
175
176 /* ARGSUSED */
177 static void
178 dnode_dest(void *arg, void *unused)
179 {
180 int i;
181 dnode_t *dn = arg;
182
183 rw_destroy(&dn->dn_struct_rwlock);
184 mutex_destroy(&dn->dn_mtx);
185 mutex_destroy(&dn->dn_dbufs_mtx);
186 cv_destroy(&dn->dn_notxholds);
187 zfs_refcount_destroy(&dn->dn_holds);
188 zfs_refcount_destroy(&dn->dn_tx_holds);
189 ASSERT(!list_link_active(&dn->dn_link));
190
191 for (i = 0; i < TXG_SIZE; i++) {
192 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
193 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
194 list_destroy(&dn->dn_dirty_records[i]);
195 ASSERT0(dn->dn_next_nblkptr[i]);
196 ASSERT0(dn->dn_next_nlevels[i]);
197 ASSERT0(dn->dn_next_indblkshift[i]);
198 ASSERT0(dn->dn_next_bonustype[i]);
199 ASSERT0(dn->dn_rm_spillblk[i]);
200 ASSERT0(dn->dn_next_bonuslen[i]);
201 ASSERT0(dn->dn_next_blksz[i]);
202 ASSERT0(dn->dn_next_maxblkid[i]);
203 }
204
205 ASSERT0(dn->dn_allocated_txg);
206 ASSERT0(dn->dn_free_txg);
207 ASSERT0(dn->dn_assigned_txg);
208 ASSERT0(dn->dn_dirty_txg);
209 ASSERT0(dn->dn_dirtyctx);
210 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
211 ASSERT3P(dn->dn_bonus, ==, NULL);
212 ASSERT(!dn->dn_have_spill);
213 ASSERT3P(dn->dn_zio, ==, NULL);
214 ASSERT0(dn->dn_oldused);
215 ASSERT0(dn->dn_oldflags);
216 ASSERT0(dn->dn_olduid);
217 ASSERT0(dn->dn_oldgid);
218 ASSERT0(dn->dn_oldprojid);
219 ASSERT0(dn->dn_newuid);
220 ASSERT0(dn->dn_newgid);
221 ASSERT0(dn->dn_newprojid);
222 ASSERT0(dn->dn_id_flags);
223
224 ASSERT0(dn->dn_dbufs_count);
225 avl_destroy(&dn->dn_dbufs);
226 }
227
228 void
229 dnode_init(void)
230 {
231 ASSERT(dnode_cache == NULL);
232 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
233 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
234 kmem_cache_set_move(dnode_cache, dnode_move);
235
236 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
237 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
238 KSTAT_FLAG_VIRTUAL);
239 if (dnode_ksp != NULL) {
240 dnode_ksp->ks_data = &dnode_stats;
241 kstat_install(dnode_ksp);
242 }
243 }
244
245 void
246 dnode_fini(void)
247 {
248 if (dnode_ksp != NULL) {
249 kstat_delete(dnode_ksp);
250 dnode_ksp = NULL;
251 }
252
253 kmem_cache_destroy(dnode_cache);
254 dnode_cache = NULL;
255 }
256
257
258 #ifdef ZFS_DEBUG
259 void
260 dnode_verify(dnode_t *dn)
261 {
262 int drop_struct_lock = FALSE;
263
264 ASSERT(dn->dn_phys);
265 ASSERT(dn->dn_objset);
266 ASSERT(dn->dn_handle->dnh_dnode == dn);
267
268 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
269
270 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
271 return;
272
273 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
274 rw_enter(&dn->dn_struct_rwlock, RW_READER);
275 drop_struct_lock = TRUE;
276 }
277 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
278 int i;
279 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
280 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
281 if (dn->dn_datablkshift) {
282 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
283 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
284 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
285 }
286 ASSERT3U(dn->dn_nlevels, <=, 30);
287 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
288 ASSERT3U(dn->dn_nblkptr, >=, 1);
289 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
290 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
291 ASSERT3U(dn->dn_datablksz, ==,
292 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
293 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
294 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
295 dn->dn_bonuslen, <=, max_bonuslen);
296 for (i = 0; i < TXG_SIZE; i++) {
297 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
298 }
299 }
300 if (dn->dn_phys->dn_type != DMU_OT_NONE)
301 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
302 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
303 if (dn->dn_dbuf != NULL) {
304 ASSERT3P(dn->dn_phys, ==,
305 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
306 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
307 }
308 if (drop_struct_lock)
309 rw_exit(&dn->dn_struct_rwlock);
310 }
311 #endif
312
313 void
314 dnode_byteswap(dnode_phys_t *dnp)
315 {
316 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
317 int i;
318
319 if (dnp->dn_type == DMU_OT_NONE) {
320 bzero(dnp, sizeof (dnode_phys_t));
321 return;
322 }
323
324 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
325 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
326 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
327 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
328 dnp->dn_used = BSWAP_64(dnp->dn_used);
329
330 /*
331 * dn_nblkptr is only one byte, so it's OK to read it in either
332 * byte order. We can't read dn_bouslen.
333 */
334 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
335 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
336 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
337 buf64[i] = BSWAP_64(buf64[i]);
338
339 /*
340 * OK to check dn_bonuslen for zero, because it won't matter if
341 * we have the wrong byte order. This is necessary because the
342 * dnode dnode is smaller than a regular dnode.
343 */
344 if (dnp->dn_bonuslen != 0) {
345 /*
346 * Note that the bonus length calculated here may be
347 * longer than the actual bonus buffer. This is because
348 * we always put the bonus buffer after the last block
349 * pointer (instead of packing it against the end of the
350 * dnode buffer).
351 */
352 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
353 int slots = dnp->dn_extra_slots + 1;
354 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
355 dmu_object_byteswap_t byteswap;
356 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
357 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
358 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
359 }
360
361 /* Swap SPILL block if we have one */
362 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
363 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
364 }
365
366 void
367 dnode_buf_byteswap(void *vbuf, size_t size)
368 {
369 int i = 0;
370
371 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
372 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
373
374 while (i < size) {
375 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
376 dnode_byteswap(dnp);
377
378 i += DNODE_MIN_SIZE;
379 if (dnp->dn_type != DMU_OT_NONE)
380 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
381 }
382 }
383
384 void
385 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
386 {
387 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
388
389 dnode_setdirty(dn, tx);
390 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
391 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
392 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
393 dn->dn_bonuslen = newsize;
394 if (newsize == 0)
395 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
396 else
397 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
398 rw_exit(&dn->dn_struct_rwlock);
399 }
400
401 void
402 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
403 {
404 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
405 dnode_setdirty(dn, tx);
406 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
407 dn->dn_bonustype = newtype;
408 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
409 rw_exit(&dn->dn_struct_rwlock);
410 }
411
412 void
413 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
414 {
415 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
416 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
417 dnode_setdirty(dn, tx);
418 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
419 dn->dn_have_spill = B_FALSE;
420 }
421
422 static void
423 dnode_setdblksz(dnode_t *dn, int size)
424 {
425 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
426 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
427 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
428 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
429 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
430 dn->dn_datablksz = size;
431 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
432 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
433 }
434
435 static dnode_t *
436 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
437 uint64_t object, dnode_handle_t *dnh)
438 {
439 dnode_t *dn;
440
441 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
442 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
443 dn->dn_moved = 0;
444
445 /*
446 * Defer setting dn_objset until the dnode is ready to be a candidate
447 * for the dnode_move() callback.
448 */
449 dn->dn_object = object;
450 dn->dn_dbuf = db;
451 dn->dn_handle = dnh;
452 dn->dn_phys = dnp;
453
454 if (dnp->dn_datablkszsec) {
455 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
456 } else {
457 dn->dn_datablksz = 0;
458 dn->dn_datablkszsec = 0;
459 dn->dn_datablkshift = 0;
460 }
461 dn->dn_indblkshift = dnp->dn_indblkshift;
462 dn->dn_nlevels = dnp->dn_nlevels;
463 dn->dn_type = dnp->dn_type;
464 dn->dn_nblkptr = dnp->dn_nblkptr;
465 dn->dn_checksum = dnp->dn_checksum;
466 dn->dn_compress = dnp->dn_compress;
467 dn->dn_bonustype = dnp->dn_bonustype;
468 dn->dn_bonuslen = dnp->dn_bonuslen;
469 dn->dn_num_slots = dnp->dn_extra_slots + 1;
470 dn->dn_maxblkid = dnp->dn_maxblkid;
471 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
472 dn->dn_id_flags = 0;
473
474 dmu_zfetch_init(&dn->dn_zfetch, dn);
475
476 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
477 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
478 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
479
480 mutex_enter(&os->os_lock);
481
482 /*
483 * Exclude special dnodes from os_dnodes so an empty os_dnodes
484 * signifies that the special dnodes have no references from
485 * their children (the entries in os_dnodes). This allows
486 * dnode_destroy() to easily determine if the last child has
487 * been removed and then complete eviction of the objset.
488 */
489 if (!DMU_OBJECT_IS_SPECIAL(object))
490 list_insert_head(&os->os_dnodes, dn);
491 membar_producer();
492
493 /*
494 * Everything else must be valid before assigning dn_objset
495 * makes the dnode eligible for dnode_move().
496 */
497 dn->dn_objset = os;
498
499 dnh->dnh_dnode = dn;
500 mutex_exit(&os->os_lock);
501
502 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
503
504 return (dn);
505 }
506
507 /*
508 * Caller must be holding the dnode handle, which is released upon return.
509 */
510 static void
511 dnode_destroy(dnode_t *dn)
512 {
513 objset_t *os = dn->dn_objset;
514 boolean_t complete_os_eviction = B_FALSE;
515
516 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
517
518 mutex_enter(&os->os_lock);
519 POINTER_INVALIDATE(&dn->dn_objset);
520 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
521 list_remove(&os->os_dnodes, dn);
522 complete_os_eviction =
523 list_is_empty(&os->os_dnodes) &&
524 list_link_active(&os->os_evicting_node);
525 }
526 mutex_exit(&os->os_lock);
527
528 /* the dnode can no longer move, so we can release the handle */
529 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
530 zrl_remove(&dn->dn_handle->dnh_zrlock);
531
532 dn->dn_allocated_txg = 0;
533 dn->dn_free_txg = 0;
534 dn->dn_assigned_txg = 0;
535 dn->dn_dirty_txg = 0;
536
537 dn->dn_dirtyctx = 0;
538 if (dn->dn_dirtyctx_firstset != NULL) {
539 kmem_free(dn->dn_dirtyctx_firstset, 1);
540 dn->dn_dirtyctx_firstset = NULL;
541 }
542 if (dn->dn_bonus != NULL) {
543 mutex_enter(&dn->dn_bonus->db_mtx);
544 dbuf_destroy(dn->dn_bonus);
545 dn->dn_bonus = NULL;
546 }
547 dn->dn_zio = NULL;
548
549 dn->dn_have_spill = B_FALSE;
550 dn->dn_oldused = 0;
551 dn->dn_oldflags = 0;
552 dn->dn_olduid = 0;
553 dn->dn_oldgid = 0;
554 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
555 dn->dn_newuid = 0;
556 dn->dn_newgid = 0;
557 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
558 dn->dn_id_flags = 0;
559
560 dmu_zfetch_fini(&dn->dn_zfetch);
561 kmem_cache_free(dnode_cache, dn);
562 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
563
564 if (complete_os_eviction)
565 dmu_objset_evict_done(os);
566 }
567
568 void
569 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
570 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
571 {
572 int i;
573
574 ASSERT3U(dn_slots, >, 0);
575 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
576 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
577 ASSERT3U(blocksize, <=,
578 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
579 if (blocksize == 0)
580 blocksize = 1 << zfs_default_bs;
581 else
582 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
583
584 if (ibs == 0)
585 ibs = zfs_default_ibs;
586
587 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
588
589 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
590 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
591 DNODE_STAT_BUMP(dnode_allocate);
592
593 ASSERT(dn->dn_type == DMU_OT_NONE);
594 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
595 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
596 ASSERT(ot != DMU_OT_NONE);
597 ASSERT(DMU_OT_IS_VALID(ot));
598 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
599 (bonustype == DMU_OT_SA && bonuslen == 0) ||
600 (bonustype != DMU_OT_NONE && bonuslen != 0));
601 ASSERT(DMU_OT_IS_VALID(bonustype));
602 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
603 ASSERT(dn->dn_type == DMU_OT_NONE);
604 ASSERT0(dn->dn_maxblkid);
605 ASSERT0(dn->dn_allocated_txg);
606 ASSERT0(dn->dn_assigned_txg);
607 ASSERT0(dn->dn_dirty_txg);
608 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
609 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
610 ASSERT(avl_is_empty(&dn->dn_dbufs));
611
612 for (i = 0; i < TXG_SIZE; i++) {
613 ASSERT0(dn->dn_next_nblkptr[i]);
614 ASSERT0(dn->dn_next_nlevels[i]);
615 ASSERT0(dn->dn_next_indblkshift[i]);
616 ASSERT0(dn->dn_next_bonuslen[i]);
617 ASSERT0(dn->dn_next_bonustype[i]);
618 ASSERT0(dn->dn_rm_spillblk[i]);
619 ASSERT0(dn->dn_next_blksz[i]);
620 ASSERT0(dn->dn_next_maxblkid[i]);
621 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
622 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
623 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
624 }
625
626 dn->dn_type = ot;
627 dnode_setdblksz(dn, blocksize);
628 dn->dn_indblkshift = ibs;
629 dn->dn_nlevels = 1;
630 dn->dn_num_slots = dn_slots;
631 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
632 dn->dn_nblkptr = 1;
633 else {
634 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
635 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
636 SPA_BLKPTRSHIFT));
637 }
638
639 dn->dn_bonustype = bonustype;
640 dn->dn_bonuslen = bonuslen;
641 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
642 dn->dn_compress = ZIO_COMPRESS_INHERIT;
643 dn->dn_dirtyctx = 0;
644
645 dn->dn_free_txg = 0;
646 if (dn->dn_dirtyctx_firstset) {
647 kmem_free(dn->dn_dirtyctx_firstset, 1);
648 dn->dn_dirtyctx_firstset = NULL;
649 }
650
651 dn->dn_allocated_txg = tx->tx_txg;
652 dn->dn_id_flags = 0;
653
654 dnode_setdirty(dn, tx);
655 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
656 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
657 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
658 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
659 }
660
661 void
662 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
663 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
664 {
665 int nblkptr;
666
667 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
668 ASSERT3U(blocksize, <=,
669 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
670 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
671 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
672 ASSERT(tx->tx_txg != 0);
673 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
674 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
675 (bonustype == DMU_OT_SA && bonuslen == 0));
676 ASSERT(DMU_OT_IS_VALID(bonustype));
677 ASSERT3U(bonuslen, <=,
678 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
679 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
680
681 dnode_free_interior_slots(dn);
682 DNODE_STAT_BUMP(dnode_reallocate);
683
684 /* clean up any unreferenced dbufs */
685 dnode_evict_dbufs(dn);
686
687 dn->dn_id_flags = 0;
688
689 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
690 dnode_setdirty(dn, tx);
691 if (dn->dn_datablksz != blocksize) {
692 /* change blocksize */
693 ASSERT(dn->dn_maxblkid == 0 &&
694 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
695 dnode_block_freed(dn, 0)));
696 dnode_setdblksz(dn, blocksize);
697 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
698 }
699 if (dn->dn_bonuslen != bonuslen)
700 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
701
702 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
703 nblkptr = 1;
704 else
705 nblkptr = MIN(DN_MAX_NBLKPTR,
706 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
707 SPA_BLKPTRSHIFT));
708 if (dn->dn_bonustype != bonustype)
709 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
710 if (dn->dn_nblkptr != nblkptr)
711 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
712 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
713 dbuf_rm_spill(dn, tx);
714 dnode_rm_spill(dn, tx);
715 }
716 rw_exit(&dn->dn_struct_rwlock);
717
718 /* change type */
719 dn->dn_type = ot;
720
721 /* change bonus size and type */
722 mutex_enter(&dn->dn_mtx);
723 dn->dn_bonustype = bonustype;
724 dn->dn_bonuslen = bonuslen;
725 dn->dn_num_slots = dn_slots;
726 dn->dn_nblkptr = nblkptr;
727 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
728 dn->dn_compress = ZIO_COMPRESS_INHERIT;
729 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
730
731 /* fix up the bonus db_size */
732 if (dn->dn_bonus) {
733 dn->dn_bonus->db.db_size =
734 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
735 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
736 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
737 }
738
739 dn->dn_allocated_txg = tx->tx_txg;
740 mutex_exit(&dn->dn_mtx);
741 }
742
743 #ifdef _KERNEL
744 static void
745 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
746 {
747 int i;
748
749 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
750 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
751 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
752 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
753
754 /* Copy fields. */
755 ndn->dn_objset = odn->dn_objset;
756 ndn->dn_object = odn->dn_object;
757 ndn->dn_dbuf = odn->dn_dbuf;
758 ndn->dn_handle = odn->dn_handle;
759 ndn->dn_phys = odn->dn_phys;
760 ndn->dn_type = odn->dn_type;
761 ndn->dn_bonuslen = odn->dn_bonuslen;
762 ndn->dn_bonustype = odn->dn_bonustype;
763 ndn->dn_nblkptr = odn->dn_nblkptr;
764 ndn->dn_checksum = odn->dn_checksum;
765 ndn->dn_compress = odn->dn_compress;
766 ndn->dn_nlevels = odn->dn_nlevels;
767 ndn->dn_indblkshift = odn->dn_indblkshift;
768 ndn->dn_datablkshift = odn->dn_datablkshift;
769 ndn->dn_datablkszsec = odn->dn_datablkszsec;
770 ndn->dn_datablksz = odn->dn_datablksz;
771 ndn->dn_maxblkid = odn->dn_maxblkid;
772 ndn->dn_num_slots = odn->dn_num_slots;
773 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
774 sizeof (odn->dn_next_type));
775 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
776 sizeof (odn->dn_next_nblkptr));
777 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
778 sizeof (odn->dn_next_nlevels));
779 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
780 sizeof (odn->dn_next_indblkshift));
781 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
782 sizeof (odn->dn_next_bonustype));
783 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
784 sizeof (odn->dn_rm_spillblk));
785 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
786 sizeof (odn->dn_next_bonuslen));
787 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
788 sizeof (odn->dn_next_blksz));
789 bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
790 sizeof (odn->dn_next_maxblkid));
791 for (i = 0; i < TXG_SIZE; i++) {
792 list_move_tail(&ndn->dn_dirty_records[i],
793 &odn->dn_dirty_records[i]);
794 }
795 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
796 sizeof (odn->dn_free_ranges));
797 ndn->dn_allocated_txg = odn->dn_allocated_txg;
798 ndn->dn_free_txg = odn->dn_free_txg;
799 ndn->dn_assigned_txg = odn->dn_assigned_txg;
800 ndn->dn_dirty_txg = odn->dn_dirty_txg;
801 ndn->dn_dirtyctx = odn->dn_dirtyctx;
802 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
803 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
804 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
805 ASSERT(avl_is_empty(&ndn->dn_dbufs));
806 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
807 ndn->dn_dbufs_count = odn->dn_dbufs_count;
808 ndn->dn_bonus = odn->dn_bonus;
809 ndn->dn_have_spill = odn->dn_have_spill;
810 ndn->dn_zio = odn->dn_zio;
811 ndn->dn_oldused = odn->dn_oldused;
812 ndn->dn_oldflags = odn->dn_oldflags;
813 ndn->dn_olduid = odn->dn_olduid;
814 ndn->dn_oldgid = odn->dn_oldgid;
815 ndn->dn_oldprojid = odn->dn_oldprojid;
816 ndn->dn_newuid = odn->dn_newuid;
817 ndn->dn_newgid = odn->dn_newgid;
818 ndn->dn_newprojid = odn->dn_newprojid;
819 ndn->dn_id_flags = odn->dn_id_flags;
820 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
821 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
822 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
823
824 /*
825 * Update back pointers. Updating the handle fixes the back pointer of
826 * every descendant dbuf as well as the bonus dbuf.
827 */
828 ASSERT(ndn->dn_handle->dnh_dnode == odn);
829 ndn->dn_handle->dnh_dnode = ndn;
830 if (ndn->dn_zfetch.zf_dnode == odn) {
831 ndn->dn_zfetch.zf_dnode = ndn;
832 }
833
834 /*
835 * Invalidate the original dnode by clearing all of its back pointers.
836 */
837 odn->dn_dbuf = NULL;
838 odn->dn_handle = NULL;
839 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
840 offsetof(dmu_buf_impl_t, db_link));
841 odn->dn_dbufs_count = 0;
842 odn->dn_bonus = NULL;
843 odn->dn_zfetch.zf_dnode = NULL;
844
845 /*
846 * Set the low bit of the objset pointer to ensure that dnode_move()
847 * recognizes the dnode as invalid in any subsequent callback.
848 */
849 POINTER_INVALIDATE(&odn->dn_objset);
850
851 /*
852 * Satisfy the destructor.
853 */
854 for (i = 0; i < TXG_SIZE; i++) {
855 list_create(&odn->dn_dirty_records[i],
856 sizeof (dbuf_dirty_record_t),
857 offsetof(dbuf_dirty_record_t, dr_dirty_node));
858 odn->dn_free_ranges[i] = NULL;
859 odn->dn_next_nlevels[i] = 0;
860 odn->dn_next_indblkshift[i] = 0;
861 odn->dn_next_bonustype[i] = 0;
862 odn->dn_rm_spillblk[i] = 0;
863 odn->dn_next_bonuslen[i] = 0;
864 odn->dn_next_blksz[i] = 0;
865 }
866 odn->dn_allocated_txg = 0;
867 odn->dn_free_txg = 0;
868 odn->dn_assigned_txg = 0;
869 odn->dn_dirty_txg = 0;
870 odn->dn_dirtyctx = 0;
871 odn->dn_dirtyctx_firstset = NULL;
872 odn->dn_have_spill = B_FALSE;
873 odn->dn_zio = NULL;
874 odn->dn_oldused = 0;
875 odn->dn_oldflags = 0;
876 odn->dn_olduid = 0;
877 odn->dn_oldgid = 0;
878 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
879 odn->dn_newuid = 0;
880 odn->dn_newgid = 0;
881 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
882 odn->dn_id_flags = 0;
883
884 /*
885 * Mark the dnode.
886 */
887 ndn->dn_moved = 1;
888 odn->dn_moved = (uint8_t)-1;
889 }
890
891 /*ARGSUSED*/
892 static kmem_cbrc_t
893 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
894 {
895 dnode_t *odn = buf, *ndn = newbuf;
896 objset_t *os;
897 int64_t refcount;
898 uint32_t dbufs;
899
900 /*
901 * The dnode is on the objset's list of known dnodes if the objset
902 * pointer is valid. We set the low bit of the objset pointer when
903 * freeing the dnode to invalidate it, and the memory patterns written
904 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
905 * A newly created dnode sets the objset pointer last of all to indicate
906 * that the dnode is known and in a valid state to be moved by this
907 * function.
908 */
909 os = odn->dn_objset;
910 if (!POINTER_IS_VALID(os)) {
911 DNODE_STAT_BUMP(dnode_move_invalid);
912 return (KMEM_CBRC_DONT_KNOW);
913 }
914
915 /*
916 * Ensure that the objset does not go away during the move.
917 */
918 rw_enter(&os_lock, RW_WRITER);
919 if (os != odn->dn_objset) {
920 rw_exit(&os_lock);
921 DNODE_STAT_BUMP(dnode_move_recheck1);
922 return (KMEM_CBRC_DONT_KNOW);
923 }
924
925 /*
926 * If the dnode is still valid, then so is the objset. We know that no
927 * valid objset can be freed while we hold os_lock, so we can safely
928 * ensure that the objset remains in use.
929 */
930 mutex_enter(&os->os_lock);
931
932 /*
933 * Recheck the objset pointer in case the dnode was removed just before
934 * acquiring the lock.
935 */
936 if (os != odn->dn_objset) {
937 mutex_exit(&os->os_lock);
938 rw_exit(&os_lock);
939 DNODE_STAT_BUMP(dnode_move_recheck2);
940 return (KMEM_CBRC_DONT_KNOW);
941 }
942
943 /*
944 * At this point we know that as long as we hold os->os_lock, the dnode
945 * cannot be freed and fields within the dnode can be safely accessed.
946 * The objset listing this dnode cannot go away as long as this dnode is
947 * on its list.
948 */
949 rw_exit(&os_lock);
950 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
951 mutex_exit(&os->os_lock);
952 DNODE_STAT_BUMP(dnode_move_special);
953 return (KMEM_CBRC_NO);
954 }
955 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
956
957 /*
958 * Lock the dnode handle to prevent the dnode from obtaining any new
959 * holds. This also prevents the descendant dbufs and the bonus dbuf
960 * from accessing the dnode, so that we can discount their holds. The
961 * handle is safe to access because we know that while the dnode cannot
962 * go away, neither can its handle. Once we hold dnh_zrlock, we can
963 * safely move any dnode referenced only by dbufs.
964 */
965 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
966 mutex_exit(&os->os_lock);
967 DNODE_STAT_BUMP(dnode_move_handle);
968 return (KMEM_CBRC_LATER);
969 }
970
971 /*
972 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
973 * We need to guarantee that there is a hold for every dbuf in order to
974 * determine whether the dnode is actively referenced. Falsely matching
975 * a dbuf to an active hold would lead to an unsafe move. It's possible
976 * that a thread already having an active dnode hold is about to add a
977 * dbuf, and we can't compare hold and dbuf counts while the add is in
978 * progress.
979 */
980 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
981 zrl_exit(&odn->dn_handle->dnh_zrlock);
982 mutex_exit(&os->os_lock);
983 DNODE_STAT_BUMP(dnode_move_rwlock);
984 return (KMEM_CBRC_LATER);
985 }
986
987 /*
988 * A dbuf may be removed (evicted) without an active dnode hold. In that
989 * case, the dbuf count is decremented under the handle lock before the
990 * dbuf's hold is released. This order ensures that if we count the hold
991 * after the dbuf is removed but before its hold is released, we will
992 * treat the unmatched hold as active and exit safely. If we count the
993 * hold before the dbuf is removed, the hold is discounted, and the
994 * removal is blocked until the move completes.
995 */
996 refcount = zfs_refcount_count(&odn->dn_holds);
997 ASSERT(refcount >= 0);
998 dbufs = odn->dn_dbufs_count;
999
1000 /* We can't have more dbufs than dnode holds. */
1001 ASSERT3U(dbufs, <=, refcount);
1002 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1003 uint32_t, dbufs);
1004
1005 if (refcount > dbufs) {
1006 rw_exit(&odn->dn_struct_rwlock);
1007 zrl_exit(&odn->dn_handle->dnh_zrlock);
1008 mutex_exit(&os->os_lock);
1009 DNODE_STAT_BUMP(dnode_move_active);
1010 return (KMEM_CBRC_LATER);
1011 }
1012
1013 rw_exit(&odn->dn_struct_rwlock);
1014
1015 /*
1016 * At this point we know that anyone with a hold on the dnode is not
1017 * actively referencing it. The dnode is known and in a valid state to
1018 * move. We're holding the locks needed to execute the critical section.
1019 */
1020 dnode_move_impl(odn, ndn);
1021
1022 list_link_replace(&odn->dn_link, &ndn->dn_link);
1023 /* If the dnode was safe to move, the refcount cannot have changed. */
1024 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1025 ASSERT(dbufs == ndn->dn_dbufs_count);
1026 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1027 mutex_exit(&os->os_lock);
1028
1029 return (KMEM_CBRC_YES);
1030 }
1031 #endif /* _KERNEL */
1032
1033 static void
1034 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1035 {
1036 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1037
1038 for (int i = idx; i < idx + slots; i++) {
1039 dnode_handle_t *dnh = &children->dnc_children[i];
1040 zrl_add(&dnh->dnh_zrlock);
1041 }
1042 }
1043
1044 static void
1045 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1046 {
1047 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1048
1049 for (int i = idx; i < idx + slots; i++) {
1050 dnode_handle_t *dnh = &children->dnc_children[i];
1051
1052 if (zrl_is_locked(&dnh->dnh_zrlock))
1053 zrl_exit(&dnh->dnh_zrlock);
1054 else
1055 zrl_remove(&dnh->dnh_zrlock);
1056 }
1057 }
1058
1059 static int
1060 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1061 {
1062 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1063
1064 for (int i = idx; i < idx + slots; i++) {
1065 dnode_handle_t *dnh = &children->dnc_children[i];
1066
1067 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1068 for (int j = idx; j < i; j++) {
1069 dnh = &children->dnc_children[j];
1070 zrl_exit(&dnh->dnh_zrlock);
1071 }
1072
1073 return (0);
1074 }
1075 }
1076
1077 return (1);
1078 }
1079
1080 static void
1081 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1082 {
1083 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1084
1085 for (int i = idx; i < idx + slots; i++) {
1086 dnode_handle_t *dnh = &children->dnc_children[i];
1087 dnh->dnh_dnode = ptr;
1088 }
1089 }
1090
1091 static boolean_t
1092 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1093 {
1094 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1095
1096 /*
1097 * If all dnode slots are either already free or
1098 * evictable return B_TRUE.
1099 */
1100 for (int i = idx; i < idx + slots; i++) {
1101 dnode_handle_t *dnh = &children->dnc_children[i];
1102 dnode_t *dn = dnh->dnh_dnode;
1103
1104 if (dn == DN_SLOT_FREE) {
1105 continue;
1106 } else if (DN_SLOT_IS_PTR(dn)) {
1107 mutex_enter(&dn->dn_mtx);
1108 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1109 zfs_refcount_is_zero(&dn->dn_holds) &&
1110 !DNODE_IS_DIRTY(dn));
1111 mutex_exit(&dn->dn_mtx);
1112
1113 if (!can_free)
1114 return (B_FALSE);
1115 else
1116 continue;
1117 } else {
1118 return (B_FALSE);
1119 }
1120 }
1121
1122 return (B_TRUE);
1123 }
1124
1125 static void
1126 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1127 {
1128 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1129
1130 for (int i = idx; i < idx + slots; i++) {
1131 dnode_handle_t *dnh = &children->dnc_children[i];
1132
1133 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1134
1135 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1136 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1137 dnode_destroy(dnh->dnh_dnode);
1138 dnh->dnh_dnode = DN_SLOT_FREE;
1139 }
1140 }
1141 }
1142
1143 void
1144 dnode_free_interior_slots(dnode_t *dn)
1145 {
1146 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1147 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1148 int idx = (dn->dn_object & (epb - 1)) + 1;
1149 int slots = dn->dn_num_slots - 1;
1150
1151 if (slots == 0)
1152 return;
1153
1154 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1155
1156 while (!dnode_slots_tryenter(children, idx, slots))
1157 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1158
1159 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1160 dnode_slots_rele(children, idx, slots);
1161 }
1162
1163 void
1164 dnode_special_close(dnode_handle_t *dnh)
1165 {
1166 dnode_t *dn = dnh->dnh_dnode;
1167
1168 /*
1169 * Wait for final references to the dnode to clear. This can
1170 * only happen if the arc is asynchronously evicting state that
1171 * has a hold on this dnode while we are trying to evict this
1172 * dnode.
1173 */
1174 while (zfs_refcount_count(&dn->dn_holds) > 0)
1175 delay(1);
1176 ASSERT(dn->dn_dbuf == NULL ||
1177 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1178 zrl_add(&dnh->dnh_zrlock);
1179 dnode_destroy(dn); /* implicit zrl_remove() */
1180 zrl_destroy(&dnh->dnh_zrlock);
1181 dnh->dnh_dnode = NULL;
1182 }
1183
1184 void
1185 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1186 dnode_handle_t *dnh)
1187 {
1188 dnode_t *dn;
1189
1190 zrl_init(&dnh->dnh_zrlock);
1191 zrl_tryenter(&dnh->dnh_zrlock);
1192
1193 dn = dnode_create(os, dnp, NULL, object, dnh);
1194 DNODE_VERIFY(dn);
1195
1196 zrl_exit(&dnh->dnh_zrlock);
1197 }
1198
1199 static void
1200 dnode_buf_evict_async(void *dbu)
1201 {
1202 dnode_children_t *dnc = dbu;
1203
1204 DNODE_STAT_BUMP(dnode_buf_evict);
1205
1206 for (int i = 0; i < dnc->dnc_count; i++) {
1207 dnode_handle_t *dnh = &dnc->dnc_children[i];
1208 dnode_t *dn;
1209
1210 /*
1211 * The dnode handle lock guards against the dnode moving to
1212 * another valid address, so there is no need here to guard
1213 * against changes to or from NULL.
1214 */
1215 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1216 zrl_destroy(&dnh->dnh_zrlock);
1217 dnh->dnh_dnode = DN_SLOT_UNINIT;
1218 continue;
1219 }
1220
1221 zrl_add(&dnh->dnh_zrlock);
1222 dn = dnh->dnh_dnode;
1223 /*
1224 * If there are holds on this dnode, then there should
1225 * be holds on the dnode's containing dbuf as well; thus
1226 * it wouldn't be eligible for eviction and this function
1227 * would not have been called.
1228 */
1229 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1230 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1231
1232 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1233 zrl_destroy(&dnh->dnh_zrlock);
1234 dnh->dnh_dnode = DN_SLOT_UNINIT;
1235 }
1236 kmem_free(dnc, sizeof (dnode_children_t) +
1237 dnc->dnc_count * sizeof (dnode_handle_t));
1238 }
1239
1240 /*
1241 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1242 * to ensure the hole at the specified object offset is large enough to
1243 * hold the dnode being created. The slots parameter is also used to ensure
1244 * a dnode does not span multiple dnode blocks. In both of these cases, if
1245 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1246 * are only possible when using DNODE_MUST_BE_FREE.
1247 *
1248 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1249 * dnode_hold_impl() will check if the requested dnode is already consumed
1250 * as an extra dnode slot by an large dnode, in which case it returns
1251 * ENOENT.
1252 *
1253 * errors:
1254 * EINVAL - Invalid object number or flags.
1255 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1256 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1257 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1258 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1259 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1260 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1261 * EIO - I/O error when reading the meta dnode dbuf.
1262 *
1263 * succeeds even for free dnodes.
1264 */
1265 int
1266 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1267 void *tag, dnode_t **dnp)
1268 {
1269 int epb, idx, err;
1270 int drop_struct_lock = FALSE;
1271 int type;
1272 uint64_t blk;
1273 dnode_t *mdn, *dn;
1274 dmu_buf_impl_t *db;
1275 dnode_children_t *dnc;
1276 dnode_phys_t *dn_block;
1277 dnode_handle_t *dnh;
1278
1279 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1280 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1281
1282 /*
1283 * If you are holding the spa config lock as writer, you shouldn't
1284 * be asking the DMU to do *anything* unless it's the root pool
1285 * which may require us to read from the root filesystem while
1286 * holding some (not all) of the locks as writer.
1287 */
1288 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1289 (spa_is_root(os->os_spa) &&
1290 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1291
1292 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1293
1294 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1295 object == DMU_PROJECTUSED_OBJECT) {
1296 if (object == DMU_USERUSED_OBJECT)
1297 dn = DMU_USERUSED_DNODE(os);
1298 else if (object == DMU_GROUPUSED_OBJECT)
1299 dn = DMU_GROUPUSED_DNODE(os);
1300 else
1301 dn = DMU_PROJECTUSED_DNODE(os);
1302 if (dn == NULL)
1303 return (SET_ERROR(ENOENT));
1304 type = dn->dn_type;
1305 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1306 return (SET_ERROR(ENOENT));
1307 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1308 return (SET_ERROR(EEXIST));
1309 DNODE_VERIFY(dn);
1310 (void) zfs_refcount_add(&dn->dn_holds, tag);
1311 *dnp = dn;
1312 return (0);
1313 }
1314
1315 if (object == 0 || object >= DN_MAX_OBJECT)
1316 return (SET_ERROR(EINVAL));
1317
1318 mdn = DMU_META_DNODE(os);
1319 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1320
1321 DNODE_VERIFY(mdn);
1322
1323 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1324 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1325 drop_struct_lock = TRUE;
1326 }
1327
1328 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1329
1330 db = dbuf_hold(mdn, blk, FTAG);
1331 if (drop_struct_lock)
1332 rw_exit(&mdn->dn_struct_rwlock);
1333 if (db == NULL) {
1334 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1335 return (SET_ERROR(EIO));
1336 }
1337
1338 /*
1339 * We do not need to decrypt to read the dnode so it doesn't matter
1340 * if we get the encrypted or decrypted version.
1341 */
1342 err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1343 if (err) {
1344 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1345 dbuf_rele(db, FTAG);
1346 return (err);
1347 }
1348
1349 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1350 epb = db->db.db_size >> DNODE_SHIFT;
1351
1352 idx = object & (epb - 1);
1353 dn_block = (dnode_phys_t *)db->db.db_data;
1354
1355 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1356 dnc = dmu_buf_get_user(&db->db);
1357 dnh = NULL;
1358 if (dnc == NULL) {
1359 dnode_children_t *winner;
1360 int skip = 0;
1361
1362 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1363 epb * sizeof (dnode_handle_t), KM_SLEEP);
1364 dnc->dnc_count = epb;
1365 dnh = &dnc->dnc_children[0];
1366
1367 /* Initialize dnode slot status from dnode_phys_t */
1368 for (int i = 0; i < epb; i++) {
1369 zrl_init(&dnh[i].dnh_zrlock);
1370
1371 if (skip) {
1372 skip--;
1373 continue;
1374 }
1375
1376 if (dn_block[i].dn_type != DMU_OT_NONE) {
1377 int interior = dn_block[i].dn_extra_slots;
1378
1379 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1380 dnode_set_slots(dnc, i + 1, interior,
1381 DN_SLOT_INTERIOR);
1382 skip = interior;
1383 } else {
1384 dnh[i].dnh_dnode = DN_SLOT_FREE;
1385 skip = 0;
1386 }
1387 }
1388
1389 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1390 dnode_buf_evict_async, NULL);
1391 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1392 if (winner != NULL) {
1393
1394 for (int i = 0; i < epb; i++)
1395 zrl_destroy(&dnh[i].dnh_zrlock);
1396
1397 kmem_free(dnc, sizeof (dnode_children_t) +
1398 epb * sizeof (dnode_handle_t));
1399 dnc = winner;
1400 }
1401 }
1402
1403 ASSERT(dnc->dnc_count == epb);
1404 dn = DN_SLOT_UNINIT;
1405
1406 if (flag & DNODE_MUST_BE_ALLOCATED) {
1407 slots = 1;
1408
1409 while (dn == DN_SLOT_UNINIT) {
1410 dnode_slots_hold(dnc, idx, slots);
1411 dnh = &dnc->dnc_children[idx];
1412
1413 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1414 dn = dnh->dnh_dnode;
1415 break;
1416 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1417 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1418 dnode_slots_rele(dnc, idx, slots);
1419 dbuf_rele(db, FTAG);
1420 return (SET_ERROR(EEXIST));
1421 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1422 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1423 dnode_slots_rele(dnc, idx, slots);
1424 dbuf_rele(db, FTAG);
1425 return (SET_ERROR(ENOENT));
1426 }
1427
1428 dnode_slots_rele(dnc, idx, slots);
1429 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1430 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1431 continue;
1432 }
1433
1434 /*
1435 * Someone else won the race and called dnode_create()
1436 * after we checked DN_SLOT_IS_PTR() above but before
1437 * we acquired the lock.
1438 */
1439 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1440 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1441 dn = dnh->dnh_dnode;
1442 } else {
1443 dn = dnode_create(os, dn_block + idx, db,
1444 object, dnh);
1445 }
1446 }
1447
1448 mutex_enter(&dn->dn_mtx);
1449 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1450 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1451 mutex_exit(&dn->dn_mtx);
1452 dnode_slots_rele(dnc, idx, slots);
1453 dbuf_rele(db, FTAG);
1454 return (SET_ERROR(ENOENT));
1455 }
1456
1457 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1458 } else if (flag & DNODE_MUST_BE_FREE) {
1459
1460 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1461 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1462 dbuf_rele(db, FTAG);
1463 return (SET_ERROR(ENOSPC));
1464 }
1465
1466 while (dn == DN_SLOT_UNINIT) {
1467 dnode_slots_hold(dnc, idx, slots);
1468
1469 if (!dnode_check_slots_free(dnc, idx, slots)) {
1470 DNODE_STAT_BUMP(dnode_hold_free_misses);
1471 dnode_slots_rele(dnc, idx, slots);
1472 dbuf_rele(db, FTAG);
1473 return (SET_ERROR(ENOSPC));
1474 }
1475
1476 dnode_slots_rele(dnc, idx, slots);
1477 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1478 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1479 continue;
1480 }
1481
1482 if (!dnode_check_slots_free(dnc, idx, slots)) {
1483 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1484 dnode_slots_rele(dnc, idx, slots);
1485 dbuf_rele(db, FTAG);
1486 return (SET_ERROR(ENOSPC));
1487 }
1488
1489 /*
1490 * Allocated but otherwise free dnodes which would
1491 * be in the interior of a multi-slot dnodes need
1492 * to be freed. Single slot dnodes can be safely
1493 * re-purposed as a performance optimization.
1494 */
1495 if (slots > 1)
1496 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1497
1498 dnh = &dnc->dnc_children[idx];
1499 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1500 dn = dnh->dnh_dnode;
1501 } else {
1502 dn = dnode_create(os, dn_block + idx, db,
1503 object, dnh);
1504 }
1505 }
1506
1507 mutex_enter(&dn->dn_mtx);
1508 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1509 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1510 mutex_exit(&dn->dn_mtx);
1511 dnode_slots_rele(dnc, idx, slots);
1512 dbuf_rele(db, FTAG);
1513 return (SET_ERROR(EEXIST));
1514 }
1515
1516 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1517 DNODE_STAT_BUMP(dnode_hold_free_hits);
1518 } else {
1519 dbuf_rele(db, FTAG);
1520 return (SET_ERROR(EINVAL));
1521 }
1522
1523 if (dn->dn_free_txg) {
1524 DNODE_STAT_BUMP(dnode_hold_free_txg);
1525 type = dn->dn_type;
1526 mutex_exit(&dn->dn_mtx);
1527 dnode_slots_rele(dnc, idx, slots);
1528 dbuf_rele(db, FTAG);
1529 return (SET_ERROR((flag & DNODE_MUST_BE_ALLOCATED) ?
1530 ENOENT : EEXIST));
1531 }
1532
1533 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1534 dbuf_add_ref(db, dnh);
1535
1536 mutex_exit(&dn->dn_mtx);
1537
1538 /* Now we can rely on the hold to prevent the dnode from moving. */
1539 dnode_slots_rele(dnc, idx, slots);
1540
1541 DNODE_VERIFY(dn);
1542 ASSERT3P(dn->dn_dbuf, ==, db);
1543 ASSERT3U(dn->dn_object, ==, object);
1544 dbuf_rele(db, FTAG);
1545
1546 *dnp = dn;
1547 return (0);
1548 }
1549
1550 /*
1551 * Return held dnode if the object is allocated, NULL if not.
1552 */
1553 int
1554 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1555 {
1556 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1557 dnp));
1558 }
1559
1560 /*
1561 * Can only add a reference if there is already at least one
1562 * reference on the dnode. Returns FALSE if unable to add a
1563 * new reference.
1564 */
1565 boolean_t
1566 dnode_add_ref(dnode_t *dn, void *tag)
1567 {
1568 mutex_enter(&dn->dn_mtx);
1569 if (zfs_refcount_is_zero(&dn->dn_holds)) {
1570 mutex_exit(&dn->dn_mtx);
1571 return (FALSE);
1572 }
1573 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1574 mutex_exit(&dn->dn_mtx);
1575 return (TRUE);
1576 }
1577
1578 void
1579 dnode_rele(dnode_t *dn, void *tag)
1580 {
1581 mutex_enter(&dn->dn_mtx);
1582 dnode_rele_and_unlock(dn, tag, B_FALSE);
1583 }
1584
1585 void
1586 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1587 {
1588 uint64_t refs;
1589 /* Get while the hold prevents the dnode from moving. */
1590 dmu_buf_impl_t *db = dn->dn_dbuf;
1591 dnode_handle_t *dnh = dn->dn_handle;
1592
1593 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1594 mutex_exit(&dn->dn_mtx);
1595
1596 /*
1597 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1598 * indirectly by dbuf_rele() while relying on the dnode handle to
1599 * prevent the dnode from moving, since releasing the last hold could
1600 * result in the dnode's parent dbuf evicting its dnode handles. For
1601 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1602 * other direct or indirect hold on the dnode must first drop the dnode
1603 * handle.
1604 */
1605 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1606
1607 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1608 if (refs == 0 && db != NULL) {
1609 /*
1610 * Another thread could add a hold to the dnode handle in
1611 * dnode_hold_impl() while holding the parent dbuf. Since the
1612 * hold on the parent dbuf prevents the handle from being
1613 * destroyed, the hold on the handle is OK. We can't yet assert
1614 * that the handle has zero references, but that will be
1615 * asserted anyway when the handle gets destroyed.
1616 */
1617 mutex_enter(&db->db_mtx);
1618 dbuf_rele_and_unlock(db, dnh, evicting);
1619 }
1620 }
1621
1622 void
1623 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1624 {
1625 objset_t *os = dn->dn_objset;
1626 uint64_t txg = tx->tx_txg;
1627
1628 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1629 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1630 return;
1631 }
1632
1633 DNODE_VERIFY(dn);
1634
1635 #ifdef ZFS_DEBUG
1636 mutex_enter(&dn->dn_mtx);
1637 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1638 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1639 mutex_exit(&dn->dn_mtx);
1640 #endif
1641
1642 /*
1643 * Determine old uid/gid when necessary
1644 */
1645 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1646
1647 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1648 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1649
1650 /*
1651 * If we are already marked dirty, we're done.
1652 */
1653 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1654 multilist_sublist_unlock(mls);
1655 return;
1656 }
1657
1658 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1659 !avl_is_empty(&dn->dn_dbufs));
1660 ASSERT(dn->dn_datablksz != 0);
1661 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1662 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1663 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1664
1665 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1666 dn->dn_object, txg);
1667
1668 multilist_sublist_insert_head(mls, dn);
1669
1670 multilist_sublist_unlock(mls);
1671
1672 /*
1673 * The dnode maintains a hold on its containing dbuf as
1674 * long as there are holds on it. Each instantiated child
1675 * dbuf maintains a hold on the dnode. When the last child
1676 * drops its hold, the dnode will drop its hold on the
1677 * containing dbuf. We add a "dirty hold" here so that the
1678 * dnode will hang around after we finish processing its
1679 * children.
1680 */
1681 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1682
1683 (void) dbuf_dirty(dn->dn_dbuf, tx);
1684
1685 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1686 }
1687
1688 void
1689 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1690 {
1691 mutex_enter(&dn->dn_mtx);
1692 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1693 mutex_exit(&dn->dn_mtx);
1694 return;
1695 }
1696 dn->dn_free_txg = tx->tx_txg;
1697 mutex_exit(&dn->dn_mtx);
1698
1699 dnode_setdirty(dn, tx);
1700 }
1701
1702 /*
1703 * Try to change the block size for the indicated dnode. This can only
1704 * succeed if there are no blocks allocated or dirty beyond first block
1705 */
1706 int
1707 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1708 {
1709 dmu_buf_impl_t *db;
1710 int err;
1711
1712 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1713 if (size == 0)
1714 size = SPA_MINBLOCKSIZE;
1715 else
1716 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1717
1718 if (ibs == dn->dn_indblkshift)
1719 ibs = 0;
1720
1721 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1722 return (0);
1723
1724 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1725
1726 /* Check for any allocated blocks beyond the first */
1727 if (dn->dn_maxblkid != 0)
1728 goto fail;
1729
1730 mutex_enter(&dn->dn_dbufs_mtx);
1731 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1732 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1733 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1734 db->db_blkid != DMU_SPILL_BLKID) {
1735 mutex_exit(&dn->dn_dbufs_mtx);
1736 goto fail;
1737 }
1738 }
1739 mutex_exit(&dn->dn_dbufs_mtx);
1740
1741 if (ibs && dn->dn_nlevels != 1)
1742 goto fail;
1743
1744 /* resize the old block */
1745 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1746 if (err == 0)
1747 dbuf_new_size(db, size, tx);
1748 else if (err != ENOENT)
1749 goto fail;
1750
1751 dnode_setdblksz(dn, size);
1752 dnode_setdirty(dn, tx);
1753 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1754 if (ibs) {
1755 dn->dn_indblkshift = ibs;
1756 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1757 }
1758 /* rele after we have fixed the blocksize in the dnode */
1759 if (db)
1760 dbuf_rele(db, FTAG);
1761
1762 rw_exit(&dn->dn_struct_rwlock);
1763 return (0);
1764
1765 fail:
1766 rw_exit(&dn->dn_struct_rwlock);
1767 return (SET_ERROR(ENOTSUP));
1768 }
1769
1770 static void
1771 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1772 {
1773 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1774 int old_nlevels = dn->dn_nlevels;
1775 dmu_buf_impl_t *db;
1776 list_t *list;
1777 dbuf_dirty_record_t *new, *dr, *dr_next;
1778
1779 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1780
1781 dn->dn_nlevels = new_nlevels;
1782
1783 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1784 dn->dn_next_nlevels[txgoff] = new_nlevels;
1785
1786 /* dirty the left indirects */
1787 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1788 ASSERT(db != NULL);
1789 new = dbuf_dirty(db, tx);
1790 dbuf_rele(db, FTAG);
1791
1792 /* transfer the dirty records to the new indirect */
1793 mutex_enter(&dn->dn_mtx);
1794 mutex_enter(&new->dt.di.dr_mtx);
1795 list = &dn->dn_dirty_records[txgoff];
1796 for (dr = list_head(list); dr; dr = dr_next) {
1797 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1798 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1799 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1800 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1801 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1802 list_remove(&dn->dn_dirty_records[txgoff], dr);
1803 list_insert_tail(&new->dt.di.dr_children, dr);
1804 dr->dr_parent = new;
1805 }
1806 }
1807 mutex_exit(&new->dt.di.dr_mtx);
1808 mutex_exit(&dn->dn_mtx);
1809 }
1810
1811 int
1812 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1813 {
1814 int ret = 0;
1815
1816 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1817
1818 if (dn->dn_nlevels == nlevels) {
1819 ret = 0;
1820 goto out;
1821 } else if (nlevels < dn->dn_nlevels) {
1822 ret = SET_ERROR(EINVAL);
1823 goto out;
1824 }
1825
1826 dnode_set_nlevels_impl(dn, nlevels, tx);
1827
1828 out:
1829 rw_exit(&dn->dn_struct_rwlock);
1830 return (ret);
1831 }
1832
1833 /* read-holding callers must not rely on the lock being continuously held */
1834 void
1835 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1836 {
1837 int epbs, new_nlevels;
1838 uint64_t sz;
1839
1840 ASSERT(blkid != DMU_BONUS_BLKID);
1841
1842 ASSERT(have_read ?
1843 RW_READ_HELD(&dn->dn_struct_rwlock) :
1844 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1845
1846 /*
1847 * if we have a read-lock, check to see if we need to do any work
1848 * before upgrading to a write-lock.
1849 */
1850 if (have_read) {
1851 if (blkid <= dn->dn_maxblkid)
1852 return;
1853
1854 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1855 rw_exit(&dn->dn_struct_rwlock);
1856 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1857 }
1858 }
1859
1860 if (blkid <= dn->dn_maxblkid)
1861 goto out;
1862
1863 dn->dn_maxblkid = blkid;
1864 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] = blkid;
1865
1866 /*
1867 * Compute the number of levels necessary to support the new maxblkid.
1868 */
1869 new_nlevels = 1;
1870 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1871 for (sz = dn->dn_nblkptr;
1872 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1873 new_nlevels++;
1874
1875 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1876
1877 if (new_nlevels > dn->dn_nlevels)
1878 dnode_set_nlevels_impl(dn, new_nlevels, tx);
1879
1880 out:
1881 if (have_read)
1882 rw_downgrade(&dn->dn_struct_rwlock);
1883 }
1884
1885 static void
1886 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1887 {
1888 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1889 if (db != NULL) {
1890 dmu_buf_will_dirty(&db->db, tx);
1891 dbuf_rele(db, FTAG);
1892 }
1893 }
1894
1895 /*
1896 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1897 * and end_blkid.
1898 */
1899 static void
1900 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1901 dmu_tx_t *tx)
1902 {
1903 dmu_buf_impl_t db_search;
1904 dmu_buf_impl_t *db;
1905 avl_index_t where;
1906
1907 mutex_enter(&dn->dn_dbufs_mtx);
1908
1909 db_search.db_level = 1;
1910 db_search.db_blkid = start_blkid + 1;
1911 db_search.db_state = DB_SEARCH;
1912 for (;;) {
1913
1914 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1915 if (db == NULL)
1916 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1917
1918 if (db == NULL || db->db_level != 1 ||
1919 db->db_blkid >= end_blkid) {
1920 break;
1921 }
1922
1923 /*
1924 * Setup the next blkid we want to search for.
1925 */
1926 db_search.db_blkid = db->db_blkid + 1;
1927 ASSERT3U(db->db_blkid, >=, start_blkid);
1928
1929 /*
1930 * If the dbuf transitions to DB_EVICTING while we're trying
1931 * to dirty it, then we will be unable to discover it in
1932 * the dbuf hash table. This will result in a call to
1933 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1934 * lock. To avoid a deadlock, we drop the lock before
1935 * dirtying the level-1 dbuf.
1936 */
1937 mutex_exit(&dn->dn_dbufs_mtx);
1938 dnode_dirty_l1(dn, db->db_blkid, tx);
1939 mutex_enter(&dn->dn_dbufs_mtx);
1940 }
1941
1942 #ifdef ZFS_DEBUG
1943 /*
1944 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1945 */
1946 db_search.db_level = 1;
1947 db_search.db_blkid = start_blkid + 1;
1948 db_search.db_state = DB_SEARCH;
1949 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1950 if (db == NULL)
1951 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1952 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
1953 if (db->db_level != 1 || db->db_blkid >= end_blkid)
1954 break;
1955 ASSERT(db->db_dirtycnt > 0);
1956 }
1957 #endif
1958 mutex_exit(&dn->dn_dbufs_mtx);
1959 }
1960
1961 void
1962 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1963 {
1964 dmu_buf_impl_t *db;
1965 uint64_t blkoff, blkid, nblks;
1966 int blksz, blkshift, head, tail;
1967 int trunc = FALSE;
1968 int epbs;
1969
1970 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1971 blksz = dn->dn_datablksz;
1972 blkshift = dn->dn_datablkshift;
1973 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1974
1975 if (len == DMU_OBJECT_END) {
1976 len = UINT64_MAX - off;
1977 trunc = TRUE;
1978 }
1979
1980 /*
1981 * First, block align the region to free:
1982 */
1983 if (ISP2(blksz)) {
1984 head = P2NPHASE(off, blksz);
1985 blkoff = P2PHASE(off, blksz);
1986 if ((off >> blkshift) > dn->dn_maxblkid)
1987 goto out;
1988 } else {
1989 ASSERT(dn->dn_maxblkid == 0);
1990 if (off == 0 && len >= blksz) {
1991 /*
1992 * Freeing the whole block; fast-track this request.
1993 */
1994 blkid = 0;
1995 nblks = 1;
1996 if (dn->dn_nlevels > 1)
1997 dnode_dirty_l1(dn, 0, tx);
1998 goto done;
1999 } else if (off >= blksz) {
2000 /* Freeing past end-of-data */
2001 goto out;
2002 } else {
2003 /* Freeing part of the block. */
2004 head = blksz - off;
2005 ASSERT3U(head, >, 0);
2006 }
2007 blkoff = off;
2008 }
2009 /* zero out any partial block data at the start of the range */
2010 if (head) {
2011 ASSERT3U(blkoff + head, ==, blksz);
2012 if (len < head)
2013 head = len;
2014 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2015 TRUE, FALSE, FTAG, &db) == 0) {
2016 caddr_t data;
2017
2018 /* don't dirty if it isn't on disk and isn't dirty */
2019 if (db->db_last_dirty ||
2020 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2021 rw_exit(&dn->dn_struct_rwlock);
2022 dmu_buf_will_dirty(&db->db, tx);
2023 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2024 data = db->db.db_data;
2025 bzero(data + blkoff, head);
2026 }
2027 dbuf_rele(db, FTAG);
2028 }
2029 off += head;
2030 len -= head;
2031 }
2032
2033 /* If the range was less than one block, we're done */
2034 if (len == 0)
2035 goto out;
2036
2037 /* If the remaining range is past end of file, we're done */
2038 if ((off >> blkshift) > dn->dn_maxblkid)
2039 goto out;
2040
2041 ASSERT(ISP2(blksz));
2042 if (trunc)
2043 tail = 0;
2044 else
2045 tail = P2PHASE(len, blksz);
2046
2047 ASSERT0(P2PHASE(off, blksz));
2048 /* zero out any partial block data at the end of the range */
2049 if (tail) {
2050 if (len < tail)
2051 tail = len;
2052 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2053 TRUE, FALSE, FTAG, &db) == 0) {
2054 /* don't dirty if not on disk and not dirty */
2055 if (db->db_last_dirty ||
2056 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2057 rw_exit(&dn->dn_struct_rwlock);
2058 dmu_buf_will_dirty(&db->db, tx);
2059 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2060 bzero(db->db.db_data, tail);
2061 }
2062 dbuf_rele(db, FTAG);
2063 }
2064 len -= tail;
2065 }
2066
2067 /* If the range did not include a full block, we are done */
2068 if (len == 0)
2069 goto out;
2070
2071 ASSERT(IS_P2ALIGNED(off, blksz));
2072 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2073 blkid = off >> blkshift;
2074 nblks = len >> blkshift;
2075 if (trunc)
2076 nblks += 1;
2077
2078 /*
2079 * Dirty all the indirect blocks in this range. Note that only
2080 * the first and last indirect blocks can actually be written
2081 * (if they were partially freed) -- they must be dirtied, even if
2082 * they do not exist on disk yet. The interior blocks will
2083 * be freed by free_children(), so they will not actually be written.
2084 * Even though these interior blocks will not be written, we
2085 * dirty them for two reasons:
2086 *
2087 * - It ensures that the indirect blocks remain in memory until
2088 * syncing context. (They have already been prefetched by
2089 * dmu_tx_hold_free(), so we don't have to worry about reading
2090 * them serially here.)
2091 *
2092 * - The dirty space accounting will put pressure on the txg sync
2093 * mechanism to begin syncing, and to delay transactions if there
2094 * is a large amount of freeing. Even though these indirect
2095 * blocks will not be written, we could need to write the same
2096 * amount of space if we copy the freed BPs into deadlists.
2097 */
2098 if (dn->dn_nlevels > 1) {
2099 uint64_t first, last;
2100
2101 first = blkid >> epbs;
2102 dnode_dirty_l1(dn, first, tx);
2103 if (trunc)
2104 last = dn->dn_maxblkid >> epbs;
2105 else
2106 last = (blkid + nblks - 1) >> epbs;
2107 if (last != first)
2108 dnode_dirty_l1(dn, last, tx);
2109
2110 dnode_dirty_l1range(dn, first, last, tx);
2111
2112 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2113 SPA_BLKPTRSHIFT;
2114 for (uint64_t i = first + 1; i < last; i++) {
2115 /*
2116 * Set i to the blockid of the next non-hole
2117 * level-1 indirect block at or after i. Note
2118 * that dnode_next_offset() operates in terms of
2119 * level-0-equivalent bytes.
2120 */
2121 uint64_t ibyte = i << shift;
2122 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2123 &ibyte, 2, 1, 0);
2124 i = ibyte >> shift;
2125 if (i >= last)
2126 break;
2127
2128 /*
2129 * Normally we should not see an error, either
2130 * from dnode_next_offset() or dbuf_hold_level()
2131 * (except for ESRCH from dnode_next_offset).
2132 * If there is an i/o error, then when we read
2133 * this block in syncing context, it will use
2134 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2135 * to the "failmode" property. dnode_next_offset()
2136 * doesn't have a flag to indicate MUSTSUCCEED.
2137 */
2138 if (err != 0)
2139 break;
2140
2141 dnode_dirty_l1(dn, i, tx);
2142 }
2143 }
2144
2145 done:
2146 /*
2147 * Add this range to the dnode range list.
2148 * We will finish up this free operation in the syncing phase.
2149 */
2150 mutex_enter(&dn->dn_mtx);
2151 {
2152 int txgoff = tx->tx_txg & TXG_MASK;
2153 if (dn->dn_free_ranges[txgoff] == NULL) {
2154 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2155 }
2156 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2157 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2158 }
2159 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2160 blkid, nblks, tx->tx_txg);
2161 mutex_exit(&dn->dn_mtx);
2162
2163 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2164 dnode_setdirty(dn, tx);
2165 out:
2166
2167 rw_exit(&dn->dn_struct_rwlock);
2168 }
2169
2170 static boolean_t
2171 dnode_spill_freed(dnode_t *dn)
2172 {
2173 int i;
2174
2175 mutex_enter(&dn->dn_mtx);
2176 for (i = 0; i < TXG_SIZE; i++) {
2177 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2178 break;
2179 }
2180 mutex_exit(&dn->dn_mtx);
2181 return (i < TXG_SIZE);
2182 }
2183
2184 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2185 uint64_t
2186 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2187 {
2188 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2189 int i;
2190
2191 if (blkid == DMU_BONUS_BLKID)
2192 return (FALSE);
2193
2194 /*
2195 * If we're in the process of opening the pool, dp will not be
2196 * set yet, but there shouldn't be anything dirty.
2197 */
2198 if (dp == NULL)
2199 return (FALSE);
2200
2201 if (dn->dn_free_txg)
2202 return (TRUE);
2203
2204 if (blkid == DMU_SPILL_BLKID)
2205 return (dnode_spill_freed(dn));
2206
2207 mutex_enter(&dn->dn_mtx);
2208 for (i = 0; i < TXG_SIZE; i++) {
2209 if (dn->dn_free_ranges[i] != NULL &&
2210 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2211 break;
2212 }
2213 mutex_exit(&dn->dn_mtx);
2214 return (i < TXG_SIZE);
2215 }
2216
2217 /* call from syncing context when we actually write/free space for this dnode */
2218 void
2219 dnode_diduse_space(dnode_t *dn, int64_t delta)
2220 {
2221 uint64_t space;
2222 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2223 dn, dn->dn_phys,
2224 (u_longlong_t)dn->dn_phys->dn_used,
2225 (longlong_t)delta);
2226
2227 mutex_enter(&dn->dn_mtx);
2228 space = DN_USED_BYTES(dn->dn_phys);
2229 if (delta > 0) {
2230 ASSERT3U(space + delta, >=, space); /* no overflow */
2231 } else {
2232 ASSERT3U(space, >=, -delta); /* no underflow */
2233 }
2234 space += delta;
2235 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2236 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2237 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2238 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2239 } else {
2240 dn->dn_phys->dn_used = space;
2241 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2242 }
2243 mutex_exit(&dn->dn_mtx);
2244 }
2245
2246 /*
2247 * Scans a block at the indicated "level" looking for a hole or data,
2248 * depending on 'flags'.
2249 *
2250 * If level > 0, then we are scanning an indirect block looking at its
2251 * pointers. If level == 0, then we are looking at a block of dnodes.
2252 *
2253 * If we don't find what we are looking for in the block, we return ESRCH.
2254 * Otherwise, return with *offset pointing to the beginning (if searching
2255 * forwards) or end (if searching backwards) of the range covered by the
2256 * block pointer we matched on (or dnode).
2257 *
2258 * The basic search algorithm used below by dnode_next_offset() is to
2259 * use this function to search up the block tree (widen the search) until
2260 * we find something (i.e., we don't return ESRCH) and then search back
2261 * down the tree (narrow the search) until we reach our original search
2262 * level.
2263 */
2264 static int
2265 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2266 int lvl, uint64_t blkfill, uint64_t txg)
2267 {
2268 dmu_buf_impl_t *db = NULL;
2269 void *data = NULL;
2270 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2271 uint64_t epb = 1ULL << epbs;
2272 uint64_t minfill, maxfill;
2273 boolean_t hole;
2274 int i, inc, error, span;
2275
2276 hole = ((flags & DNODE_FIND_HOLE) != 0);
2277 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2278 ASSERT(txg == 0 || !hole);
2279
2280 if (lvl == dn->dn_phys->dn_nlevels) {
2281 error = 0;
2282 epb = dn->dn_phys->dn_nblkptr;
2283 data = dn->dn_phys->dn_blkptr;
2284 } else {
2285 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2286 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2287 if (error) {
2288 if (error != ENOENT)
2289 return (error);
2290 if (hole)
2291 return (0);
2292 /*
2293 * This can only happen when we are searching up
2294 * the block tree for data. We don't really need to
2295 * adjust the offset, as we will just end up looking
2296 * at the pointer to this block in its parent, and its
2297 * going to be unallocated, so we will skip over it.
2298 */
2299 return (SET_ERROR(ESRCH));
2300 }
2301 error = dbuf_read(db, NULL,
2302 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2303 if (error) {
2304 dbuf_rele(db, FTAG);
2305 return (error);
2306 }
2307 data = db->db.db_data;
2308 }
2309
2310
2311 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2312 db->db_blkptr->blk_birth <= txg ||
2313 BP_IS_HOLE(db->db_blkptr))) {
2314 /*
2315 * This can only happen when we are searching up the tree
2316 * and these conditions mean that we need to keep climbing.
2317 */
2318 error = SET_ERROR(ESRCH);
2319 } else if (lvl == 0) {
2320 dnode_phys_t *dnp = data;
2321
2322 ASSERT(dn->dn_type == DMU_OT_DNODE);
2323 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2324
2325 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2326 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2327 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2328 break;
2329 }
2330
2331 if (i == blkfill)
2332 error = SET_ERROR(ESRCH);
2333
2334 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2335 (i << DNODE_SHIFT);
2336 } else {
2337 blkptr_t *bp = data;
2338 uint64_t start = *offset;
2339 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2340 minfill = 0;
2341 maxfill = blkfill << ((lvl - 1) * epbs);
2342
2343 if (hole)
2344 maxfill--;
2345 else
2346 minfill++;
2347
2348 if (span >= 8 * sizeof (*offset)) {
2349 /* This only happens on the highest indirection level */
2350 ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2351 *offset = 0;
2352 } else {
2353 *offset = *offset >> span;
2354 }
2355
2356 for (i = BF64_GET(*offset, 0, epbs);
2357 i >= 0 && i < epb; i += inc) {
2358 if (BP_GET_FILL(&bp[i]) >= minfill &&
2359 BP_GET_FILL(&bp[i]) <= maxfill &&
2360 (hole || bp[i].blk_birth > txg))
2361 break;
2362 if (inc > 0 || *offset > 0)
2363 *offset += inc;
2364 }
2365
2366 if (span >= 8 * sizeof (*offset)) {
2367 *offset = start;
2368 } else {
2369 *offset = *offset << span;
2370 }
2371
2372 if (inc < 0) {
2373 /* traversing backwards; position offset at the end */
2374 ASSERT3U(*offset, <=, start);
2375 *offset = MIN(*offset + (1ULL << span) - 1, start);
2376 } else if (*offset < start) {
2377 *offset = start;
2378 }
2379 if (i < 0 || i >= epb)
2380 error = SET_ERROR(ESRCH);
2381 }
2382
2383 if (db)
2384 dbuf_rele(db, FTAG);
2385
2386 return (error);
2387 }
2388
2389 /*
2390 * Find the next hole, data, or sparse region at or after *offset.
2391 * The value 'blkfill' tells us how many items we expect to find
2392 * in an L0 data block; this value is 1 for normal objects,
2393 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2394 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2395 *
2396 * Examples:
2397 *
2398 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2399 * Finds the next/previous hole/data in a file.
2400 * Used in dmu_offset_next().
2401 *
2402 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2403 * Finds the next free/allocated dnode an objset's meta-dnode.
2404 * Only finds objects that have new contents since txg (ie.
2405 * bonus buffer changes and content removal are ignored).
2406 * Used in dmu_object_next().
2407 *
2408 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2409 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2410 * Used in dmu_object_alloc().
2411 */
2412 int
2413 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2414 int minlvl, uint64_t blkfill, uint64_t txg)
2415 {
2416 uint64_t initial_offset = *offset;
2417 int lvl, maxlvl;
2418 int error = 0;
2419
2420 if (!(flags & DNODE_FIND_HAVELOCK))
2421 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2422
2423 if (dn->dn_phys->dn_nlevels == 0) {
2424 error = SET_ERROR(ESRCH);
2425 goto out;
2426 }
2427
2428 if (dn->dn_datablkshift == 0) {
2429 if (*offset < dn->dn_datablksz) {
2430 if (flags & DNODE_FIND_HOLE)
2431 *offset = dn->dn_datablksz;
2432 } else {
2433 error = SET_ERROR(ESRCH);
2434 }
2435 goto out;
2436 }
2437
2438 maxlvl = dn->dn_phys->dn_nlevels;
2439
2440 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2441 error = dnode_next_offset_level(dn,
2442 flags, offset, lvl, blkfill, txg);
2443 if (error != ESRCH)
2444 break;
2445 }
2446
2447 while (error == 0 && --lvl >= minlvl) {
2448 error = dnode_next_offset_level(dn,
2449 flags, offset, lvl, blkfill, txg);
2450 }
2451
2452 /*
2453 * There's always a "virtual hole" at the end of the object, even
2454 * if all BP's which physically exist are non-holes.
2455 */
2456 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2457 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2458 error = 0;
2459 }
2460
2461 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2462 initial_offset < *offset : initial_offset > *offset))
2463 error = SET_ERROR(ESRCH);
2464 out:
2465 if (!(flags & DNODE_FIND_HAVELOCK))
2466 rw_exit(&dn->dn_struct_rwlock);
2467
2468 return (error);
2469 }
ZFS On Linux mirror for PVE