1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * This file and its contents are supplied under the terms of the
4 * Common Development and Distribution License ("CDDL"), version 1.0.
5 * You may only use this file in accordance with the terms of version
6 * 1.0 of the CDDL.
7 *
8 * A full copy of the text of the CDDL should have accompanied this
9 * source. A copy of the CDDL is also available via the Internet at
10 * https://opensource.org/license/CDDL-1.0.
11 */
12 /*
13 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
14 * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
15 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
16 */
17
18 #include <sys/zfs_context.h>
19 #include <sys/dbuf.h>
20 #include <sys/dnode.h>
21 #include <sys/dmu.h>
22 #include <sys/dmu_impl.h>
23 #include <sys/dmu_tx.h>
24 #include <sys/dmu_objset.h>
25 #include <sys/dsl_dir.h>
26 #include <sys/dsl_dataset.h>
27 #include <sys/spa.h>
28 #include <sys/zio.h>
29 #include <sys/dmu_zfetch.h>
30 #include <sys/range_tree.h>
31 #include <sys/trace_zfs.h>
32 #include <sys/zfs_project.h>
33
34 dnode_stats_t dnode_stats = {
35 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64 },
36 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64 },
37 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64 },
38 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64 },
39 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64 },
40 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64 },
41 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64 },
42 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64 },
43 { "dnode_hold_free_hits", KSTAT_DATA_UINT64 },
44 { "dnode_hold_free_misses", KSTAT_DATA_UINT64 },
45 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64 },
46 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64 },
47 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64 },
48 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64 },
49 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 },
50 { "dnode_allocate", KSTAT_DATA_UINT64 },
51 { "dnode_reallocate", KSTAT_DATA_UINT64 },
52 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
53 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
54 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
55 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
56 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
57 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
58 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
59 { "dnode_move_special", KSTAT_DATA_UINT64 },
60 { "dnode_move_handle", KSTAT_DATA_UINT64 },
61 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
62 { "dnode_move_active", KSTAT_DATA_UINT64 },
63 };
64
65 dnode_sums_t dnode_sums;
66
67 static kstat_t *dnode_ksp;
68 static kmem_cache_t *dnode_cache;
69
70 static dnode_phys_t dnode_phys_zero __maybe_unused;
71
72 int zfs_default_bs = SPA_MINBLOCKSHIFT;
73 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
74
75 #ifdef _KERNEL
76 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
77 #endif /* _KERNEL */
78
79 static char *
rt_name(dnode_t * dn,const char * name)80 rt_name(dnode_t *dn, const char *name)
81 {
82 struct objset *os = dn->dn_objset;
83
84 return (kmem_asprintf("{spa=%s objset=%llu obj=%llu %s}",
85 spa_name(os->os_spa),
86 (u_longlong_t)(os->os_dsl_dataset ?
87 os->os_dsl_dataset->ds_object : DMU_META_OBJSET),
88 (u_longlong_t)dn->dn_object,
89 name));
90 }
91
92 static int
dbuf_compare(const void * x1,const void * x2)93 dbuf_compare(const void *x1, const void *x2)
94 {
95 const dmu_buf_impl_t *d1 = x1;
96 const dmu_buf_impl_t *d2 = x2;
97
98 int cmp = TREE_CMP(d1->db_level, d2->db_level);
99 if (likely(cmp))
100 return (cmp);
101
102 cmp = TREE_CMP(d1->db_blkid, d2->db_blkid);
103 if (likely(cmp))
104 return (cmp);
105
106 if (d1->db_state == DB_MARKER) {
107 ASSERT3S(d2->db_state, !=, DB_MARKER);
108 return (TREE_PCMP(d1->db_parent, d2));
109 } else if (d2->db_state == DB_MARKER) {
110 ASSERT3S(d1->db_state, !=, DB_MARKER);
111 return (TREE_PCMP(d1, d2->db_parent));
112 }
113
114 if (d1->db_state == DB_SEARCH) {
115 ASSERT3S(d2->db_state, !=, DB_SEARCH);
116 return (-1);
117 } else if (d2->db_state == DB_SEARCH) {
118 ASSERT3S(d1->db_state, !=, DB_SEARCH);
119 return (1);
120 }
121
122 return (TREE_PCMP(d1, d2));
123 }
124
125 static int
dnode_cons(void * arg,void * unused,int kmflag)126 dnode_cons(void *arg, void *unused, int kmflag)
127 {
128 (void) unused, (void) kmflag;
129 dnode_t *dn = arg;
130
131 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
132 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
133 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
134 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
135 cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
136
137 /*
138 * Every dbuf has a reference, and dropping a tracked reference is
139 * O(number of references), so don't track dn_holds.
140 */
141 zfs_refcount_create_untracked(&dn->dn_holds);
142 zfs_refcount_create(&dn->dn_tx_holds);
143 list_link_init(&dn->dn_link);
144
145 memset(dn->dn_next_type, 0, sizeof (dn->dn_next_type));
146 memset(dn->dn_next_nblkptr, 0, sizeof (dn->dn_next_nblkptr));
147 memset(dn->dn_next_nlevels, 0, sizeof (dn->dn_next_nlevels));
148 memset(dn->dn_next_indblkshift, 0, sizeof (dn->dn_next_indblkshift));
149 memset(dn->dn_next_bonustype, 0, sizeof (dn->dn_next_bonustype));
150 memset(dn->dn_rm_spillblk, 0, sizeof (dn->dn_rm_spillblk));
151 memset(dn->dn_next_bonuslen, 0, sizeof (dn->dn_next_bonuslen));
152 memset(dn->dn_next_blksz, 0, sizeof (dn->dn_next_blksz));
153 memset(dn->dn_next_maxblkid, 0, sizeof (dn->dn_next_maxblkid));
154
155 for (int i = 0; i < TXG_SIZE; i++) {
156 multilist_link_init(&dn->dn_dirty_link[i]);
157 dn->dn_free_ranges[i] = NULL;
158 list_create(&dn->dn_dirty_records[i],
159 sizeof (dbuf_dirty_record_t),
160 offsetof(dbuf_dirty_record_t, dr_dirty_node));
161 }
162
163 dn->dn_allocated_txg = 0;
164 dn->dn_free_txg = 0;
165 dn->dn_assigned_txg = 0;
166 dn->dn_dirtycnt = 0;
167 dn->dn_bonus = NULL;
168 dn->dn_have_spill = B_FALSE;
169 dn->dn_zio = NULL;
170 dn->dn_oldused = 0;
171 dn->dn_oldflags = 0;
172 dn->dn_olduid = 0;
173 dn->dn_oldgid = 0;
174 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
175 dn->dn_newuid = 0;
176 dn->dn_newgid = 0;
177 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
178 dn->dn_id_flags = 0;
179
180 dn->dn_dbufs_count = 0;
181 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
182 offsetof(dmu_buf_impl_t, db_link));
183
184 dn->dn_moved = 0;
185 return (0);
186 }
187
188 static void
dnode_dest(void * arg,void * unused)189 dnode_dest(void *arg, void *unused)
190 {
191 (void) unused;
192 dnode_t *dn = arg;
193
194 rw_destroy(&dn->dn_struct_rwlock);
195 mutex_destroy(&dn->dn_mtx);
196 mutex_destroy(&dn->dn_dbufs_mtx);
197 cv_destroy(&dn->dn_notxholds);
198 cv_destroy(&dn->dn_nodnholds);
199 zfs_refcount_destroy(&dn->dn_holds);
200 zfs_refcount_destroy(&dn->dn_tx_holds);
201 ASSERT(!list_link_active(&dn->dn_link));
202
203 for (int i = 0; i < TXG_SIZE; i++) {
204 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
205 ASSERT0P(dn->dn_free_ranges[i]);
206 list_destroy(&dn->dn_dirty_records[i]);
207 ASSERT0(dn->dn_next_nblkptr[i]);
208 ASSERT0(dn->dn_next_nlevels[i]);
209 ASSERT0(dn->dn_next_indblkshift[i]);
210 ASSERT0(dn->dn_next_bonustype[i]);
211 ASSERT0(dn->dn_rm_spillblk[i]);
212 ASSERT0(dn->dn_next_bonuslen[i]);
213 ASSERT0(dn->dn_next_blksz[i]);
214 ASSERT0(dn->dn_next_maxblkid[i]);
215 }
216
217 ASSERT0(dn->dn_allocated_txg);
218 ASSERT0(dn->dn_free_txg);
219 ASSERT0(dn->dn_assigned_txg);
220 ASSERT0(dn->dn_dirtycnt);
221 ASSERT0P(dn->dn_bonus);
222 ASSERT(!dn->dn_have_spill);
223 ASSERT0P(dn->dn_zio);
224 ASSERT0(dn->dn_oldused);
225 ASSERT0(dn->dn_oldflags);
226 ASSERT0(dn->dn_olduid);
227 ASSERT0(dn->dn_oldgid);
228 ASSERT0(dn->dn_oldprojid);
229 ASSERT0(dn->dn_newuid);
230 ASSERT0(dn->dn_newgid);
231 ASSERT0(dn->dn_newprojid);
232 ASSERT0(dn->dn_id_flags);
233
234 ASSERT0(dn->dn_dbufs_count);
235 avl_destroy(&dn->dn_dbufs);
236 }
237
238 static int
dnode_kstats_update(kstat_t * ksp,int rw)239 dnode_kstats_update(kstat_t *ksp, int rw)
240 {
241 dnode_stats_t *ds = ksp->ks_data;
242
243 if (rw == KSTAT_WRITE)
244 return (EACCES);
245 ds->dnode_hold_dbuf_hold.value.ui64 =
246 wmsum_value(&dnode_sums.dnode_hold_dbuf_hold);
247 ds->dnode_hold_dbuf_read.value.ui64 =
248 wmsum_value(&dnode_sums.dnode_hold_dbuf_read);
249 ds->dnode_hold_alloc_hits.value.ui64 =
250 wmsum_value(&dnode_sums.dnode_hold_alloc_hits);
251 ds->dnode_hold_alloc_misses.value.ui64 =
252 wmsum_value(&dnode_sums.dnode_hold_alloc_misses);
253 ds->dnode_hold_alloc_interior.value.ui64 =
254 wmsum_value(&dnode_sums.dnode_hold_alloc_interior);
255 ds->dnode_hold_alloc_lock_retry.value.ui64 =
256 wmsum_value(&dnode_sums.dnode_hold_alloc_lock_retry);
257 ds->dnode_hold_alloc_lock_misses.value.ui64 =
258 wmsum_value(&dnode_sums.dnode_hold_alloc_lock_misses);
259 ds->dnode_hold_alloc_type_none.value.ui64 =
260 wmsum_value(&dnode_sums.dnode_hold_alloc_type_none);
261 ds->dnode_hold_free_hits.value.ui64 =
262 wmsum_value(&dnode_sums.dnode_hold_free_hits);
263 ds->dnode_hold_free_misses.value.ui64 =
264 wmsum_value(&dnode_sums.dnode_hold_free_misses);
265 ds->dnode_hold_free_lock_misses.value.ui64 =
266 wmsum_value(&dnode_sums.dnode_hold_free_lock_misses);
267 ds->dnode_hold_free_lock_retry.value.ui64 =
268 wmsum_value(&dnode_sums.dnode_hold_free_lock_retry);
269 ds->dnode_hold_free_refcount.value.ui64 =
270 wmsum_value(&dnode_sums.dnode_hold_free_refcount);
271 ds->dnode_hold_free_overflow.value.ui64 =
272 wmsum_value(&dnode_sums.dnode_hold_free_overflow);
273 ds->dnode_free_interior_lock_retry.value.ui64 =
274 wmsum_value(&dnode_sums.dnode_free_interior_lock_retry);
275 ds->dnode_allocate.value.ui64 =
276 wmsum_value(&dnode_sums.dnode_allocate);
277 ds->dnode_reallocate.value.ui64 =
278 wmsum_value(&dnode_sums.dnode_reallocate);
279 ds->dnode_buf_evict.value.ui64 =
280 wmsum_value(&dnode_sums.dnode_buf_evict);
281 ds->dnode_alloc_next_chunk.value.ui64 =
282 wmsum_value(&dnode_sums.dnode_alloc_next_chunk);
283 ds->dnode_alloc_race.value.ui64 =
284 wmsum_value(&dnode_sums.dnode_alloc_race);
285 ds->dnode_alloc_next_block.value.ui64 =
286 wmsum_value(&dnode_sums.dnode_alloc_next_block);
287 ds->dnode_move_invalid.value.ui64 =
288 wmsum_value(&dnode_sums.dnode_move_invalid);
289 ds->dnode_move_recheck1.value.ui64 =
290 wmsum_value(&dnode_sums.dnode_move_recheck1);
291 ds->dnode_move_recheck2.value.ui64 =
292 wmsum_value(&dnode_sums.dnode_move_recheck2);
293 ds->dnode_move_special.value.ui64 =
294 wmsum_value(&dnode_sums.dnode_move_special);
295 ds->dnode_move_handle.value.ui64 =
296 wmsum_value(&dnode_sums.dnode_move_handle);
297 ds->dnode_move_rwlock.value.ui64 =
298 wmsum_value(&dnode_sums.dnode_move_rwlock);
299 ds->dnode_move_active.value.ui64 =
300 wmsum_value(&dnode_sums.dnode_move_active);
301 return (0);
302 }
303
304 void
dnode_init(void)305 dnode_init(void)
306 {
307 ASSERT0P(dnode_cache);
308 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
309 0, dnode_cons, dnode_dest, NULL, NULL, NULL, KMC_RECLAIMABLE);
310 kmem_cache_set_move(dnode_cache, dnode_move);
311
312 wmsum_init(&dnode_sums.dnode_hold_dbuf_hold, 0);
313 wmsum_init(&dnode_sums.dnode_hold_dbuf_read, 0);
314 wmsum_init(&dnode_sums.dnode_hold_alloc_hits, 0);
315 wmsum_init(&dnode_sums.dnode_hold_alloc_misses, 0);
316 wmsum_init(&dnode_sums.dnode_hold_alloc_interior, 0);
317 wmsum_init(&dnode_sums.dnode_hold_alloc_lock_retry, 0);
318 wmsum_init(&dnode_sums.dnode_hold_alloc_lock_misses, 0);
319 wmsum_init(&dnode_sums.dnode_hold_alloc_type_none, 0);
320 wmsum_init(&dnode_sums.dnode_hold_free_hits, 0);
321 wmsum_init(&dnode_sums.dnode_hold_free_misses, 0);
322 wmsum_init(&dnode_sums.dnode_hold_free_lock_misses, 0);
323 wmsum_init(&dnode_sums.dnode_hold_free_lock_retry, 0);
324 wmsum_init(&dnode_sums.dnode_hold_free_refcount, 0);
325 wmsum_init(&dnode_sums.dnode_hold_free_overflow, 0);
326 wmsum_init(&dnode_sums.dnode_free_interior_lock_retry, 0);
327 wmsum_init(&dnode_sums.dnode_allocate, 0);
328 wmsum_init(&dnode_sums.dnode_reallocate, 0);
329 wmsum_init(&dnode_sums.dnode_buf_evict, 0);
330 wmsum_init(&dnode_sums.dnode_alloc_next_chunk, 0);
331 wmsum_init(&dnode_sums.dnode_alloc_race, 0);
332 wmsum_init(&dnode_sums.dnode_alloc_next_block, 0);
333 wmsum_init(&dnode_sums.dnode_move_invalid, 0);
334 wmsum_init(&dnode_sums.dnode_move_recheck1, 0);
335 wmsum_init(&dnode_sums.dnode_move_recheck2, 0);
336 wmsum_init(&dnode_sums.dnode_move_special, 0);
337 wmsum_init(&dnode_sums.dnode_move_handle, 0);
338 wmsum_init(&dnode_sums.dnode_move_rwlock, 0);
339 wmsum_init(&dnode_sums.dnode_move_active, 0);
340
341 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
342 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
343 KSTAT_FLAG_VIRTUAL);
344 if (dnode_ksp != NULL) {
345 dnode_ksp->ks_data = &dnode_stats;
346 dnode_ksp->ks_update = dnode_kstats_update;
347 kstat_install(dnode_ksp);
348 }
349 }
350
351 void
dnode_fini(void)352 dnode_fini(void)
353 {
354 if (dnode_ksp != NULL) {
355 kstat_delete(dnode_ksp);
356 dnode_ksp = NULL;
357 }
358
359 wmsum_fini(&dnode_sums.dnode_hold_dbuf_hold);
360 wmsum_fini(&dnode_sums.dnode_hold_dbuf_read);
361 wmsum_fini(&dnode_sums.dnode_hold_alloc_hits);
362 wmsum_fini(&dnode_sums.dnode_hold_alloc_misses);
363 wmsum_fini(&dnode_sums.dnode_hold_alloc_interior);
364 wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_retry);
365 wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_misses);
366 wmsum_fini(&dnode_sums.dnode_hold_alloc_type_none);
367 wmsum_fini(&dnode_sums.dnode_hold_free_hits);
368 wmsum_fini(&dnode_sums.dnode_hold_free_misses);
369 wmsum_fini(&dnode_sums.dnode_hold_free_lock_misses);
370 wmsum_fini(&dnode_sums.dnode_hold_free_lock_retry);
371 wmsum_fini(&dnode_sums.dnode_hold_free_refcount);
372 wmsum_fini(&dnode_sums.dnode_hold_free_overflow);
373 wmsum_fini(&dnode_sums.dnode_free_interior_lock_retry);
374 wmsum_fini(&dnode_sums.dnode_allocate);
375 wmsum_fini(&dnode_sums.dnode_reallocate);
376 wmsum_fini(&dnode_sums.dnode_buf_evict);
377 wmsum_fini(&dnode_sums.dnode_alloc_next_chunk);
378 wmsum_fini(&dnode_sums.dnode_alloc_race);
379 wmsum_fini(&dnode_sums.dnode_alloc_next_block);
380 wmsum_fini(&dnode_sums.dnode_move_invalid);
381 wmsum_fini(&dnode_sums.dnode_move_recheck1);
382 wmsum_fini(&dnode_sums.dnode_move_recheck2);
383 wmsum_fini(&dnode_sums.dnode_move_special);
384 wmsum_fini(&dnode_sums.dnode_move_handle);
385 wmsum_fini(&dnode_sums.dnode_move_rwlock);
386 wmsum_fini(&dnode_sums.dnode_move_active);
387
388 kmem_cache_destroy(dnode_cache);
389 dnode_cache = NULL;
390 }
391
392
393 #ifdef ZFS_DEBUG
394 void
dnode_verify(dnode_t * dn)395 dnode_verify(dnode_t *dn)
396 {
397 int drop_struct_lock = FALSE;
398
399 ASSERT(dn->dn_phys);
400 ASSERT(dn->dn_objset);
401 ASSERT(dn->dn_handle->dnh_dnode == dn);
402
403 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
404
405 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
406 return;
407
408 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
409 rw_enter(&dn->dn_struct_rwlock, RW_READER);
410 drop_struct_lock = TRUE;
411 }
412 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
413 int i;
414 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
415 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
416 if (dn->dn_datablkshift) {
417 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
418 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
419 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
420 }
421 ASSERT3U(dn->dn_nlevels, <=, 30);
422 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
423 ASSERT3U(dn->dn_nblkptr, >=, 1);
424 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
425 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
426 ASSERT3U(dn->dn_datablksz, ==,
427 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
428 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
429 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
430 dn->dn_bonuslen, <=, max_bonuslen);
431 for (i = 0; i < TXG_SIZE; i++) {
432 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
433 }
434 }
435 if (dn->dn_phys->dn_type != DMU_OT_NONE)
436 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
437 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
438 if (dn->dn_dbuf != NULL) {
439 ASSERT3P(dn->dn_phys, ==,
440 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
441 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
442 }
443 if (drop_struct_lock)
444 rw_exit(&dn->dn_struct_rwlock);
445 }
446 #endif
447
448 void
dnode_byteswap(dnode_phys_t * dnp)449 dnode_byteswap(dnode_phys_t *dnp)
450 {
451 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
452 int i;
453
454 if (dnp->dn_type == DMU_OT_NONE) {
455 memset(dnp, 0, sizeof (dnode_phys_t));
456 return;
457 }
458
459 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
460 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
461 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
462 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
463 dnp->dn_used = BSWAP_64(dnp->dn_used);
464
465 /*
466 * dn_nblkptr is only one byte, so it's OK to read it in either
467 * byte order. We can't read dn_bouslen.
468 */
469 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
470 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
471 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
472 buf64[i] = BSWAP_64(buf64[i]);
473
474 /*
475 * OK to check dn_bonuslen for zero, because it won't matter if
476 * we have the wrong byte order. This is necessary because the
477 * dnode dnode is smaller than a regular dnode.
478 */
479 if (dnp->dn_bonuslen != 0) {
480 dmu_object_byteswap_t byteswap;
481 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
482 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
483 dmu_ot_byteswap[byteswap].ob_func(DN_BONUS(dnp),
484 DN_MAX_BONUS_LEN(dnp));
485 }
486
487 /* Swap SPILL block if we have one */
488 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
489 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
490 }
491
492 void
dnode_buf_byteswap(void * vbuf,size_t size)493 dnode_buf_byteswap(void *vbuf, size_t size)
494 {
495 int i = 0;
496
497 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
498 ASSERT0((size & (sizeof (dnode_phys_t)-1)));
499
500 while (i < size) {
501 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
502 dnode_byteswap(dnp);
503
504 i += DNODE_MIN_SIZE;
505 if (dnp->dn_type != DMU_OT_NONE)
506 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
507 }
508 }
509
510 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)511 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
512 {
513 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
514
515 dnode_setdirty(dn, tx);
516 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
517 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
518 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
519
520 if (newsize < dn->dn_bonuslen) {
521 /* clear any data after the end of the new size */
522 size_t diff = dn->dn_bonuslen - newsize;
523 char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
524 memset(data_end, 0, diff);
525 }
526
527 dn->dn_bonuslen = newsize;
528 if (newsize == 0)
529 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
530 else
531 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
532 rw_exit(&dn->dn_struct_rwlock);
533 }
534
535 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)536 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
537 {
538 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
539 dnode_setdirty(dn, tx);
540 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
541 dn->dn_bonustype = newtype;
542 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
543 rw_exit(&dn->dn_struct_rwlock);
544 }
545
546 void
dnode_set_storage_type(dnode_t * dn,dmu_object_type_t newtype)547 dnode_set_storage_type(dnode_t *dn, dmu_object_type_t newtype)
548 {
549 /*
550 * This is not in the dnode_phys, but it should be, and perhaps one day
551 * will. For now we require it be set after taking a hold.
552 */
553 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
554 dn->dn_storage_type = newtype;
555 }
556
557 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)558 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
559 {
560 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
561 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
562 dnode_setdirty(dn, tx);
563 dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
564 dn->dn_have_spill = B_FALSE;
565 }
566
567 static void
dnode_setdblksz(dnode_t * dn,int size)568 dnode_setdblksz(dnode_t *dn, int size)
569 {
570 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
571 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
572 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
573 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
574 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
575 dn->dn_datablksz = size;
576 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
577 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
578 }
579
580 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)581 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
582 uint64_t object, dnode_handle_t *dnh)
583 {
584 dnode_t *dn;
585
586 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
587 dn->dn_moved = 0;
588
589 /*
590 * Defer setting dn_objset until the dnode is ready to be a candidate
591 * for the dnode_move() callback.
592 */
593 dn->dn_object = object;
594 dn->dn_dbuf = db;
595 dn->dn_handle = dnh;
596 dn->dn_phys = dnp;
597
598 if (dnp->dn_datablkszsec) {
599 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
600 } else {
601 dn->dn_datablksz = 0;
602 dn->dn_datablkszsec = 0;
603 dn->dn_datablkshift = 0;
604 }
605 dn->dn_indblkshift = dnp->dn_indblkshift;
606 dn->dn_nlevels = dnp->dn_nlevels;
607 dn->dn_type = dnp->dn_type;
608 dn->dn_nblkptr = dnp->dn_nblkptr;
609 dn->dn_checksum = dnp->dn_checksum;
610 dn->dn_compress = dnp->dn_compress;
611 dn->dn_bonustype = dnp->dn_bonustype;
612 dn->dn_bonuslen = dnp->dn_bonuslen;
613 dn->dn_num_slots = dnp->dn_extra_slots + 1;
614 dn->dn_maxblkid = dnp->dn_maxblkid;
615 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
616 dn->dn_id_flags = 0;
617
618 dn->dn_storage_type = DMU_OT_NONE;
619
620 dmu_zfetch_init(&dn->dn_zfetch, dn);
621
622 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
623 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
624 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
625
626 mutex_enter(&os->os_lock);
627
628 /*
629 * Exclude special dnodes from os_dnodes so an empty os_dnodes
630 * signifies that the special dnodes have no references from
631 * their children (the entries in os_dnodes). This allows
632 * dnode_destroy() to easily determine if the last child has
633 * been removed and then complete eviction of the objset.
634 */
635 if (!DMU_OBJECT_IS_SPECIAL(object))
636 list_insert_head(&os->os_dnodes, dn);
637 membar_producer();
638
639 /*
640 * Everything else must be valid before assigning dn_objset
641 * makes the dnode eligible for dnode_move().
642 */
643 dn->dn_objset = os;
644
645 dnh->dnh_dnode = dn;
646 mutex_exit(&os->os_lock);
647
648 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
649
650 return (dn);
651 }
652
653 /*
654 * Caller must be holding the dnode handle, which is released upon return.
655 */
656 static void
dnode_destroy(dnode_t * dn)657 dnode_destroy(dnode_t *dn)
658 {
659 objset_t *os = dn->dn_objset;
660 boolean_t complete_os_eviction = B_FALSE;
661
662 ASSERT0((dn->dn_id_flags & DN_ID_NEW_EXIST));
663
664 mutex_enter(&os->os_lock);
665 POINTER_INVALIDATE(&dn->dn_objset);
666 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
667 list_remove(&os->os_dnodes, dn);
668 complete_os_eviction =
669 list_is_empty(&os->os_dnodes) &&
670 list_link_active(&os->os_evicting_node);
671 }
672 mutex_exit(&os->os_lock);
673
674 /* the dnode can no longer move, so we can release the handle */
675 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
676 zrl_remove(&dn->dn_handle->dnh_zrlock);
677
678 dn->dn_allocated_txg = 0;
679 dn->dn_free_txg = 0;
680 dn->dn_assigned_txg = 0;
681 dn->dn_dirtycnt = 0;
682
683 if (dn->dn_bonus != NULL) {
684 mutex_enter(&dn->dn_bonus->db_mtx);
685 dbuf_destroy(dn->dn_bonus);
686 dn->dn_bonus = NULL;
687 }
688 dn->dn_zio = NULL;
689
690 dn->dn_have_spill = B_FALSE;
691 dn->dn_oldused = 0;
692 dn->dn_oldflags = 0;
693 dn->dn_olduid = 0;
694 dn->dn_oldgid = 0;
695 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
696 dn->dn_newuid = 0;
697 dn->dn_newgid = 0;
698 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
699 dn->dn_id_flags = 0;
700
701 dn->dn_storage_type = DMU_OT_NONE;
702
703 dmu_zfetch_fini(&dn->dn_zfetch);
704 kmem_cache_free(dnode_cache, dn);
705 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
706
707 if (complete_os_eviction)
708 dmu_objset_evict_done(os);
709 }
710
711 void
dnode_allocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,int ibs,dmu_object_type_t bonustype,int bonuslen,int dn_slots,dmu_tx_t * tx)712 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
713 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
714 {
715 int i;
716
717 ASSERT3U(dn_slots, >, 0);
718 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
719 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
720 ASSERT3U(blocksize, <=,
721 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
722 if (blocksize == 0)
723 blocksize = 1 << zfs_default_bs;
724 else
725 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
726
727 if (ibs == 0)
728 ibs = zfs_default_ibs;
729
730 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
731
732 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
733 dn->dn_objset, (u_longlong_t)dn->dn_object,
734 (u_longlong_t)tx->tx_txg, blocksize, ibs, dn_slots);
735 DNODE_STAT_BUMP(dnode_allocate);
736
737 ASSERT(dn->dn_type == DMU_OT_NONE);
738 ASSERT0(memcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)));
739 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
740 ASSERT(ot != DMU_OT_NONE);
741 ASSERT(DMU_OT_IS_VALID(ot));
742 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
743 (bonustype == DMU_OT_SA && bonuslen == 0) ||
744 (bonustype == DMU_OTN_UINT64_METADATA && bonuslen == 0) ||
745 (bonustype != DMU_OT_NONE && bonuslen != 0));
746 ASSERT(DMU_OT_IS_VALID(bonustype));
747 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
748 ASSERT(dn->dn_type == DMU_OT_NONE);
749 ASSERT0(dn->dn_maxblkid);
750 ASSERT0(dn->dn_allocated_txg);
751 ASSERT0(dn->dn_assigned_txg);
752 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
753 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
754 ASSERT(avl_is_empty(&dn->dn_dbufs));
755
756 for (i = 0; i < TXG_SIZE; i++) {
757 ASSERT0(dn->dn_next_nblkptr[i]);
758 ASSERT0(dn->dn_next_nlevels[i]);
759 ASSERT0(dn->dn_next_indblkshift[i]);
760 ASSERT0(dn->dn_next_bonuslen[i]);
761 ASSERT0(dn->dn_next_bonustype[i]);
762 ASSERT0(dn->dn_rm_spillblk[i]);
763 ASSERT0(dn->dn_next_blksz[i]);
764 ASSERT0(dn->dn_next_maxblkid[i]);
765 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
766 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
767 ASSERT0P(dn->dn_free_ranges[i]);
768 }
769
770 dn->dn_type = ot;
771 dnode_setdblksz(dn, blocksize);
772 dn->dn_indblkshift = ibs;
773 dn->dn_nlevels = 1;
774 dn->dn_num_slots = dn_slots;
775 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
776 dn->dn_nblkptr = 1;
777 else {
778 /*
779 * Keep in sync with deduce_nblkptr() in dmu_recv.c.
780 */
781 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
782 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
783 SPA_BLKPTRSHIFT));
784 }
785
786 dn->dn_bonustype = bonustype;
787 dn->dn_bonuslen = bonuslen;
788 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
789 dn->dn_compress = ZIO_COMPRESS_INHERIT;
790
791 dn->dn_free_txg = 0;
792 dn->dn_dirtycnt = 0;
793
794 dn->dn_allocated_txg = tx->tx_txg;
795 dn->dn_id_flags = 0;
796
797 dnode_setdirty(dn, tx);
798 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
799 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
800 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
801 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
802 }
803
804 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,int dn_slots,boolean_t keep_spill,dmu_tx_t * tx)805 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
806 dmu_object_type_t bonustype, int bonuslen, int dn_slots,
807 boolean_t keep_spill, dmu_tx_t *tx)
808 {
809 int nblkptr;
810
811 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
812 ASSERT3U(blocksize, <=,
813 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
814 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
815 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
816 ASSERT(tx->tx_txg != 0);
817 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
818 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
819 (bonustype == DMU_OT_SA && bonuslen == 0));
820 ASSERT(DMU_OT_IS_VALID(bonustype));
821 ASSERT3U(bonuslen, <=,
822 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
823 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
824
825 dnode_free_interior_slots(dn);
826 DNODE_STAT_BUMP(dnode_reallocate);
827
828 /* clean up any unreferenced dbufs */
829 dnode_evict_dbufs(dn);
830
831 dn->dn_id_flags = 0;
832
833 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
834 dnode_setdirty(dn, tx);
835 if (dn->dn_datablksz != blocksize) {
836 /* change blocksize */
837 ASSERT0(dn->dn_maxblkid);
838 ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
839 dnode_block_freed(dn, 0));
840
841 dnode_setdblksz(dn, blocksize);
842 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
843 }
844 if (dn->dn_bonuslen != bonuslen)
845 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
846
847 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
848 nblkptr = 1;
849 else {
850 /*
851 * Keep in sync with deduce_nblkptr() in dmu_recv.c.
852 */
853 nblkptr = MIN(DN_MAX_NBLKPTR,
854 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
855 SPA_BLKPTRSHIFT));
856 }
857 if (dn->dn_bonustype != bonustype)
858 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
859 if (dn->dn_nblkptr != nblkptr)
860 dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
861 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
862 dbuf_rm_spill(dn, tx);
863 dnode_rm_spill(dn, tx);
864 }
865
866 rw_exit(&dn->dn_struct_rwlock);
867
868 /* change type */
869 dn->dn_type = ot;
870
871 /* change bonus size and type */
872 mutex_enter(&dn->dn_mtx);
873 dn->dn_bonustype = bonustype;
874 dn->dn_bonuslen = bonuslen;
875 dn->dn_num_slots = dn_slots;
876 dn->dn_nblkptr = nblkptr;
877 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
878 dn->dn_compress = ZIO_COMPRESS_INHERIT;
879 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
880
881 /* fix up the bonus db_size */
882 if (dn->dn_bonus) {
883 dn->dn_bonus->db.db_size =
884 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
885 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
886 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
887 }
888
889 dn->dn_allocated_txg = tx->tx_txg;
890 mutex_exit(&dn->dn_mtx);
891 }
892
893 #ifdef _KERNEL
894 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)895 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
896 {
897 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
898 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
899 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
900
901 /* Copy fields. */
902 ndn->dn_objset = odn->dn_objset;
903 ndn->dn_object = odn->dn_object;
904 ndn->dn_dbuf = odn->dn_dbuf;
905 ndn->dn_handle = odn->dn_handle;
906 ndn->dn_phys = odn->dn_phys;
907 ndn->dn_type = odn->dn_type;
908 ndn->dn_bonuslen = odn->dn_bonuslen;
909 ndn->dn_bonustype = odn->dn_bonustype;
910 ndn->dn_nblkptr = odn->dn_nblkptr;
911 ndn->dn_checksum = odn->dn_checksum;
912 ndn->dn_compress = odn->dn_compress;
913 ndn->dn_nlevels = odn->dn_nlevels;
914 ndn->dn_indblkshift = odn->dn_indblkshift;
915 ndn->dn_datablkshift = odn->dn_datablkshift;
916 ndn->dn_datablkszsec = odn->dn_datablkszsec;
917 ndn->dn_datablksz = odn->dn_datablksz;
918 ndn->dn_maxblkid = odn->dn_maxblkid;
919 ndn->dn_num_slots = odn->dn_num_slots;
920 memcpy(ndn->dn_next_type, odn->dn_next_type,
921 sizeof (odn->dn_next_type));
922 memcpy(ndn->dn_next_nblkptr, odn->dn_next_nblkptr,
923 sizeof (odn->dn_next_nblkptr));
924 memcpy(ndn->dn_next_nlevels, odn->dn_next_nlevels,
925 sizeof (odn->dn_next_nlevels));
926 memcpy(ndn->dn_next_indblkshift, odn->dn_next_indblkshift,
927 sizeof (odn->dn_next_indblkshift));
928 memcpy(ndn->dn_next_bonustype, odn->dn_next_bonustype,
929 sizeof (odn->dn_next_bonustype));
930 memcpy(ndn->dn_rm_spillblk, odn->dn_rm_spillblk,
931 sizeof (odn->dn_rm_spillblk));
932 memcpy(ndn->dn_next_bonuslen, odn->dn_next_bonuslen,
933 sizeof (odn->dn_next_bonuslen));
934 memcpy(ndn->dn_next_blksz, odn->dn_next_blksz,
935 sizeof (odn->dn_next_blksz));
936 memcpy(ndn->dn_next_maxblkid, odn->dn_next_maxblkid,
937 sizeof (odn->dn_next_maxblkid));
938 for (int i = 0; i < TXG_SIZE; i++) {
939 list_move_tail(&ndn->dn_dirty_records[i],
940 &odn->dn_dirty_records[i]);
941 }
942 memcpy(ndn->dn_free_ranges, odn->dn_free_ranges,
943 sizeof (odn->dn_free_ranges));
944 ndn->dn_allocated_txg = odn->dn_allocated_txg;
945 ndn->dn_free_txg = odn->dn_free_txg;
946 ndn->dn_assigned_txg = odn->dn_assigned_txg;
947 ndn->dn_dirtycnt = odn->dn_dirtycnt;
948 ASSERT0(zfs_refcount_count(&odn->dn_tx_holds));
949 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
950 ASSERT(avl_is_empty(&ndn->dn_dbufs));
951 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
952 ndn->dn_dbufs_count = odn->dn_dbufs_count;
953 ndn->dn_bonus = odn->dn_bonus;
954 ndn->dn_have_spill = odn->dn_have_spill;
955 ndn->dn_zio = odn->dn_zio;
956 ndn->dn_oldused = odn->dn_oldused;
957 ndn->dn_oldflags = odn->dn_oldflags;
958 ndn->dn_olduid = odn->dn_olduid;
959 ndn->dn_oldgid = odn->dn_oldgid;
960 ndn->dn_oldprojid = odn->dn_oldprojid;
961 ndn->dn_newuid = odn->dn_newuid;
962 ndn->dn_newgid = odn->dn_newgid;
963 ndn->dn_newprojid = odn->dn_newprojid;
964 ndn->dn_id_flags = odn->dn_id_flags;
965 ndn->dn_storage_type = odn->dn_storage_type;
966 dmu_zfetch_init(&ndn->dn_zfetch, ndn);
967
968 /*
969 * Update back pointers. Updating the handle fixes the back pointer of
970 * every descendant dbuf as well as the bonus dbuf.
971 */
972 ASSERT(ndn->dn_handle->dnh_dnode == odn);
973 ndn->dn_handle->dnh_dnode = ndn;
974
975 /*
976 * Invalidate the original dnode by clearing all of its back pointers.
977 */
978 odn->dn_dbuf = NULL;
979 odn->dn_handle = NULL;
980 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
981 offsetof(dmu_buf_impl_t, db_link));
982 odn->dn_dbufs_count = 0;
983 odn->dn_bonus = NULL;
984 dmu_zfetch_fini(&odn->dn_zfetch);
985
986 /*
987 * Set the low bit of the objset pointer to ensure that dnode_move()
988 * recognizes the dnode as invalid in any subsequent callback.
989 */
990 POINTER_INVALIDATE(&odn->dn_objset);
991
992 /*
993 * Satisfy the destructor.
994 */
995 for (int i = 0; i < TXG_SIZE; i++) {
996 list_create(&odn->dn_dirty_records[i],
997 sizeof (dbuf_dirty_record_t),
998 offsetof(dbuf_dirty_record_t, dr_dirty_node));
999 odn->dn_free_ranges[i] = NULL;
1000 odn->dn_next_nlevels[i] = 0;
1001 odn->dn_next_indblkshift[i] = 0;
1002 odn->dn_next_bonustype[i] = 0;
1003 odn->dn_rm_spillblk[i] = 0;
1004 odn->dn_next_bonuslen[i] = 0;
1005 odn->dn_next_blksz[i] = 0;
1006 }
1007 odn->dn_allocated_txg = 0;
1008 odn->dn_free_txg = 0;
1009 odn->dn_assigned_txg = 0;
1010 odn->dn_dirtycnt = 0;
1011 odn->dn_have_spill = B_FALSE;
1012 odn->dn_zio = NULL;
1013 odn->dn_oldused = 0;
1014 odn->dn_oldflags = 0;
1015 odn->dn_olduid = 0;
1016 odn->dn_oldgid = 0;
1017 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
1018 odn->dn_newuid = 0;
1019 odn->dn_newgid = 0;
1020 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
1021 odn->dn_id_flags = 0;
1022 odn->dn_storage_type = DMU_OT_NONE;
1023
1024 /*
1025 * Mark the dnode.
1026 */
1027 ndn->dn_moved = 1;
1028 odn->dn_moved = (uint8_t)-1;
1029 }
1030
1031 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)1032 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
1033 {
1034 dnode_t *odn = buf, *ndn = newbuf;
1035 objset_t *os;
1036 int64_t refcount;
1037 uint32_t dbufs;
1038
1039 #ifndef USE_DNODE_HANDLE
1040 /*
1041 * We can't move dnodes if dbufs reference them directly without
1042 * using handles and respecitve locking. Unless USE_DNODE_HANDLE
1043 * is defined the code below is only to make sure it still builds,
1044 * but it should never be used, since it is unsafe.
1045 */
1046 #ifdef ZFS_DEBUG
1047 PANIC("dnode_move() called without USE_DNODE_HANDLE");
1048 #endif
1049 return (KMEM_CBRC_NO);
1050 #endif
1051
1052 /*
1053 * The dnode is on the objset's list of known dnodes if the objset
1054 * pointer is valid. We set the low bit of the objset pointer when
1055 * freeing the dnode to invalidate it, and the memory patterns written
1056 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
1057 * A newly created dnode sets the objset pointer last of all to indicate
1058 * that the dnode is known and in a valid state to be moved by this
1059 * function.
1060 */
1061 os = odn->dn_objset;
1062 if (!POINTER_IS_VALID(os)) {
1063 DNODE_STAT_BUMP(dnode_move_invalid);
1064 return (KMEM_CBRC_DONT_KNOW);
1065 }
1066
1067 /*
1068 * Ensure that the objset does not go away during the move.
1069 */
1070 rw_enter(&os_lock, RW_WRITER);
1071 if (os != odn->dn_objset) {
1072 rw_exit(&os_lock);
1073 DNODE_STAT_BUMP(dnode_move_recheck1);
1074 return (KMEM_CBRC_DONT_KNOW);
1075 }
1076
1077 /*
1078 * If the dnode is still valid, then so is the objset. We know that no
1079 * valid objset can be freed while we hold os_lock, so we can safely
1080 * ensure that the objset remains in use.
1081 */
1082 mutex_enter(&os->os_lock);
1083
1084 /*
1085 * Recheck the objset pointer in case the dnode was removed just before
1086 * acquiring the lock.
1087 */
1088 if (os != odn->dn_objset) {
1089 mutex_exit(&os->os_lock);
1090 rw_exit(&os_lock);
1091 DNODE_STAT_BUMP(dnode_move_recheck2);
1092 return (KMEM_CBRC_DONT_KNOW);
1093 }
1094
1095 /*
1096 * At this point we know that as long as we hold os->os_lock, the dnode
1097 * cannot be freed and fields within the dnode can be safely accessed.
1098 * The objset listing this dnode cannot go away as long as this dnode is
1099 * on its list.
1100 */
1101 rw_exit(&os_lock);
1102 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
1103 mutex_exit(&os->os_lock);
1104 DNODE_STAT_BUMP(dnode_move_special);
1105 return (KMEM_CBRC_NO);
1106 }
1107 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
1108
1109 /*
1110 * Lock the dnode handle to prevent the dnode from obtaining any new
1111 * holds. This also prevents the descendant dbufs and the bonus dbuf
1112 * from accessing the dnode, so that we can discount their holds. The
1113 * handle is safe to access because we know that while the dnode cannot
1114 * go away, neither can its handle. Once we hold dnh_zrlock, we can
1115 * safely move any dnode referenced only by dbufs.
1116 */
1117 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
1118 mutex_exit(&os->os_lock);
1119 DNODE_STAT_BUMP(dnode_move_handle);
1120 return (KMEM_CBRC_LATER);
1121 }
1122
1123 /*
1124 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
1125 * We need to guarantee that there is a hold for every dbuf in order to
1126 * determine whether the dnode is actively referenced. Falsely matching
1127 * a dbuf to an active hold would lead to an unsafe move. It's possible
1128 * that a thread already having an active dnode hold is about to add a
1129 * dbuf, and we can't compare hold and dbuf counts while the add is in
1130 * progress.
1131 */
1132 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
1133 zrl_exit(&odn->dn_handle->dnh_zrlock);
1134 mutex_exit(&os->os_lock);
1135 DNODE_STAT_BUMP(dnode_move_rwlock);
1136 return (KMEM_CBRC_LATER);
1137 }
1138
1139 /*
1140 * A dbuf may be removed (evicted) without an active dnode hold. In that
1141 * case, the dbuf count is decremented under the handle lock before the
1142 * dbuf's hold is released. This order ensures that if we count the hold
1143 * after the dbuf is removed but before its hold is released, we will
1144 * treat the unmatched hold as active and exit safely. If we count the
1145 * hold before the dbuf is removed, the hold is discounted, and the
1146 * removal is blocked until the move completes.
1147 */
1148 refcount = zfs_refcount_count(&odn->dn_holds);
1149 ASSERT(refcount >= 0);
1150 dbufs = DN_DBUFS_COUNT(odn);
1151
1152 /* We can't have more dbufs than dnode holds. */
1153 ASSERT3U(dbufs, <=, refcount);
1154 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1155 uint32_t, dbufs);
1156
1157 if (refcount > dbufs) {
1158 rw_exit(&odn->dn_struct_rwlock);
1159 zrl_exit(&odn->dn_handle->dnh_zrlock);
1160 mutex_exit(&os->os_lock);
1161 DNODE_STAT_BUMP(dnode_move_active);
1162 return (KMEM_CBRC_LATER);
1163 }
1164
1165 rw_exit(&odn->dn_struct_rwlock);
1166
1167 /*
1168 * At this point we know that anyone with a hold on the dnode is not
1169 * actively referencing it. The dnode is known and in a valid state to
1170 * move. We're holding the locks needed to execute the critical section.
1171 */
1172 dnode_move_impl(odn, ndn);
1173
1174 list_link_replace(&odn->dn_link, &ndn->dn_link);
1175 /* If the dnode was safe to move, the refcount cannot have changed. */
1176 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1177 ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1178 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1179 mutex_exit(&os->os_lock);
1180
1181 return (KMEM_CBRC_YES);
1182 }
1183 #endif /* _KERNEL */
1184
1185 static void
dnode_slots_hold(dnode_children_t * children,int idx,int slots)1186 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1187 {
1188 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1189
1190 for (int i = idx; i < idx + slots; i++) {
1191 dnode_handle_t *dnh = &children->dnc_children[i];
1192 zrl_add(&dnh->dnh_zrlock);
1193 }
1194 }
1195
1196 static void
dnode_slots_rele(dnode_children_t * children,int idx,int slots)1197 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1198 {
1199 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1200
1201 for (int i = idx; i < idx + slots; i++) {
1202 dnode_handle_t *dnh = &children->dnc_children[i];
1203
1204 if (zrl_is_locked(&dnh->dnh_zrlock))
1205 zrl_exit(&dnh->dnh_zrlock);
1206 else
1207 zrl_remove(&dnh->dnh_zrlock);
1208 }
1209 }
1210
1211 static int
dnode_slots_tryenter(dnode_children_t * children,int idx,int slots)1212 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1213 {
1214 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1215
1216 for (int i = idx; i < idx + slots; i++) {
1217 dnode_handle_t *dnh = &children->dnc_children[i];
1218
1219 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1220 for (int j = idx; j < i; j++) {
1221 dnh = &children->dnc_children[j];
1222 zrl_exit(&dnh->dnh_zrlock);
1223 }
1224
1225 return (0);
1226 }
1227 }
1228
1229 return (1);
1230 }
1231
1232 static void
dnode_set_slots(dnode_children_t * children,int idx,int slots,void * ptr)1233 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1234 {
1235 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1236
1237 for (int i = idx; i < idx + slots; i++) {
1238 dnode_handle_t *dnh = &children->dnc_children[i];
1239 dnh->dnh_dnode = ptr;
1240 }
1241 }
1242
1243 static boolean_t
dnode_check_slots_free(dnode_children_t * children,int idx,int slots)1244 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1245 {
1246 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1247
1248 /*
1249 * If all dnode slots are either already free or
1250 * evictable return B_TRUE.
1251 */
1252 for (int i = idx; i < idx + slots; i++) {
1253 dnode_handle_t *dnh = &children->dnc_children[i];
1254 dnode_t *dn = dnh->dnh_dnode;
1255
1256 if (dn == DN_SLOT_FREE) {
1257 continue;
1258 } else if (DN_SLOT_IS_PTR(dn)) {
1259 mutex_enter(&dn->dn_mtx);
1260 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1261 dn->dn_dirtycnt == 0 &&
1262 zfs_refcount_is_zero(&dn->dn_holds));
1263 mutex_exit(&dn->dn_mtx);
1264
1265 if (!can_free)
1266 return (B_FALSE);
1267 else
1268 continue;
1269 } else {
1270 return (B_FALSE);
1271 }
1272 }
1273
1274 return (B_TRUE);
1275 }
1276
1277 static uint_t
dnode_reclaim_slots(dnode_children_t * children,int idx,int slots)1278 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1279 {
1280 uint_t reclaimed = 0;
1281
1282 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1283
1284 for (int i = idx; i < idx + slots; i++) {
1285 dnode_handle_t *dnh = &children->dnc_children[i];
1286
1287 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1288
1289 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1290 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1291 dnode_destroy(dnh->dnh_dnode);
1292 dnh->dnh_dnode = DN_SLOT_FREE;
1293 reclaimed++;
1294 }
1295 }
1296
1297 return (reclaimed);
1298 }
1299
1300 void
dnode_free_interior_slots(dnode_t * dn)1301 dnode_free_interior_slots(dnode_t *dn)
1302 {
1303 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1304 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1305 int idx = (dn->dn_object & (epb - 1)) + 1;
1306 int slots = dn->dn_num_slots - 1;
1307
1308 if (slots == 0)
1309 return;
1310
1311 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1312
1313 while (!dnode_slots_tryenter(children, idx, slots)) {
1314 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1315 kpreempt(KPREEMPT_SYNC);
1316 }
1317
1318 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1319 dnode_slots_rele(children, idx, slots);
1320 }
1321
1322 void
dnode_special_close(dnode_handle_t * dnh)1323 dnode_special_close(dnode_handle_t *dnh)
1324 {
1325 dnode_t *dn = dnh->dnh_dnode;
1326
1327 /*
1328 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1329 * zfs_refcount_remove()
1330 */
1331 mutex_enter(&dn->dn_mtx);
1332 if (zfs_refcount_count(&dn->dn_holds) > 0)
1333 cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1334 mutex_exit(&dn->dn_mtx);
1335 ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1336
1337 ASSERT(dn->dn_dbuf == NULL ||
1338 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1339 zrl_add(&dnh->dnh_zrlock);
1340 dnode_destroy(dn); /* implicit zrl_remove() */
1341 zrl_destroy(&dnh->dnh_zrlock);
1342 dnh->dnh_dnode = NULL;
1343 }
1344
1345 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1346 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1347 dnode_handle_t *dnh)
1348 {
1349 dnode_t *dn;
1350
1351 zrl_init(&dnh->dnh_zrlock);
1352 VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1353
1354 dn = dnode_create(os, dnp, NULL, object, dnh);
1355 DNODE_VERIFY(dn);
1356
1357 zrl_exit(&dnh->dnh_zrlock);
1358 }
1359
1360 static void
dnode_buf_evict_async(void * dbu)1361 dnode_buf_evict_async(void *dbu)
1362 {
1363 dnode_children_t *dnc = dbu;
1364
1365 DNODE_STAT_BUMP(dnode_buf_evict);
1366
1367 for (int i = 0; i < dnc->dnc_count; i++) {
1368 dnode_handle_t *dnh = &dnc->dnc_children[i];
1369 dnode_t *dn;
1370
1371 /*
1372 * The dnode handle lock guards against the dnode moving to
1373 * another valid address, so there is no need here to guard
1374 * against changes to or from NULL.
1375 */
1376 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1377 zrl_destroy(&dnh->dnh_zrlock);
1378 dnh->dnh_dnode = DN_SLOT_UNINIT;
1379 continue;
1380 }
1381
1382 zrl_add(&dnh->dnh_zrlock);
1383 dn = dnh->dnh_dnode;
1384 /*
1385 * If there are holds on this dnode, then there should
1386 * be holds on the dnode's containing dbuf as well; thus
1387 * it wouldn't be eligible for eviction and this function
1388 * would not have been called.
1389 */
1390 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1391 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1392
1393 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1394 zrl_destroy(&dnh->dnh_zrlock);
1395 dnh->dnh_dnode = DN_SLOT_UNINIT;
1396 }
1397 kmem_free(dnc, sizeof (dnode_children_t) +
1398 dnc->dnc_count * sizeof (dnode_handle_t));
1399 }
1400
1401 /*
1402 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1403 * to ensure the hole at the specified object offset is large enough to
1404 * hold the dnode being created. The slots parameter is also used to ensure
1405 * a dnode does not span multiple dnode blocks. In both of these cases, if
1406 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1407 * are only possible when using DNODE_MUST_BE_FREE.
1408 *
1409 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1410 * dnode_hold_impl() will check if the requested dnode is already consumed
1411 * as an extra dnode slot by an large dnode, in which case it returns
1412 * ENOENT.
1413 *
1414 * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1415 * return whether the hold would succeed or not. tag and dnp should set to
1416 * NULL in this case.
1417 *
1418 * errors:
1419 * EINVAL - Invalid object number or flags.
1420 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1421 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1422 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1423 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1424 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1425 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1426 * EIO - I/O error when reading the meta dnode dbuf.
1427 *
1428 * succeeds even for free dnodes.
1429 */
1430 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,int slots,const void * tag,dnode_t ** dnp)1431 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1432 const void *tag, dnode_t **dnp)
1433 {
1434 int epb, idx, err;
1435 int drop_struct_lock = FALSE;
1436 int type;
1437 uint64_t blk;
1438 dnode_t *mdn, *dn;
1439 dmu_buf_impl_t *db;
1440 dnode_children_t *dnc;
1441 dnode_phys_t *dn_block;
1442 dnode_handle_t *dnh;
1443
1444 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1445 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1446 IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1447
1448 /*
1449 * If you are holding the spa config lock as writer, you shouldn't
1450 * be asking the DMU to do *anything* unless it's the root pool
1451 * which may require us to read from the root filesystem while
1452 * holding some (not all) of the locks as writer.
1453 */
1454 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1455 (spa_is_root(os->os_spa) &&
1456 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1457
1458 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1459
1460 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1461 object == DMU_PROJECTUSED_OBJECT) {
1462 if (object == DMU_USERUSED_OBJECT)
1463 dn = DMU_USERUSED_DNODE(os);
1464 else if (object == DMU_GROUPUSED_OBJECT)
1465 dn = DMU_GROUPUSED_DNODE(os);
1466 else
1467 dn = DMU_PROJECTUSED_DNODE(os);
1468 if (dn == NULL)
1469 return (SET_ERROR(ENOENT));
1470 type = dn->dn_type;
1471 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1472 return (SET_ERROR(ENOENT));
1473 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1474 return (SET_ERROR(EEXIST));
1475 DNODE_VERIFY(dn);
1476 /* Don't actually hold if dry run, just return 0 */
1477 if (!(flag & DNODE_DRY_RUN)) {
1478 (void) zfs_refcount_add(&dn->dn_holds, tag);
1479 *dnp = dn;
1480 }
1481 return (0);
1482 }
1483
1484 if (object == 0 || object >= DN_MAX_OBJECT)
1485 return (SET_ERROR(EINVAL));
1486
1487 mdn = DMU_META_DNODE(os);
1488 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1489
1490 DNODE_VERIFY(mdn);
1491
1492 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1493 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1494 drop_struct_lock = TRUE;
1495 }
1496
1497 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1498 db = dbuf_hold(mdn, blk, FTAG);
1499 if (drop_struct_lock)
1500 rw_exit(&mdn->dn_struct_rwlock);
1501 if (db == NULL) {
1502 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1503 return (SET_ERROR(EIO));
1504 }
1505
1506 /*
1507 * We do not need to decrypt to read the dnode so it doesn't matter
1508 * if we get the encrypted or decrypted version.
1509 */
1510 err = dbuf_read(db, NULL, DB_RF_CANFAIL |
1511 DMU_READ_NO_PREFETCH | DMU_READ_NO_DECRYPT);
1512 if (err) {
1513 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1514 dbuf_rele(db, FTAG);
1515 return (err);
1516 }
1517
1518 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1519 epb = db->db.db_size >> DNODE_SHIFT;
1520
1521 idx = object & (epb - 1);
1522 dn_block = (dnode_phys_t *)db->db.db_data;
1523
1524 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1525 dnc = dmu_buf_get_user(&db->db);
1526 dnh = NULL;
1527 if (dnc == NULL) {
1528 dnode_children_t *winner;
1529 int skip = 0;
1530
1531 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1532 epb * sizeof (dnode_handle_t), KM_SLEEP);
1533 dnc->dnc_count = epb;
1534 dnh = &dnc->dnc_children[0];
1535
1536 /* Initialize dnode slot status from dnode_phys_t */
1537 for (int i = 0; i < epb; i++) {
1538 zrl_init(&dnh[i].dnh_zrlock);
1539
1540 if (skip) {
1541 skip--;
1542 continue;
1543 }
1544
1545 if (dn_block[i].dn_type != DMU_OT_NONE) {
1546 int interior = dn_block[i].dn_extra_slots;
1547
1548 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1549 dnode_set_slots(dnc, i + 1, interior,
1550 DN_SLOT_INTERIOR);
1551 skip = interior;
1552 } else {
1553 dnh[i].dnh_dnode = DN_SLOT_FREE;
1554 skip = 0;
1555 }
1556 }
1557
1558 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1559 dnode_buf_evict_async, NULL);
1560 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1561 if (winner != NULL) {
1562
1563 for (int i = 0; i < epb; i++)
1564 zrl_destroy(&dnh[i].dnh_zrlock);
1565
1566 kmem_free(dnc, sizeof (dnode_children_t) +
1567 epb * sizeof (dnode_handle_t));
1568 dnc = winner;
1569 }
1570 }
1571
1572 ASSERT(dnc->dnc_count == epb);
1573
1574 if (flag & DNODE_MUST_BE_ALLOCATED) {
1575 slots = 1;
1576
1577 dnode_slots_hold(dnc, idx, slots);
1578 dnh = &dnc->dnc_children[idx];
1579
1580 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1581 dn = dnh->dnh_dnode;
1582 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1583 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1584 dnode_slots_rele(dnc, idx, slots);
1585 dbuf_rele(db, FTAG);
1586 return (SET_ERROR(EEXIST));
1587 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1588 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1589 dnode_slots_rele(dnc, idx, slots);
1590 dbuf_rele(db, FTAG);
1591 return (SET_ERROR(ENOENT));
1592 } else {
1593 dnode_slots_rele(dnc, idx, slots);
1594 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1595 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1596 kpreempt(KPREEMPT_SYNC);
1597 }
1598
1599 /*
1600 * Someone else won the race and called dnode_create()
1601 * after we checked DN_SLOT_IS_PTR() above but before
1602 * we acquired the lock.
1603 */
1604 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1605 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1606 dn = dnh->dnh_dnode;
1607 } else {
1608 dn = dnode_create(os, dn_block + idx, db,
1609 object, dnh);
1610 dmu_buf_add_user_size(&db->db,
1611 sizeof (dnode_t));
1612 }
1613 }
1614
1615 mutex_enter(&dn->dn_mtx);
1616 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1617 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1618 mutex_exit(&dn->dn_mtx);
1619 dnode_slots_rele(dnc, idx, slots);
1620 dbuf_rele(db, FTAG);
1621 return (SET_ERROR(ENOENT));
1622 }
1623
1624 /* Don't actually hold if dry run, just return 0 */
1625 if (flag & DNODE_DRY_RUN) {
1626 mutex_exit(&dn->dn_mtx);
1627 dnode_slots_rele(dnc, idx, slots);
1628 dbuf_rele(db, FTAG);
1629 return (0);
1630 }
1631
1632 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1633 } else if (flag & DNODE_MUST_BE_FREE) {
1634
1635 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1636 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1637 dbuf_rele(db, FTAG);
1638 return (SET_ERROR(ENOSPC));
1639 }
1640
1641 dnode_slots_hold(dnc, idx, slots);
1642
1643 if (!dnode_check_slots_free(dnc, idx, slots)) {
1644 DNODE_STAT_BUMP(dnode_hold_free_misses);
1645 dnode_slots_rele(dnc, idx, slots);
1646 dbuf_rele(db, FTAG);
1647 return (SET_ERROR(ENOSPC));
1648 }
1649
1650 dnode_slots_rele(dnc, idx, slots);
1651 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1652 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1653 kpreempt(KPREEMPT_SYNC);
1654 }
1655
1656 if (!dnode_check_slots_free(dnc, idx, slots)) {
1657 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1658 dnode_slots_rele(dnc, idx, slots);
1659 dbuf_rele(db, FTAG);
1660 return (SET_ERROR(ENOSPC));
1661 }
1662
1663 /*
1664 * Allocated but otherwise free dnodes which would
1665 * be in the interior of a multi-slot dnodes need
1666 * to be freed. Single slot dnodes can be safely
1667 * re-purposed as a performance optimization.
1668 */
1669 if (slots > 1) {
1670 uint_t reclaimed =
1671 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1672 if (reclaimed > 0)
1673 dmu_buf_sub_user_size(&db->db,
1674 reclaimed * sizeof (dnode_t));
1675 }
1676
1677 dnh = &dnc->dnc_children[idx];
1678 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1679 dn = dnh->dnh_dnode;
1680 } else {
1681 dn = dnode_create(os, dn_block + idx, db,
1682 object, dnh);
1683 dmu_buf_add_user_size(&db->db, sizeof (dnode_t));
1684 }
1685
1686 mutex_enter(&dn->dn_mtx);
1687 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1688 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1689 mutex_exit(&dn->dn_mtx);
1690 dnode_slots_rele(dnc, idx, slots);
1691 dbuf_rele(db, FTAG);
1692 return (SET_ERROR(EEXIST));
1693 }
1694
1695 /* Don't actually hold if dry run, just return 0 */
1696 if (flag & DNODE_DRY_RUN) {
1697 mutex_exit(&dn->dn_mtx);
1698 dnode_slots_rele(dnc, idx, slots);
1699 dbuf_rele(db, FTAG);
1700 return (0);
1701 }
1702
1703 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1704 DNODE_STAT_BUMP(dnode_hold_free_hits);
1705 } else {
1706 dbuf_rele(db, FTAG);
1707 return (SET_ERROR(EINVAL));
1708 }
1709
1710 ASSERT0(dn->dn_free_txg);
1711
1712 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1713 dbuf_add_ref(db, dnh);
1714
1715 mutex_exit(&dn->dn_mtx);
1716
1717 /* Now we can rely on the hold to prevent the dnode from moving. */
1718 dnode_slots_rele(dnc, idx, slots);
1719
1720 DNODE_VERIFY(dn);
1721 ASSERT3P(dnp, !=, NULL);
1722 ASSERT3P(dn->dn_dbuf, ==, db);
1723 ASSERT3U(dn->dn_object, ==, object);
1724 dbuf_rele(db, FTAG);
1725
1726 *dnp = dn;
1727 return (0);
1728 }
1729
1730 /*
1731 * Return held dnode if the object is allocated, NULL if not.
1732 */
1733 int
dnode_hold(objset_t * os,uint64_t object,const void * tag,dnode_t ** dnp)1734 dnode_hold(objset_t *os, uint64_t object, const void *tag, dnode_t **dnp)
1735 {
1736 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1737 dnp));
1738 }
1739
1740 /*
1741 * Can only add a reference if there is already at least one
1742 * reference on the dnode. Returns FALSE if unable to add a
1743 * new reference.
1744 */
1745 static boolean_t
dnode_add_ref_locked(dnode_t * dn,const void * tag)1746 dnode_add_ref_locked(dnode_t *dn, const void *tag)
1747 {
1748 ASSERT(MUTEX_HELD(&dn->dn_mtx));
1749 if (zfs_refcount_is_zero(&dn->dn_holds))
1750 return (FALSE);
1751 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1752 return (TRUE);
1753 }
1754
1755 boolean_t
dnode_add_ref(dnode_t * dn,const void * tag)1756 dnode_add_ref(dnode_t *dn, const void *tag)
1757 {
1758 mutex_enter(&dn->dn_mtx);
1759 boolean_t r = dnode_add_ref_locked(dn, tag);
1760 mutex_exit(&dn->dn_mtx);
1761 return (r);
1762 }
1763
1764 void
dnode_rele(dnode_t * dn,const void * tag)1765 dnode_rele(dnode_t *dn, const void *tag)
1766 {
1767 mutex_enter(&dn->dn_mtx);
1768 dnode_rele_and_unlock(dn, tag, B_FALSE);
1769 }
1770
1771 void
dnode_rele_and_unlock(dnode_t * dn,const void * tag,boolean_t evicting)1772 dnode_rele_and_unlock(dnode_t *dn, const void *tag, boolean_t evicting)
1773 {
1774 uint64_t refs;
1775 /* Get while the hold prevents the dnode from moving. */
1776 dmu_buf_impl_t *db = dn->dn_dbuf;
1777 dnode_handle_t *dnh = dn->dn_handle;
1778
1779 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1780 if (refs == 0)
1781 cv_broadcast(&dn->dn_nodnholds);
1782
1783 /*
1784 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1785 * indirectly by dbuf_rele() while relying on the dnode handle to
1786 * prevent the dnode from moving, since releasing the last hold could
1787 * result in the dnode's parent dbuf evicting its dnode handles. For
1788 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1789 * other direct or indirect hold on the dnode must first drop the dnode
1790 * handle.
1791 */
1792 #ifdef ZFS_DEBUG
1793 ASSERT(refs > 0 || zrl_owner(&dnh->dnh_zrlock) != curthread);
1794 #endif
1795
1796 mutex_exit(&dn->dn_mtx);
1797 /*
1798 * After the dn_mtx is released the dn and dnh may be destroyed,
1799 * they are no longer safe to use after this point.
1800 */
1801
1802 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1803 if (refs == 0 && db != NULL) {
1804 /*
1805 * Another thread could add a hold to the dnode handle in
1806 * dnode_hold_impl() while holding the parent dbuf. Since the
1807 * hold on the parent dbuf prevents the handle from being
1808 * destroyed, the hold on the handle is OK. We can't yet assert
1809 * that the handle has zero references, but that will be
1810 * asserted anyway when the handle gets destroyed.
1811 */
1812 mutex_enter(&db->db_mtx);
1813 dbuf_rele_and_unlock(db, dnh, evicting);
1814 }
1815 }
1816
1817 /*
1818 * Test whether we can create a dnode at the specified location.
1819 */
1820 int
dnode_try_claim(objset_t * os,uint64_t object,int slots)1821 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1822 {
1823 return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1824 slots, NULL, NULL));
1825 }
1826
1827 /*
1828 * Test if the dnode is dirty, or carrying uncommitted records.
1829 *
1830 * dn_dirtycnt is the number of txgs this dnode is dirty on. It's incremented
1831 * in dnode_setdirty() the first time the dnode is dirtied on a txg, and
1832 * decremented in either dnode_rele_task() or userquota_updates_task() when the
1833 * txg is synced out.
1834 */
1835 boolean_t
dnode_is_dirty(dnode_t * dn)1836 dnode_is_dirty(dnode_t *dn)
1837 {
1838 mutex_enter(&dn->dn_mtx);
1839 boolean_t dirty = (dn->dn_dirtycnt != 0);
1840 mutex_exit(&dn->dn_mtx);
1841 return (dirty);
1842 }
1843
1844 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1845 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1846 {
1847 objset_t *os = dn->dn_objset;
1848 uint64_t txg = tx->tx_txg;
1849
1850 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1851 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1852 return;
1853 }
1854
1855 DNODE_VERIFY(dn);
1856
1857 #ifdef ZFS_DEBUG
1858 mutex_enter(&dn->dn_mtx);
1859 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1860 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1861 mutex_exit(&dn->dn_mtx);
1862 #endif
1863
1864 /*
1865 * Determine old uid/gid when necessary
1866 */
1867 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1868
1869 multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK];
1870 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1871
1872 /*
1873 * If we are already marked dirty, we're done.
1874 */
1875 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1876 multilist_sublist_unlock(mls);
1877 return;
1878 }
1879
1880 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1881 !avl_is_empty(&dn->dn_dbufs));
1882 ASSERT(dn->dn_datablksz != 0);
1883 ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1884 ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1885 ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1886
1887 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1888 (u_longlong_t)dn->dn_object, (u_longlong_t)txg);
1889
1890 multilist_sublist_insert_head(mls, dn);
1891
1892 multilist_sublist_unlock(mls);
1893
1894 /*
1895 * The dnode maintains a hold on its containing dbuf as
1896 * long as there are holds on it. Each instantiated child
1897 * dbuf maintains a hold on the dnode. When the last child
1898 * drops its hold, the dnode will drop its hold on the
1899 * containing dbuf. We add a "dirty hold" here so that the
1900 * dnode will hang around after we finish processing its
1901 * children.
1902 */
1903 mutex_enter(&dn->dn_mtx);
1904 VERIFY(dnode_add_ref_locked(dn, (void *)(uintptr_t)tx->tx_txg));
1905 dn->dn_dirtycnt++;
1906 ASSERT3U(dn->dn_dirtycnt, <=, 3);
1907 mutex_exit(&dn->dn_mtx);
1908
1909 (void) dbuf_dirty(dn->dn_dbuf, tx);
1910
1911 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1912 }
1913
1914 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1915 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1916 {
1917 mutex_enter(&dn->dn_mtx);
1918 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1919 mutex_exit(&dn->dn_mtx);
1920 return;
1921 }
1922 dn->dn_free_txg = tx->tx_txg;
1923 mutex_exit(&dn->dn_mtx);
1924
1925 dnode_setdirty(dn, tx);
1926 }
1927
1928 /*
1929 * Try to change the block size for the indicated dnode. This can only
1930 * succeed if there are no blocks allocated or dirty beyond first block
1931 */
1932 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1933 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1934 {
1935 dmu_buf_impl_t *db;
1936 int err;
1937
1938 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1939 if (size == 0)
1940 size = SPA_MINBLOCKSIZE;
1941 else
1942 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1943
1944 if (ibs == dn->dn_indblkshift)
1945 ibs = 0;
1946
1947 if (size == dn->dn_datablksz && ibs == 0)
1948 return (0);
1949
1950 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1951
1952 /* Check for any allocated blocks beyond the first */
1953 if (dn->dn_maxblkid != 0)
1954 goto fail;
1955
1956 mutex_enter(&dn->dn_dbufs_mtx);
1957 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1958 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1959 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1960 db->db_blkid != DMU_SPILL_BLKID) {
1961 mutex_exit(&dn->dn_dbufs_mtx);
1962 goto fail;
1963 }
1964 }
1965 mutex_exit(&dn->dn_dbufs_mtx);
1966
1967 if (ibs && dn->dn_nlevels != 1)
1968 goto fail;
1969
1970 dnode_setdirty(dn, tx);
1971 if (size != dn->dn_datablksz) {
1972 /* resize the old block */
1973 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1974 if (err == 0) {
1975 dbuf_new_size(db, size, tx);
1976 } else if (err != ENOENT) {
1977 goto fail;
1978 }
1979
1980 dnode_setdblksz(dn, size);
1981 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = size;
1982 if (db)
1983 dbuf_rele(db, FTAG);
1984 }
1985 if (ibs) {
1986 dn->dn_indblkshift = ibs;
1987 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
1988 }
1989
1990 rw_exit(&dn->dn_struct_rwlock);
1991 return (0);
1992
1993 fail:
1994 rw_exit(&dn->dn_struct_rwlock);
1995 return (SET_ERROR(ENOTSUP));
1996 }
1997
1998 static void
dnode_set_nlevels_impl(dnode_t * dn,int new_nlevels,dmu_tx_t * tx)1999 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
2000 {
2001 uint64_t txgoff = tx->tx_txg & TXG_MASK;
2002 int old_nlevels = dn->dn_nlevels;
2003 dmu_buf_impl_t *db;
2004 list_t *list;
2005 dbuf_dirty_record_t *new, *dr, *dr_next;
2006
2007 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2008
2009 ASSERT3U(new_nlevels, >, dn->dn_nlevels);
2010 dn->dn_nlevels = new_nlevels;
2011
2012 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
2013 dn->dn_next_nlevels[txgoff] = new_nlevels;
2014
2015 /* dirty the left indirects */
2016 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
2017 ASSERT(db != NULL);
2018 new = dbuf_dirty(db, tx);
2019 dbuf_rele(db, FTAG);
2020
2021 /* transfer the dirty records to the new indirect */
2022 mutex_enter(&dn->dn_mtx);
2023 mutex_enter(&new->dt.di.dr_mtx);
2024 list = &dn->dn_dirty_records[txgoff];
2025 for (dr = list_head(list); dr; dr = dr_next) {
2026 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
2027
2028 IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1);
2029 if (dr->dr_dbuf == NULL ||
2030 (dr->dr_dbuf->db_level == old_nlevels - 1 &&
2031 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2032 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) {
2033 list_remove(&dn->dn_dirty_records[txgoff], dr);
2034 list_insert_tail(&new->dt.di.dr_children, dr);
2035 dr->dr_parent = new;
2036 }
2037 }
2038 mutex_exit(&new->dt.di.dr_mtx);
2039 mutex_exit(&dn->dn_mtx);
2040 }
2041
2042 int
dnode_set_nlevels(dnode_t * dn,int nlevels,dmu_tx_t * tx)2043 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
2044 {
2045 int ret = 0;
2046
2047 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2048
2049 if (dn->dn_nlevels == nlevels) {
2050 ret = 0;
2051 goto out;
2052 } else if (nlevels < dn->dn_nlevels) {
2053 ret = SET_ERROR(EINVAL);
2054 goto out;
2055 }
2056
2057 dnode_set_nlevels_impl(dn, nlevels, tx);
2058
2059 out:
2060 rw_exit(&dn->dn_struct_rwlock);
2061 return (ret);
2062 }
2063
2064 /* read-holding callers must not rely on the lock being continuously held */
2065 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read,boolean_t force)2066 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
2067 boolean_t force)
2068 {
2069 int epbs, new_nlevels;
2070 uint64_t sz;
2071
2072 ASSERT(blkid != DMU_BONUS_BLKID);
2073
2074 ASSERT(have_read ?
2075 RW_READ_HELD(&dn->dn_struct_rwlock) :
2076 RW_WRITE_HELD(&dn->dn_struct_rwlock));
2077
2078 /*
2079 * if we have a read-lock, check to see if we need to do any work
2080 * before upgrading to a write-lock.
2081 */
2082 if (have_read) {
2083 if (blkid <= dn->dn_maxblkid)
2084 return;
2085
2086 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
2087 rw_exit(&dn->dn_struct_rwlock);
2088 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2089 }
2090 }
2091
2092 /*
2093 * Raw sends (indicated by the force flag) require that we take the
2094 * given blkid even if the value is lower than the current value.
2095 */
2096 if (!force && blkid <= dn->dn_maxblkid)
2097 goto out;
2098
2099 /*
2100 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
2101 * to indicate that this field is set. This allows us to set the
2102 * maxblkid to 0 on an existing object in dnode_sync().
2103 */
2104 dn->dn_maxblkid = blkid;
2105 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
2106 blkid | DMU_NEXT_MAXBLKID_SET;
2107
2108 /*
2109 * Compute the number of levels necessary to support the new maxblkid.
2110 * Raw sends will ensure nlevels is set correctly for us.
2111 */
2112 new_nlevels = 1;
2113 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2114 for (sz = dn->dn_nblkptr;
2115 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
2116 new_nlevels++;
2117
2118 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
2119
2120 if (!force) {
2121 if (new_nlevels > dn->dn_nlevels)
2122 dnode_set_nlevels_impl(dn, new_nlevels, tx);
2123 } else {
2124 ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
2125 }
2126
2127 out:
2128 if (have_read)
2129 rw_downgrade(&dn->dn_struct_rwlock);
2130 }
2131
2132 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)2133 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
2134 {
2135 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
2136 if (db != NULL) {
2137 dmu_buf_will_dirty(&db->db, tx);
2138 dbuf_rele(db, FTAG);
2139 }
2140 }
2141
2142 /*
2143 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
2144 * and end_blkid.
2145 */
2146 static void
dnode_dirty_l1range(dnode_t * dn,uint64_t start_blkid,uint64_t end_blkid,dmu_tx_t * tx)2147 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
2148 dmu_tx_t *tx)
2149 {
2150 dmu_buf_impl_t *db_search;
2151 dmu_buf_impl_t *db;
2152 avl_index_t where;
2153
2154 db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
2155
2156 mutex_enter(&dn->dn_dbufs_mtx);
2157
2158 db_search->db_level = 1;
2159 db_search->db_blkid = start_blkid + 1;
2160 db_search->db_state = DB_SEARCH;
2161 for (;;) {
2162
2163 db = avl_find(&dn->dn_dbufs, db_search, &where);
2164 if (db == NULL)
2165 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2166
2167 if (db == NULL || db->db_level != 1 ||
2168 db->db_blkid >= end_blkid) {
2169 break;
2170 }
2171
2172 /*
2173 * Setup the next blkid we want to search for.
2174 */
2175 db_search->db_blkid = db->db_blkid + 1;
2176 ASSERT3U(db->db_blkid, >=, start_blkid);
2177
2178 /*
2179 * If the dbuf transitions to DB_EVICTING while we're trying
2180 * to dirty it, then we will be unable to discover it in
2181 * the dbuf hash table. This will result in a call to
2182 * dbuf_create() which needs to acquire the dn_dbufs_mtx
2183 * lock. To avoid a deadlock, we drop the lock before
2184 * dirtying the level-1 dbuf.
2185 */
2186 mutex_exit(&dn->dn_dbufs_mtx);
2187 dnode_dirty_l1(dn, db->db_blkid, tx);
2188 mutex_enter(&dn->dn_dbufs_mtx);
2189 }
2190
2191 #ifdef ZFS_DEBUG
2192 /*
2193 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
2194 */
2195 db_search->db_level = 1;
2196 db_search->db_blkid = start_blkid + 1;
2197 db_search->db_state = DB_SEARCH;
2198 db = avl_find(&dn->dn_dbufs, db_search, &where);
2199 if (db == NULL)
2200 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2201 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2202 if (db->db_level != 1 || db->db_blkid >= end_blkid)
2203 break;
2204 if (db->db_state != DB_EVICTING)
2205 ASSERT(db->db_dirtycnt > 0);
2206 }
2207 #endif
2208 kmem_free(db_search, sizeof (dmu_buf_impl_t));
2209 mutex_exit(&dn->dn_dbufs_mtx);
2210 }
2211
2212 static void
dnode_partial_zero(dnode_t * dn,uint64_t off,uint64_t blkoff,uint64_t len,dmu_tx_t * tx)2213 dnode_partial_zero(dnode_t *dn, uint64_t off, uint64_t blkoff, uint64_t len,
2214 dmu_tx_t *tx)
2215 {
2216 dmu_buf_impl_t *db;
2217 int res;
2218
2219 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2220 res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), TRUE, FALSE,
2221 FTAG, &db);
2222 rw_exit(&dn->dn_struct_rwlock);
2223 if (res == 0) {
2224 db_lock_type_t dblt;
2225 boolean_t dirty;
2226
2227 dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2228 /* don't dirty if not on disk and not dirty */
2229 dirty = !list_is_empty(&db->db_dirty_records) ||
2230 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2231 dmu_buf_unlock_parent(db, dblt, FTAG);
2232 if (dirty) {
2233 caddr_t data;
2234
2235 dmu_buf_will_dirty(&db->db, tx);
2236 data = db->db.db_data;
2237 memset(data + blkoff, 0, len);
2238 }
2239 dbuf_rele(db, FTAG);
2240 }
2241 }
2242
2243 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)2244 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2245 {
2246 uint64_t blkoff, blkid, nblks;
2247 int blksz, blkshift, head, tail;
2248 int trunc = FALSE;
2249 int epbs;
2250
2251 blksz = dn->dn_datablksz;
2252 blkshift = dn->dn_datablkshift;
2253 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2254
2255 if (len == DMU_OBJECT_END) {
2256 len = UINT64_MAX - off;
2257 trunc = TRUE;
2258 }
2259
2260 /*
2261 * First, block align the region to free:
2262 */
2263 if (ISP2(blksz)) {
2264 head = P2NPHASE(off, blksz);
2265 blkoff = P2PHASE(off, blksz);
2266 if ((off >> blkshift) > dn->dn_maxblkid)
2267 return;
2268 } else {
2269 ASSERT0(dn->dn_maxblkid);
2270 if (off == 0 && len >= blksz) {
2271 /*
2272 * Freeing the whole block; fast-track this request.
2273 */
2274 blkid = 0;
2275 nblks = 1;
2276 if (dn->dn_nlevels > 1) {
2277 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2278 dnode_dirty_l1(dn, 0, tx);
2279 rw_exit(&dn->dn_struct_rwlock);
2280 }
2281 goto done;
2282 } else if (off >= blksz) {
2283 /* Freeing past end-of-data */
2284 return;
2285 } else {
2286 /* Freeing part of the block. */
2287 head = blksz - off;
2288 ASSERT3U(head, >, 0);
2289 }
2290 blkoff = off;
2291 }
2292 /* zero out any partial block data at the start of the range */
2293 if (head) {
2294 ASSERT3U(blkoff + head, ==, blksz);
2295 if (len < head)
2296 head = len;
2297 dnode_partial_zero(dn, off, blkoff, head, tx);
2298 off += head;
2299 len -= head;
2300 }
2301
2302 /* If the range was less than one block, we're done */
2303 if (len == 0)
2304 return;
2305
2306 /* If the remaining range is past end of file, we're done */
2307 if ((off >> blkshift) > dn->dn_maxblkid)
2308 return;
2309
2310 ASSERT(ISP2(blksz));
2311 if (trunc)
2312 tail = 0;
2313 else
2314 tail = P2PHASE(len, blksz);
2315
2316 ASSERT0(P2PHASE(off, blksz));
2317 /* zero out any partial block data at the end of the range */
2318 if (tail) {
2319 if (len < tail)
2320 tail = len;
2321 dnode_partial_zero(dn, off + len, 0, tail, tx);
2322 len -= tail;
2323 }
2324
2325 /* If the range did not include a full block, we are done */
2326 if (len == 0)
2327 return;
2328
2329 ASSERT(IS_P2ALIGNED(off, blksz));
2330 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2331 blkid = off >> blkshift;
2332 nblks = len >> blkshift;
2333 if (trunc)
2334 nblks += 1;
2335
2336 /*
2337 * Dirty all the indirect blocks in this range. Note that only
2338 * the first and last indirect blocks can actually be written
2339 * (if they were partially freed) -- they must be dirtied, even if
2340 * they do not exist on disk yet. The interior blocks will
2341 * be freed by free_children(), so they will not actually be written.
2342 * Even though these interior blocks will not be written, we
2343 * dirty them for two reasons:
2344 *
2345 * - It ensures that the indirect blocks remain in memory until
2346 * syncing context. (They have already been prefetched by
2347 * dmu_tx_hold_free(), so we don't have to worry about reading
2348 * them serially here.)
2349 *
2350 * - The dirty space accounting will put pressure on the txg sync
2351 * mechanism to begin syncing, and to delay transactions if there
2352 * is a large amount of freeing. Even though these indirect
2353 * blocks will not be written, we could need to write the same
2354 * amount of space if we copy the freed BPs into deadlists.
2355 */
2356 if (dn->dn_nlevels > 1) {
2357 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2358 uint64_t first, last;
2359
2360 first = blkid >> epbs;
2361 dnode_dirty_l1(dn, first, tx);
2362 if (trunc)
2363 last = dn->dn_maxblkid >> epbs;
2364 else
2365 last = (blkid + nblks - 1) >> epbs;
2366 if (last != first)
2367 dnode_dirty_l1(dn, last, tx);
2368
2369 dnode_dirty_l1range(dn, first, last, tx);
2370
2371 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2372 SPA_BLKPTRSHIFT;
2373 for (uint64_t i = first + 1; i < last; i++) {
2374 /*
2375 * Set i to the blockid of the next non-hole
2376 * level-1 indirect block at or after i. Note
2377 * that dnode_next_offset() operates in terms of
2378 * level-0-equivalent bytes.
2379 */
2380 uint64_t ibyte = i << shift;
2381 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2382 &ibyte, 2, 1, 0);
2383 i = ibyte >> shift;
2384 if (i >= last)
2385 break;
2386
2387 /*
2388 * Normally we should not see an error, either
2389 * from dnode_next_offset() or dbuf_hold_level()
2390 * (except for ESRCH from dnode_next_offset).
2391 * If there is an i/o error, then when we read
2392 * this block in syncing context, it will use
2393 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2394 * to the "failmode" property. dnode_next_offset()
2395 * doesn't have a flag to indicate MUSTSUCCEED.
2396 */
2397 if (err != 0)
2398 break;
2399
2400 dnode_dirty_l1(dn, i, tx);
2401 }
2402 rw_exit(&dn->dn_struct_rwlock);
2403 }
2404
2405 done:
2406 /*
2407 * Add this range to the dnode range list.
2408 * We will finish up this free operation in the syncing phase.
2409 */
2410 mutex_enter(&dn->dn_mtx);
2411 {
2412 int txgoff = tx->tx_txg & TXG_MASK;
2413
2414 FREE_RANGE_VERIFY(tx, dn);
2415 if (dn->dn_free_ranges[txgoff] == NULL) {
2416 dn->dn_free_ranges[txgoff] =
2417 zfs_range_tree_create_flags(
2418 NULL, ZFS_RANGE_SEG64, NULL, 0, 0,
2419 ZFS_RT_F_DYN_NAME, rt_name(dn, "dn_free_ranges"));
2420 }
2421 zfs_range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2422 zfs_range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2423 }
2424 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2425 (u_longlong_t)blkid, (u_longlong_t)nblks,
2426 (u_longlong_t)tx->tx_txg);
2427 mutex_exit(&dn->dn_mtx);
2428
2429 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2430 dnode_setdirty(dn, tx);
2431 }
2432
2433 static boolean_t
dnode_spill_freed(dnode_t * dn)2434 dnode_spill_freed(dnode_t *dn)
2435 {
2436 int i;
2437
2438 mutex_enter(&dn->dn_mtx);
2439 for (i = 0; i < TXG_SIZE; i++) {
2440 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2441 break;
2442 }
2443 mutex_exit(&dn->dn_mtx);
2444 return (i < TXG_SIZE);
2445 }
2446
2447 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2448 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)2449 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2450 {
2451 int i;
2452
2453 if (blkid == DMU_BONUS_BLKID)
2454 return (FALSE);
2455
2456 if (dn->dn_free_txg)
2457 return (TRUE);
2458
2459 if (blkid == DMU_SPILL_BLKID)
2460 return (dnode_spill_freed(dn));
2461
2462 mutex_enter(&dn->dn_mtx);
2463 for (i = 0; i < TXG_SIZE; i++) {
2464 if (dn->dn_free_ranges[i] != NULL &&
2465 zfs_range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2466 break;
2467 }
2468 mutex_exit(&dn->dn_mtx);
2469 return (i < TXG_SIZE);
2470 }
2471
2472 /*
2473 * Check if a level-0 block was freed in a TXG after override_txg.
2474 *
2475 * When a block has been overridden (e.g., by block clone or direct I/O),
2476 * we can't use dnode_block_freed() because it checks all active TXGs.
2477 * A free from a TXG *before* the override should not make the block appear
2478 * freed.
2479 */
2480 uint64_t
dnode_block_freed_after(dnode_t * dn,uint64_t blkid,uint64_t override_txg)2481 dnode_block_freed_after(dnode_t *dn, uint64_t blkid, uint64_t override_txg)
2482 {
2483 ASSERT(blkid != DMU_BONUS_BLKID);
2484 ASSERT(blkid != DMU_SPILL_BLKID);
2485
2486 if (dn->dn_free_txg)
2487 return (TRUE);
2488
2489 mutex_enter(&dn->dn_mtx);
2490 uint64_t open = spa_open_txg(dmu_objset_spa(dn->dn_objset));
2491 for (uint64_t txg = override_txg + 1; txg <= open; txg++) {
2492 int i = txg & TXG_MASK;
2493 if (dn->dn_free_ranges[i] != NULL &&
2494 zfs_range_tree_contains(dn->dn_free_ranges[i], blkid, 1)) {
2495 mutex_exit(&dn->dn_mtx);
2496 return (TRUE);
2497 }
2498 }
2499 mutex_exit(&dn->dn_mtx);
2500 return (FALSE);
2501 }
2502
2503 /* call from syncing context when we actually write/free space for this dnode */
2504 void
dnode_diduse_space(dnode_t * dn,int64_t delta)2505 dnode_diduse_space(dnode_t *dn, int64_t delta)
2506 {
2507 uint64_t space;
2508 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2509 dn, dn->dn_phys,
2510 (u_longlong_t)dn->dn_phys->dn_used,
2511 (longlong_t)delta);
2512
2513 mutex_enter(&dn->dn_mtx);
2514 space = DN_USED_BYTES(dn->dn_phys);
2515 if (delta > 0) {
2516 ASSERT3U(space + delta, >=, space); /* no overflow */
2517 } else {
2518 ASSERT3U(space, >=, -delta); /* no underflow */
2519 }
2520 space += delta;
2521 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2522 ASSERT0((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES));
2523 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2524 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2525 } else {
2526 dn->dn_phys->dn_used = space;
2527 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2528 }
2529 mutex_exit(&dn->dn_mtx);
2530 }
2531
2532 /*
2533 * Scans the block at the indicated "level" looking for a hole or data,
2534 * depending on 'flags' starting from array position given by *index.
2535 *
2536 * If lvl > 0, then we are scanning an indirect block looking at its
2537 * pointers. If lvl == 0, then we are looking at a block of dnodes.
2538 *
2539 * If we don't find what we are looking for in the block, we return ESRCH.
2540 * Otherwise, return with *index set to the matching array position.
2541 *
2542 * In both cases, *offset is updated to point at the matched BP/dnode or
2543 * the next offset to search (unless at the limit of possible offsets).
2544 *
2545 * The basic search algorithm used below by dnode_next_offset() uses this
2546 * function to perform a block-order tree traversal. We search up the block
2547 * tree (widen the search) until we find something (i.e., we don't return
2548 * ESRCH) and then search back down the tree (narrow the search) until we
2549 * reach our original search level or backtrack up because nothing matches.
2550 */
2551 static int
dnode_next_offset_level(dnode_t * dn,int flags,int lvl,uint64_t blkid,int * index,uint64_t blkfill,uint64_t txg,uint64_t * offset)2552 dnode_next_offset_level(dnode_t *dn, int flags, int lvl, uint64_t blkid,
2553 int *index, uint64_t blkfill, uint64_t txg, uint64_t *offset)
2554 {
2555 dmu_buf_impl_t *db = NULL;
2556 void *data = NULL;
2557 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2558 uint64_t epb = 1ULL << epbs;
2559 uint64_t minfill, maxfill;
2560 boolean_t hole;
2561 int i, inc, error, span;
2562
2563 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2564
2565 hole = ((flags & DNODE_FIND_HOLE) != 0);
2566 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2567 ASSERT(txg == 0 || !hole);
2568
2569 if (lvl == dn->dn_phys->dn_nlevels) {
2570 error = 0;
2571 epb = dn->dn_phys->dn_nblkptr;
2572 data = dn->dn_phys->dn_blkptr;
2573 if (dn->dn_dbuf != NULL)
2574 rw_enter(&dn->dn_dbuf->db_rwlock, RW_READER);
2575 else if (dmu_objset_ds(dn->dn_objset) != NULL)
2576 rrw_enter(&dmu_objset_ds(dn->dn_objset)->ds_bp_rwlock,
2577 RW_READER, FTAG);
2578 } else {
2579 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2580 if (error) {
2581 if (error != ENOENT)
2582 return (error);
2583 if (hole)
2584 return (0);
2585 return (SET_ERROR(ESRCH));
2586 }
2587 error = dbuf_read(db, NULL,
2588 DB_RF_CANFAIL | DB_RF_HAVESTRUCT |
2589 DMU_READ_NO_PREFETCH | DMU_READ_NO_DECRYPT);
2590 if (error) {
2591 dbuf_rele(db, FTAG);
2592 return (error);
2593 }
2594 data = db->db.db_data;
2595 rw_enter(&db->db_rwlock, RW_READER);
2596 }
2597
2598 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2599 BP_GET_LOGICAL_BIRTH(db->db_blkptr) <= txg ||
2600 BP_IS_HOLE(db->db_blkptr))) {
2601 /*
2602 * This can only happen when we are searching up the tree
2603 * and these conditions mean that we need to keep climbing.
2604 */
2605 error = SET_ERROR(ESRCH);
2606 } else if (lvl == 0) {
2607 dnode_phys_t *dnp = data;
2608
2609 ASSERT(dn->dn_type == DMU_OT_DNODE);
2610 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2611
2612 for (i = *index; i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2613 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2614 break;
2615 }
2616
2617 if (i == blkfill)
2618 error = SET_ERROR(ESRCH);
2619
2620 *index = i;
2621 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2622 (i << DNODE_SHIFT);
2623 } else {
2624 blkptr_t *bp = data;
2625 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2626 minfill = 0;
2627 maxfill = blkfill << ((lvl - 1) * epbs);
2628
2629 if (hole)
2630 maxfill--;
2631 else
2632 minfill++;
2633
2634 for (i = *index; i >= 0 && i < epb; i += inc) {
2635 if (BP_GET_FILL(&bp[i]) >= minfill &&
2636 BP_GET_FILL(&bp[i]) <= maxfill &&
2637 (hole || BP_GET_LOGICAL_BIRTH(&bp[i]) > txg))
2638 break;
2639 }
2640
2641 if (i < 0 || i >= epb)
2642 error = SET_ERROR(ESRCH);
2643
2644 *index = i;
2645 if (span < 8 * sizeof (*offset)) {
2646 uint64_t nblk = blkid << epbs;
2647 if (i >= 0 || blkid != 0)
2648 nblk += i;
2649 if ((nblk >> (8 * sizeof (*offset) - span)) == 0)
2650 *offset = (flags & DNODE_FIND_BACKWARDS) ?
2651 /* backwards: position offset at the end */
2652 MIN(*offset, ((nblk + 1) << span) - 1) :
2653 MAX(*offset, nblk << span);
2654 }
2655 }
2656
2657 if (db != NULL) {
2658 rw_exit(&db->db_rwlock);
2659 dbuf_rele(db, FTAG);
2660 } else {
2661 if (dn->dn_dbuf != NULL)
2662 rw_exit(&dn->dn_dbuf->db_rwlock);
2663 else if (dmu_objset_ds(dn->dn_objset) != NULL)
2664 rrw_exit(&dmu_objset_ds(dn->dn_objset)->ds_bp_rwlock,
2665 FTAG);
2666 }
2667
2668 return (error);
2669 }
2670
2671 /*
2672 * Find the next hole, data, or sparse region at or after *offset.
2673 * The value 'blkfill' tells us how many items we expect to find
2674 * in an L0 data block; this value is 1 for normal objects,
2675 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2676 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2677 *
2678 * If minlvl == 0, this searches for dnodes or unallocated dnodes.
2679 * If found, *offset points to the first offset of the matched dnode.
2680 * Backwards search is not allowed for dnodes.
2681 *
2682 * If minlvl > 0, this searches for blocks at the given level.
2683 * If found, *offset points to the first L0 offset of the block
2684 * (or for backwards search, the last offset, inclusive).
2685 *
2686 * If not found, in both cases, *offset is set to the first (or last)
2687 * offset of the unallocated indirect block where the search ended or
2688 * the initial offset if no such block was encountered.
2689 *
2690 * Examples:
2691 *
2692 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2693 * Finds the next/previous hole/data in a file.
2694 * Used in dmu_offset_next().
2695 *
2696 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2697 * Finds the next free/allocated dnode an objset's meta-dnode.
2698 * Only finds objects that have new contents since txg (ie.
2699 * bonus buffer changes and content removal are ignored).
2700 * Used in dmu_object_next().
2701 *
2702 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2703 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2704 * Used in dmu_object_alloc().
2705 */
2706 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)2707 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2708 int minlvl, uint64_t blkfill, uint64_t txg)
2709 {
2710 uint64_t blkid;
2711 int index, epbs;
2712 int lvl, maxlvl;
2713 int error = 0;
2714
2715 if (!(flags & DNODE_FIND_HAVELOCK))
2716 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2717
2718 if (dn->dn_phys->dn_nlevels == 0) {
2719 error = SET_ERROR(ESRCH);
2720 goto out;
2721 }
2722
2723 if (dn->dn_datablkshift == 0) {
2724 if (*offset < dn->dn_datablksz) {
2725 if (flags & DNODE_FIND_HOLE)
2726 *offset = dn->dn_datablksz;
2727 } else {
2728 error = SET_ERROR(ESRCH);
2729 }
2730 goto out;
2731 }
2732
2733 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2734 maxlvl = dn->dn_phys->dn_nlevels;
2735
2736 if (minlvl > 0) {
2737 uint64_t n = dbuf_whichblock(dn, minlvl - 1, *offset);
2738 blkid = n >> epbs;
2739 index = BF64_GET(n, 0, epbs);
2740 } else {
2741 blkid = dbuf_whichblock(dn, 0, *offset);
2742 index = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2743 ASSERT3U(BF64_GET(*offset, 0, DNODE_SHIFT), ==, 0);
2744 }
2745
2746 for (lvl = minlvl; lvl <= maxlvl; ) {
2747 error = dnode_next_offset_level(dn,
2748 flags, lvl, blkid, &index, blkfill, txg, offset);
2749
2750 if (error == 0 && lvl > minlvl) {
2751 /* Continue search at matched block in lvl-1. */
2752 blkid = (blkid << epbs) + index;
2753 index = 0;
2754 --lvl;
2755 } else if (error == ESRCH && lvl < maxlvl) {
2756 /*
2757 * Continue search at next/prev index in lvl+1 block.
2758 *
2759 * Usually we only search upwards at the start of the
2760 * search as higher level blocks point at a matching
2761 * minlvl block in most cases, but we backtrack if not.
2762 *
2763 * This can happen for txg > 0 searches if the block
2764 * contains only BPs/dnodes freed at that txg. It also
2765 * happens if we are still syncing out the tree, and
2766 * some BP's at higher levels are not updated yet.
2767 *
2768 * We must adjust index to avoid coming back to the
2769 * same offset and getting stuck looping forever. The
2770 * next loop goes up again if index is -1 or (1<<epbs).
2771 */
2772 index = BF64_GET(blkid, 0, epbs) +
2773 ((flags & DNODE_FIND_BACKWARDS) ? -1 : 1);
2774 blkid = blkid >> epbs;
2775 ++lvl;
2776 } else {
2777 break;
2778 }
2779 }
2780
2781 /*
2782 * There's always a "virtual hole" at the end of the object, even
2783 * if all BP's which physically exist are non-holes.
2784 */
2785 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2786 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2787 error = 0;
2788 }
2789
2790 out:
2791 if (!(flags & DNODE_FIND_HAVELOCK))
2792 rw_exit(&dn->dn_struct_rwlock);
2793
2794 return (error);
2795 }
2796
2797 #if defined(_KERNEL)
2798 EXPORT_SYMBOL(dnode_hold);
2799 EXPORT_SYMBOL(dnode_rele);
2800 EXPORT_SYMBOL(dnode_set_nlevels);
2801 EXPORT_SYMBOL(dnode_set_blksz);
2802 EXPORT_SYMBOL(dnode_free_range);
2803 EXPORT_SYMBOL(dnode_evict_dbufs);
2804 EXPORT_SYMBOL(dnode_evict_bonus);
2805 #endif
2806
2807 ZFS_MODULE_PARAM(zfs, zfs_, default_bs, INT, ZMOD_RW,
2808 "Default dnode block shift");
2809 ZFS_MODULE_PARAM(zfs, zfs_, default_ibs, INT, ZMOD_RW,
2810 "Default dnode indirect block shift");
2811