]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dnode.c
projects / mirror_zfs.git / blob
? search:


78a90f68fbd3bb4768b4477e487bf60484b2644b
[mirror_zfs.git] / module / zfs / dnode.c
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40 #include <sys/trace_dnode.h>
41 #include <sys/zfs_project.h>
42
43 dnode_stats_t dnode_stats = {
44 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64 },
45 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64 },
46 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64 },
47 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64 },
48 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64 },
49 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64 },
50 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64 },
51 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64 },
52 { "dnode_hold_free_hits", KSTAT_DATA_UINT64 },
53 { "dnode_hold_free_misses", KSTAT_DATA_UINT64 },
54 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64 },
55 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64 },
56 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64 },
57 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64 },
58 { "dnode_hold_free_txg", KSTAT_DATA_UINT64 },
59 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 },
60 { "dnode_allocate", KSTAT_DATA_UINT64 },
61 { "dnode_reallocate", KSTAT_DATA_UINT64 },
62 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
63 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
64 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
65 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
66 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
67 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
68 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
69 { "dnode_move_special", KSTAT_DATA_UINT64 },
70 { "dnode_move_handle", KSTAT_DATA_UINT64 },
71 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
72 { "dnode_move_active", KSTAT_DATA_UINT64 },
73 };
74
75 static kstat_t *dnode_ksp;
76 static kmem_cache_t *dnode_cache;
77
78 ASSERTV(static dnode_phys_t dnode_phys_zero);
79
80 int zfs_default_bs = SPA_MINBLOCKSHIFT;
81 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
82
83 #ifdef _KERNEL
84 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
85 #endif /* _KERNEL */
86
87 static int
88 dbuf_compare(const void *x1, const void *x2)
89 {
90 const dmu_buf_impl_t *d1 = x1;
91 const dmu_buf_impl_t *d2 = x2;
92
93 int cmp = AVL_CMP(d1->db_level, d2->db_level);
94 if (likely(cmp))
95 return (cmp);
96
97 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
98 if (likely(cmp))
99 return (cmp);
100
101 if (d1->db_state == DB_SEARCH) {
102 ASSERT3S(d2->db_state, !=, DB_SEARCH);
103 return (-1);
104 } else if (d2->db_state == DB_SEARCH) {
105 ASSERT3S(d1->db_state, !=, DB_SEARCH);
106 return (1);
107 }
108
109 return (AVL_PCMP(d1, d2));
110 }
111
112 /* ARGSUSED */
113 static int
114 dnode_cons(void *arg, void *unused, int kmflag)
115 {
116 dnode_t *dn = arg;
117 int i;
118
119 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
120 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
121 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
122 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
123
124 /*
125 * Every dbuf has a reference, and dropping a tracked reference is
126 * O(number of references), so don't track dn_holds.
127 */
128 zfs_refcount_create_untracked(&dn->dn_holds);
129 zfs_refcount_create(&dn->dn_tx_holds);
130 list_link_init(&dn->dn_link);
131
132 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
133 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
134 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
135 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
136 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
137 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
138 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
139 bzero(&dn->dn_next_maxblkid[0], sizeof (dn->dn_next_maxblkid));
140
141 for (i = 0; i < TXG_SIZE; i++) {
142 multilist_link_init(&dn->dn_dirty_link[i]);
143 dn->dn_free_ranges[i] = NULL;
144 list_create(&dn->dn_dirty_records[i],
145 sizeof (dbuf_dirty_record_t),
146 offsetof(dbuf_dirty_record_t, dr_dirty_node));
147 }
148
149 dn->dn_allocated_txg = 0;
150 dn->dn_free_txg = 0;
151 dn->dn_assigned_txg = 0;
152 dn->dn_dirty_txg = 0;
153 dn->dn_dirtyctx = 0;
154 dn->dn_dirtyctx_firstset = NULL;
155 dn->dn_bonus = NULL;
156 dn->dn_have_spill = B_FALSE;
157 dn->dn_zio = NULL;
158 dn->dn_oldused = 0;
159 dn->dn_oldflags = 0;
160 dn->dn_olduid = 0;
161 dn->dn_oldgid = 0;
162 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
163 dn->dn_newuid = 0;
164 dn->dn_newgid = 0;
165 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
166 dn->dn_id_flags = 0;
167
168 dn->dn_dbufs_count = 0;
169 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
170 offsetof(dmu_buf_impl_t, db_link));
171
172 dn->dn_moved = 0;
173 return (0);
174 }
175
176 /* ARGSUSED */
177 static void
178 dnode_dest(void *arg, void *unused)
179 {
180 int i;
181 dnode_t *dn = arg;
182
183 rw_destroy(&dn->dn_struct_rwlock);
184 mutex_destroy(&dn->dn_mtx);
185 mutex_destroy(&dn->dn_dbufs_mtx);
186 cv_destroy(&dn->dn_notxholds);
187 zfs_refcount_destroy(&dn->dn_holds);
188 zfs_refcount_destroy(&dn->dn_tx_holds);
189 ASSERT(!list_link_active(&dn->dn_link));
190
191 for (i = 0; i < TXG_SIZE; i++) {
192 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
193 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
194 list_destroy(&dn->dn_dirty_records[i]);
195 ASSERT0(dn->dn_next_nblkptr[i]);
196 ASSERT0(dn->dn_next_nlevels[i]);
197 ASSERT0(dn->dn_next_indblkshift[i]);
198 ASSERT0(dn->dn_next_bonustype[i]);
199 ASSERT0(dn->dn_rm_spillblk[i]);
200 ASSERT0(dn->dn_next_bonuslen[i]);
201 ASSERT0(dn->dn_next_blksz[i]);
202 ASSERT0(dn->dn_next_maxblkid[i]);
203 }
204
205 ASSERT0(dn->dn_allocated_txg);
206 ASSERT0(dn->dn_free_txg);
207 ASSERT0(dn->dn_assigned_txg);
208 ASSERT0(dn->dn_dirty_txg);
209 ASSERT0(dn->dn_dirtyctx);
210 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
211 ASSERT3P(dn->dn_bonus, ==, NULL);
212 ASSERT(!dn->dn_have_spill);
213 ASSERT3P(dn->dn_zio, ==, NULL);
214 ASSERT0(dn->dn_oldused);
215 ASSERT0(dn->dn_oldflags);
216 ASSERT0(dn->dn_olduid);
217 ASSERT0(dn->dn_oldgid);
218 ASSERT0(dn->dn_oldprojid);
219 ASSERT0(dn->dn_newuid);
220 ASSERT0(dn->dn_newgid);
221 ASSERT0(dn->dn_newprojid);
222 ASSERT0(dn->dn_id_flags);
223
224 ASSERT0(dn->dn_dbufs_count);
225 avl_destroy(&dn->dn_dbufs);
226 }
227
228 void
229 dnode_init(void)
230 {
231 ASSERT(dnode_cache == NULL);
232 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
233 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
234 kmem_cache_set_move(dnode_cache, dnode_move);
235
236 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
237 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
238 KSTAT_FLAG_VIRTUAL);
239 if (dnode_ksp != NULL) {
240 dnode_ksp->ks_data = &dnode_stats;
241 kstat_install(dnode_ksp);
242 }
243 }
244
245 void
246 dnode_fini(void)
247 {
248 if (dnode_ksp != NULL) {
249 kstat_delete(dnode_ksp);
250 dnode_ksp = NULL;
251 }
252
253 kmem_cache_destroy(dnode_cache);
254 dnode_cache = NULL;
255 }
256
257
258 #ifdef ZFS_DEBUG
259 void
260 dnode_verify(dnode_t *dn)
261 {
262 int drop_struct_lock = FALSE;
263
264 ASSERT(dn->dn_phys);
265 ASSERT(dn->dn_objset);
266 ASSERT(dn->dn_handle->dnh_dnode == dn);
267
268 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
269
270 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
271 return;
272
273 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
274 rw_enter(&dn->dn_struct_rwlock, RW_READER);
275 drop_struct_lock = TRUE;
276 }
277 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
278 int i;
279 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
280 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
281 if (dn->dn_datablkshift) {
282 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
283 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
284 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
285 }
286 ASSERT3U(dn->dn_nlevels, <=, 30);
287 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
288 ASSERT3U(dn->dn_nblkptr, >=, 1);
289 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
290 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
291 ASSERT3U(dn->dn_datablksz, ==,
292 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
293 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
294 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
295 dn->dn_bonuslen, <=, max_bonuslen);
296 for (i = 0; i < TXG_SIZE; i++) {
297 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
298 }
299 }
300 if (dn->dn_phys->dn_type != DMU_OT_NONE)
301 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
302 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
303 if (dn->dn_dbuf != NULL) {
304 ASSERT3P(dn->dn_phys, ==,
305 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
306 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
307 }
308 if (drop_struct_lock)
309 rw_exit(&dn->dn_struct_rwlock);
310 }
311 #endif
312
313 void
314 dnode_byteswap(dnode_phys_t *dnp)
315 {
316 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
317 int i;
318
319 if (dnp->dn_type == DMU_OT_NONE) {
320 bzero(dnp, sizeof (dnode_phys_t));
321 return;
322 }
323
324 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
325 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
326 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
327 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
328 dnp->dn_used = BSWAP_64(dnp->dn_used);
329
330 /*
331 * dn_nblkptr is only one byte, so it's OK to read it in either
332 * byte order. We can't read dn_bouslen.
333 */
334 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
335 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
336 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
337 buf64[i] = BSWAP_64(buf64[i]);
338
339 /*
340 * OK to check dn_bonuslen for zero, because it won't matter if
341 * we have the wrong byte order. This is necessary because the
342 * dnode dnode is smaller than a regular dnode.
343 */
344 if (dnp->dn_bonuslen != 0) {
345 /*
346 * Note that the bonus length calculated here may be
347 * longer than the actual bonus buffer. This is because
348 * we always put the bonus buffer after the last block
349 * pointer (instead of packing it against the end of the
350 * dnode buffer).
351 */
352 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
353 int slots = dnp->dn_extra_slots + 1;
354 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
355 dmu_object_byteswap_t byteswap;
356 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
357 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
358 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
359 }
360
361 /* Swap SPILL block if we have one */
362 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
363 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
364 }
365
366 void
367 dnode_buf_byteswap(void *vbuf, size_t size)
368 {
369 int i = 0;
370
371 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
372 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
373
374 while (i < size) {
375 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
376 dnode_byteswap(dnp);
377
378 i += DNODE_MIN_SIZE;
379 if (dnp->dn_type != DMU_OT_NONE)
380 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
381 }
382 }
383
384 void
385 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
386 {
387 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
388
389 dnode_setdirty(dn, tx);
390 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
391 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
392 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
393 dn->dn_bonuslen = newsize;
394 if (newsize == 0)
395 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
396 else
397 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
398 rw_exit(&dn->dn_struct_rwlock);
399 }
400
401 void
402 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
403 {
404 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
405 dnode_setdirty(dn, tx);
406 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
407 dn->dn_bonustype = newtype;
408 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
409 rw_exit(&dn->dn_struct_rwlock);
410 }
411
412 void
413 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
414 {
415 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
416 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
417 dnode_setdirty(dn, tx);
418 dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK;
419 dn->dn_have_spill = B_FALSE;
420 }
421
422 static void
423 dnode_setdblksz(dnode_t *dn, int size)
424 {
425 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
426 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
427 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
428 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
429 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
430 dn->dn_datablksz = size;
431 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
432 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
433 }
434
435 static dnode_t *
436 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
437 uint64_t object, dnode_handle_t *dnh)
438 {
439 dnode_t *dn;
440
441 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
442 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
443 dn->dn_moved = 0;
444
445 /*
446 * Defer setting dn_objset until the dnode is ready to be a candidate
447 * for the dnode_move() callback.
448 */
449 dn->dn_object = object;
450 dn->dn_dbuf = db;
451 dn->dn_handle = dnh;
452 dn->dn_phys = dnp;
453
454 if (dnp->dn_datablkszsec) {
455 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
456 } else {
457 dn->dn_datablksz = 0;
458 dn->dn_datablkszsec = 0;
459 dn->dn_datablkshift = 0;
460 }
461 dn->dn_indblkshift = dnp->dn_indblkshift;
462 dn->dn_nlevels = dnp->dn_nlevels;
463 dn->dn_type = dnp->dn_type;
464 dn->dn_nblkptr = dnp->dn_nblkptr;
465 dn->dn_checksum = dnp->dn_checksum;
466 dn->dn_compress = dnp->dn_compress;
467 dn->dn_bonustype = dnp->dn_bonustype;
468 dn->dn_bonuslen = dnp->dn_bonuslen;
469 dn->dn_num_slots = dnp->dn_extra_slots + 1;
470 dn->dn_maxblkid = dnp->dn_maxblkid;
471 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
472 dn->dn_id_flags = 0;
473
474 dmu_zfetch_init(&dn->dn_zfetch, dn);
475
476 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
477 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
478 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
479
480 mutex_enter(&os->os_lock);
481
482 /*
483 * Exclude special dnodes from os_dnodes so an empty os_dnodes
484 * signifies that the special dnodes have no references from
485 * their children (the entries in os_dnodes). This allows
486 * dnode_destroy() to easily determine if the last child has
487 * been removed and then complete eviction of the objset.
488 */
489 if (!DMU_OBJECT_IS_SPECIAL(object))
490 list_insert_head(&os->os_dnodes, dn);
491 membar_producer();
492
493 /*
494 * Everything else must be valid before assigning dn_objset
495 * makes the dnode eligible for dnode_move().
496 */
497 dn->dn_objset = os;
498
499 dnh->dnh_dnode = dn;
500 mutex_exit(&os->os_lock);
501
502 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
503
504 return (dn);
505 }
506
507 /*
508 * Caller must be holding the dnode handle, which is released upon return.
509 */
510 static void
511 dnode_destroy(dnode_t *dn)
512 {
513 objset_t *os = dn->dn_objset;
514 boolean_t complete_os_eviction = B_FALSE;
515
516 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
517
518 mutex_enter(&os->os_lock);
519 POINTER_INVALIDATE(&dn->dn_objset);
520 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
521 list_remove(&os->os_dnodes, dn);
522 complete_os_eviction =
523 list_is_empty(&os->os_dnodes) &&
524 list_link_active(&os->os_evicting_node);
525 }
526 mutex_exit(&os->os_lock);
527
528 /* the dnode can no longer move, so we can release the handle */
529 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
530 zrl_remove(&dn->dn_handle->dnh_zrlock);
531
532 dn->dn_allocated_txg = 0;
533 dn->dn_free_txg = 0;
534 dn->dn_assigned_txg = 0;
535 dn->dn_dirty_txg = 0;
536
537 dn->dn_dirtyctx = 0;
538 if (dn->dn_dirtyctx_firstset != NULL) {
539 kmem_free(dn->dn_dirtyctx_firstset, 1);
540 dn->dn_dirtyctx_firstset = NULL;
541 }
542 if (dn->dn_bonus != NULL) {
543 mutex_enter(&dn->dn_bonus->db_mtx);
544 dbuf_destroy(dn->dn_bonus);
545 dn->dn_bonus = NULL;
546 }
547 dn->dn_zio = NULL;
548
549 dn->dn_have_spill = B_FALSE;
550 dn->dn_oldused = 0;
551 dn->dn_oldflags = 0;
552 dn->dn_olduid = 0;
553 dn->dn_oldgid = 0;
554 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
555 dn->dn_newuid = 0;
556 dn->dn_newgid = 0;
557 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
558 dn->dn_id_flags = 0;
559
560 dmu_zfetch_fini(&dn->dn_zfetch);
561 kmem_cache_free(dnode_cache, dn);
562 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
563
564 if (complete_os_eviction)
565 dmu_objset_evict_done(os);
566 }
567
568 void
569 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
570 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
571 {
572 int i;
573
574 ASSERT3U(dn_slots, >, 0);
575 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
576 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
577 ASSERT3U(blocksize, <=,
578 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
579 if (blocksize == 0)
580 blocksize = 1 << zfs_default_bs;
581 else
582 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
583
584 if (ibs == 0)
585 ibs = zfs_default_ibs;
586
587 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
588
589 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
590 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
591 DNODE_STAT_BUMP(dnode_allocate);
592
593 ASSERT(dn->dn_type == DMU_OT_NONE);
594 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
595 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
596 ASSERT(ot != DMU_OT_NONE);
597 ASSERT(DMU_OT_IS_VALID(ot));
598 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
599 (bonustype == DMU_OT_SA && bonuslen == 0) ||
600 (bonustype != DMU_OT_NONE && bonuslen != 0));
601 ASSERT(DMU_OT_IS_VALID(bonustype));
602 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
603 ASSERT(dn->dn_type == DMU_OT_NONE);
604 ASSERT0(dn->dn_maxblkid);
605 ASSERT0(dn->dn_allocated_txg);
606 ASSERT0(dn->dn_assigned_txg);
607 ASSERT0(dn->dn_dirty_txg);
608 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
609 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
610 ASSERT(avl_is_empty(&dn->dn_dbufs));
611
612 for (i = 0; i < TXG_SIZE; i++) {
613 ASSERT0(dn->dn_next_nblkptr[i]);
614 ASSERT0(dn->dn_next_nlevels[i]);
615 ASSERT0(dn->dn_next_indblkshift[i]);
616 ASSERT0(dn->dn_next_bonuslen[i]);
617 ASSERT0(dn->dn_next_bonustype[i]);
618 ASSERT0(dn->dn_rm_spillblk[i]);
619 ASSERT0(dn->dn_next_blksz[i]);
620 ASSERT0(dn->dn_next_maxblkid[i]);
621 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
622 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
623 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
624 }
625
626 dn->dn_type = ot;
627 dnode_setdblksz(dn, blocksize);
628 dn->dn_indblkshift = ibs;
629 dn->dn_nlevels = 1;
630 dn->dn_num_slots = dn_slots;
631 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
632 dn->dn_nblkptr = 1;
633 else {
634 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
635 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
636 SPA_BLKPTRSHIFT));
637 }
638
639 dn->dn_bonustype = bonustype;
640 dn->dn_bonuslen = bonuslen;
641 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
642 dn->dn_compress = ZIO_COMPRESS_INHERIT;
643 dn->dn_dirtyctx = 0;
644
645 dn->dn_free_txg = 0;
646 if (dn->dn_dirtyctx_firstset) {
647 kmem_free(dn->dn_dirtyctx_firstset, 1);
648 dn->dn_dirtyctx_firstset = NULL;
649 }
650
651 dn->dn_allocated_txg = tx->tx_txg;
652 dn->dn_id_flags = 0;
653
654 dnode_setdirty(dn, tx);
655 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
656 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
657 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
658 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
659 }
660
661 void
662 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
663 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
664 {
665 int nblkptr;
666
667 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
668 ASSERT3U(blocksize, <=,
669 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
670 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
671 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
672 ASSERT(tx->tx_txg != 0);
673 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
674 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
675 (bonustype == DMU_OT_SA && bonuslen == 0));
676 ASSERT(DMU_OT_IS_VALID(bonustype));
677 ASSERT3U(bonuslen, <=,
678 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
679 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
680
681 dnode_free_interior_slots(dn);
682 DNODE_STAT_BUMP(dnode_reallocate);
683
684 /* clean up any unreferenced dbufs */
685 dnode_evict_dbufs(dn);
686
687 dn->dn_id_flags = 0;
688
689 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
690 dnode_setdirty(dn, tx);
691 if (dn->dn_datablksz != blocksize) {
692 /* change blocksize */
693 ASSERT0(dn->dn_maxblkid);
694 ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
695 dnode_block_freed(dn, 0));
696
697 dnode_setdblksz(dn, blocksize);
698 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize;
699 }
700 if (dn->dn_bonuslen != bonuslen)
701 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen;
702
703 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
704 nblkptr = 1;
705 else
706 nblkptr = MIN(DN_MAX_NBLKPTR,
707 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
708 SPA_BLKPTRSHIFT));
709 if (dn->dn_bonustype != bonustype)
710 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype;
711 if (dn->dn_nblkptr != nblkptr)
712 dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr;
713 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
714 dbuf_rm_spill(dn, tx);
715 dnode_rm_spill(dn, tx);
716 }
717
718 rw_exit(&dn->dn_struct_rwlock);
719
720 /* change type */
721 dn->dn_type = ot;
722
723 /* change bonus size and type */
724 mutex_enter(&dn->dn_mtx);
725 dn->dn_bonustype = bonustype;
726 dn->dn_bonuslen = bonuslen;
727 dn->dn_num_slots = dn_slots;
728 dn->dn_nblkptr = nblkptr;
729 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
730 dn->dn_compress = ZIO_COMPRESS_INHERIT;
731 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
732
733 /* fix up the bonus db_size */
734 if (dn->dn_bonus) {
735 dn->dn_bonus->db.db_size =
736 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
737 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
738 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
739 }
740
741 dn->dn_allocated_txg = tx->tx_txg;
742 mutex_exit(&dn->dn_mtx);
743 }
744
745 #ifdef _KERNEL
746 static void
747 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
748 {
749 int i;
750
751 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
752 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
753 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
754 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
755
756 /* Copy fields. */
757 ndn->dn_objset = odn->dn_objset;
758 ndn->dn_object = odn->dn_object;
759 ndn->dn_dbuf = odn->dn_dbuf;
760 ndn->dn_handle = odn->dn_handle;
761 ndn->dn_phys = odn->dn_phys;
762 ndn->dn_type = odn->dn_type;
763 ndn->dn_bonuslen = odn->dn_bonuslen;
764 ndn->dn_bonustype = odn->dn_bonustype;
765 ndn->dn_nblkptr = odn->dn_nblkptr;
766 ndn->dn_checksum = odn->dn_checksum;
767 ndn->dn_compress = odn->dn_compress;
768 ndn->dn_nlevels = odn->dn_nlevels;
769 ndn->dn_indblkshift = odn->dn_indblkshift;
770 ndn->dn_datablkshift = odn->dn_datablkshift;
771 ndn->dn_datablkszsec = odn->dn_datablkszsec;
772 ndn->dn_datablksz = odn->dn_datablksz;
773 ndn->dn_maxblkid = odn->dn_maxblkid;
774 ndn->dn_num_slots = odn->dn_num_slots;
775 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
776 sizeof (odn->dn_next_type));
777 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
778 sizeof (odn->dn_next_nblkptr));
779 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
780 sizeof (odn->dn_next_nlevels));
781 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
782 sizeof (odn->dn_next_indblkshift));
783 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
784 sizeof (odn->dn_next_bonustype));
785 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
786 sizeof (odn->dn_rm_spillblk));
787 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
788 sizeof (odn->dn_next_bonuslen));
789 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
790 sizeof (odn->dn_next_blksz));
791 bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
792 sizeof (odn->dn_next_maxblkid));
793 for (i = 0; i < TXG_SIZE; i++) {
794 list_move_tail(&ndn->dn_dirty_records[i],
795 &odn->dn_dirty_records[i]);
796 }
797 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
798 sizeof (odn->dn_free_ranges));
799 ndn->dn_allocated_txg = odn->dn_allocated_txg;
800 ndn->dn_free_txg = odn->dn_free_txg;
801 ndn->dn_assigned_txg = odn->dn_assigned_txg;
802 ndn->dn_dirty_txg = odn->dn_dirty_txg;
803 ndn->dn_dirtyctx = odn->dn_dirtyctx;
804 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
805 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
806 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
807 ASSERT(avl_is_empty(&ndn->dn_dbufs));
808 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
809 ndn->dn_dbufs_count = odn->dn_dbufs_count;
810 ndn->dn_bonus = odn->dn_bonus;
811 ndn->dn_have_spill = odn->dn_have_spill;
812 ndn->dn_zio = odn->dn_zio;
813 ndn->dn_oldused = odn->dn_oldused;
814 ndn->dn_oldflags = odn->dn_oldflags;
815 ndn->dn_olduid = odn->dn_olduid;
816 ndn->dn_oldgid = odn->dn_oldgid;
817 ndn->dn_oldprojid = odn->dn_oldprojid;
818 ndn->dn_newuid = odn->dn_newuid;
819 ndn->dn_newgid = odn->dn_newgid;
820 ndn->dn_newprojid = odn->dn_newprojid;
821 ndn->dn_id_flags = odn->dn_id_flags;
822 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
823 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
824 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
825
826 /*
827 * Update back pointers. Updating the handle fixes the back pointer of
828 * every descendant dbuf as well as the bonus dbuf.
829 */
830 ASSERT(ndn->dn_handle->dnh_dnode == odn);
831 ndn->dn_handle->dnh_dnode = ndn;
832 if (ndn->dn_zfetch.zf_dnode == odn) {
833 ndn->dn_zfetch.zf_dnode = ndn;
834 }
835
836 /*
837 * Invalidate the original dnode by clearing all of its back pointers.
838 */
839 odn->dn_dbuf = NULL;
840 odn->dn_handle = NULL;
841 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
842 offsetof(dmu_buf_impl_t, db_link));
843 odn->dn_dbufs_count = 0;
844 odn->dn_bonus = NULL;
845 dmu_zfetch_fini(&odn->dn_zfetch);
846
847 /*
848 * Set the low bit of the objset pointer to ensure that dnode_move()
849 * recognizes the dnode as invalid in any subsequent callback.
850 */
851 POINTER_INVALIDATE(&odn->dn_objset);
852
853 /*
854 * Satisfy the destructor.
855 */
856 for (i = 0; i < TXG_SIZE; i++) {
857 list_create(&odn->dn_dirty_records[i],
858 sizeof (dbuf_dirty_record_t),
859 offsetof(dbuf_dirty_record_t, dr_dirty_node));
860 odn->dn_free_ranges[i] = NULL;
861 odn->dn_next_nlevels[i] = 0;
862 odn->dn_next_indblkshift[i] = 0;
863 odn->dn_next_bonustype[i] = 0;
864 odn->dn_rm_spillblk[i] = 0;
865 odn->dn_next_bonuslen[i] = 0;
866 odn->dn_next_blksz[i] = 0;
867 }
868 odn->dn_allocated_txg = 0;
869 odn->dn_free_txg = 0;
870 odn->dn_assigned_txg = 0;
871 odn->dn_dirty_txg = 0;
872 odn->dn_dirtyctx = 0;
873 odn->dn_dirtyctx_firstset = NULL;
874 odn->dn_have_spill = B_FALSE;
875 odn->dn_zio = NULL;
876 odn->dn_oldused = 0;
877 odn->dn_oldflags = 0;
878 odn->dn_olduid = 0;
879 odn->dn_oldgid = 0;
880 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
881 odn->dn_newuid = 0;
882 odn->dn_newgid = 0;
883 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
884 odn->dn_id_flags = 0;
885
886 /*
887 * Mark the dnode.
888 */
889 ndn->dn_moved = 1;
890 odn->dn_moved = (uint8_t)-1;
891 }
892
893 /*ARGSUSED*/
894 static kmem_cbrc_t
895 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
896 {
897 dnode_t *odn = buf, *ndn = newbuf;
898 objset_t *os;
899 int64_t refcount;
900 uint32_t dbufs;
901
902 /*
903 * The dnode is on the objset's list of known dnodes if the objset
904 * pointer is valid. We set the low bit of the objset pointer when
905 * freeing the dnode to invalidate it, and the memory patterns written
906 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
907 * A newly created dnode sets the objset pointer last of all to indicate
908 * that the dnode is known and in a valid state to be moved by this
909 * function.
910 */
911 os = odn->dn_objset;
912 if (!POINTER_IS_VALID(os)) {
913 DNODE_STAT_BUMP(dnode_move_invalid);
914 return (KMEM_CBRC_DONT_KNOW);
915 }
916
917 /*
918 * Ensure that the objset does not go away during the move.
919 */
920 rw_enter(&os_lock, RW_WRITER);
921 if (os != odn->dn_objset) {
922 rw_exit(&os_lock);
923 DNODE_STAT_BUMP(dnode_move_recheck1);
924 return (KMEM_CBRC_DONT_KNOW);
925 }
926
927 /*
928 * If the dnode is still valid, then so is the objset. We know that no
929 * valid objset can be freed while we hold os_lock, so we can safely
930 * ensure that the objset remains in use.
931 */
932 mutex_enter(&os->os_lock);
933
934 /*
935 * Recheck the objset pointer in case the dnode was removed just before
936 * acquiring the lock.
937 */
938 if (os != odn->dn_objset) {
939 mutex_exit(&os->os_lock);
940 rw_exit(&os_lock);
941 DNODE_STAT_BUMP(dnode_move_recheck2);
942 return (KMEM_CBRC_DONT_KNOW);
943 }
944
945 /*
946 * At this point we know that as long as we hold os->os_lock, the dnode
947 * cannot be freed and fields within the dnode can be safely accessed.
948 * The objset listing this dnode cannot go away as long as this dnode is
949 * on its list.
950 */
951 rw_exit(&os_lock);
952 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
953 mutex_exit(&os->os_lock);
954 DNODE_STAT_BUMP(dnode_move_special);
955 return (KMEM_CBRC_NO);
956 }
957 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
958
959 /*
960 * Lock the dnode handle to prevent the dnode from obtaining any new
961 * holds. This also prevents the descendant dbufs and the bonus dbuf
962 * from accessing the dnode, so that we can discount their holds. The
963 * handle is safe to access because we know that while the dnode cannot
964 * go away, neither can its handle. Once we hold dnh_zrlock, we can
965 * safely move any dnode referenced only by dbufs.
966 */
967 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
968 mutex_exit(&os->os_lock);
969 DNODE_STAT_BUMP(dnode_move_handle);
970 return (KMEM_CBRC_LATER);
971 }
972
973 /*
974 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
975 * We need to guarantee that there is a hold for every dbuf in order to
976 * determine whether the dnode is actively referenced. Falsely matching
977 * a dbuf to an active hold would lead to an unsafe move. It's possible
978 * that a thread already having an active dnode hold is about to add a
979 * dbuf, and we can't compare hold and dbuf counts while the add is in
980 * progress.
981 */
982 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
983 zrl_exit(&odn->dn_handle->dnh_zrlock);
984 mutex_exit(&os->os_lock);
985 DNODE_STAT_BUMP(dnode_move_rwlock);
986 return (KMEM_CBRC_LATER);
987 }
988
989 /*
990 * A dbuf may be removed (evicted) without an active dnode hold. In that
991 * case, the dbuf count is decremented under the handle lock before the
992 * dbuf's hold is released. This order ensures that if we count the hold
993 * after the dbuf is removed but before its hold is released, we will
994 * treat the unmatched hold as active and exit safely. If we count the
995 * hold before the dbuf is removed, the hold is discounted, and the
996 * removal is blocked until the move completes.
997 */
998 refcount = zfs_refcount_count(&odn->dn_holds);
999 ASSERT(refcount >= 0);
1000 dbufs = odn->dn_dbufs_count;
1001
1002 /* We can't have more dbufs than dnode holds. */
1003 ASSERT3U(dbufs, <=, refcount);
1004 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1005 uint32_t, dbufs);
1006
1007 if (refcount > dbufs) {
1008 rw_exit(&odn->dn_struct_rwlock);
1009 zrl_exit(&odn->dn_handle->dnh_zrlock);
1010 mutex_exit(&os->os_lock);
1011 DNODE_STAT_BUMP(dnode_move_active);
1012 return (KMEM_CBRC_LATER);
1013 }
1014
1015 rw_exit(&odn->dn_struct_rwlock);
1016
1017 /*
1018 * At this point we know that anyone with a hold on the dnode is not
1019 * actively referencing it. The dnode is known and in a valid state to
1020 * move. We're holding the locks needed to execute the critical section.
1021 */
1022 dnode_move_impl(odn, ndn);
1023
1024 list_link_replace(&odn->dn_link, &ndn->dn_link);
1025 /* If the dnode was safe to move, the refcount cannot have changed. */
1026 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1027 ASSERT(dbufs == ndn->dn_dbufs_count);
1028 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1029 mutex_exit(&os->os_lock);
1030
1031 return (KMEM_CBRC_YES);
1032 }
1033 #endif /* _KERNEL */
1034
1035 static void
1036 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1037 {
1038 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1039
1040 for (int i = idx; i < idx + slots; i++) {
1041 dnode_handle_t *dnh = &children->dnc_children[i];
1042 zrl_add(&dnh->dnh_zrlock);
1043 }
1044 }
1045
1046 static void
1047 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1048 {
1049 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1050
1051 for (int i = idx; i < idx + slots; i++) {
1052 dnode_handle_t *dnh = &children->dnc_children[i];
1053
1054 if (zrl_is_locked(&dnh->dnh_zrlock))
1055 zrl_exit(&dnh->dnh_zrlock);
1056 else
1057 zrl_remove(&dnh->dnh_zrlock);
1058 }
1059 }
1060
1061 static int
1062 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1063 {
1064 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1065
1066 for (int i = idx; i < idx + slots; i++) {
1067 dnode_handle_t *dnh = &children->dnc_children[i];
1068
1069 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1070 for (int j = idx; j < i; j++) {
1071 dnh = &children->dnc_children[j];
1072 zrl_exit(&dnh->dnh_zrlock);
1073 }
1074
1075 return (0);
1076 }
1077 }
1078
1079 return (1);
1080 }
1081
1082 static void
1083 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1084 {
1085 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1086
1087 for (int i = idx; i < idx + slots; i++) {
1088 dnode_handle_t *dnh = &children->dnc_children[i];
1089 dnh->dnh_dnode = ptr;
1090 }
1091 }
1092
1093 static boolean_t
1094 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1095 {
1096 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1097
1098 /*
1099 * If all dnode slots are either already free or
1100 * evictable return B_TRUE.
1101 */
1102 for (int i = idx; i < idx + slots; i++) {
1103 dnode_handle_t *dnh = &children->dnc_children[i];
1104 dnode_t *dn = dnh->dnh_dnode;
1105
1106 if (dn == DN_SLOT_FREE) {
1107 continue;
1108 } else if (DN_SLOT_IS_PTR(dn)) {
1109 mutex_enter(&dn->dn_mtx);
1110 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1111 zfs_refcount_is_zero(&dn->dn_holds) &&
1112 !DNODE_IS_DIRTY(dn));
1113 mutex_exit(&dn->dn_mtx);
1114
1115 if (!can_free)
1116 return (B_FALSE);
1117 else
1118 continue;
1119 } else {
1120 return (B_FALSE);
1121 }
1122 }
1123
1124 return (B_TRUE);
1125 }
1126
1127 static void
1128 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1129 {
1130 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1131
1132 for (int i = idx; i < idx + slots; i++) {
1133 dnode_handle_t *dnh = &children->dnc_children[i];
1134
1135 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1136
1137 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1138 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1139 dnode_destroy(dnh->dnh_dnode);
1140 dnh->dnh_dnode = DN_SLOT_FREE;
1141 }
1142 }
1143 }
1144
1145 void
1146 dnode_free_interior_slots(dnode_t *dn)
1147 {
1148 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1149 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1150 int idx = (dn->dn_object & (epb - 1)) + 1;
1151 int slots = dn->dn_num_slots - 1;
1152
1153 if (slots == 0)
1154 return;
1155
1156 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1157
1158 while (!dnode_slots_tryenter(children, idx, slots)) {
1159 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1160 cond_resched();
1161 }
1162
1163 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1164 dnode_slots_rele(children, idx, slots);
1165 }
1166
1167 void
1168 dnode_special_close(dnode_handle_t *dnh)
1169 {
1170 dnode_t *dn = dnh->dnh_dnode;
1171
1172 /*
1173 * Wait for final references to the dnode to clear. This can
1174 * only happen if the arc is asynchronously evicting state that
1175 * has a hold on this dnode while we are trying to evict this
1176 * dnode.
1177 */
1178 while (zfs_refcount_count(&dn->dn_holds) > 0)
1179 delay(1);
1180 ASSERT(dn->dn_dbuf == NULL ||
1181 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1182 zrl_add(&dnh->dnh_zrlock);
1183 dnode_destroy(dn); /* implicit zrl_remove() */
1184 zrl_destroy(&dnh->dnh_zrlock);
1185 dnh->dnh_dnode = NULL;
1186 }
1187
1188 void
1189 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1190 dnode_handle_t *dnh)
1191 {
1192 dnode_t *dn;
1193
1194 zrl_init(&dnh->dnh_zrlock);
1195 zrl_tryenter(&dnh->dnh_zrlock);
1196
1197 dn = dnode_create(os, dnp, NULL, object, dnh);
1198 DNODE_VERIFY(dn);
1199
1200 zrl_exit(&dnh->dnh_zrlock);
1201 }
1202
1203 static void
1204 dnode_buf_evict_async(void *dbu)
1205 {
1206 dnode_children_t *dnc = dbu;
1207
1208 DNODE_STAT_BUMP(dnode_buf_evict);
1209
1210 for (int i = 0; i < dnc->dnc_count; i++) {
1211 dnode_handle_t *dnh = &dnc->dnc_children[i];
1212 dnode_t *dn;
1213
1214 /*
1215 * The dnode handle lock guards against the dnode moving to
1216 * another valid address, so there is no need here to guard
1217 * against changes to or from NULL.
1218 */
1219 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1220 zrl_destroy(&dnh->dnh_zrlock);
1221 dnh->dnh_dnode = DN_SLOT_UNINIT;
1222 continue;
1223 }
1224
1225 zrl_add(&dnh->dnh_zrlock);
1226 dn = dnh->dnh_dnode;
1227 /*
1228 * If there are holds on this dnode, then there should
1229 * be holds on the dnode's containing dbuf as well; thus
1230 * it wouldn't be eligible for eviction and this function
1231 * would not have been called.
1232 */
1233 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1234 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1235
1236 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1237 zrl_destroy(&dnh->dnh_zrlock);
1238 dnh->dnh_dnode = DN_SLOT_UNINIT;
1239 }
1240 kmem_free(dnc, sizeof (dnode_children_t) +
1241 dnc->dnc_count * sizeof (dnode_handle_t));
1242 }
1243
1244 /*
1245 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1246 * to ensure the hole at the specified object offset is large enough to
1247 * hold the dnode being created. The slots parameter is also used to ensure
1248 * a dnode does not span multiple dnode blocks. In both of these cases, if
1249 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1250 * are only possible when using DNODE_MUST_BE_FREE.
1251 *
1252 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1253 * dnode_hold_impl() will check if the requested dnode is already consumed
1254 * as an extra dnode slot by an large dnode, in which case it returns
1255 * ENOENT.
1256 *
1257 * errors:
1258 * EINVAL - Invalid object number or flags.
1259 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1260 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1261 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1262 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1263 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1264 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1265 * EIO - I/O error when reading the meta dnode dbuf.
1266 *
1267 * succeeds even for free dnodes.
1268 */
1269 int
1270 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1271 void *tag, dnode_t **dnp)
1272 {
1273 int epb, idx, err;
1274 int drop_struct_lock = FALSE;
1275 int type;
1276 uint64_t blk;
1277 dnode_t *mdn, *dn;
1278 dmu_buf_impl_t *db;
1279 dnode_children_t *dnc;
1280 dnode_phys_t *dn_block;
1281 dnode_handle_t *dnh;
1282
1283 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1284 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1285
1286 /*
1287 * If you are holding the spa config lock as writer, you shouldn't
1288 * be asking the DMU to do *anything* unless it's the root pool
1289 * which may require us to read from the root filesystem while
1290 * holding some (not all) of the locks as writer.
1291 */
1292 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1293 (spa_is_root(os->os_spa) &&
1294 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1295
1296 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1297
1298 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1299 object == DMU_PROJECTUSED_OBJECT) {
1300 if (object == DMU_USERUSED_OBJECT)
1301 dn = DMU_USERUSED_DNODE(os);
1302 else if (object == DMU_GROUPUSED_OBJECT)
1303 dn = DMU_GROUPUSED_DNODE(os);
1304 else
1305 dn = DMU_PROJECTUSED_DNODE(os);
1306 if (dn == NULL)
1307 return (SET_ERROR(ENOENT));
1308 type = dn->dn_type;
1309 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1310 return (SET_ERROR(ENOENT));
1311 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1312 return (SET_ERROR(EEXIST));
1313 DNODE_VERIFY(dn);
1314 (void) zfs_refcount_add(&dn->dn_holds, tag);
1315 *dnp = dn;
1316 return (0);
1317 }
1318
1319 if (object == 0 || object >= DN_MAX_OBJECT)
1320 return (SET_ERROR(EINVAL));
1321
1322 mdn = DMU_META_DNODE(os);
1323 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1324
1325 DNODE_VERIFY(mdn);
1326
1327 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1328 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1329 drop_struct_lock = TRUE;
1330 }
1331
1332 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1333
1334 db = dbuf_hold(mdn, blk, FTAG);
1335 if (drop_struct_lock)
1336 rw_exit(&mdn->dn_struct_rwlock);
1337 if (db == NULL) {
1338 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1339 return (SET_ERROR(EIO));
1340 }
1341
1342 /*
1343 * We do not need to decrypt to read the dnode so it doesn't matter
1344 * if we get the encrypted or decrypted version.
1345 */
1346 err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1347 if (err) {
1348 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1349 dbuf_rele(db, FTAG);
1350 return (err);
1351 }
1352
1353 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1354 epb = db->db.db_size >> DNODE_SHIFT;
1355
1356 idx = object & (epb - 1);
1357 dn_block = (dnode_phys_t *)db->db.db_data;
1358
1359 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1360 dnc = dmu_buf_get_user(&db->db);
1361 dnh = NULL;
1362 if (dnc == NULL) {
1363 dnode_children_t *winner;
1364 int skip = 0;
1365
1366 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1367 epb * sizeof (dnode_handle_t), KM_SLEEP);
1368 dnc->dnc_count = epb;
1369 dnh = &dnc->dnc_children[0];
1370
1371 /* Initialize dnode slot status from dnode_phys_t */
1372 for (int i = 0; i < epb; i++) {
1373 zrl_init(&dnh[i].dnh_zrlock);
1374
1375 if (skip) {
1376 skip--;
1377 continue;
1378 }
1379
1380 if (dn_block[i].dn_type != DMU_OT_NONE) {
1381 int interior = dn_block[i].dn_extra_slots;
1382
1383 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1384 dnode_set_slots(dnc, i + 1, interior,
1385 DN_SLOT_INTERIOR);
1386 skip = interior;
1387 } else {
1388 dnh[i].dnh_dnode = DN_SLOT_FREE;
1389 skip = 0;
1390 }
1391 }
1392
1393 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1394 dnode_buf_evict_async, NULL);
1395 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1396 if (winner != NULL) {
1397
1398 for (int i = 0; i < epb; i++)
1399 zrl_destroy(&dnh[i].dnh_zrlock);
1400
1401 kmem_free(dnc, sizeof (dnode_children_t) +
1402 epb * sizeof (dnode_handle_t));
1403 dnc = winner;
1404 }
1405 }
1406
1407 ASSERT(dnc->dnc_count == epb);
1408
1409 if (flag & DNODE_MUST_BE_ALLOCATED) {
1410 slots = 1;
1411
1412 dnode_slots_hold(dnc, idx, slots);
1413 dnh = &dnc->dnc_children[idx];
1414
1415 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1416 dn = dnh->dnh_dnode;
1417 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1418 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1419 dnode_slots_rele(dnc, idx, slots);
1420 dbuf_rele(db, FTAG);
1421 return (SET_ERROR(EEXIST));
1422 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1423 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1424 dnode_slots_rele(dnc, idx, slots);
1425 dbuf_rele(db, FTAG);
1426 return (SET_ERROR(ENOENT));
1427 } else {
1428 dnode_slots_rele(dnc, idx, slots);
1429 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1430 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1431 cond_resched();
1432 }
1433
1434 /*
1435 * Someone else won the race and called dnode_create()
1436 * after we checked DN_SLOT_IS_PTR() above but before
1437 * we acquired the lock.
1438 */
1439 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1440 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1441 dn = dnh->dnh_dnode;
1442 } else {
1443 dn = dnode_create(os, dn_block + idx, db,
1444 object, dnh);
1445 }
1446 }
1447
1448 mutex_enter(&dn->dn_mtx);
1449 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1450 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1451 mutex_exit(&dn->dn_mtx);
1452 dnode_slots_rele(dnc, idx, slots);
1453 dbuf_rele(db, FTAG);
1454 return (SET_ERROR(ENOENT));
1455 }
1456
1457 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1458 } else if (flag & DNODE_MUST_BE_FREE) {
1459
1460 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1461 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1462 dbuf_rele(db, FTAG);
1463 return (SET_ERROR(ENOSPC));
1464 }
1465
1466 dnode_slots_hold(dnc, idx, slots);
1467
1468 if (!dnode_check_slots_free(dnc, idx, slots)) {
1469 DNODE_STAT_BUMP(dnode_hold_free_misses);
1470 dnode_slots_rele(dnc, idx, slots);
1471 dbuf_rele(db, FTAG);
1472 return (SET_ERROR(ENOSPC));
1473 }
1474
1475 dnode_slots_rele(dnc, idx, slots);
1476 while (!dnode_slots_tryenter(dnc, idx, slots)) {
1477 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1478 cond_resched();
1479 }
1480
1481 if (!dnode_check_slots_free(dnc, idx, slots)) {
1482 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1483 dnode_slots_rele(dnc, idx, slots);
1484 dbuf_rele(db, FTAG);
1485 return (SET_ERROR(ENOSPC));
1486 }
1487
1488 /*
1489 * Allocated but otherwise free dnodes which would
1490 * be in the interior of a multi-slot dnodes need
1491 * to be freed. Single slot dnodes can be safely
1492 * re-purposed as a performance optimization.
1493 */
1494 if (slots > 1)
1495 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1496
1497 dnh = &dnc->dnc_children[idx];
1498 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1499 dn = dnh->dnh_dnode;
1500 } else {
1501 dn = dnode_create(os, dn_block + idx, db,
1502 object, dnh);
1503 }
1504
1505 mutex_enter(&dn->dn_mtx);
1506 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1507 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1508 mutex_exit(&dn->dn_mtx);
1509 dnode_slots_rele(dnc, idx, slots);
1510 dbuf_rele(db, FTAG);
1511 return (SET_ERROR(EEXIST));
1512 }
1513
1514 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1515 DNODE_STAT_BUMP(dnode_hold_free_hits);
1516 } else {
1517 dbuf_rele(db, FTAG);
1518 return (SET_ERROR(EINVAL));
1519 }
1520
1521 if (dn->dn_free_txg) {
1522 DNODE_STAT_BUMP(dnode_hold_free_txg);
1523 type = dn->dn_type;
1524 mutex_exit(&dn->dn_mtx);
1525 dnode_slots_rele(dnc, idx, slots);
1526 dbuf_rele(db, FTAG);
1527 return (SET_ERROR((flag & DNODE_MUST_BE_ALLOCATED) ?
1528 ENOENT : EEXIST));
1529 }
1530
1531 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1532 dbuf_add_ref(db, dnh);
1533
1534 mutex_exit(&dn->dn_mtx);
1535
1536 /* Now we can rely on the hold to prevent the dnode from moving. */
1537 dnode_slots_rele(dnc, idx, slots);
1538
1539 DNODE_VERIFY(dn);
1540 ASSERT3P(dn->dn_dbuf, ==, db);
1541 ASSERT3U(dn->dn_object, ==, object);
1542 dbuf_rele(db, FTAG);
1543
1544 *dnp = dn;
1545 return (0);
1546 }
1547
1548 /*
1549 * Return held dnode if the object is allocated, NULL if not.
1550 */
1551 int
1552 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1553 {
1554 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1555 dnp));
1556 }
1557
1558 /*
1559 * Can only add a reference if there is already at least one
1560 * reference on the dnode. Returns FALSE if unable to add a
1561 * new reference.
1562 */
1563 boolean_t
1564 dnode_add_ref(dnode_t *dn, void *tag)
1565 {
1566 mutex_enter(&dn->dn_mtx);
1567 if (zfs_refcount_is_zero(&dn->dn_holds)) {
1568 mutex_exit(&dn->dn_mtx);
1569 return (FALSE);
1570 }
1571 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1572 mutex_exit(&dn->dn_mtx);
1573 return (TRUE);
1574 }
1575
1576 void
1577 dnode_rele(dnode_t *dn, void *tag)
1578 {
1579 mutex_enter(&dn->dn_mtx);
1580 dnode_rele_and_unlock(dn, tag, B_FALSE);
1581 }
1582
1583 void
1584 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1585 {
1586 uint64_t refs;
1587 /* Get while the hold prevents the dnode from moving. */
1588 dmu_buf_impl_t *db = dn->dn_dbuf;
1589 dnode_handle_t *dnh = dn->dn_handle;
1590
1591 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1592 mutex_exit(&dn->dn_mtx);
1593
1594 /*
1595 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1596 * indirectly by dbuf_rele() while relying on the dnode handle to
1597 * prevent the dnode from moving, since releasing the last hold could
1598 * result in the dnode's parent dbuf evicting its dnode handles. For
1599 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1600 * other direct or indirect hold on the dnode must first drop the dnode
1601 * handle.
1602 */
1603 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1604
1605 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1606 if (refs == 0 && db != NULL) {
1607 /*
1608 * Another thread could add a hold to the dnode handle in
1609 * dnode_hold_impl() while holding the parent dbuf. Since the
1610 * hold on the parent dbuf prevents the handle from being
1611 * destroyed, the hold on the handle is OK. We can't yet assert
1612 * that the handle has zero references, but that will be
1613 * asserted anyway when the handle gets destroyed.
1614 */
1615 mutex_enter(&db->db_mtx);
1616 dbuf_rele_and_unlock(db, dnh, evicting);
1617 }
1618 }
1619
1620 void
1621 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1622 {
1623 objset_t *os = dn->dn_objset;
1624 uint64_t txg = tx->tx_txg;
1625
1626 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1627 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1628 return;
1629 }
1630
1631 DNODE_VERIFY(dn);
1632
1633 #ifdef ZFS_DEBUG
1634 mutex_enter(&dn->dn_mtx);
1635 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1636 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1637 mutex_exit(&dn->dn_mtx);
1638 #endif
1639
1640 /*
1641 * Determine old uid/gid when necessary
1642 */
1643 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1644
1645 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1646 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1647
1648 /*
1649 * If we are already marked dirty, we're done.
1650 */
1651 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1652 multilist_sublist_unlock(mls);
1653 return;
1654 }
1655
1656 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1657 !avl_is_empty(&dn->dn_dbufs));
1658 ASSERT(dn->dn_datablksz != 0);
1659 ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]);
1660 ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]);
1661 ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]);
1662
1663 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1664 dn->dn_object, txg);
1665
1666 multilist_sublist_insert_head(mls, dn);
1667
1668 multilist_sublist_unlock(mls);
1669
1670 /*
1671 * The dnode maintains a hold on its containing dbuf as
1672 * long as there are holds on it. Each instantiated child
1673 * dbuf maintains a hold on the dnode. When the last child
1674 * drops its hold, the dnode will drop its hold on the
1675 * containing dbuf. We add a "dirty hold" here so that the
1676 * dnode will hang around after we finish processing its
1677 * children.
1678 */
1679 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1680
1681 (void) dbuf_dirty(dn->dn_dbuf, tx);
1682
1683 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1684 }
1685
1686 void
1687 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1688 {
1689 mutex_enter(&dn->dn_mtx);
1690 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1691 mutex_exit(&dn->dn_mtx);
1692 return;
1693 }
1694 dn->dn_free_txg = tx->tx_txg;
1695 mutex_exit(&dn->dn_mtx);
1696
1697 dnode_setdirty(dn, tx);
1698 }
1699
1700 /*
1701 * Try to change the block size for the indicated dnode. This can only
1702 * succeed if there are no blocks allocated or dirty beyond first block
1703 */
1704 int
1705 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1706 {
1707 dmu_buf_impl_t *db;
1708 int err;
1709
1710 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1711 if (size == 0)
1712 size = SPA_MINBLOCKSIZE;
1713 else
1714 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1715
1716 if (ibs == dn->dn_indblkshift)
1717 ibs = 0;
1718
1719 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1720 return (0);
1721
1722 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1723
1724 /* Check for any allocated blocks beyond the first */
1725 if (dn->dn_maxblkid != 0)
1726 goto fail;
1727
1728 mutex_enter(&dn->dn_dbufs_mtx);
1729 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1730 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1731 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1732 db->db_blkid != DMU_SPILL_BLKID) {
1733 mutex_exit(&dn->dn_dbufs_mtx);
1734 goto fail;
1735 }
1736 }
1737 mutex_exit(&dn->dn_dbufs_mtx);
1738
1739 if (ibs && dn->dn_nlevels != 1)
1740 goto fail;
1741
1742 /* resize the old block */
1743 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1744 if (err == 0)
1745 dbuf_new_size(db, size, tx);
1746 else if (err != ENOENT)
1747 goto fail;
1748
1749 dnode_setdblksz(dn, size);
1750 dnode_setdirty(dn, tx);
1751 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1752 if (ibs) {
1753 dn->dn_indblkshift = ibs;
1754 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1755 }
1756 /* rele after we have fixed the blocksize in the dnode */
1757 if (db)
1758 dbuf_rele(db, FTAG);
1759
1760 rw_exit(&dn->dn_struct_rwlock);
1761 return (0);
1762
1763 fail:
1764 rw_exit(&dn->dn_struct_rwlock);
1765 return (SET_ERROR(ENOTSUP));
1766 }
1767
1768 static void
1769 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1770 {
1771 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1772 int old_nlevels = dn->dn_nlevels;
1773 dmu_buf_impl_t *db;
1774 list_t *list;
1775 dbuf_dirty_record_t *new, *dr, *dr_next;
1776
1777 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1778
1779 dn->dn_nlevels = new_nlevels;
1780
1781 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1782 dn->dn_next_nlevels[txgoff] = new_nlevels;
1783
1784 /* dirty the left indirects */
1785 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1786 ASSERT(db != NULL);
1787 new = dbuf_dirty(db, tx);
1788 dbuf_rele(db, FTAG);
1789
1790 /* transfer the dirty records to the new indirect */
1791 mutex_enter(&dn->dn_mtx);
1792 mutex_enter(&new->dt.di.dr_mtx);
1793 list = &dn->dn_dirty_records[txgoff];
1794 for (dr = list_head(list); dr; dr = dr_next) {
1795 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1796 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1797 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1798 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1799 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1800 list_remove(&dn->dn_dirty_records[txgoff], dr);
1801 list_insert_tail(&new->dt.di.dr_children, dr);
1802 dr->dr_parent = new;
1803 }
1804 }
1805 mutex_exit(&new->dt.di.dr_mtx);
1806 mutex_exit(&dn->dn_mtx);
1807 }
1808
1809 int
1810 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1811 {
1812 int ret = 0;
1813
1814 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1815
1816 if (dn->dn_nlevels == nlevels) {
1817 ret = 0;
1818 goto out;
1819 } else if (nlevels < dn->dn_nlevels) {
1820 ret = SET_ERROR(EINVAL);
1821 goto out;
1822 }
1823
1824 dnode_set_nlevels_impl(dn, nlevels, tx);
1825
1826 out:
1827 rw_exit(&dn->dn_struct_rwlock);
1828 return (ret);
1829 }
1830
1831 /* read-holding callers must not rely on the lock being continuously held */
1832 void
1833 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1834 boolean_t force)
1835 {
1836 int epbs, new_nlevels;
1837 uint64_t sz;
1838
1839 ASSERT(blkid != DMU_BONUS_BLKID);
1840
1841 ASSERT(have_read ?
1842 RW_READ_HELD(&dn->dn_struct_rwlock) :
1843 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1844
1845 /*
1846 * if we have a read-lock, check to see if we need to do any work
1847 * before upgrading to a write-lock.
1848 */
1849 if (have_read) {
1850 if (blkid <= dn->dn_maxblkid)
1851 return;
1852
1853 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1854 rw_exit(&dn->dn_struct_rwlock);
1855 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1856 }
1857 }
1858
1859 /*
1860 * Raw sends (indicated by the force flag) require that we take the
1861 * given blkid even if the value is lower than the current value.
1862 */
1863 if (!force && blkid <= dn->dn_maxblkid)
1864 goto out;
1865
1866 /*
1867 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1868 * to indicate that this field is set. This allows us to set the
1869 * maxblkid to 0 on an existing object in dnode_sync().
1870 */
1871 dn->dn_maxblkid = blkid;
1872 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1873 blkid | DMU_NEXT_MAXBLKID_SET;
1874
1875 /*
1876 * Compute the number of levels necessary to support the new maxblkid.
1877 * Raw sends will ensure nlevels is set correctly for us.
1878 */
1879 new_nlevels = 1;
1880 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1881 for (sz = dn->dn_nblkptr;
1882 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1883 new_nlevels++;
1884
1885 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
1886
1887 if (!force) {
1888 if (new_nlevels > dn->dn_nlevels)
1889 dnode_set_nlevels_impl(dn, new_nlevels, tx);
1890 } else {
1891 ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1892 }
1893
1894 out:
1895 if (have_read)
1896 rw_downgrade(&dn->dn_struct_rwlock);
1897 }
1898
1899 static void
1900 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1901 {
1902 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1903 if (db != NULL) {
1904 dmu_buf_will_dirty(&db->db, tx);
1905 dbuf_rele(db, FTAG);
1906 }
1907 }
1908
1909 /*
1910 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1911 * and end_blkid.
1912 */
1913 static void
1914 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1915 dmu_tx_t *tx)
1916 {
1917 dmu_buf_impl_t db_search;
1918 dmu_buf_impl_t *db;
1919 avl_index_t where;
1920
1921 mutex_enter(&dn->dn_dbufs_mtx);
1922
1923 db_search.db_level = 1;
1924 db_search.db_blkid = start_blkid + 1;
1925 db_search.db_state = DB_SEARCH;
1926 for (;;) {
1927
1928 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1929 if (db == NULL)
1930 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1931
1932 if (db == NULL || db->db_level != 1 ||
1933 db->db_blkid >= end_blkid) {
1934 break;
1935 }
1936
1937 /*
1938 * Setup the next blkid we want to search for.
1939 */
1940 db_search.db_blkid = db->db_blkid + 1;
1941 ASSERT3U(db->db_blkid, >=, start_blkid);
1942
1943 /*
1944 * If the dbuf transitions to DB_EVICTING while we're trying
1945 * to dirty it, then we will be unable to discover it in
1946 * the dbuf hash table. This will result in a call to
1947 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1948 * lock. To avoid a deadlock, we drop the lock before
1949 * dirtying the level-1 dbuf.
1950 */
1951 mutex_exit(&dn->dn_dbufs_mtx);
1952 dnode_dirty_l1(dn, db->db_blkid, tx);
1953 mutex_enter(&dn->dn_dbufs_mtx);
1954 }
1955
1956 #ifdef ZFS_DEBUG
1957 /*
1958 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1959 */
1960 db_search.db_level = 1;
1961 db_search.db_blkid = start_blkid + 1;
1962 db_search.db_state = DB_SEARCH;
1963 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1964 if (db == NULL)
1965 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1966 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
1967 if (db->db_level != 1 || db->db_blkid >= end_blkid)
1968 break;
1969 ASSERT(db->db_dirtycnt > 0);
1970 }
1971 #endif
1972 mutex_exit(&dn->dn_dbufs_mtx);
1973 }
1974
1975 void
1976 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1977 {
1978 dmu_buf_impl_t *db;
1979 uint64_t blkoff, blkid, nblks;
1980 int blksz, blkshift, head, tail;
1981 int trunc = FALSE;
1982 int epbs;
1983
1984 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1985 blksz = dn->dn_datablksz;
1986 blkshift = dn->dn_datablkshift;
1987 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1988
1989 if (len == DMU_OBJECT_END) {
1990 len = UINT64_MAX - off;
1991 trunc = TRUE;
1992 }
1993
1994 /*
1995 * First, block align the region to free:
1996 */
1997 if (ISP2(blksz)) {
1998 head = P2NPHASE(off, blksz);
1999 blkoff = P2PHASE(off, blksz);
2000 if ((off >> blkshift) > dn->dn_maxblkid)
2001 goto out;
2002 } else {
2003 ASSERT(dn->dn_maxblkid == 0);
2004 if (off == 0 && len >= blksz) {
2005 /*
2006 * Freeing the whole block; fast-track this request.
2007 */
2008 blkid = 0;
2009 nblks = 1;
2010 if (dn->dn_nlevels > 1)
2011 dnode_dirty_l1(dn, 0, tx);
2012 goto done;
2013 } else if (off >= blksz) {
2014 /* Freeing past end-of-data */
2015 goto out;
2016 } else {
2017 /* Freeing part of the block. */
2018 head = blksz - off;
2019 ASSERT3U(head, >, 0);
2020 }
2021 blkoff = off;
2022 }
2023 /* zero out any partial block data at the start of the range */
2024 if (head) {
2025 ASSERT3U(blkoff + head, ==, blksz);
2026 if (len < head)
2027 head = len;
2028 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2029 TRUE, FALSE, FTAG, &db) == 0) {
2030 caddr_t data;
2031
2032 /* don't dirty if it isn't on disk and isn't dirty */
2033 if (db->db_last_dirty ||
2034 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2035 rw_exit(&dn->dn_struct_rwlock);
2036 dmu_buf_will_dirty(&db->db, tx);
2037 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2038 data = db->db.db_data;
2039 bzero(data + blkoff, head);
2040 }
2041 dbuf_rele(db, FTAG);
2042 }
2043 off += head;
2044 len -= head;
2045 }
2046
2047 /* If the range was less than one block, we're done */
2048 if (len == 0)
2049 goto out;
2050
2051 /* If the remaining range is past end of file, we're done */
2052 if ((off >> blkshift) > dn->dn_maxblkid)
2053 goto out;
2054
2055 ASSERT(ISP2(blksz));
2056 if (trunc)
2057 tail = 0;
2058 else
2059 tail = P2PHASE(len, blksz);
2060
2061 ASSERT0(P2PHASE(off, blksz));
2062 /* zero out any partial block data at the end of the range */
2063 if (tail) {
2064 if (len < tail)
2065 tail = len;
2066 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2067 TRUE, FALSE, FTAG, &db) == 0) {
2068 /* don't dirty if not on disk and not dirty */
2069 if (db->db_last_dirty ||
2070 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2071 rw_exit(&dn->dn_struct_rwlock);
2072 dmu_buf_will_dirty(&db->db, tx);
2073 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2074 bzero(db->db.db_data, tail);
2075 }
2076 dbuf_rele(db, FTAG);
2077 }
2078 len -= tail;
2079 }
2080
2081 /* If the range did not include a full block, we are done */
2082 if (len == 0)
2083 goto out;
2084
2085 ASSERT(IS_P2ALIGNED(off, blksz));
2086 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2087 blkid = off >> blkshift;
2088 nblks = len >> blkshift;
2089 if (trunc)
2090 nblks += 1;
2091
2092 /*
2093 * Dirty all the indirect blocks in this range. Note that only
2094 * the first and last indirect blocks can actually be written
2095 * (if they were partially freed) -- they must be dirtied, even if
2096 * they do not exist on disk yet. The interior blocks will
2097 * be freed by free_children(), so they will not actually be written.
2098 * Even though these interior blocks will not be written, we
2099 * dirty them for two reasons:
2100 *
2101 * - It ensures that the indirect blocks remain in memory until
2102 * syncing context. (They have already been prefetched by
2103 * dmu_tx_hold_free(), so we don't have to worry about reading
2104 * them serially here.)
2105 *
2106 * - The dirty space accounting will put pressure on the txg sync
2107 * mechanism to begin syncing, and to delay transactions if there
2108 * is a large amount of freeing. Even though these indirect
2109 * blocks will not be written, we could need to write the same
2110 * amount of space if we copy the freed BPs into deadlists.
2111 */
2112 if (dn->dn_nlevels > 1) {
2113 uint64_t first, last;
2114
2115 first = blkid >> epbs;
2116 dnode_dirty_l1(dn, first, tx);
2117 if (trunc)
2118 last = dn->dn_maxblkid >> epbs;
2119 else
2120 last = (blkid + nblks - 1) >> epbs;
2121 if (last != first)
2122 dnode_dirty_l1(dn, last, tx);
2123
2124 dnode_dirty_l1range(dn, first, last, tx);
2125
2126 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2127 SPA_BLKPTRSHIFT;
2128 for (uint64_t i = first + 1; i < last; i++) {
2129 /*
2130 * Set i to the blockid of the next non-hole
2131 * level-1 indirect block at or after i. Note
2132 * that dnode_next_offset() operates in terms of
2133 * level-0-equivalent bytes.
2134 */
2135 uint64_t ibyte = i << shift;
2136 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2137 &ibyte, 2, 1, 0);
2138 i = ibyte >> shift;
2139 if (i >= last)
2140 break;
2141
2142 /*
2143 * Normally we should not see an error, either
2144 * from dnode_next_offset() or dbuf_hold_level()
2145 * (except for ESRCH from dnode_next_offset).
2146 * If there is an i/o error, then when we read
2147 * this block in syncing context, it will use
2148 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2149 * to the "failmode" property. dnode_next_offset()
2150 * doesn't have a flag to indicate MUSTSUCCEED.
2151 */
2152 if (err != 0)
2153 break;
2154
2155 dnode_dirty_l1(dn, i, tx);
2156 }
2157 }
2158
2159 done:
2160 /*
2161 * Add this range to the dnode range list.
2162 * We will finish up this free operation in the syncing phase.
2163 */
2164 mutex_enter(&dn->dn_mtx);
2165 {
2166 int txgoff = tx->tx_txg & TXG_MASK;
2167 if (dn->dn_free_ranges[txgoff] == NULL) {
2168 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2169 }
2170 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2171 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2172 }
2173 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2174 blkid, nblks, tx->tx_txg);
2175 mutex_exit(&dn->dn_mtx);
2176
2177 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2178 dnode_setdirty(dn, tx);
2179 out:
2180
2181 rw_exit(&dn->dn_struct_rwlock);
2182 }
2183
2184 static boolean_t
2185 dnode_spill_freed(dnode_t *dn)
2186 {
2187 int i;
2188
2189 mutex_enter(&dn->dn_mtx);
2190 for (i = 0; i < TXG_SIZE; i++) {
2191 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2192 break;
2193 }
2194 mutex_exit(&dn->dn_mtx);
2195 return (i < TXG_SIZE);
2196 }
2197
2198 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2199 uint64_t
2200 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2201 {
2202 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2203 int i;
2204
2205 if (blkid == DMU_BONUS_BLKID)
2206 return (FALSE);
2207
2208 /*
2209 * If we're in the process of opening the pool, dp will not be
2210 * set yet, but there shouldn't be anything dirty.
2211 */
2212 if (dp == NULL)
2213 return (FALSE);
2214
2215 if (dn->dn_free_txg)
2216 return (TRUE);
2217
2218 if (blkid == DMU_SPILL_BLKID)
2219 return (dnode_spill_freed(dn));
2220
2221 mutex_enter(&dn->dn_mtx);
2222 for (i = 0; i < TXG_SIZE; i++) {
2223 if (dn->dn_free_ranges[i] != NULL &&
2224 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2225 break;
2226 }
2227 mutex_exit(&dn->dn_mtx);
2228 return (i < TXG_SIZE);
2229 }
2230
2231 /* call from syncing context when we actually write/free space for this dnode */
2232 void
2233 dnode_diduse_space(dnode_t *dn, int64_t delta)
2234 {
2235 uint64_t space;
2236 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2237 dn, dn->dn_phys,
2238 (u_longlong_t)dn->dn_phys->dn_used,
2239 (longlong_t)delta);
2240
2241 mutex_enter(&dn->dn_mtx);
2242 space = DN_USED_BYTES(dn->dn_phys);
2243 if (delta > 0) {
2244 ASSERT3U(space + delta, >=, space); /* no overflow */
2245 } else {
2246 ASSERT3U(space, >=, -delta); /* no underflow */
2247 }
2248 space += delta;
2249 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2250 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2251 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2252 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2253 } else {
2254 dn->dn_phys->dn_used = space;
2255 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2256 }
2257 mutex_exit(&dn->dn_mtx);
2258 }
2259
2260 /*
2261 * Scans a block at the indicated "level" looking for a hole or data,
2262 * depending on 'flags'.
2263 *
2264 * If level > 0, then we are scanning an indirect block looking at its
2265 * pointers. If level == 0, then we are looking at a block of dnodes.
2266 *
2267 * If we don't find what we are looking for in the block, we return ESRCH.
2268 * Otherwise, return with *offset pointing to the beginning (if searching
2269 * forwards) or end (if searching backwards) of the range covered by the
2270 * block pointer we matched on (or dnode).
2271 *
2272 * The basic search algorithm used below by dnode_next_offset() is to
2273 * use this function to search up the block tree (widen the search) until
2274 * we find something (i.e., we don't return ESRCH) and then search back
2275 * down the tree (narrow the search) until we reach our original search
2276 * level.
2277 */
2278 static int
2279 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2280 int lvl, uint64_t blkfill, uint64_t txg)
2281 {
2282 dmu_buf_impl_t *db = NULL;
2283 void *data = NULL;
2284 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2285 uint64_t epb = 1ULL << epbs;
2286 uint64_t minfill, maxfill;
2287 boolean_t hole;
2288 int i, inc, error, span;
2289
2290 hole = ((flags & DNODE_FIND_HOLE) != 0);
2291 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2292 ASSERT(txg == 0 || !hole);
2293
2294 if (lvl == dn->dn_phys->dn_nlevels) {
2295 error = 0;
2296 epb = dn->dn_phys->dn_nblkptr;
2297 data = dn->dn_phys->dn_blkptr;
2298 } else {
2299 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2300 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2301 if (error) {
2302 if (error != ENOENT)
2303 return (error);
2304 if (hole)
2305 return (0);
2306 /*
2307 * This can only happen when we are searching up
2308 * the block tree for data. We don't really need to
2309 * adjust the offset, as we will just end up looking
2310 * at the pointer to this block in its parent, and its
2311 * going to be unallocated, so we will skip over it.
2312 */
2313 return (SET_ERROR(ESRCH));
2314 }
2315 error = dbuf_read(db, NULL,
2316 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2317 if (error) {
2318 dbuf_rele(db, FTAG);
2319 return (error);
2320 }
2321 data = db->db.db_data;
2322 }
2323
2324
2325 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2326 db->db_blkptr->blk_birth <= txg ||
2327 BP_IS_HOLE(db->db_blkptr))) {
2328 /*
2329 * This can only happen when we are searching up the tree
2330 * and these conditions mean that we need to keep climbing.
2331 */
2332 error = SET_ERROR(ESRCH);
2333 } else if (lvl == 0) {
2334 dnode_phys_t *dnp = data;
2335
2336 ASSERT(dn->dn_type == DMU_OT_DNODE);
2337 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2338
2339 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2340 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2341 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2342 break;
2343 }
2344
2345 if (i == blkfill)
2346 error = SET_ERROR(ESRCH);
2347
2348 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2349 (i << DNODE_SHIFT);
2350 } else {
2351 blkptr_t *bp = data;
2352 uint64_t start = *offset;
2353 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2354 minfill = 0;
2355 maxfill = blkfill << ((lvl - 1) * epbs);
2356
2357 if (hole)
2358 maxfill--;
2359 else
2360 minfill++;
2361
2362 if (span >= 8 * sizeof (*offset)) {
2363 /* This only happens on the highest indirection level */
2364 ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
2365 *offset = 0;
2366 } else {
2367 *offset = *offset >> span;
2368 }
2369
2370 for (i = BF64_GET(*offset, 0, epbs);
2371 i >= 0 && i < epb; i += inc) {
2372 if (BP_GET_FILL(&bp[i]) >= minfill &&
2373 BP_GET_FILL(&bp[i]) <= maxfill &&
2374 (hole || bp[i].blk_birth > txg))
2375 break;
2376 if (inc > 0 || *offset > 0)
2377 *offset += inc;
2378 }
2379
2380 if (span >= 8 * sizeof (*offset)) {
2381 *offset = start;
2382 } else {
2383 *offset = *offset << span;
2384 }
2385
2386 if (inc < 0) {
2387 /* traversing backwards; position offset at the end */
2388 ASSERT3U(*offset, <=, start);
2389 *offset = MIN(*offset + (1ULL << span) - 1, start);
2390 } else if (*offset < start) {
2391 *offset = start;
2392 }
2393 if (i < 0 || i >= epb)
2394 error = SET_ERROR(ESRCH);
2395 }
2396
2397 if (db)
2398 dbuf_rele(db, FTAG);
2399
2400 return (error);
2401 }
2402
2403 /*
2404 * Find the next hole, data, or sparse region at or after *offset.
2405 * The value 'blkfill' tells us how many items we expect to find
2406 * in an L0 data block; this value is 1 for normal objects,
2407 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2408 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2409 *
2410 * Examples:
2411 *
2412 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2413 * Finds the next/previous hole/data in a file.
2414 * Used in dmu_offset_next().
2415 *
2416 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2417 * Finds the next free/allocated dnode an objset's meta-dnode.
2418 * Only finds objects that have new contents since txg (ie.
2419 * bonus buffer changes and content removal are ignored).
2420 * Used in dmu_object_next().
2421 *
2422 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2423 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2424 * Used in dmu_object_alloc().
2425 */
2426 int
2427 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2428 int minlvl, uint64_t blkfill, uint64_t txg)
2429 {
2430 uint64_t initial_offset = *offset;
2431 int lvl, maxlvl;
2432 int error = 0;
2433
2434 if (!(flags & DNODE_FIND_HAVELOCK))
2435 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2436
2437 if (dn->dn_phys->dn_nlevels == 0) {
2438 error = SET_ERROR(ESRCH);
2439 goto out;
2440 }
2441
2442 if (dn->dn_datablkshift == 0) {
2443 if (*offset < dn->dn_datablksz) {
2444 if (flags & DNODE_FIND_HOLE)
2445 *offset = dn->dn_datablksz;
2446 } else {
2447 error = SET_ERROR(ESRCH);
2448 }
2449 goto out;
2450 }
2451
2452 maxlvl = dn->dn_phys->dn_nlevels;
2453
2454 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2455 error = dnode_next_offset_level(dn,
2456 flags, offset, lvl, blkfill, txg);
2457 if (error != ESRCH)
2458 break;
2459 }
2460
2461 while (error == 0 && --lvl >= minlvl) {
2462 error = dnode_next_offset_level(dn,
2463 flags, offset, lvl, blkfill, txg);
2464 }
2465
2466 /*
2467 * There's always a "virtual hole" at the end of the object, even
2468 * if all BP's which physically exist are non-holes.
2469 */
2470 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2471 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2472 error = 0;
2473 }
2474
2475 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2476 initial_offset < *offset : initial_offset > *offset))
2477 error = SET_ERROR(ESRCH);
2478 out:
2479 if (!(flags & DNODE_FIND_HAVELOCK))
2480 rw_exit(&dn->dn_struct_rwlock);
2481
2482 return (error);
2483 }
ZFS On Linux mirror for PVE