xref: /freebsd/sys/contrib/openzfs/module/zfs/dnode.c (revision 5e2fc2c3c59455ddd6354e765ca474182d90ba28)
1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or https://opensource.org/licenses/CDDL-1.0.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2012, 2020 by Delphix. All rights reserved.
25  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
26  */
27 
28 #include <sys/zfs_context.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/spa.h>
38 #include <sys/zio.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/range_tree.h>
41 #include <sys/trace_zfs.h>
42 #include <sys/zfs_project.h>
43 
44 dnode_stats_t dnode_stats = {
45 	{ "dnode_hold_dbuf_hold",		KSTAT_DATA_UINT64 },
46 	{ "dnode_hold_dbuf_read",		KSTAT_DATA_UINT64 },
47 	{ "dnode_hold_alloc_hits",		KSTAT_DATA_UINT64 },
48 	{ "dnode_hold_alloc_misses",		KSTAT_DATA_UINT64 },
49 	{ "dnode_hold_alloc_interior",		KSTAT_DATA_UINT64 },
50 	{ "dnode_hold_alloc_lock_retry",	KSTAT_DATA_UINT64 },
51 	{ "dnode_hold_alloc_lock_misses",	KSTAT_DATA_UINT64 },
52 	{ "dnode_hold_alloc_type_none",		KSTAT_DATA_UINT64 },
53 	{ "dnode_hold_free_hits",		KSTAT_DATA_UINT64 },
54 	{ "dnode_hold_free_misses",		KSTAT_DATA_UINT64 },
55 	{ "dnode_hold_free_lock_misses",	KSTAT_DATA_UINT64 },
56 	{ "dnode_hold_free_lock_retry",		KSTAT_DATA_UINT64 },
57 	{ "dnode_hold_free_overflow",		KSTAT_DATA_UINT64 },
58 	{ "dnode_hold_free_refcount",		KSTAT_DATA_UINT64 },
59 	{ "dnode_free_interior_lock_retry",	KSTAT_DATA_UINT64 },
60 	{ "dnode_allocate",			KSTAT_DATA_UINT64 },
61 	{ "dnode_reallocate",			KSTAT_DATA_UINT64 },
62 	{ "dnode_buf_evict",			KSTAT_DATA_UINT64 },
63 	{ "dnode_alloc_next_chunk",		KSTAT_DATA_UINT64 },
64 	{ "dnode_alloc_race",			KSTAT_DATA_UINT64 },
65 	{ "dnode_alloc_next_block",		KSTAT_DATA_UINT64 },
66 	{ "dnode_move_invalid",			KSTAT_DATA_UINT64 },
67 	{ "dnode_move_recheck1",		KSTAT_DATA_UINT64 },
68 	{ "dnode_move_recheck2",		KSTAT_DATA_UINT64 },
69 	{ "dnode_move_special",			KSTAT_DATA_UINT64 },
70 	{ "dnode_move_handle",			KSTAT_DATA_UINT64 },
71 	{ "dnode_move_rwlock",			KSTAT_DATA_UINT64 },
72 	{ "dnode_move_active",			KSTAT_DATA_UINT64 },
73 };
74 
75 dnode_sums_t dnode_sums;
76 
77 static kstat_t *dnode_ksp;
78 static kmem_cache_t *dnode_cache;
79 
80 static dnode_phys_t dnode_phys_zero __maybe_unused;
81 
82 int zfs_default_bs = SPA_MINBLOCKSHIFT;
83 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
84 
85 #ifdef	_KERNEL
86 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
87 #endif /* _KERNEL */
88 
89 static char *
rt_name(dnode_t * dn,const char * name)90 rt_name(dnode_t *dn, const char *name)
91 {
92 	struct objset *os = dn->dn_objset;
93 
94 	return (kmem_asprintf("{spa=%s objset=%llu obj=%llu %s}",
95 	    spa_name(os->os_spa),
96 	    (u_longlong_t)(os->os_dsl_dataset ?
97 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET),
98 	    (u_longlong_t)dn->dn_object,
99 	    name));
100 }
101 
102 static int
dbuf_compare(const void * x1,const void * x2)103 dbuf_compare(const void *x1, const void *x2)
104 {
105 	const dmu_buf_impl_t *d1 = x1;
106 	const dmu_buf_impl_t *d2 = x2;
107 
108 	int cmp = TREE_CMP(d1->db_level, d2->db_level);
109 	if (likely(cmp))
110 		return (cmp);
111 
112 	cmp = TREE_CMP(d1->db_blkid, d2->db_blkid);
113 	if (likely(cmp))
114 		return (cmp);
115 
116 	if (d1->db_state == DB_MARKER) {
117 		ASSERT3S(d2->db_state, !=, DB_MARKER);
118 		return (TREE_PCMP(d1->db_parent, d2));
119 	} else if (d2->db_state == DB_MARKER) {
120 		ASSERT3S(d1->db_state, !=, DB_MARKER);
121 		return (TREE_PCMP(d1, d2->db_parent));
122 	}
123 
124 	if (d1->db_state == DB_SEARCH) {
125 		ASSERT3S(d2->db_state, !=, DB_SEARCH);
126 		return (-1);
127 	} else if (d2->db_state == DB_SEARCH) {
128 		ASSERT3S(d1->db_state, !=, DB_SEARCH);
129 		return (1);
130 	}
131 
132 	return (TREE_PCMP(d1, d2));
133 }
134 
135 static int
dnode_cons(void * arg,void * unused,int kmflag)136 dnode_cons(void *arg, void *unused, int kmflag)
137 {
138 	(void) unused, (void) kmflag;
139 	dnode_t *dn = arg;
140 
141 	rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
142 	mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
143 	mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
144 	cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
145 	cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL);
146 
147 	/*
148 	 * Every dbuf has a reference, and dropping a tracked reference is
149 	 * O(number of references), so don't track dn_holds.
150 	 */
151 	zfs_refcount_create_untracked(&dn->dn_holds);
152 	zfs_refcount_create(&dn->dn_tx_holds);
153 	list_link_init(&dn->dn_link);
154 
155 	memset(dn->dn_next_type, 0, sizeof (dn->dn_next_type));
156 	memset(dn->dn_next_nblkptr, 0, sizeof (dn->dn_next_nblkptr));
157 	memset(dn->dn_next_nlevels, 0, sizeof (dn->dn_next_nlevels));
158 	memset(dn->dn_next_indblkshift, 0, sizeof (dn->dn_next_indblkshift));
159 	memset(dn->dn_next_bonustype, 0, sizeof (dn->dn_next_bonustype));
160 	memset(dn->dn_rm_spillblk, 0, sizeof (dn->dn_rm_spillblk));
161 	memset(dn->dn_next_bonuslen, 0, sizeof (dn->dn_next_bonuslen));
162 	memset(dn->dn_next_blksz, 0, sizeof (dn->dn_next_blksz));
163 	memset(dn->dn_next_maxblkid, 0, sizeof (dn->dn_next_maxblkid));
164 
165 	for (int i = 0; i < TXG_SIZE; i++) {
166 		multilist_link_init(&dn->dn_dirty_link[i]);
167 		dn->dn_free_ranges[i] = NULL;
168 		list_create(&dn->dn_dirty_records[i],
169 		    sizeof (dbuf_dirty_record_t),
170 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
171 	}
172 
173 	dn->dn_allocated_txg = 0;
174 	dn->dn_free_txg = 0;
175 	dn->dn_assigned_txg = 0;
176 	dn->dn_dirtycnt = 0;
177 	dn->dn_bonus = NULL;
178 	dn->dn_have_spill = B_FALSE;
179 	dn->dn_zio = NULL;
180 	dn->dn_oldused = 0;
181 	dn->dn_oldflags = 0;
182 	dn->dn_olduid = 0;
183 	dn->dn_oldgid = 0;
184 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
185 	dn->dn_newuid = 0;
186 	dn->dn_newgid = 0;
187 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
188 	dn->dn_id_flags = 0;
189 
190 	dn->dn_dbufs_count = 0;
191 	avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
192 	    offsetof(dmu_buf_impl_t, db_link));
193 
194 	dn->dn_moved = 0;
195 	return (0);
196 }
197 
198 static void
dnode_dest(void * arg,void * unused)199 dnode_dest(void *arg, void *unused)
200 {
201 	(void) unused;
202 	dnode_t *dn = arg;
203 
204 	rw_destroy(&dn->dn_struct_rwlock);
205 	mutex_destroy(&dn->dn_mtx);
206 	mutex_destroy(&dn->dn_dbufs_mtx);
207 	cv_destroy(&dn->dn_notxholds);
208 	cv_destroy(&dn->dn_nodnholds);
209 	zfs_refcount_destroy(&dn->dn_holds);
210 	zfs_refcount_destroy(&dn->dn_tx_holds);
211 	ASSERT(!list_link_active(&dn->dn_link));
212 
213 	for (int i = 0; i < TXG_SIZE; i++) {
214 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
215 		ASSERT0P(dn->dn_free_ranges[i]);
216 		list_destroy(&dn->dn_dirty_records[i]);
217 		ASSERT0(dn->dn_next_nblkptr[i]);
218 		ASSERT0(dn->dn_next_nlevels[i]);
219 		ASSERT0(dn->dn_next_indblkshift[i]);
220 		ASSERT0(dn->dn_next_bonustype[i]);
221 		ASSERT0(dn->dn_rm_spillblk[i]);
222 		ASSERT0(dn->dn_next_bonuslen[i]);
223 		ASSERT0(dn->dn_next_blksz[i]);
224 		ASSERT0(dn->dn_next_maxblkid[i]);
225 	}
226 
227 	ASSERT0(dn->dn_allocated_txg);
228 	ASSERT0(dn->dn_free_txg);
229 	ASSERT0(dn->dn_assigned_txg);
230 	ASSERT0(dn->dn_dirtycnt);
231 	ASSERT0P(dn->dn_bonus);
232 	ASSERT(!dn->dn_have_spill);
233 	ASSERT0P(dn->dn_zio);
234 	ASSERT0(dn->dn_oldused);
235 	ASSERT0(dn->dn_oldflags);
236 	ASSERT0(dn->dn_olduid);
237 	ASSERT0(dn->dn_oldgid);
238 	ASSERT0(dn->dn_oldprojid);
239 	ASSERT0(dn->dn_newuid);
240 	ASSERT0(dn->dn_newgid);
241 	ASSERT0(dn->dn_newprojid);
242 	ASSERT0(dn->dn_id_flags);
243 
244 	ASSERT0(dn->dn_dbufs_count);
245 	avl_destroy(&dn->dn_dbufs);
246 }
247 
248 static int
dnode_kstats_update(kstat_t * ksp,int rw)249 dnode_kstats_update(kstat_t *ksp, int rw)
250 {
251 	dnode_stats_t *ds = ksp->ks_data;
252 
253 	if (rw == KSTAT_WRITE)
254 		return (EACCES);
255 	ds->dnode_hold_dbuf_hold.value.ui64 =
256 	    wmsum_value(&dnode_sums.dnode_hold_dbuf_hold);
257 	ds->dnode_hold_dbuf_read.value.ui64 =
258 	    wmsum_value(&dnode_sums.dnode_hold_dbuf_read);
259 	ds->dnode_hold_alloc_hits.value.ui64 =
260 	    wmsum_value(&dnode_sums.dnode_hold_alloc_hits);
261 	ds->dnode_hold_alloc_misses.value.ui64 =
262 	    wmsum_value(&dnode_sums.dnode_hold_alloc_misses);
263 	ds->dnode_hold_alloc_interior.value.ui64 =
264 	    wmsum_value(&dnode_sums.dnode_hold_alloc_interior);
265 	ds->dnode_hold_alloc_lock_retry.value.ui64 =
266 	    wmsum_value(&dnode_sums.dnode_hold_alloc_lock_retry);
267 	ds->dnode_hold_alloc_lock_misses.value.ui64 =
268 	    wmsum_value(&dnode_sums.dnode_hold_alloc_lock_misses);
269 	ds->dnode_hold_alloc_type_none.value.ui64 =
270 	    wmsum_value(&dnode_sums.dnode_hold_alloc_type_none);
271 	ds->dnode_hold_free_hits.value.ui64 =
272 	    wmsum_value(&dnode_sums.dnode_hold_free_hits);
273 	ds->dnode_hold_free_misses.value.ui64 =
274 	    wmsum_value(&dnode_sums.dnode_hold_free_misses);
275 	ds->dnode_hold_free_lock_misses.value.ui64 =
276 	    wmsum_value(&dnode_sums.dnode_hold_free_lock_misses);
277 	ds->dnode_hold_free_lock_retry.value.ui64 =
278 	    wmsum_value(&dnode_sums.dnode_hold_free_lock_retry);
279 	ds->dnode_hold_free_refcount.value.ui64 =
280 	    wmsum_value(&dnode_sums.dnode_hold_free_refcount);
281 	ds->dnode_hold_free_overflow.value.ui64 =
282 	    wmsum_value(&dnode_sums.dnode_hold_free_overflow);
283 	ds->dnode_free_interior_lock_retry.value.ui64 =
284 	    wmsum_value(&dnode_sums.dnode_free_interior_lock_retry);
285 	ds->dnode_allocate.value.ui64 =
286 	    wmsum_value(&dnode_sums.dnode_allocate);
287 	ds->dnode_reallocate.value.ui64 =
288 	    wmsum_value(&dnode_sums.dnode_reallocate);
289 	ds->dnode_buf_evict.value.ui64 =
290 	    wmsum_value(&dnode_sums.dnode_buf_evict);
291 	ds->dnode_alloc_next_chunk.value.ui64 =
292 	    wmsum_value(&dnode_sums.dnode_alloc_next_chunk);
293 	ds->dnode_alloc_race.value.ui64 =
294 	    wmsum_value(&dnode_sums.dnode_alloc_race);
295 	ds->dnode_alloc_next_block.value.ui64 =
296 	    wmsum_value(&dnode_sums.dnode_alloc_next_block);
297 	ds->dnode_move_invalid.value.ui64 =
298 	    wmsum_value(&dnode_sums.dnode_move_invalid);
299 	ds->dnode_move_recheck1.value.ui64 =
300 	    wmsum_value(&dnode_sums.dnode_move_recheck1);
301 	ds->dnode_move_recheck2.value.ui64 =
302 	    wmsum_value(&dnode_sums.dnode_move_recheck2);
303 	ds->dnode_move_special.value.ui64 =
304 	    wmsum_value(&dnode_sums.dnode_move_special);
305 	ds->dnode_move_handle.value.ui64 =
306 	    wmsum_value(&dnode_sums.dnode_move_handle);
307 	ds->dnode_move_rwlock.value.ui64 =
308 	    wmsum_value(&dnode_sums.dnode_move_rwlock);
309 	ds->dnode_move_active.value.ui64 =
310 	    wmsum_value(&dnode_sums.dnode_move_active);
311 	return (0);
312 }
313 
314 void
dnode_init(void)315 dnode_init(void)
316 {
317 	ASSERT0P(dnode_cache);
318 	dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
319 	    0, dnode_cons, dnode_dest, NULL, NULL, NULL, KMC_RECLAIMABLE);
320 	kmem_cache_set_move(dnode_cache, dnode_move);
321 
322 	wmsum_init(&dnode_sums.dnode_hold_dbuf_hold, 0);
323 	wmsum_init(&dnode_sums.dnode_hold_dbuf_read, 0);
324 	wmsum_init(&dnode_sums.dnode_hold_alloc_hits, 0);
325 	wmsum_init(&dnode_sums.dnode_hold_alloc_misses, 0);
326 	wmsum_init(&dnode_sums.dnode_hold_alloc_interior, 0);
327 	wmsum_init(&dnode_sums.dnode_hold_alloc_lock_retry, 0);
328 	wmsum_init(&dnode_sums.dnode_hold_alloc_lock_misses, 0);
329 	wmsum_init(&dnode_sums.dnode_hold_alloc_type_none, 0);
330 	wmsum_init(&dnode_sums.dnode_hold_free_hits, 0);
331 	wmsum_init(&dnode_sums.dnode_hold_free_misses, 0);
332 	wmsum_init(&dnode_sums.dnode_hold_free_lock_misses, 0);
333 	wmsum_init(&dnode_sums.dnode_hold_free_lock_retry, 0);
334 	wmsum_init(&dnode_sums.dnode_hold_free_refcount, 0);
335 	wmsum_init(&dnode_sums.dnode_hold_free_overflow, 0);
336 	wmsum_init(&dnode_sums.dnode_free_interior_lock_retry, 0);
337 	wmsum_init(&dnode_sums.dnode_allocate, 0);
338 	wmsum_init(&dnode_sums.dnode_reallocate, 0);
339 	wmsum_init(&dnode_sums.dnode_buf_evict, 0);
340 	wmsum_init(&dnode_sums.dnode_alloc_next_chunk, 0);
341 	wmsum_init(&dnode_sums.dnode_alloc_race, 0);
342 	wmsum_init(&dnode_sums.dnode_alloc_next_block, 0);
343 	wmsum_init(&dnode_sums.dnode_move_invalid, 0);
344 	wmsum_init(&dnode_sums.dnode_move_recheck1, 0);
345 	wmsum_init(&dnode_sums.dnode_move_recheck2, 0);
346 	wmsum_init(&dnode_sums.dnode_move_special, 0);
347 	wmsum_init(&dnode_sums.dnode_move_handle, 0);
348 	wmsum_init(&dnode_sums.dnode_move_rwlock, 0);
349 	wmsum_init(&dnode_sums.dnode_move_active, 0);
350 
351 	dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
352 	    KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
353 	    KSTAT_FLAG_VIRTUAL);
354 	if (dnode_ksp != NULL) {
355 		dnode_ksp->ks_data = &dnode_stats;
356 		dnode_ksp->ks_update = dnode_kstats_update;
357 		kstat_install(dnode_ksp);
358 	}
359 }
360 
361 void
dnode_fini(void)362 dnode_fini(void)
363 {
364 	if (dnode_ksp != NULL) {
365 		kstat_delete(dnode_ksp);
366 		dnode_ksp = NULL;
367 	}
368 
369 	wmsum_fini(&dnode_sums.dnode_hold_dbuf_hold);
370 	wmsum_fini(&dnode_sums.dnode_hold_dbuf_read);
371 	wmsum_fini(&dnode_sums.dnode_hold_alloc_hits);
372 	wmsum_fini(&dnode_sums.dnode_hold_alloc_misses);
373 	wmsum_fini(&dnode_sums.dnode_hold_alloc_interior);
374 	wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_retry);
375 	wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_misses);
376 	wmsum_fini(&dnode_sums.dnode_hold_alloc_type_none);
377 	wmsum_fini(&dnode_sums.dnode_hold_free_hits);
378 	wmsum_fini(&dnode_sums.dnode_hold_free_misses);
379 	wmsum_fini(&dnode_sums.dnode_hold_free_lock_misses);
380 	wmsum_fini(&dnode_sums.dnode_hold_free_lock_retry);
381 	wmsum_fini(&dnode_sums.dnode_hold_free_refcount);
382 	wmsum_fini(&dnode_sums.dnode_hold_free_overflow);
383 	wmsum_fini(&dnode_sums.dnode_free_interior_lock_retry);
384 	wmsum_fini(&dnode_sums.dnode_allocate);
385 	wmsum_fini(&dnode_sums.dnode_reallocate);
386 	wmsum_fini(&dnode_sums.dnode_buf_evict);
387 	wmsum_fini(&dnode_sums.dnode_alloc_next_chunk);
388 	wmsum_fini(&dnode_sums.dnode_alloc_race);
389 	wmsum_fini(&dnode_sums.dnode_alloc_next_block);
390 	wmsum_fini(&dnode_sums.dnode_move_invalid);
391 	wmsum_fini(&dnode_sums.dnode_move_recheck1);
392 	wmsum_fini(&dnode_sums.dnode_move_recheck2);
393 	wmsum_fini(&dnode_sums.dnode_move_special);
394 	wmsum_fini(&dnode_sums.dnode_move_handle);
395 	wmsum_fini(&dnode_sums.dnode_move_rwlock);
396 	wmsum_fini(&dnode_sums.dnode_move_active);
397 
398 	kmem_cache_destroy(dnode_cache);
399 	dnode_cache = NULL;
400 }
401 
402 
403 #ifdef ZFS_DEBUG
404 void
dnode_verify(dnode_t * dn)405 dnode_verify(dnode_t *dn)
406 {
407 	int drop_struct_lock = FALSE;
408 
409 	ASSERT(dn->dn_phys);
410 	ASSERT(dn->dn_objset);
411 	ASSERT(dn->dn_handle->dnh_dnode == dn);
412 
413 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
414 
415 	if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
416 		return;
417 
418 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
419 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
420 		drop_struct_lock = TRUE;
421 	}
422 	if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
423 		int i;
424 		int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
425 		ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
426 		if (dn->dn_datablkshift) {
427 			ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
428 			ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
429 			ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
430 		}
431 		ASSERT3U(dn->dn_nlevels, <=, 30);
432 		ASSERT(DMU_OT_IS_VALID(dn->dn_type));
433 		ASSERT3U(dn->dn_nblkptr, >=, 1);
434 		ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
435 		ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
436 		ASSERT3U(dn->dn_datablksz, ==,
437 		    dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
438 		ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
439 		ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
440 		    dn->dn_bonuslen, <=, max_bonuslen);
441 		for (i = 0; i < TXG_SIZE; i++) {
442 			ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
443 		}
444 	}
445 	if (dn->dn_phys->dn_type != DMU_OT_NONE)
446 		ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
447 	ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
448 	if (dn->dn_dbuf != NULL) {
449 		ASSERT3P(dn->dn_phys, ==,
450 		    (dnode_phys_t *)dn->dn_dbuf->db.db_data +
451 		    (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
452 	}
453 	if (drop_struct_lock)
454 		rw_exit(&dn->dn_struct_rwlock);
455 }
456 #endif
457 
458 void
dnode_byteswap(dnode_phys_t * dnp)459 dnode_byteswap(dnode_phys_t *dnp)
460 {
461 	uint64_t *buf64 = (void*)&dnp->dn_blkptr;
462 	int i;
463 
464 	if (dnp->dn_type == DMU_OT_NONE) {
465 		memset(dnp, 0, sizeof (dnode_phys_t));
466 		return;
467 	}
468 
469 	dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
470 	dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
471 	dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
472 	dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
473 	dnp->dn_used = BSWAP_64(dnp->dn_used);
474 
475 	/*
476 	 * dn_nblkptr is only one byte, so it's OK to read it in either
477 	 * byte order.  We can't read dn_bouslen.
478 	 */
479 	ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
480 	ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
481 	for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
482 		buf64[i] = BSWAP_64(buf64[i]);
483 
484 	/*
485 	 * OK to check dn_bonuslen for zero, because it won't matter if
486 	 * we have the wrong byte order.  This is necessary because the
487 	 * dnode dnode is smaller than a regular dnode.
488 	 */
489 	if (dnp->dn_bonuslen != 0) {
490 		dmu_object_byteswap_t byteswap;
491 		ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
492 		byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
493 		dmu_ot_byteswap[byteswap].ob_func(DN_BONUS(dnp),
494 		    DN_MAX_BONUS_LEN(dnp));
495 	}
496 
497 	/* Swap SPILL block if we have one */
498 	if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
499 		byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
500 }
501 
502 void
dnode_buf_byteswap(void * vbuf,size_t size)503 dnode_buf_byteswap(void *vbuf, size_t size)
504 {
505 	int i = 0;
506 
507 	ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
508 	ASSERT0((size & (sizeof (dnode_phys_t)-1)));
509 
510 	while (i < size) {
511 		dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
512 		dnode_byteswap(dnp);
513 
514 		i += DNODE_MIN_SIZE;
515 		if (dnp->dn_type != DMU_OT_NONE)
516 			i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
517 	}
518 }
519 
520 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)521 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
522 {
523 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
524 
525 	dnode_setdirty(dn, tx);
526 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
527 	ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
528 	    (dn->dn_nblkptr-1) * sizeof (blkptr_t));
529 
530 	if (newsize < dn->dn_bonuslen) {
531 		/* clear any data after the end of the new size */
532 		size_t diff = dn->dn_bonuslen - newsize;
533 		char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize;
534 		memset(data_end, 0, diff);
535 	}
536 
537 	dn->dn_bonuslen = newsize;
538 	if (newsize == 0)
539 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
540 	else
541 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
542 	rw_exit(&dn->dn_struct_rwlock);
543 }
544 
545 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)546 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
547 {
548 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
549 	dnode_setdirty(dn, tx);
550 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
551 	dn->dn_bonustype = newtype;
552 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
553 	rw_exit(&dn->dn_struct_rwlock);
554 }
555 
556 void
dnode_set_storage_type(dnode_t * dn,dmu_object_type_t newtype)557 dnode_set_storage_type(dnode_t *dn, dmu_object_type_t newtype)
558 {
559 	/*
560 	 * This is not in the dnode_phys, but it should be, and perhaps one day
561 	 * will. For now we require it be set after taking a hold.
562 	 */
563 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
564 	dn->dn_storage_type = newtype;
565 }
566 
567 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)568 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
569 {
570 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
571 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
572 	dnode_setdirty(dn, tx);
573 	dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
574 	dn->dn_have_spill = B_FALSE;
575 }
576 
577 static void
dnode_setdblksz(dnode_t * dn,int size)578 dnode_setdblksz(dnode_t *dn, int size)
579 {
580 	ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
581 	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
582 	ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
583 	ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
584 	    1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
585 	dn->dn_datablksz = size;
586 	dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
587 	dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
588 }
589 
590 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)591 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
592     uint64_t object, dnode_handle_t *dnh)
593 {
594 	dnode_t *dn;
595 
596 	dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
597 	dn->dn_moved = 0;
598 
599 	/*
600 	 * Defer setting dn_objset until the dnode is ready to be a candidate
601 	 * for the dnode_move() callback.
602 	 */
603 	dn->dn_object = object;
604 	dn->dn_dbuf = db;
605 	dn->dn_handle = dnh;
606 	dn->dn_phys = dnp;
607 
608 	if (dnp->dn_datablkszsec) {
609 		dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
610 	} else {
611 		dn->dn_datablksz = 0;
612 		dn->dn_datablkszsec = 0;
613 		dn->dn_datablkshift = 0;
614 	}
615 	dn->dn_indblkshift = dnp->dn_indblkshift;
616 	dn->dn_nlevels = dnp->dn_nlevels;
617 	dn->dn_type = dnp->dn_type;
618 	dn->dn_nblkptr = dnp->dn_nblkptr;
619 	dn->dn_checksum = dnp->dn_checksum;
620 	dn->dn_compress = dnp->dn_compress;
621 	dn->dn_bonustype = dnp->dn_bonustype;
622 	dn->dn_bonuslen = dnp->dn_bonuslen;
623 	dn->dn_num_slots = dnp->dn_extra_slots + 1;
624 	dn->dn_maxblkid = dnp->dn_maxblkid;
625 	dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
626 	dn->dn_id_flags = 0;
627 
628 	dn->dn_storage_type = DMU_OT_NONE;
629 
630 	dmu_zfetch_init(&dn->dn_zfetch, dn);
631 
632 	ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
633 	ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
634 	ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
635 
636 	mutex_enter(&os->os_lock);
637 
638 	/*
639 	 * Exclude special dnodes from os_dnodes so an empty os_dnodes
640 	 * signifies that the special dnodes have no references from
641 	 * their children (the entries in os_dnodes).  This allows
642 	 * dnode_destroy() to easily determine if the last child has
643 	 * been removed and then complete eviction of the objset.
644 	 */
645 	if (!DMU_OBJECT_IS_SPECIAL(object))
646 		list_insert_head(&os->os_dnodes, dn);
647 	membar_producer();
648 
649 	/*
650 	 * Everything else must be valid before assigning dn_objset
651 	 * makes the dnode eligible for dnode_move().
652 	 */
653 	dn->dn_objset = os;
654 
655 	dnh->dnh_dnode = dn;
656 	mutex_exit(&os->os_lock);
657 
658 	arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
659 
660 	return (dn);
661 }
662 
663 /*
664  * Caller must be holding the dnode handle, which is released upon return.
665  */
666 static void
dnode_destroy(dnode_t * dn)667 dnode_destroy(dnode_t *dn)
668 {
669 	objset_t *os = dn->dn_objset;
670 	boolean_t complete_os_eviction = B_FALSE;
671 
672 	ASSERT0((dn->dn_id_flags & DN_ID_NEW_EXIST));
673 
674 	mutex_enter(&os->os_lock);
675 	POINTER_INVALIDATE(&dn->dn_objset);
676 	if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
677 		list_remove(&os->os_dnodes, dn);
678 		complete_os_eviction =
679 		    list_is_empty(&os->os_dnodes) &&
680 		    list_link_active(&os->os_evicting_node);
681 	}
682 	mutex_exit(&os->os_lock);
683 
684 	/* the dnode can no longer move, so we can release the handle */
685 	if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
686 		zrl_remove(&dn->dn_handle->dnh_zrlock);
687 
688 	dn->dn_allocated_txg = 0;
689 	dn->dn_free_txg = 0;
690 	dn->dn_assigned_txg = 0;
691 	dn->dn_dirtycnt = 0;
692 
693 	if (dn->dn_bonus != NULL) {
694 		mutex_enter(&dn->dn_bonus->db_mtx);
695 		dbuf_destroy(dn->dn_bonus);
696 		dn->dn_bonus = NULL;
697 	}
698 	dn->dn_zio = NULL;
699 
700 	dn->dn_have_spill = B_FALSE;
701 	dn->dn_oldused = 0;
702 	dn->dn_oldflags = 0;
703 	dn->dn_olduid = 0;
704 	dn->dn_oldgid = 0;
705 	dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
706 	dn->dn_newuid = 0;
707 	dn->dn_newgid = 0;
708 	dn->dn_newprojid = ZFS_DEFAULT_PROJID;
709 	dn->dn_id_flags = 0;
710 
711 	dn->dn_storage_type = DMU_OT_NONE;
712 
713 	dmu_zfetch_fini(&dn->dn_zfetch);
714 	kmem_cache_free(dnode_cache, dn);
715 	arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
716 
717 	if (complete_os_eviction)
718 		dmu_objset_evict_done(os);
719 }
720 
721 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)722 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
723     dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
724 {
725 	int i;
726 
727 	ASSERT3U(dn_slots, >, 0);
728 	ASSERT3U(dn_slots << DNODE_SHIFT, <=,
729 	    spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
730 	ASSERT3U(blocksize, <=,
731 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
732 	if (blocksize == 0)
733 		blocksize = 1 << zfs_default_bs;
734 	else
735 		blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
736 
737 	if (ibs == 0)
738 		ibs = zfs_default_ibs;
739 
740 	ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
741 
742 	dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
743 	    dn->dn_objset, (u_longlong_t)dn->dn_object,
744 	    (u_longlong_t)tx->tx_txg, blocksize, ibs, dn_slots);
745 	DNODE_STAT_BUMP(dnode_allocate);
746 
747 	ASSERT(dn->dn_type == DMU_OT_NONE);
748 	ASSERT0(memcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)));
749 	ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
750 	ASSERT(ot != DMU_OT_NONE);
751 	ASSERT(DMU_OT_IS_VALID(ot));
752 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
753 	    (bonustype == DMU_OT_SA && bonuslen == 0) ||
754 	    (bonustype == DMU_OTN_UINT64_METADATA && bonuslen == 0) ||
755 	    (bonustype != DMU_OT_NONE && bonuslen != 0));
756 	ASSERT(DMU_OT_IS_VALID(bonustype));
757 	ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
758 	ASSERT(dn->dn_type == DMU_OT_NONE);
759 	ASSERT0(dn->dn_maxblkid);
760 	ASSERT0(dn->dn_allocated_txg);
761 	ASSERT0(dn->dn_assigned_txg);
762 	ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
763 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
764 	ASSERT(avl_is_empty(&dn->dn_dbufs));
765 
766 	for (i = 0; i < TXG_SIZE; i++) {
767 		ASSERT0(dn->dn_next_nblkptr[i]);
768 		ASSERT0(dn->dn_next_nlevels[i]);
769 		ASSERT0(dn->dn_next_indblkshift[i]);
770 		ASSERT0(dn->dn_next_bonuslen[i]);
771 		ASSERT0(dn->dn_next_bonustype[i]);
772 		ASSERT0(dn->dn_rm_spillblk[i]);
773 		ASSERT0(dn->dn_next_blksz[i]);
774 		ASSERT0(dn->dn_next_maxblkid[i]);
775 		ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
776 		ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
777 		ASSERT0P(dn->dn_free_ranges[i]);
778 	}
779 
780 	dn->dn_type = ot;
781 	dnode_setdblksz(dn, blocksize);
782 	dn->dn_indblkshift = ibs;
783 	dn->dn_nlevels = 1;
784 	dn->dn_num_slots = dn_slots;
785 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
786 		dn->dn_nblkptr = 1;
787 	else {
788 		dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
789 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
790 		    SPA_BLKPTRSHIFT));
791 	}
792 
793 	dn->dn_bonustype = bonustype;
794 	dn->dn_bonuslen = bonuslen;
795 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
796 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
797 
798 	dn->dn_free_txg = 0;
799 	dn->dn_dirtycnt = 0;
800 
801 	dn->dn_allocated_txg = tx->tx_txg;
802 	dn->dn_id_flags = 0;
803 
804 	dnode_setdirty(dn, tx);
805 	dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
806 	dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
807 	dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
808 	dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
809 }
810 
811 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)812 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
813     dmu_object_type_t bonustype, int bonuslen, int dn_slots,
814     boolean_t keep_spill, dmu_tx_t *tx)
815 {
816 	int nblkptr;
817 
818 	ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
819 	ASSERT3U(blocksize, <=,
820 	    spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
821 	ASSERT0(blocksize % SPA_MINBLOCKSIZE);
822 	ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
823 	ASSERT(tx->tx_txg != 0);
824 	ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
825 	    (bonustype != DMU_OT_NONE && bonuslen != 0) ||
826 	    (bonustype == DMU_OT_SA && bonuslen == 0));
827 	ASSERT(DMU_OT_IS_VALID(bonustype));
828 	ASSERT3U(bonuslen, <=,
829 	    DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
830 	ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
831 
832 	dnode_free_interior_slots(dn);
833 	DNODE_STAT_BUMP(dnode_reallocate);
834 
835 	/* clean up any unreferenced dbufs */
836 	dnode_evict_dbufs(dn);
837 
838 	dn->dn_id_flags = 0;
839 
840 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
841 	dnode_setdirty(dn, tx);
842 	if (dn->dn_datablksz != blocksize) {
843 		/* change blocksize */
844 		ASSERT0(dn->dn_maxblkid);
845 		ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
846 		    dnode_block_freed(dn, 0));
847 
848 		dnode_setdblksz(dn, blocksize);
849 		dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
850 	}
851 	if (dn->dn_bonuslen != bonuslen)
852 		dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
853 
854 	if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
855 		nblkptr = 1;
856 	else
857 		nblkptr = MIN(DN_MAX_NBLKPTR,
858 		    1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
859 		    SPA_BLKPTRSHIFT));
860 	if (dn->dn_bonustype != bonustype)
861 		dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
862 	if (dn->dn_nblkptr != nblkptr)
863 		dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
864 	if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
865 		dbuf_rm_spill(dn, tx);
866 		dnode_rm_spill(dn, tx);
867 	}
868 
869 	rw_exit(&dn->dn_struct_rwlock);
870 
871 	/* change type */
872 	dn->dn_type = ot;
873 
874 	/* change bonus size and type */
875 	mutex_enter(&dn->dn_mtx);
876 	dn->dn_bonustype = bonustype;
877 	dn->dn_bonuslen = bonuslen;
878 	dn->dn_num_slots = dn_slots;
879 	dn->dn_nblkptr = nblkptr;
880 	dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
881 	dn->dn_compress = ZIO_COMPRESS_INHERIT;
882 	ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
883 
884 	/* fix up the bonus db_size */
885 	if (dn->dn_bonus) {
886 		dn->dn_bonus->db.db_size =
887 		    DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
888 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
889 		ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
890 	}
891 
892 	dn->dn_allocated_txg = tx->tx_txg;
893 	mutex_exit(&dn->dn_mtx);
894 }
895 
896 #ifdef	_KERNEL
897 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)898 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
899 {
900 	ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
901 	ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
902 	ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
903 
904 	/* Copy fields. */
905 	ndn->dn_objset = odn->dn_objset;
906 	ndn->dn_object = odn->dn_object;
907 	ndn->dn_dbuf = odn->dn_dbuf;
908 	ndn->dn_handle = odn->dn_handle;
909 	ndn->dn_phys = odn->dn_phys;
910 	ndn->dn_type = odn->dn_type;
911 	ndn->dn_bonuslen = odn->dn_bonuslen;
912 	ndn->dn_bonustype = odn->dn_bonustype;
913 	ndn->dn_nblkptr = odn->dn_nblkptr;
914 	ndn->dn_checksum = odn->dn_checksum;
915 	ndn->dn_compress = odn->dn_compress;
916 	ndn->dn_nlevels = odn->dn_nlevels;
917 	ndn->dn_indblkshift = odn->dn_indblkshift;
918 	ndn->dn_datablkshift = odn->dn_datablkshift;
919 	ndn->dn_datablkszsec = odn->dn_datablkszsec;
920 	ndn->dn_datablksz = odn->dn_datablksz;
921 	ndn->dn_maxblkid = odn->dn_maxblkid;
922 	ndn->dn_num_slots = odn->dn_num_slots;
923 	memcpy(ndn->dn_next_type, odn->dn_next_type,
924 	    sizeof (odn->dn_next_type));
925 	memcpy(ndn->dn_next_nblkptr, odn->dn_next_nblkptr,
926 	    sizeof (odn->dn_next_nblkptr));
927 	memcpy(ndn->dn_next_nlevels, odn->dn_next_nlevels,
928 	    sizeof (odn->dn_next_nlevels));
929 	memcpy(ndn->dn_next_indblkshift, odn->dn_next_indblkshift,
930 	    sizeof (odn->dn_next_indblkshift));
931 	memcpy(ndn->dn_next_bonustype, odn->dn_next_bonustype,
932 	    sizeof (odn->dn_next_bonustype));
933 	memcpy(ndn->dn_rm_spillblk, odn->dn_rm_spillblk,
934 	    sizeof (odn->dn_rm_spillblk));
935 	memcpy(ndn->dn_next_bonuslen, odn->dn_next_bonuslen,
936 	    sizeof (odn->dn_next_bonuslen));
937 	memcpy(ndn->dn_next_blksz, odn->dn_next_blksz,
938 	    sizeof (odn->dn_next_blksz));
939 	memcpy(ndn->dn_next_maxblkid, odn->dn_next_maxblkid,
940 	    sizeof (odn->dn_next_maxblkid));
941 	for (int i = 0; i < TXG_SIZE; i++) {
942 		list_move_tail(&ndn->dn_dirty_records[i],
943 		    &odn->dn_dirty_records[i]);
944 	}
945 	memcpy(ndn->dn_free_ranges, odn->dn_free_ranges,
946 	    sizeof (odn->dn_free_ranges));
947 	ndn->dn_allocated_txg = odn->dn_allocated_txg;
948 	ndn->dn_free_txg = odn->dn_free_txg;
949 	ndn->dn_assigned_txg = odn->dn_assigned_txg;
950 	ndn->dn_dirtycnt = odn->dn_dirtycnt;
951 	ASSERT0(zfs_refcount_count(&odn->dn_tx_holds));
952 	zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
953 	ASSERT(avl_is_empty(&ndn->dn_dbufs));
954 	avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
955 	ndn->dn_dbufs_count = odn->dn_dbufs_count;
956 	ndn->dn_bonus = odn->dn_bonus;
957 	ndn->dn_have_spill = odn->dn_have_spill;
958 	ndn->dn_zio = odn->dn_zio;
959 	ndn->dn_oldused = odn->dn_oldused;
960 	ndn->dn_oldflags = odn->dn_oldflags;
961 	ndn->dn_olduid = odn->dn_olduid;
962 	ndn->dn_oldgid = odn->dn_oldgid;
963 	ndn->dn_oldprojid = odn->dn_oldprojid;
964 	ndn->dn_newuid = odn->dn_newuid;
965 	ndn->dn_newgid = odn->dn_newgid;
966 	ndn->dn_newprojid = odn->dn_newprojid;
967 	ndn->dn_id_flags = odn->dn_id_flags;
968 	ndn->dn_storage_type = odn->dn_storage_type;
969 	dmu_zfetch_init(&ndn->dn_zfetch, ndn);
970 
971 	/*
972 	 * Update back pointers. Updating the handle fixes the back pointer of
973 	 * every descendant dbuf as well as the bonus dbuf.
974 	 */
975 	ASSERT(ndn->dn_handle->dnh_dnode == odn);
976 	ndn->dn_handle->dnh_dnode = ndn;
977 
978 	/*
979 	 * Invalidate the original dnode by clearing all of its back pointers.
980 	 */
981 	odn->dn_dbuf = NULL;
982 	odn->dn_handle = NULL;
983 	avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
984 	    offsetof(dmu_buf_impl_t, db_link));
985 	odn->dn_dbufs_count = 0;
986 	odn->dn_bonus = NULL;
987 	dmu_zfetch_fini(&odn->dn_zfetch);
988 
989 	/*
990 	 * Set the low bit of the objset pointer to ensure that dnode_move()
991 	 * recognizes the dnode as invalid in any subsequent callback.
992 	 */
993 	POINTER_INVALIDATE(&odn->dn_objset);
994 
995 	/*
996 	 * Satisfy the destructor.
997 	 */
998 	for (int i = 0; i < TXG_SIZE; i++) {
999 		list_create(&odn->dn_dirty_records[i],
1000 		    sizeof (dbuf_dirty_record_t),
1001 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1002 		odn->dn_free_ranges[i] = NULL;
1003 		odn->dn_next_nlevels[i] = 0;
1004 		odn->dn_next_indblkshift[i] = 0;
1005 		odn->dn_next_bonustype[i] = 0;
1006 		odn->dn_rm_spillblk[i] = 0;
1007 		odn->dn_next_bonuslen[i] = 0;
1008 		odn->dn_next_blksz[i] = 0;
1009 	}
1010 	odn->dn_allocated_txg = 0;
1011 	odn->dn_free_txg = 0;
1012 	odn->dn_assigned_txg = 0;
1013 	odn->dn_dirtycnt = 0;
1014 	odn->dn_have_spill = B_FALSE;
1015 	odn->dn_zio = NULL;
1016 	odn->dn_oldused = 0;
1017 	odn->dn_oldflags = 0;
1018 	odn->dn_olduid = 0;
1019 	odn->dn_oldgid = 0;
1020 	odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
1021 	odn->dn_newuid = 0;
1022 	odn->dn_newgid = 0;
1023 	odn->dn_newprojid = ZFS_DEFAULT_PROJID;
1024 	odn->dn_id_flags = 0;
1025 	odn->dn_storage_type = DMU_OT_NONE;
1026 
1027 	/*
1028 	 * Mark the dnode.
1029 	 */
1030 	ndn->dn_moved = 1;
1031 	odn->dn_moved = (uint8_t)-1;
1032 }
1033 
1034 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)1035 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
1036 {
1037 	dnode_t *odn = buf, *ndn = newbuf;
1038 	objset_t *os;
1039 	int64_t refcount;
1040 	uint32_t dbufs;
1041 
1042 #ifndef USE_DNODE_HANDLE
1043 	/*
1044 	 * We can't move dnodes if dbufs reference them directly without
1045 	 * using handles and respecitve locking.  Unless USE_DNODE_HANDLE
1046 	 * is defined the code below is only to make sure it still builds,
1047 	 * but it should never be used, since it is unsafe.
1048 	 */
1049 #ifdef ZFS_DEBUG
1050 	PANIC("dnode_move() called without USE_DNODE_HANDLE");
1051 #endif
1052 	return (KMEM_CBRC_NO);
1053 #endif
1054 
1055 	/*
1056 	 * The dnode is on the objset's list of known dnodes if the objset
1057 	 * pointer is valid. We set the low bit of the objset pointer when
1058 	 * freeing the dnode to invalidate it, and the memory patterns written
1059 	 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
1060 	 * A newly created dnode sets the objset pointer last of all to indicate
1061 	 * that the dnode is known and in a valid state to be moved by this
1062 	 * function.
1063 	 */
1064 	os = odn->dn_objset;
1065 	if (!POINTER_IS_VALID(os)) {
1066 		DNODE_STAT_BUMP(dnode_move_invalid);
1067 		return (KMEM_CBRC_DONT_KNOW);
1068 	}
1069 
1070 	/*
1071 	 * Ensure that the objset does not go away during the move.
1072 	 */
1073 	rw_enter(&os_lock, RW_WRITER);
1074 	if (os != odn->dn_objset) {
1075 		rw_exit(&os_lock);
1076 		DNODE_STAT_BUMP(dnode_move_recheck1);
1077 		return (KMEM_CBRC_DONT_KNOW);
1078 	}
1079 
1080 	/*
1081 	 * If the dnode is still valid, then so is the objset. We know that no
1082 	 * valid objset can be freed while we hold os_lock, so we can safely
1083 	 * ensure that the objset remains in use.
1084 	 */
1085 	mutex_enter(&os->os_lock);
1086 
1087 	/*
1088 	 * Recheck the objset pointer in case the dnode was removed just before
1089 	 * acquiring the lock.
1090 	 */
1091 	if (os != odn->dn_objset) {
1092 		mutex_exit(&os->os_lock);
1093 		rw_exit(&os_lock);
1094 		DNODE_STAT_BUMP(dnode_move_recheck2);
1095 		return (KMEM_CBRC_DONT_KNOW);
1096 	}
1097 
1098 	/*
1099 	 * At this point we know that as long as we hold os->os_lock, the dnode
1100 	 * cannot be freed and fields within the dnode can be safely accessed.
1101 	 * The objset listing this dnode cannot go away as long as this dnode is
1102 	 * on its list.
1103 	 */
1104 	rw_exit(&os_lock);
1105 	if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
1106 		mutex_exit(&os->os_lock);
1107 		DNODE_STAT_BUMP(dnode_move_special);
1108 		return (KMEM_CBRC_NO);
1109 	}
1110 	ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
1111 
1112 	/*
1113 	 * Lock the dnode handle to prevent the dnode from obtaining any new
1114 	 * holds. This also prevents the descendant dbufs and the bonus dbuf
1115 	 * from accessing the dnode, so that we can discount their holds. The
1116 	 * handle is safe to access because we know that while the dnode cannot
1117 	 * go away, neither can its handle. Once we hold dnh_zrlock, we can
1118 	 * safely move any dnode referenced only by dbufs.
1119 	 */
1120 	if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
1121 		mutex_exit(&os->os_lock);
1122 		DNODE_STAT_BUMP(dnode_move_handle);
1123 		return (KMEM_CBRC_LATER);
1124 	}
1125 
1126 	/*
1127 	 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
1128 	 * We need to guarantee that there is a hold for every dbuf in order to
1129 	 * determine whether the dnode is actively referenced. Falsely matching
1130 	 * a dbuf to an active hold would lead to an unsafe move. It's possible
1131 	 * that a thread already having an active dnode hold is about to add a
1132 	 * dbuf, and we can't compare hold and dbuf counts while the add is in
1133 	 * progress.
1134 	 */
1135 	if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
1136 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1137 		mutex_exit(&os->os_lock);
1138 		DNODE_STAT_BUMP(dnode_move_rwlock);
1139 		return (KMEM_CBRC_LATER);
1140 	}
1141 
1142 	/*
1143 	 * A dbuf may be removed (evicted) without an active dnode hold. In that
1144 	 * case, the dbuf count is decremented under the handle lock before the
1145 	 * dbuf's hold is released. This order ensures that if we count the hold
1146 	 * after the dbuf is removed but before its hold is released, we will
1147 	 * treat the unmatched hold as active and exit safely. If we count the
1148 	 * hold before the dbuf is removed, the hold is discounted, and the
1149 	 * removal is blocked until the move completes.
1150 	 */
1151 	refcount = zfs_refcount_count(&odn->dn_holds);
1152 	ASSERT(refcount >= 0);
1153 	dbufs = DN_DBUFS_COUNT(odn);
1154 
1155 	/* We can't have more dbufs than dnode holds. */
1156 	ASSERT3U(dbufs, <=, refcount);
1157 	DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1158 	    uint32_t, dbufs);
1159 
1160 	if (refcount > dbufs) {
1161 		rw_exit(&odn->dn_struct_rwlock);
1162 		zrl_exit(&odn->dn_handle->dnh_zrlock);
1163 		mutex_exit(&os->os_lock);
1164 		DNODE_STAT_BUMP(dnode_move_active);
1165 		return (KMEM_CBRC_LATER);
1166 	}
1167 
1168 	rw_exit(&odn->dn_struct_rwlock);
1169 
1170 	/*
1171 	 * At this point we know that anyone with a hold on the dnode is not
1172 	 * actively referencing it. The dnode is known and in a valid state to
1173 	 * move. We're holding the locks needed to execute the critical section.
1174 	 */
1175 	dnode_move_impl(odn, ndn);
1176 
1177 	list_link_replace(&odn->dn_link, &ndn->dn_link);
1178 	/* If the dnode was safe to move, the refcount cannot have changed. */
1179 	ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1180 	ASSERT(dbufs == DN_DBUFS_COUNT(ndn));
1181 	zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1182 	mutex_exit(&os->os_lock);
1183 
1184 	return (KMEM_CBRC_YES);
1185 }
1186 #endif	/* _KERNEL */
1187 
1188 static void
dnode_slots_hold(dnode_children_t * children,int idx,int slots)1189 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1190 {
1191 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1192 
1193 	for (int i = idx; i < idx + slots; i++) {
1194 		dnode_handle_t *dnh = &children->dnc_children[i];
1195 		zrl_add(&dnh->dnh_zrlock);
1196 	}
1197 }
1198 
1199 static void
dnode_slots_rele(dnode_children_t * children,int idx,int slots)1200 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1201 {
1202 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1203 
1204 	for (int i = idx; i < idx + slots; i++) {
1205 		dnode_handle_t *dnh = &children->dnc_children[i];
1206 
1207 		if (zrl_is_locked(&dnh->dnh_zrlock))
1208 			zrl_exit(&dnh->dnh_zrlock);
1209 		else
1210 			zrl_remove(&dnh->dnh_zrlock);
1211 	}
1212 }
1213 
1214 static int
dnode_slots_tryenter(dnode_children_t * children,int idx,int slots)1215 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1216 {
1217 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1218 
1219 	for (int i = idx; i < idx + slots; i++) {
1220 		dnode_handle_t *dnh = &children->dnc_children[i];
1221 
1222 		if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1223 			for (int j = idx; j < i; j++) {
1224 				dnh = &children->dnc_children[j];
1225 				zrl_exit(&dnh->dnh_zrlock);
1226 			}
1227 
1228 			return (0);
1229 		}
1230 	}
1231 
1232 	return (1);
1233 }
1234 
1235 static void
dnode_set_slots(dnode_children_t * children,int idx,int slots,void * ptr)1236 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1237 {
1238 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1239 
1240 	for (int i = idx; i < idx + slots; i++) {
1241 		dnode_handle_t *dnh = &children->dnc_children[i];
1242 		dnh->dnh_dnode = ptr;
1243 	}
1244 }
1245 
1246 static boolean_t
dnode_check_slots_free(dnode_children_t * children,int idx,int slots)1247 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1248 {
1249 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1250 
1251 	/*
1252 	 * If all dnode slots are either already free or
1253 	 * evictable return B_TRUE.
1254 	 */
1255 	for (int i = idx; i < idx + slots; i++) {
1256 		dnode_handle_t *dnh = &children->dnc_children[i];
1257 		dnode_t *dn = dnh->dnh_dnode;
1258 
1259 		if (dn == DN_SLOT_FREE) {
1260 			continue;
1261 		} else if (DN_SLOT_IS_PTR(dn)) {
1262 			mutex_enter(&dn->dn_mtx);
1263 			boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1264 			    dn->dn_dirtycnt == 0 &&
1265 			    zfs_refcount_is_zero(&dn->dn_holds));
1266 			mutex_exit(&dn->dn_mtx);
1267 
1268 			if (!can_free)
1269 				return (B_FALSE);
1270 			else
1271 				continue;
1272 		} else {
1273 			return (B_FALSE);
1274 		}
1275 	}
1276 
1277 	return (B_TRUE);
1278 }
1279 
1280 static uint_t
dnode_reclaim_slots(dnode_children_t * children,int idx,int slots)1281 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1282 {
1283 	uint_t reclaimed = 0;
1284 
1285 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1286 
1287 	for (int i = idx; i < idx + slots; i++) {
1288 		dnode_handle_t *dnh = &children->dnc_children[i];
1289 
1290 		ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1291 
1292 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1293 			ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1294 			dnode_destroy(dnh->dnh_dnode);
1295 			dnh->dnh_dnode = DN_SLOT_FREE;
1296 			reclaimed++;
1297 		}
1298 	}
1299 
1300 	return (reclaimed);
1301 }
1302 
1303 void
dnode_free_interior_slots(dnode_t * dn)1304 dnode_free_interior_slots(dnode_t *dn)
1305 {
1306 	dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1307 	int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1308 	int idx = (dn->dn_object & (epb - 1)) + 1;
1309 	int slots = dn->dn_num_slots - 1;
1310 
1311 	if (slots == 0)
1312 		return;
1313 
1314 	ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1315 
1316 	while (!dnode_slots_tryenter(children, idx, slots)) {
1317 		DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1318 		kpreempt(KPREEMPT_SYNC);
1319 	}
1320 
1321 	dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1322 	dnode_slots_rele(children, idx, slots);
1323 }
1324 
1325 void
dnode_special_close(dnode_handle_t * dnh)1326 dnode_special_close(dnode_handle_t *dnh)
1327 {
1328 	dnode_t *dn = dnh->dnh_dnode;
1329 
1330 	/*
1331 	 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final
1332 	 * zfs_refcount_remove()
1333 	 */
1334 	mutex_enter(&dn->dn_mtx);
1335 	if (zfs_refcount_count(&dn->dn_holds) > 0)
1336 		cv_wait(&dn->dn_nodnholds, &dn->dn_mtx);
1337 	mutex_exit(&dn->dn_mtx);
1338 	ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0);
1339 
1340 	ASSERT(dn->dn_dbuf == NULL ||
1341 	    dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1342 	zrl_add(&dnh->dnh_zrlock);
1343 	dnode_destroy(dn); /* implicit zrl_remove() */
1344 	zrl_destroy(&dnh->dnh_zrlock);
1345 	dnh->dnh_dnode = NULL;
1346 }
1347 
1348 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1349 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1350     dnode_handle_t *dnh)
1351 {
1352 	dnode_t *dn;
1353 
1354 	zrl_init(&dnh->dnh_zrlock);
1355 	VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock));
1356 
1357 	dn = dnode_create(os, dnp, NULL, object, dnh);
1358 	DNODE_VERIFY(dn);
1359 
1360 	zrl_exit(&dnh->dnh_zrlock);
1361 }
1362 
1363 static void
dnode_buf_evict_async(void * dbu)1364 dnode_buf_evict_async(void *dbu)
1365 {
1366 	dnode_children_t *dnc = dbu;
1367 
1368 	DNODE_STAT_BUMP(dnode_buf_evict);
1369 
1370 	for (int i = 0; i < dnc->dnc_count; i++) {
1371 		dnode_handle_t *dnh = &dnc->dnc_children[i];
1372 		dnode_t *dn;
1373 
1374 		/*
1375 		 * The dnode handle lock guards against the dnode moving to
1376 		 * another valid address, so there is no need here to guard
1377 		 * against changes to or from NULL.
1378 		 */
1379 		if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1380 			zrl_destroy(&dnh->dnh_zrlock);
1381 			dnh->dnh_dnode = DN_SLOT_UNINIT;
1382 			continue;
1383 		}
1384 
1385 		zrl_add(&dnh->dnh_zrlock);
1386 		dn = dnh->dnh_dnode;
1387 		/*
1388 		 * If there are holds on this dnode, then there should
1389 		 * be holds on the dnode's containing dbuf as well; thus
1390 		 * it wouldn't be eligible for eviction and this function
1391 		 * would not have been called.
1392 		 */
1393 		ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1394 		ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1395 
1396 		dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1397 		zrl_destroy(&dnh->dnh_zrlock);
1398 		dnh->dnh_dnode = DN_SLOT_UNINIT;
1399 	}
1400 	kmem_free(dnc, sizeof (dnode_children_t) +
1401 	    dnc->dnc_count * sizeof (dnode_handle_t));
1402 }
1403 
1404 /*
1405  * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1406  * to ensure the hole at the specified object offset is large enough to
1407  * hold the dnode being created. The slots parameter is also used to ensure
1408  * a dnode does not span multiple dnode blocks. In both of these cases, if
1409  * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1410  * are only possible when using DNODE_MUST_BE_FREE.
1411  *
1412  * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1413  * dnode_hold_impl() will check if the requested dnode is already consumed
1414  * as an extra dnode slot by an large dnode, in which case it returns
1415  * ENOENT.
1416  *
1417  * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just
1418  * return whether the hold would succeed or not. tag and dnp should set to
1419  * NULL in this case.
1420  *
1421  * errors:
1422  * EINVAL - Invalid object number or flags.
1423  * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1424  * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1425  *        - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1426  *        - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1427  * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1428  *        - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1429  * EIO    - I/O error when reading the meta dnode dbuf.
1430  *
1431  * succeeds even for free dnodes.
1432  */
1433 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,int slots,const void * tag,dnode_t ** dnp)1434 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1435     const void *tag, dnode_t **dnp)
1436 {
1437 	int epb, idx, err;
1438 	int drop_struct_lock = FALSE;
1439 	int type;
1440 	uint64_t blk;
1441 	dnode_t *mdn, *dn;
1442 	dmu_buf_impl_t *db;
1443 	dnode_children_t *dnc;
1444 	dnode_phys_t *dn_block;
1445 	dnode_handle_t *dnh;
1446 
1447 	ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1448 	ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1449 	IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL));
1450 
1451 	/*
1452 	 * If you are holding the spa config lock as writer, you shouldn't
1453 	 * be asking the DMU to do *anything* unless it's the root pool
1454 	 * which may require us to read from the root filesystem while
1455 	 * holding some (not all) of the locks as writer.
1456 	 */
1457 	ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1458 	    (spa_is_root(os->os_spa) &&
1459 	    spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1460 
1461 	ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1462 
1463 	if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1464 	    object == DMU_PROJECTUSED_OBJECT) {
1465 		if (object == DMU_USERUSED_OBJECT)
1466 			dn = DMU_USERUSED_DNODE(os);
1467 		else if (object == DMU_GROUPUSED_OBJECT)
1468 			dn = DMU_GROUPUSED_DNODE(os);
1469 		else
1470 			dn = DMU_PROJECTUSED_DNODE(os);
1471 		if (dn == NULL)
1472 			return (SET_ERROR(ENOENT));
1473 		type = dn->dn_type;
1474 		if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1475 			return (SET_ERROR(ENOENT));
1476 		if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1477 			return (SET_ERROR(EEXIST));
1478 		DNODE_VERIFY(dn);
1479 		/* Don't actually hold if dry run, just return 0 */
1480 		if (!(flag & DNODE_DRY_RUN)) {
1481 			(void) zfs_refcount_add(&dn->dn_holds, tag);
1482 			*dnp = dn;
1483 		}
1484 		return (0);
1485 	}
1486 
1487 	if (object == 0 || object >= DN_MAX_OBJECT)
1488 		return (SET_ERROR(EINVAL));
1489 
1490 	mdn = DMU_META_DNODE(os);
1491 	ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1492 
1493 	DNODE_VERIFY(mdn);
1494 
1495 	if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1496 		rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1497 		drop_struct_lock = TRUE;
1498 	}
1499 
1500 	blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1501 	db = dbuf_hold(mdn, blk, FTAG);
1502 	if (drop_struct_lock)
1503 		rw_exit(&mdn->dn_struct_rwlock);
1504 	if (db == NULL) {
1505 		DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1506 		return (SET_ERROR(EIO));
1507 	}
1508 
1509 	/*
1510 	 * We do not need to decrypt to read the dnode so it doesn't matter
1511 	 * if we get the encrypted or decrypted version.
1512 	 */
1513 	err = dbuf_read(db, NULL, DB_RF_CANFAIL |
1514 	    DMU_READ_NO_PREFETCH | DMU_READ_NO_DECRYPT);
1515 	if (err) {
1516 		DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1517 		dbuf_rele(db, FTAG);
1518 		return (err);
1519 	}
1520 
1521 	ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1522 	epb = db->db.db_size >> DNODE_SHIFT;
1523 
1524 	idx = object & (epb - 1);
1525 	dn_block = (dnode_phys_t *)db->db.db_data;
1526 
1527 	ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1528 	dnc = dmu_buf_get_user(&db->db);
1529 	dnh = NULL;
1530 	if (dnc == NULL) {
1531 		dnode_children_t *winner;
1532 		int skip = 0;
1533 
1534 		dnc = kmem_zalloc(sizeof (dnode_children_t) +
1535 		    epb * sizeof (dnode_handle_t), KM_SLEEP);
1536 		dnc->dnc_count = epb;
1537 		dnh = &dnc->dnc_children[0];
1538 
1539 		/* Initialize dnode slot status from dnode_phys_t */
1540 		for (int i = 0; i < epb; i++) {
1541 			zrl_init(&dnh[i].dnh_zrlock);
1542 
1543 			if (skip) {
1544 				skip--;
1545 				continue;
1546 			}
1547 
1548 			if (dn_block[i].dn_type != DMU_OT_NONE) {
1549 				int interior = dn_block[i].dn_extra_slots;
1550 
1551 				dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1552 				dnode_set_slots(dnc, i + 1, interior,
1553 				    DN_SLOT_INTERIOR);
1554 				skip = interior;
1555 			} else {
1556 				dnh[i].dnh_dnode = DN_SLOT_FREE;
1557 				skip = 0;
1558 			}
1559 		}
1560 
1561 		dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1562 		    dnode_buf_evict_async, NULL);
1563 		winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1564 		if (winner != NULL) {
1565 
1566 			for (int i = 0; i < epb; i++)
1567 				zrl_destroy(&dnh[i].dnh_zrlock);
1568 
1569 			kmem_free(dnc, sizeof (dnode_children_t) +
1570 			    epb * sizeof (dnode_handle_t));
1571 			dnc = winner;
1572 		}
1573 	}
1574 
1575 	ASSERT(dnc->dnc_count == epb);
1576 
1577 	if (flag & DNODE_MUST_BE_ALLOCATED) {
1578 		slots = 1;
1579 
1580 		dnode_slots_hold(dnc, idx, slots);
1581 		dnh = &dnc->dnc_children[idx];
1582 
1583 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1584 			dn = dnh->dnh_dnode;
1585 		} else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1586 			DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1587 			dnode_slots_rele(dnc, idx, slots);
1588 			dbuf_rele(db, FTAG);
1589 			return (SET_ERROR(EEXIST));
1590 		} else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1591 			DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1592 			dnode_slots_rele(dnc, idx, slots);
1593 			dbuf_rele(db, FTAG);
1594 			return (SET_ERROR(ENOENT));
1595 		} else {
1596 			dnode_slots_rele(dnc, idx, slots);
1597 			while (!dnode_slots_tryenter(dnc, idx, slots)) {
1598 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1599 				kpreempt(KPREEMPT_SYNC);
1600 			}
1601 
1602 			/*
1603 			 * Someone else won the race and called dnode_create()
1604 			 * after we checked DN_SLOT_IS_PTR() above but before
1605 			 * we acquired the lock.
1606 			 */
1607 			if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1608 				DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1609 				dn = dnh->dnh_dnode;
1610 			} else {
1611 				dn = dnode_create(os, dn_block + idx, db,
1612 				    object, dnh);
1613 				dmu_buf_add_user_size(&db->db,
1614 				    sizeof (dnode_t));
1615 			}
1616 		}
1617 
1618 		mutex_enter(&dn->dn_mtx);
1619 		if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1620 			DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1621 			mutex_exit(&dn->dn_mtx);
1622 			dnode_slots_rele(dnc, idx, slots);
1623 			dbuf_rele(db, FTAG);
1624 			return (SET_ERROR(ENOENT));
1625 		}
1626 
1627 		/* Don't actually hold if dry run, just return 0 */
1628 		if (flag & DNODE_DRY_RUN) {
1629 			mutex_exit(&dn->dn_mtx);
1630 			dnode_slots_rele(dnc, idx, slots);
1631 			dbuf_rele(db, FTAG);
1632 			return (0);
1633 		}
1634 
1635 		DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1636 	} else if (flag & DNODE_MUST_BE_FREE) {
1637 
1638 		if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1639 			DNODE_STAT_BUMP(dnode_hold_free_overflow);
1640 			dbuf_rele(db, FTAG);
1641 			return (SET_ERROR(ENOSPC));
1642 		}
1643 
1644 		dnode_slots_hold(dnc, idx, slots);
1645 
1646 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1647 			DNODE_STAT_BUMP(dnode_hold_free_misses);
1648 			dnode_slots_rele(dnc, idx, slots);
1649 			dbuf_rele(db, FTAG);
1650 			return (SET_ERROR(ENOSPC));
1651 		}
1652 
1653 		dnode_slots_rele(dnc, idx, slots);
1654 		while (!dnode_slots_tryenter(dnc, idx, slots)) {
1655 			DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1656 			kpreempt(KPREEMPT_SYNC);
1657 		}
1658 
1659 		if (!dnode_check_slots_free(dnc, idx, slots)) {
1660 			DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1661 			dnode_slots_rele(dnc, idx, slots);
1662 			dbuf_rele(db, FTAG);
1663 			return (SET_ERROR(ENOSPC));
1664 		}
1665 
1666 		/*
1667 		 * Allocated but otherwise free dnodes which would
1668 		 * be in the interior of a multi-slot dnodes need
1669 		 * to be freed.  Single slot dnodes can be safely
1670 		 * re-purposed as a performance optimization.
1671 		 */
1672 		if (slots > 1) {
1673 			uint_t reclaimed =
1674 			    dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1675 			if (reclaimed > 0)
1676 				dmu_buf_sub_user_size(&db->db,
1677 				    reclaimed * sizeof (dnode_t));
1678 		}
1679 
1680 		dnh = &dnc->dnc_children[idx];
1681 		if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1682 			dn = dnh->dnh_dnode;
1683 		} else {
1684 			dn = dnode_create(os, dn_block + idx, db,
1685 			    object, dnh);
1686 			dmu_buf_add_user_size(&db->db, sizeof (dnode_t));
1687 		}
1688 
1689 		mutex_enter(&dn->dn_mtx);
1690 		if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1691 			DNODE_STAT_BUMP(dnode_hold_free_refcount);
1692 			mutex_exit(&dn->dn_mtx);
1693 			dnode_slots_rele(dnc, idx, slots);
1694 			dbuf_rele(db, FTAG);
1695 			return (SET_ERROR(EEXIST));
1696 		}
1697 
1698 		/* Don't actually hold if dry run, just return 0 */
1699 		if (flag & DNODE_DRY_RUN) {
1700 			mutex_exit(&dn->dn_mtx);
1701 			dnode_slots_rele(dnc, idx, slots);
1702 			dbuf_rele(db, FTAG);
1703 			return (0);
1704 		}
1705 
1706 		dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1707 		DNODE_STAT_BUMP(dnode_hold_free_hits);
1708 	} else {
1709 		dbuf_rele(db, FTAG);
1710 		return (SET_ERROR(EINVAL));
1711 	}
1712 
1713 	ASSERT0(dn->dn_free_txg);
1714 
1715 	if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1716 		dbuf_add_ref(db, dnh);
1717 
1718 	mutex_exit(&dn->dn_mtx);
1719 
1720 	/* Now we can rely on the hold to prevent the dnode from moving. */
1721 	dnode_slots_rele(dnc, idx, slots);
1722 
1723 	DNODE_VERIFY(dn);
1724 	ASSERT3P(dnp, !=, NULL);
1725 	ASSERT3P(dn->dn_dbuf, ==, db);
1726 	ASSERT3U(dn->dn_object, ==, object);
1727 	dbuf_rele(db, FTAG);
1728 
1729 	*dnp = dn;
1730 	return (0);
1731 }
1732 
1733 /*
1734  * Return held dnode if the object is allocated, NULL if not.
1735  */
1736 int
dnode_hold(objset_t * os,uint64_t object,const void * tag,dnode_t ** dnp)1737 dnode_hold(objset_t *os, uint64_t object, const void *tag, dnode_t **dnp)
1738 {
1739 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1740 	    dnp));
1741 }
1742 
1743 /*
1744  * Can only add a reference if there is already at least one
1745  * reference on the dnode.  Returns FALSE if unable to add a
1746  * new reference.
1747  */
1748 static boolean_t
dnode_add_ref_locked(dnode_t * dn,const void * tag)1749 dnode_add_ref_locked(dnode_t *dn, const void *tag)
1750 {
1751 	ASSERT(MUTEX_HELD(&dn->dn_mtx));
1752 	if (zfs_refcount_is_zero(&dn->dn_holds))
1753 		return (FALSE);
1754 	VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1755 	return (TRUE);
1756 }
1757 
1758 boolean_t
dnode_add_ref(dnode_t * dn,const void * tag)1759 dnode_add_ref(dnode_t *dn, const void *tag)
1760 {
1761 	mutex_enter(&dn->dn_mtx);
1762 	boolean_t r = dnode_add_ref_locked(dn, tag);
1763 	mutex_exit(&dn->dn_mtx);
1764 	return (r);
1765 }
1766 
1767 void
dnode_rele(dnode_t * dn,const void * tag)1768 dnode_rele(dnode_t *dn, const void *tag)
1769 {
1770 	mutex_enter(&dn->dn_mtx);
1771 	dnode_rele_and_unlock(dn, tag, B_FALSE);
1772 }
1773 
1774 void
dnode_rele_and_unlock(dnode_t * dn,const void * tag,boolean_t evicting)1775 dnode_rele_and_unlock(dnode_t *dn, const void *tag, boolean_t evicting)
1776 {
1777 	uint64_t refs;
1778 	/* Get while the hold prevents the dnode from moving. */
1779 	dmu_buf_impl_t *db = dn->dn_dbuf;
1780 	dnode_handle_t *dnh = dn->dn_handle;
1781 
1782 	refs = zfs_refcount_remove(&dn->dn_holds, tag);
1783 	if (refs == 0)
1784 		cv_broadcast(&dn->dn_nodnholds);
1785 	mutex_exit(&dn->dn_mtx);
1786 	/* dnode could get destroyed at this point, so don't use it anymore */
1787 
1788 	/*
1789 	 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1790 	 * indirectly by dbuf_rele() while relying on the dnode handle to
1791 	 * prevent the dnode from moving, since releasing the last hold could
1792 	 * result in the dnode's parent dbuf evicting its dnode handles. For
1793 	 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1794 	 * other direct or indirect hold on the dnode must first drop the dnode
1795 	 * handle.
1796 	 */
1797 #ifdef ZFS_DEBUG
1798 	ASSERT(refs > 0 || zrl_owner(&dnh->dnh_zrlock) != curthread);
1799 #endif
1800 
1801 	/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1802 	if (refs == 0 && db != NULL) {
1803 		/*
1804 		 * Another thread could add a hold to the dnode handle in
1805 		 * dnode_hold_impl() while holding the parent dbuf. Since the
1806 		 * hold on the parent dbuf prevents the handle from being
1807 		 * destroyed, the hold on the handle is OK. We can't yet assert
1808 		 * that the handle has zero references, but that will be
1809 		 * asserted anyway when the handle gets destroyed.
1810 		 */
1811 		mutex_enter(&db->db_mtx);
1812 		dbuf_rele_and_unlock(db, dnh, evicting);
1813 	}
1814 }
1815 
1816 /*
1817  * Test whether we can create a dnode at the specified location.
1818  */
1819 int
dnode_try_claim(objset_t * os,uint64_t object,int slots)1820 dnode_try_claim(objset_t *os, uint64_t object, int slots)
1821 {
1822 	return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN,
1823 	    slots, NULL, NULL));
1824 }
1825 
1826 /*
1827  * Test if the dnode is dirty, or carrying uncommitted records.
1828  *
1829  * dn_dirtycnt is the number of txgs this dnode is dirty on. It's incremented
1830  * in dnode_setdirty() the first time the dnode is dirtied on a txg, and
1831  * decremented in either dnode_rele_task() or userquota_updates_task() when the
1832  * txg is synced out.
1833  */
1834 boolean_t
dnode_is_dirty(dnode_t * dn)1835 dnode_is_dirty(dnode_t *dn)
1836 {
1837 	mutex_enter(&dn->dn_mtx);
1838 	boolean_t dirty = (dn->dn_dirtycnt != 0);
1839 	mutex_exit(&dn->dn_mtx);
1840 	return (dirty);
1841 }
1842 
1843 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1844 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1845 {
1846 	objset_t *os = dn->dn_objset;
1847 	uint64_t txg = tx->tx_txg;
1848 
1849 	if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1850 		dsl_dataset_dirty(os->os_dsl_dataset, tx);
1851 		return;
1852 	}
1853 
1854 	DNODE_VERIFY(dn);
1855 
1856 #ifdef ZFS_DEBUG
1857 	mutex_enter(&dn->dn_mtx);
1858 	ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1859 	ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1860 	mutex_exit(&dn->dn_mtx);
1861 #endif
1862 
1863 	/*
1864 	 * Determine old uid/gid when necessary
1865 	 */
1866 	dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1867 
1868 	multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK];
1869 	multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1870 
1871 	/*
1872 	 * If we are already marked dirty, we're done.
1873 	 */
1874 	if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1875 		multilist_sublist_unlock(mls);
1876 		return;
1877 	}
1878 
1879 	ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1880 	    !avl_is_empty(&dn->dn_dbufs));
1881 	ASSERT(dn->dn_datablksz != 0);
1882 	ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1883 	ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1884 	ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1885 
1886 	dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1887 	    (u_longlong_t)dn->dn_object, (u_longlong_t)txg);
1888 
1889 	multilist_sublist_insert_head(mls, dn);
1890 
1891 	multilist_sublist_unlock(mls);
1892 
1893 	/*
1894 	 * The dnode maintains a hold on its containing dbuf as
1895 	 * long as there are holds on it.  Each instantiated child
1896 	 * dbuf maintains a hold on the dnode.  When the last child
1897 	 * drops its hold, the dnode will drop its hold on the
1898 	 * containing dbuf. We add a "dirty hold" here so that the
1899 	 * dnode will hang around after we finish processing its
1900 	 * children.
1901 	 */
1902 	mutex_enter(&dn->dn_mtx);
1903 	VERIFY(dnode_add_ref_locked(dn, (void *)(uintptr_t)tx->tx_txg));
1904 	dn->dn_dirtycnt++;
1905 	ASSERT3U(dn->dn_dirtycnt, <=, 3);
1906 	mutex_exit(&dn->dn_mtx);
1907 
1908 	(void) dbuf_dirty(dn->dn_dbuf, tx);
1909 
1910 	dsl_dataset_dirty(os->os_dsl_dataset, tx);
1911 }
1912 
1913 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1914 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1915 {
1916 	mutex_enter(&dn->dn_mtx);
1917 	if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1918 		mutex_exit(&dn->dn_mtx);
1919 		return;
1920 	}
1921 	dn->dn_free_txg = tx->tx_txg;
1922 	mutex_exit(&dn->dn_mtx);
1923 
1924 	dnode_setdirty(dn, tx);
1925 }
1926 
1927 /*
1928  * Try to change the block size for the indicated dnode.  This can only
1929  * succeed if there are no blocks allocated or dirty beyond first block
1930  */
1931 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1932 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1933 {
1934 	dmu_buf_impl_t *db;
1935 	int err;
1936 
1937 	ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1938 	if (size == 0)
1939 		size = SPA_MINBLOCKSIZE;
1940 	else
1941 		size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1942 
1943 	if (ibs == dn->dn_indblkshift)
1944 		ibs = 0;
1945 
1946 	if (size == dn->dn_datablksz && ibs == 0)
1947 		return (0);
1948 
1949 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1950 
1951 	/* Check for any allocated blocks beyond the first */
1952 	if (dn->dn_maxblkid != 0)
1953 		goto fail;
1954 
1955 	mutex_enter(&dn->dn_dbufs_mtx);
1956 	for (db = avl_first(&dn->dn_dbufs); db != NULL;
1957 	    db = AVL_NEXT(&dn->dn_dbufs, db)) {
1958 		if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1959 		    db->db_blkid != DMU_SPILL_BLKID) {
1960 			mutex_exit(&dn->dn_dbufs_mtx);
1961 			goto fail;
1962 		}
1963 	}
1964 	mutex_exit(&dn->dn_dbufs_mtx);
1965 
1966 	if (ibs && dn->dn_nlevels != 1)
1967 		goto fail;
1968 
1969 	dnode_setdirty(dn, tx);
1970 	if (size != dn->dn_datablksz) {
1971 		/* resize the old block */
1972 		err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1973 		if (err == 0) {
1974 			dbuf_new_size(db, size, tx);
1975 		} else if (err != ENOENT) {
1976 			goto fail;
1977 		}
1978 
1979 		dnode_setdblksz(dn, size);
1980 		dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = size;
1981 		if (db)
1982 			dbuf_rele(db, FTAG);
1983 	}
1984 	if (ibs) {
1985 		dn->dn_indblkshift = ibs;
1986 		dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
1987 	}
1988 
1989 	rw_exit(&dn->dn_struct_rwlock);
1990 	return (0);
1991 
1992 fail:
1993 	rw_exit(&dn->dn_struct_rwlock);
1994 	return (SET_ERROR(ENOTSUP));
1995 }
1996 
1997 static void
dnode_set_nlevels_impl(dnode_t * dn,int new_nlevels,dmu_tx_t * tx)1998 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1999 {
2000 	uint64_t txgoff = tx->tx_txg & TXG_MASK;
2001 	int old_nlevels = dn->dn_nlevels;
2002 	dmu_buf_impl_t *db;
2003 	list_t *list;
2004 	dbuf_dirty_record_t *new, *dr, *dr_next;
2005 
2006 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2007 
2008 	ASSERT3U(new_nlevels, >, dn->dn_nlevels);
2009 	dn->dn_nlevels = new_nlevels;
2010 
2011 	ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
2012 	dn->dn_next_nlevels[txgoff] = new_nlevels;
2013 
2014 	/* dirty the left indirects */
2015 	db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
2016 	ASSERT(db != NULL);
2017 	new = dbuf_dirty(db, tx);
2018 	dbuf_rele(db, FTAG);
2019 
2020 	/* transfer the dirty records to the new indirect */
2021 	mutex_enter(&dn->dn_mtx);
2022 	mutex_enter(&new->dt.di.dr_mtx);
2023 	list = &dn->dn_dirty_records[txgoff];
2024 	for (dr = list_head(list); dr; dr = dr_next) {
2025 		dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
2026 
2027 		IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1);
2028 		if (dr->dr_dbuf == NULL ||
2029 		    (dr->dr_dbuf->db_level == old_nlevels - 1 &&
2030 		    dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2031 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) {
2032 			list_remove(&dn->dn_dirty_records[txgoff], dr);
2033 			list_insert_tail(&new->dt.di.dr_children, dr);
2034 			dr->dr_parent = new;
2035 		}
2036 	}
2037 	mutex_exit(&new->dt.di.dr_mtx);
2038 	mutex_exit(&dn->dn_mtx);
2039 }
2040 
2041 int
dnode_set_nlevels(dnode_t * dn,int nlevels,dmu_tx_t * tx)2042 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
2043 {
2044 	int ret = 0;
2045 
2046 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2047 
2048 	if (dn->dn_nlevels == nlevels) {
2049 		ret = 0;
2050 		goto out;
2051 	} else if (nlevels < dn->dn_nlevels) {
2052 		ret = SET_ERROR(EINVAL);
2053 		goto out;
2054 	}
2055 
2056 	dnode_set_nlevels_impl(dn, nlevels, tx);
2057 
2058 out:
2059 	rw_exit(&dn->dn_struct_rwlock);
2060 	return (ret);
2061 }
2062 
2063 /* read-holding callers must not rely on the lock being continuously held */
2064 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read,boolean_t force)2065 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
2066     boolean_t force)
2067 {
2068 	int epbs, new_nlevels;
2069 	uint64_t sz;
2070 
2071 	ASSERT(blkid != DMU_BONUS_BLKID);
2072 
2073 	ASSERT(have_read ?
2074 	    RW_READ_HELD(&dn->dn_struct_rwlock) :
2075 	    RW_WRITE_HELD(&dn->dn_struct_rwlock));
2076 
2077 	/*
2078 	 * if we have a read-lock, check to see if we need to do any work
2079 	 * before upgrading to a write-lock.
2080 	 */
2081 	if (have_read) {
2082 		if (blkid <= dn->dn_maxblkid)
2083 			return;
2084 
2085 		if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
2086 			rw_exit(&dn->dn_struct_rwlock);
2087 			rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2088 		}
2089 	}
2090 
2091 	/*
2092 	 * Raw sends (indicated by the force flag) require that we take the
2093 	 * given blkid even if the value is lower than the current value.
2094 	 */
2095 	if (!force && blkid <= dn->dn_maxblkid)
2096 		goto out;
2097 
2098 	/*
2099 	 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
2100 	 * to indicate that this field is set. This allows us to set the
2101 	 * maxblkid to 0 on an existing object in dnode_sync().
2102 	 */
2103 	dn->dn_maxblkid = blkid;
2104 	dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
2105 	    blkid | DMU_NEXT_MAXBLKID_SET;
2106 
2107 	/*
2108 	 * Compute the number of levels necessary to support the new maxblkid.
2109 	 * Raw sends will ensure nlevels is set correctly for us.
2110 	 */
2111 	new_nlevels = 1;
2112 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2113 	for (sz = dn->dn_nblkptr;
2114 	    sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
2115 		new_nlevels++;
2116 
2117 	ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
2118 
2119 	if (!force) {
2120 		if (new_nlevels > dn->dn_nlevels)
2121 			dnode_set_nlevels_impl(dn, new_nlevels, tx);
2122 	} else {
2123 		ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
2124 	}
2125 
2126 out:
2127 	if (have_read)
2128 		rw_downgrade(&dn->dn_struct_rwlock);
2129 }
2130 
2131 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)2132 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
2133 {
2134 	dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
2135 	if (db != NULL) {
2136 		dmu_buf_will_dirty(&db->db, tx);
2137 		dbuf_rele(db, FTAG);
2138 	}
2139 }
2140 
2141 /*
2142  * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
2143  * and end_blkid.
2144  */
2145 static void
dnode_dirty_l1range(dnode_t * dn,uint64_t start_blkid,uint64_t end_blkid,dmu_tx_t * tx)2146 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
2147     dmu_tx_t *tx)
2148 {
2149 	dmu_buf_impl_t *db_search;
2150 	dmu_buf_impl_t *db;
2151 	avl_index_t where;
2152 
2153 	db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
2154 
2155 	mutex_enter(&dn->dn_dbufs_mtx);
2156 
2157 	db_search->db_level = 1;
2158 	db_search->db_blkid = start_blkid + 1;
2159 	db_search->db_state = DB_SEARCH;
2160 	for (;;) {
2161 
2162 		db = avl_find(&dn->dn_dbufs, db_search, &where);
2163 		if (db == NULL)
2164 			db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2165 
2166 		if (db == NULL || db->db_level != 1 ||
2167 		    db->db_blkid >= end_blkid) {
2168 			break;
2169 		}
2170 
2171 		/*
2172 		 * Setup the next blkid we want to search for.
2173 		 */
2174 		db_search->db_blkid = db->db_blkid + 1;
2175 		ASSERT3U(db->db_blkid, >=, start_blkid);
2176 
2177 		/*
2178 		 * If the dbuf transitions to DB_EVICTING while we're trying
2179 		 * to dirty it, then we will be unable to discover it in
2180 		 * the dbuf hash table. This will result in a call to
2181 		 * dbuf_create() which needs to acquire the dn_dbufs_mtx
2182 		 * lock. To avoid a deadlock, we drop the lock before
2183 		 * dirtying the level-1 dbuf.
2184 		 */
2185 		mutex_exit(&dn->dn_dbufs_mtx);
2186 		dnode_dirty_l1(dn, db->db_blkid, tx);
2187 		mutex_enter(&dn->dn_dbufs_mtx);
2188 	}
2189 
2190 #ifdef ZFS_DEBUG
2191 	/*
2192 	 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
2193 	 */
2194 	db_search->db_level = 1;
2195 	db_search->db_blkid = start_blkid + 1;
2196 	db_search->db_state = DB_SEARCH;
2197 	db = avl_find(&dn->dn_dbufs, db_search, &where);
2198 	if (db == NULL)
2199 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
2200 	for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
2201 		if (db->db_level != 1 || db->db_blkid >= end_blkid)
2202 			break;
2203 		if (db->db_state != DB_EVICTING)
2204 			ASSERT(db->db_dirtycnt > 0);
2205 	}
2206 #endif
2207 	kmem_free(db_search, sizeof (dmu_buf_impl_t));
2208 	mutex_exit(&dn->dn_dbufs_mtx);
2209 }
2210 
2211 static void
dnode_partial_zero(dnode_t * dn,uint64_t off,uint64_t blkoff,uint64_t len,dmu_tx_t * tx)2212 dnode_partial_zero(dnode_t *dn, uint64_t off, uint64_t blkoff, uint64_t len,
2213     dmu_tx_t *tx)
2214 {
2215 	dmu_buf_impl_t *db;
2216 	int res;
2217 
2218 	rw_enter(&dn->dn_struct_rwlock, RW_READER);
2219 	res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), TRUE, FALSE,
2220 	    FTAG, &db);
2221 	rw_exit(&dn->dn_struct_rwlock);
2222 	if (res == 0) {
2223 		db_lock_type_t dblt;
2224 		boolean_t dirty;
2225 
2226 		dblt = dmu_buf_lock_parent(db, RW_READER, FTAG);
2227 		/* don't dirty if not on disk and not dirty */
2228 		dirty = !list_is_empty(&db->db_dirty_records) ||
2229 		    (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr));
2230 		dmu_buf_unlock_parent(db, dblt, FTAG);
2231 		if (dirty) {
2232 			caddr_t data;
2233 
2234 			dmu_buf_will_dirty(&db->db, tx);
2235 			data = db->db.db_data;
2236 			memset(data + blkoff, 0, len);
2237 		}
2238 		dbuf_rele(db, FTAG);
2239 	}
2240 }
2241 
2242 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)2243 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
2244 {
2245 	uint64_t blkoff, blkid, nblks;
2246 	int blksz, blkshift, head, tail;
2247 	int trunc = FALSE;
2248 	int epbs;
2249 
2250 	blksz = dn->dn_datablksz;
2251 	blkshift = dn->dn_datablkshift;
2252 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2253 
2254 	if (len == DMU_OBJECT_END) {
2255 		len = UINT64_MAX - off;
2256 		trunc = TRUE;
2257 	}
2258 
2259 	/*
2260 	 * First, block align the region to free:
2261 	 */
2262 	if (ISP2(blksz)) {
2263 		head = P2NPHASE(off, blksz);
2264 		blkoff = P2PHASE(off, blksz);
2265 		if ((off >> blkshift) > dn->dn_maxblkid)
2266 			return;
2267 	} else {
2268 		ASSERT0(dn->dn_maxblkid);
2269 		if (off == 0 && len >= blksz) {
2270 			/*
2271 			 * Freeing the whole block; fast-track this request.
2272 			 */
2273 			blkid = 0;
2274 			nblks = 1;
2275 			if (dn->dn_nlevels > 1) {
2276 				rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2277 				dnode_dirty_l1(dn, 0, tx);
2278 				rw_exit(&dn->dn_struct_rwlock);
2279 			}
2280 			goto done;
2281 		} else if (off >= blksz) {
2282 			/* Freeing past end-of-data */
2283 			return;
2284 		} else {
2285 			/* Freeing part of the block. */
2286 			head = blksz - off;
2287 			ASSERT3U(head, >, 0);
2288 		}
2289 		blkoff = off;
2290 	}
2291 	/* zero out any partial block data at the start of the range */
2292 	if (head) {
2293 		ASSERT3U(blkoff + head, ==, blksz);
2294 		if (len < head)
2295 			head = len;
2296 		dnode_partial_zero(dn, off, blkoff, head, tx);
2297 		off += head;
2298 		len -= head;
2299 	}
2300 
2301 	/* If the range was less than one block, we're done */
2302 	if (len == 0)
2303 		return;
2304 
2305 	/* If the remaining range is past end of file, we're done */
2306 	if ((off >> blkshift) > dn->dn_maxblkid)
2307 		return;
2308 
2309 	ASSERT(ISP2(blksz));
2310 	if (trunc)
2311 		tail = 0;
2312 	else
2313 		tail = P2PHASE(len, blksz);
2314 
2315 	ASSERT0(P2PHASE(off, blksz));
2316 	/* zero out any partial block data at the end of the range */
2317 	if (tail) {
2318 		if (len < tail)
2319 			tail = len;
2320 		dnode_partial_zero(dn, off + len, 0, tail, tx);
2321 		len -= tail;
2322 	}
2323 
2324 	/* If the range did not include a full block, we are done */
2325 	if (len == 0)
2326 		return;
2327 
2328 	ASSERT(IS_P2ALIGNED(off, blksz));
2329 	ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2330 	blkid = off >> blkshift;
2331 	nblks = len >> blkshift;
2332 	if (trunc)
2333 		nblks += 1;
2334 
2335 	/*
2336 	 * Dirty all the indirect blocks in this range.  Note that only
2337 	 * the first and last indirect blocks can actually be written
2338 	 * (if they were partially freed) -- they must be dirtied, even if
2339 	 * they do not exist on disk yet.  The interior blocks will
2340 	 * be freed by free_children(), so they will not actually be written.
2341 	 * Even though these interior blocks will not be written, we
2342 	 * dirty them for two reasons:
2343 	 *
2344 	 *  - It ensures that the indirect blocks remain in memory until
2345 	 *    syncing context.  (They have already been prefetched by
2346 	 *    dmu_tx_hold_free(), so we don't have to worry about reading
2347 	 *    them serially here.)
2348 	 *
2349 	 *  - The dirty space accounting will put pressure on the txg sync
2350 	 *    mechanism to begin syncing, and to delay transactions if there
2351 	 *    is a large amount of freeing.  Even though these indirect
2352 	 *    blocks will not be written, we could need to write the same
2353 	 *    amount of space if we copy the freed BPs into deadlists.
2354 	 */
2355 	if (dn->dn_nlevels > 1) {
2356 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2357 		uint64_t first, last;
2358 
2359 		first = blkid >> epbs;
2360 		dnode_dirty_l1(dn, first, tx);
2361 		if (trunc)
2362 			last = dn->dn_maxblkid >> epbs;
2363 		else
2364 			last = (blkid + nblks - 1) >> epbs;
2365 		if (last != first)
2366 			dnode_dirty_l1(dn, last, tx);
2367 
2368 		dnode_dirty_l1range(dn, first, last, tx);
2369 
2370 		int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2371 		    SPA_BLKPTRSHIFT;
2372 		for (uint64_t i = first + 1; i < last; i++) {
2373 			/*
2374 			 * Set i to the blockid of the next non-hole
2375 			 * level-1 indirect block at or after i.  Note
2376 			 * that dnode_next_offset() operates in terms of
2377 			 * level-0-equivalent bytes.
2378 			 */
2379 			uint64_t ibyte = i << shift;
2380 			int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2381 			    &ibyte, 2, 1, 0);
2382 			i = ibyte >> shift;
2383 			if (i >= last)
2384 				break;
2385 
2386 			/*
2387 			 * Normally we should not see an error, either
2388 			 * from dnode_next_offset() or dbuf_hold_level()
2389 			 * (except for ESRCH from dnode_next_offset).
2390 			 * If there is an i/o error, then when we read
2391 			 * this block in syncing context, it will use
2392 			 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2393 			 * to the "failmode" property.  dnode_next_offset()
2394 			 * doesn't have a flag to indicate MUSTSUCCEED.
2395 			 */
2396 			if (err != 0)
2397 				break;
2398 
2399 			dnode_dirty_l1(dn, i, tx);
2400 		}
2401 		rw_exit(&dn->dn_struct_rwlock);
2402 	}
2403 
2404 done:
2405 	/*
2406 	 * Add this range to the dnode range list.
2407 	 * We will finish up this free operation in the syncing phase.
2408 	 */
2409 	mutex_enter(&dn->dn_mtx);
2410 	{
2411 		int txgoff = tx->tx_txg & TXG_MASK;
2412 		if (dn->dn_free_ranges[txgoff] == NULL) {
2413 			dn->dn_free_ranges[txgoff] =
2414 			    zfs_range_tree_create_flags(
2415 			    NULL, ZFS_RANGE_SEG64, NULL, 0, 0,
2416 			    ZFS_RT_F_DYN_NAME, rt_name(dn, "dn_free_ranges"));
2417 		}
2418 		zfs_range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2419 		zfs_range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2420 	}
2421 	dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2422 	    (u_longlong_t)blkid, (u_longlong_t)nblks,
2423 	    (u_longlong_t)tx->tx_txg);
2424 	mutex_exit(&dn->dn_mtx);
2425 
2426 	dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2427 	dnode_setdirty(dn, tx);
2428 }
2429 
2430 static boolean_t
dnode_spill_freed(dnode_t * dn)2431 dnode_spill_freed(dnode_t *dn)
2432 {
2433 	int i;
2434 
2435 	mutex_enter(&dn->dn_mtx);
2436 	for (i = 0; i < TXG_SIZE; i++) {
2437 		if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2438 			break;
2439 	}
2440 	mutex_exit(&dn->dn_mtx);
2441 	return (i < TXG_SIZE);
2442 }
2443 
2444 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2445 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)2446 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2447 {
2448 	int i;
2449 
2450 	if (blkid == DMU_BONUS_BLKID)
2451 		return (FALSE);
2452 
2453 	if (dn->dn_free_txg)
2454 		return (TRUE);
2455 
2456 	if (blkid == DMU_SPILL_BLKID)
2457 		return (dnode_spill_freed(dn));
2458 
2459 	mutex_enter(&dn->dn_mtx);
2460 	for (i = 0; i < TXG_SIZE; i++) {
2461 		if (dn->dn_free_ranges[i] != NULL &&
2462 		    zfs_range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2463 			break;
2464 	}
2465 	mutex_exit(&dn->dn_mtx);
2466 	return (i < TXG_SIZE);
2467 }
2468 
2469 /* call from syncing context when we actually write/free space for this dnode */
2470 void
dnode_diduse_space(dnode_t * dn,int64_t delta)2471 dnode_diduse_space(dnode_t *dn, int64_t delta)
2472 {
2473 	uint64_t space;
2474 	dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2475 	    dn, dn->dn_phys,
2476 	    (u_longlong_t)dn->dn_phys->dn_used,
2477 	    (longlong_t)delta);
2478 
2479 	mutex_enter(&dn->dn_mtx);
2480 	space = DN_USED_BYTES(dn->dn_phys);
2481 	if (delta > 0) {
2482 		ASSERT3U(space + delta, >=, space); /* no overflow */
2483 	} else {
2484 		ASSERT3U(space, >=, -delta); /* no underflow */
2485 	}
2486 	space += delta;
2487 	if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2488 		ASSERT0((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES));
2489 		ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2490 		dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2491 	} else {
2492 		dn->dn_phys->dn_used = space;
2493 		dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2494 	}
2495 	mutex_exit(&dn->dn_mtx);
2496 }
2497 
2498 /*
2499  * Scans a block at the indicated "level" looking for a hole or data,
2500  * depending on 'flags'.
2501  *
2502  * If level > 0, then we are scanning an indirect block looking at its
2503  * pointers.  If level == 0, then we are looking at a block of dnodes.
2504  *
2505  * If we don't find what we are looking for in the block, we return ESRCH.
2506  * Otherwise, return with *offset pointing to the beginning (if searching
2507  * forwards) or end (if searching backwards) of the range covered by the
2508  * block pointer we matched on (or dnode).
2509  *
2510  * The basic search algorithm used below by dnode_next_offset() is to
2511  * use this function to search up the block tree (widen the search) until
2512  * we find something (i.e., we don't return ESRCH) and then search back
2513  * down the tree (narrow the search) until we reach our original search
2514  * level.
2515  */
2516 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)2517 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2518     int lvl, uint64_t blkfill, uint64_t txg)
2519 {
2520 	dmu_buf_impl_t *db = NULL;
2521 	void *data = NULL;
2522 	uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2523 	uint64_t epb = 1ULL << epbs;
2524 	uint64_t minfill, maxfill;
2525 	boolean_t hole;
2526 	int i, inc, error, span;
2527 
2528 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2529 
2530 	hole = ((flags & DNODE_FIND_HOLE) != 0);
2531 	inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2532 	ASSERT(txg == 0 || !hole);
2533 
2534 	if (lvl == dn->dn_phys->dn_nlevels) {
2535 		error = 0;
2536 		epb = dn->dn_phys->dn_nblkptr;
2537 		data = dn->dn_phys->dn_blkptr;
2538 		if (dn->dn_dbuf != NULL)
2539 			rw_enter(&dn->dn_dbuf->db_rwlock, RW_READER);
2540 		else if (dmu_objset_ds(dn->dn_objset) != NULL)
2541 			rrw_enter(&dmu_objset_ds(dn->dn_objset)->ds_bp_rwlock,
2542 			    RW_READER, FTAG);
2543 	} else {
2544 		uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2545 		error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2546 		if (error) {
2547 			if (error != ENOENT)
2548 				return (error);
2549 			if (hole)
2550 				return (0);
2551 			/*
2552 			 * This can only happen when we are searching up
2553 			 * the block tree for data.  We don't really need to
2554 			 * adjust the offset, as we will just end up looking
2555 			 * at the pointer to this block in its parent, and its
2556 			 * going to be unallocated, so we will skip over it.
2557 			 */
2558 			return (SET_ERROR(ESRCH));
2559 		}
2560 		error = dbuf_read(db, NULL,
2561 		    DB_RF_CANFAIL | DB_RF_HAVESTRUCT |
2562 		    DMU_READ_NO_PREFETCH | DMU_READ_NO_DECRYPT);
2563 		if (error) {
2564 			dbuf_rele(db, FTAG);
2565 			return (error);
2566 		}
2567 		data = db->db.db_data;
2568 		rw_enter(&db->db_rwlock, RW_READER);
2569 	}
2570 
2571 	if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2572 	    BP_GET_LOGICAL_BIRTH(db->db_blkptr) <= txg ||
2573 	    BP_IS_HOLE(db->db_blkptr))) {
2574 		/*
2575 		 * This can only happen when we are searching up the tree
2576 		 * and these conditions mean that we need to keep climbing.
2577 		 */
2578 		error = SET_ERROR(ESRCH);
2579 	} else if (lvl == 0) {
2580 		dnode_phys_t *dnp = data;
2581 
2582 		ASSERT(dn->dn_type == DMU_OT_DNODE);
2583 		ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2584 
2585 		for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2586 		    i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2587 			if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2588 				break;
2589 		}
2590 
2591 		if (i == blkfill)
2592 			error = SET_ERROR(ESRCH);
2593 
2594 		*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2595 		    (i << DNODE_SHIFT);
2596 	} else {
2597 		blkptr_t *bp = data;
2598 		uint64_t start = *offset;
2599 		span = (lvl - 1) * epbs + dn->dn_datablkshift;
2600 		minfill = 0;
2601 		maxfill = blkfill << ((lvl - 1) * epbs);
2602 
2603 		if (hole)
2604 			maxfill--;
2605 		else
2606 			minfill++;
2607 
2608 		if (span >= 8 * sizeof (*offset)) {
2609 			/* This only happens on the highest indirection level */
2610 			ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2611 			*offset = 0;
2612 		} else {
2613 			*offset = *offset >> span;
2614 		}
2615 
2616 		for (i = BF64_GET(*offset, 0, epbs);
2617 		    i >= 0 && i < epb; i += inc) {
2618 			if (BP_GET_FILL(&bp[i]) >= minfill &&
2619 			    BP_GET_FILL(&bp[i]) <= maxfill &&
2620 			    (hole || BP_GET_LOGICAL_BIRTH(&bp[i]) > txg))
2621 				break;
2622 			if (inc > 0 || *offset > 0)
2623 				*offset += inc;
2624 		}
2625 
2626 		if (span >= 8 * sizeof (*offset)) {
2627 			*offset = start;
2628 		} else {
2629 			*offset = *offset << span;
2630 		}
2631 
2632 		if (inc < 0) {
2633 			/* traversing backwards; position offset at the end */
2634 			if (span < 8 * sizeof (*offset))
2635 				*offset = MIN(*offset + (1ULL << span) - 1,
2636 				    start);
2637 		} else if (*offset < start) {
2638 			*offset = start;
2639 		}
2640 		if (i < 0 || i >= epb)
2641 			error = SET_ERROR(ESRCH);
2642 	}
2643 
2644 	if (db != NULL) {
2645 		rw_exit(&db->db_rwlock);
2646 		dbuf_rele(db, FTAG);
2647 	} else {
2648 		if (dn->dn_dbuf != NULL)
2649 			rw_exit(&dn->dn_dbuf->db_rwlock);
2650 		else if (dmu_objset_ds(dn->dn_objset) != NULL)
2651 			rrw_exit(&dmu_objset_ds(dn->dn_objset)->ds_bp_rwlock,
2652 			    FTAG);
2653 	}
2654 
2655 	return (error);
2656 }
2657 
2658 /*
2659  * Find the next hole, data, or sparse region at or after *offset.
2660  * The value 'blkfill' tells us how many items we expect to find
2661  * in an L0 data block; this value is 1 for normal objects,
2662  * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2663  * DNODES_PER_BLOCK when searching for sparse regions thereof.
2664  *
2665  * Examples:
2666  *
2667  * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2668  *	Finds the next/previous hole/data in a file.
2669  *	Used in dmu_offset_next().
2670  *
2671  * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2672  *	Finds the next free/allocated dnode an objset's meta-dnode.
2673  *	Only finds objects that have new contents since txg (ie.
2674  *	bonus buffer changes and content removal are ignored).
2675  *	Used in dmu_object_next().
2676  *
2677  * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2678  *	Finds the next L2 meta-dnode bp that's at most 1/4 full.
2679  *	Used in dmu_object_alloc().
2680  */
2681 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)2682 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2683     int minlvl, uint64_t blkfill, uint64_t txg)
2684 {
2685 	uint64_t initial_offset = *offset;
2686 	int lvl, maxlvl;
2687 	int error = 0;
2688 
2689 	if (!(flags & DNODE_FIND_HAVELOCK))
2690 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
2691 
2692 	if (dn->dn_phys->dn_nlevels == 0) {
2693 		error = SET_ERROR(ESRCH);
2694 		goto out;
2695 	}
2696 
2697 	if (dn->dn_datablkshift == 0) {
2698 		if (*offset < dn->dn_datablksz) {
2699 			if (flags & DNODE_FIND_HOLE)
2700 				*offset = dn->dn_datablksz;
2701 		} else {
2702 			error = SET_ERROR(ESRCH);
2703 		}
2704 		goto out;
2705 	}
2706 
2707 	maxlvl = dn->dn_phys->dn_nlevels;
2708 
2709 	for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2710 		error = dnode_next_offset_level(dn,
2711 		    flags, offset, lvl, blkfill, txg);
2712 		if (error != ESRCH)
2713 			break;
2714 	}
2715 
2716 	while (error == 0 && --lvl >= minlvl) {
2717 		error = dnode_next_offset_level(dn,
2718 		    flags, offset, lvl, blkfill, txg);
2719 	}
2720 
2721 	/*
2722 	 * There's always a "virtual hole" at the end of the object, even
2723 	 * if all BP's which physically exist are non-holes.
2724 	 */
2725 	if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2726 	    minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2727 		error = 0;
2728 	}
2729 
2730 	if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2731 	    initial_offset < *offset : initial_offset > *offset))
2732 		error = SET_ERROR(ESRCH);
2733 out:
2734 	if (!(flags & DNODE_FIND_HAVELOCK))
2735 		rw_exit(&dn->dn_struct_rwlock);
2736 
2737 	return (error);
2738 }
2739 
2740 #if defined(_KERNEL)
2741 EXPORT_SYMBOL(dnode_hold);
2742 EXPORT_SYMBOL(dnode_rele);
2743 EXPORT_SYMBOL(dnode_set_nlevels);
2744 EXPORT_SYMBOL(dnode_set_blksz);
2745 EXPORT_SYMBOL(dnode_free_range);
2746 EXPORT_SYMBOL(dnode_evict_dbufs);
2747 EXPORT_SYMBOL(dnode_evict_bonus);
2748 #endif
2749 
2750 ZFS_MODULE_PARAM(zfs, zfs_, default_bs, INT, ZMOD_RW,
2751 	"Default dnode block shift");
2752 ZFS_MODULE_PARAM(zfs, zfs_, default_ibs, INT, ZMOD_RW,
2753 	"Default dnode indirect block shift");
2754