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 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 RackTop Systems.
27 */
28
29 #include <sys/zfs_context.h>
30 #include <sys/dbuf.h>
31 #include <sys/dnode.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/spa.h>
39 #include <sys/zio.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/range_tree.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_hold_free_txg", KSTAT_DATA_UINT64 },
60 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 },
61 { "dnode_allocate", KSTAT_DATA_UINT64 },
62 { "dnode_reallocate", KSTAT_DATA_UINT64 },
63 { "dnode_buf_evict", KSTAT_DATA_UINT64 },
64 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 },
65 { "dnode_alloc_race", KSTAT_DATA_UINT64 },
66 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 },
67 { "dnode_move_invalid", KSTAT_DATA_UINT64 },
68 { "dnode_move_recheck1", KSTAT_DATA_UINT64 },
69 { "dnode_move_recheck2", KSTAT_DATA_UINT64 },
70 { "dnode_move_special", KSTAT_DATA_UINT64 },
71 { "dnode_move_handle", KSTAT_DATA_UINT64 },
72 { "dnode_move_rwlock", KSTAT_DATA_UINT64 },
73 { "dnode_move_active", KSTAT_DATA_UINT64 },
74 };
75
76 static kstat_t *dnode_ksp;
77 static kmem_cache_t *dnode_cache;
78
79 static dnode_phys_t dnode_phys_zero;
80
81 int zfs_default_bs = SPA_MINBLOCKSHIFT;
82 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
83
84 #ifdef _KERNEL
85 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
86 #endif /* _KERNEL */
87
88 static int
89 dbuf_compare(const void *x1, const void *x2)
90 {
91 const dmu_buf_impl_t *d1 = x1;
92 const dmu_buf_impl_t *d2 = x2;
93
94 int cmp = AVL_CMP(d1->db_level, d2->db_level);
95 if (likely(cmp))
96 return (cmp);
97
98 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
99 if (likely(cmp))
100 return (cmp);
101
102 if (d1->db_state == DB_SEARCH) {
103 ASSERT3S(d2->db_state, !=, DB_SEARCH);
104 return (-1);
105 } else if (d2->db_state == DB_SEARCH) {
106 ASSERT3S(d1->db_state, !=, DB_SEARCH);
107 return (1);
108 }
109
110 return (AVL_PCMP(d1, d2));
111 }
112
113 /* ARGSUSED */
114 static int
115 dnode_cons(void *arg, void *unused, int kmflag)
116 {
117 dnode_t *dn = arg;
118 int i;
119
120 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
121 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
122 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
123 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
124
125 /*
126 * Every dbuf has a reference, and dropping a tracked reference is
127 * O(number of references), so don't track dn_holds.
128 */
129 zfs_refcount_create_untracked(&dn->dn_holds);
130 zfs_refcount_create(&dn->dn_tx_holds);
131 list_link_init(&dn->dn_link);
132
133 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
134 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
135 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
136 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
137 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
138 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
139 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
140 bzero(&dn->dn_next_maxblkid[0], sizeof (dn->dn_next_maxblkid));
141
142 for (i = 0; i < TXG_SIZE; i++) {
143 multilist_link_init(&dn->dn_dirty_link[i]);
144 dn->dn_free_ranges[i] = NULL;
145 list_create(&dn->dn_dirty_records[i],
146 sizeof (dbuf_dirty_record_t),
147 offsetof(dbuf_dirty_record_t, dr_dirty_node));
148 }
149
150 dn->dn_allocated_txg = 0;
151 dn->dn_free_txg = 0;
152 dn->dn_assigned_txg = 0;
153 dn->dn_dirty_txg = 0;
154 dn->dn_dirtyctx = 0;
155 dn->dn_dirtyctx_firstset = NULL;
156 dn->dn_bonus = NULL;
157 dn->dn_have_spill = B_FALSE;
158 dn->dn_zio = NULL;
159 dn->dn_oldused = 0;
160 dn->dn_oldflags = 0;
161 dn->dn_olduid = 0;
162 dn->dn_oldgid = 0;
163 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
164 dn->dn_newuid = 0;
165 dn->dn_newgid = 0;
166 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
167 dn->dn_id_flags = 0;
168
169 dn->dn_dbufs_count = 0;
170 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
171 offsetof(dmu_buf_impl_t, db_link));
172
173 dn->dn_moved = 0;
174 return (0);
175 }
176
177 /* ARGSUSED */
178 static void
179 dnode_dest(void *arg, void *unused)
180 {
181 int i;
182 dnode_t *dn = arg;
183
184 rw_destroy(&dn->dn_struct_rwlock);
185 mutex_destroy(&dn->dn_mtx);
186 mutex_destroy(&dn->dn_dbufs_mtx);
187 cv_destroy(&dn->dn_notxholds);
188 zfs_refcount_destroy(&dn->dn_holds);
189 zfs_refcount_destroy(&dn->dn_tx_holds);
190 ASSERT(!list_link_active(&dn->dn_link));
191
192 for (i = 0; i < TXG_SIZE; i++) {
193 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
194 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
195 list_destroy(&dn->dn_dirty_records[i]);
196 ASSERT0(dn->dn_next_nblkptr[i]);
197 ASSERT0(dn->dn_next_nlevels[i]);
198 ASSERT0(dn->dn_next_indblkshift[i]);
199 ASSERT0(dn->dn_next_bonustype[i]);
200 ASSERT0(dn->dn_rm_spillblk[i]);
201 ASSERT0(dn->dn_next_bonuslen[i]);
202 ASSERT0(dn->dn_next_blksz[i]);
203 ASSERT0(dn->dn_next_maxblkid[i]);
204 }
205
206 ASSERT0(dn->dn_allocated_txg);
207 ASSERT0(dn->dn_free_txg);
208 ASSERT0(dn->dn_assigned_txg);
209 ASSERT0(dn->dn_dirty_txg);
210 ASSERT0(dn->dn_dirtyctx);
211 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
212 ASSERT3P(dn->dn_bonus, ==, NULL);
213 ASSERT(!dn->dn_have_spill);
214 ASSERT3P(dn->dn_zio, ==, NULL);
215 ASSERT0(dn->dn_oldused);
216 ASSERT0(dn->dn_oldflags);
217 ASSERT0(dn->dn_olduid);
218 ASSERT0(dn->dn_oldgid);
219 ASSERT0(dn->dn_oldprojid);
220 ASSERT0(dn->dn_newuid);
221 ASSERT0(dn->dn_newgid);
222 ASSERT0(dn->dn_newprojid);
223 ASSERT0(dn->dn_id_flags);
224
225 ASSERT0(dn->dn_dbufs_count);
226 avl_destroy(&dn->dn_dbufs);
227 }
228
229 void
230 dnode_init(void)
231 {
232 ASSERT(dnode_cache == NULL);
233 dnode_cache = kmem_cache_create("dnode_t",
234 sizeof (dnode_t),
235 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
236 #ifdef _KERNEL
237 kmem_cache_set_move(dnode_cache, dnode_move);
238
239 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc",
240 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t),
241 KSTAT_FLAG_VIRTUAL);
242 if (dnode_ksp != NULL) {
243 dnode_ksp->ks_data = &dnode_stats;
244 kstat_install(dnode_ksp);
245 }
246 #endif /* _KERNEL */
247 }
248
249 void
250 dnode_fini(void)
251 {
252 if (dnode_ksp != NULL) {
253 kstat_delete(dnode_ksp);
254 dnode_ksp = NULL;
255 }
256
257 kmem_cache_destroy(dnode_cache);
258 dnode_cache = NULL;
259 }
260
261
262 #ifdef ZFS_DEBUG
263 void
264 dnode_verify(dnode_t *dn)
265 {
266 int drop_struct_lock = FALSE;
267
268 ASSERT(dn->dn_phys);
269 ASSERT(dn->dn_objset);
270 ASSERT(dn->dn_handle->dnh_dnode == dn);
271
272 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
273
274 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
275 return;
276
277 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
278 rw_enter(&dn->dn_struct_rwlock, RW_READER);
279 drop_struct_lock = TRUE;
280 }
281 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
282 int i;
283 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
284 ASSERT3U(dn->dn_indblkshift, >=, 0);
285 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
286 if (dn->dn_datablkshift) {
287 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
288 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
289 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
290 }
291 ASSERT3U(dn->dn_nlevels, <=, 30);
292 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
293 ASSERT3U(dn->dn_nblkptr, >=, 1);
294 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
295 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
296 ASSERT3U(dn->dn_datablksz, ==,
297 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
298 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
299 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
300 dn->dn_bonuslen, <=, max_bonuslen);
301 for (i = 0; i < TXG_SIZE; i++) {
302 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
303 }
304 }
305 if (dn->dn_phys->dn_type != DMU_OT_NONE)
306 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
307 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
308 if (dn->dn_dbuf != NULL) {
309 ASSERT3P(dn->dn_phys, ==,
310 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
311 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
312 }
313 if (drop_struct_lock)
314 rw_exit(&dn->dn_struct_rwlock);
315 }
316 #endif
317
318 void
319 dnode_byteswap(dnode_phys_t *dnp)
320 {
321 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
322 int i;
323
324 if (dnp->dn_type == DMU_OT_NONE) {
325 bzero(dnp, sizeof (dnode_phys_t));
326 return;
327 }
328
329 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
330 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
331 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
332 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
333 dnp->dn_used = BSWAP_64(dnp->dn_used);
334
335 /*
336 * dn_nblkptr is only one byte, so it's OK to read it in either
337 * byte order. We can't read dn_bouslen.
338 */
339 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
340 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
341 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
342 buf64[i] = BSWAP_64(buf64[i]);
343
344 /*
345 * OK to check dn_bonuslen for zero, because it won't matter if
346 * we have the wrong byte order. This is necessary because the
347 * dnode dnode is smaller than a regular dnode.
348 */
349 if (dnp->dn_bonuslen != 0) {
350 /*
351 * Note that the bonus length calculated here may be
352 * longer than the actual bonus buffer. This is because
353 * we always put the bonus buffer after the last block
354 * pointer (instead of packing it against the end of the
355 * dnode buffer).
356 */
357 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
358 int slots = dnp->dn_extra_slots + 1;
359 size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
360 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
361 dmu_object_byteswap_t byteswap =
362 DMU_OT_BYTESWAP(dnp->dn_bonustype);
363 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
364 }
365
366 /* Swap SPILL block if we have one */
367 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
368 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
369
370 }
371
372 void
373 dnode_buf_byteswap(void *vbuf, size_t size)
374 {
375 int i = 0;
376
377 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
378 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
379
380 while (i < size) {
381 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i);
382 dnode_byteswap(dnp);
383
384 i += DNODE_MIN_SIZE;
385 if (dnp->dn_type != DMU_OT_NONE)
386 i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
387 }
388 }
389
390 void
391 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
392 {
393 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
394
395 dnode_setdirty(dn, tx);
396 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
397 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
398 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
399 dn->dn_bonuslen = newsize;
400 if (newsize == 0)
401 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
402 else
403 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
404 rw_exit(&dn->dn_struct_rwlock);
405 }
406
407 void
408 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
409 {
410 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
411 dnode_setdirty(dn, tx);
412 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
413 dn->dn_bonustype = newtype;
414 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
415 rw_exit(&dn->dn_struct_rwlock);
416 }
417
418 void
419 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
420 {
421 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1);
422 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
423 dnode_setdirty(dn, tx);
424 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
425 dn->dn_have_spill = B_FALSE;
426 }
427
428 static void
429 dnode_setdblksz(dnode_t *dn, int size)
430 {
431 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
432 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
433 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
434 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
435 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
436 dn->dn_datablksz = size;
437 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
438 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
439 }
440
441 static dnode_t *
442 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
443 uint64_t object, dnode_handle_t *dnh)
444 {
445 dnode_t *dn;
446
447 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
448 #ifdef _KERNEL
449 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
450 #endif /* _KERNEL */
451 dn->dn_moved = 0;
452
453 /*
454 * Defer setting dn_objset until the dnode is ready to be a candidate
455 * for the dnode_move() callback.
456 */
457 dn->dn_object = object;
458 dn->dn_dbuf = db;
459 dn->dn_handle = dnh;
460 dn->dn_phys = dnp;
461
462 if (dnp->dn_datablkszsec) {
463 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
464 } else {
465 dn->dn_datablksz = 0;
466 dn->dn_datablkszsec = 0;
467 dn->dn_datablkshift = 0;
468 }
469 dn->dn_indblkshift = dnp->dn_indblkshift;
470 dn->dn_nlevels = dnp->dn_nlevels;
471 dn->dn_type = dnp->dn_type;
472 dn->dn_nblkptr = dnp->dn_nblkptr;
473 dn->dn_checksum = dnp->dn_checksum;
474 dn->dn_compress = dnp->dn_compress;
475 dn->dn_bonustype = dnp->dn_bonustype;
476 dn->dn_bonuslen = dnp->dn_bonuslen;
477 dn->dn_num_slots = dnp->dn_extra_slots + 1;
478 dn->dn_maxblkid = dnp->dn_maxblkid;
479 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
480 dn->dn_id_flags = 0;
481
482 dmu_zfetch_init(&dn->dn_zfetch, dn);
483
484 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
485 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
486 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode));
487
488 mutex_enter(&os->os_lock);
489
490 /*
491 * Exclude special dnodes from os_dnodes so an empty os_dnodes
492 * signifies that the special dnodes have no references from
493 * their children (the entries in os_dnodes). This allows
494 * dnode_destroy() to easily determine if the last child has
495 * been removed and then complete eviction of the objset.
496 */
497 if (!DMU_OBJECT_IS_SPECIAL(object))
498 list_insert_head(&os->os_dnodes, dn);
499 membar_producer();
500
501 /*
502 * Everything else must be valid before assigning dn_objset
503 * makes the dnode eligible for dnode_move().
504 */
505 dn->dn_objset = os;
506
507 dnh->dnh_dnode = dn;
508 mutex_exit(&os->os_lock);
509
510 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
511
512 return (dn);
513 }
514
515 /*
516 * Caller must be holding the dnode handle, which is released upon return.
517 */
518 static void
519 dnode_destroy(dnode_t *dn)
520 {
521 objset_t *os = dn->dn_objset;
522 boolean_t complete_os_eviction = B_FALSE;
523
524 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
525
526 mutex_enter(&os->os_lock);
527 POINTER_INVALIDATE(&dn->dn_objset);
528 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
529 list_remove(&os->os_dnodes, dn);
530 complete_os_eviction =
531 list_is_empty(&os->os_dnodes) &&
532 list_link_active(&os->os_evicting_node);
533 }
534 mutex_exit(&os->os_lock);
535
536 /* the dnode can no longer move, so we can release the handle */
537 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock))
538 zrl_remove(&dn->dn_handle->dnh_zrlock);
539
540 dn->dn_allocated_txg = 0;
541 dn->dn_free_txg = 0;
542 dn->dn_assigned_txg = 0;
543 dn->dn_dirty_txg = 0;
544
545 dn->dn_dirtyctx = 0;
546 if (dn->dn_dirtyctx_firstset != NULL) {
547 kmem_free(dn->dn_dirtyctx_firstset, 1);
548 dn->dn_dirtyctx_firstset = NULL;
549 }
550 if (dn->dn_bonus != NULL) {
551 mutex_enter(&dn->dn_bonus->db_mtx);
552 dbuf_destroy(dn->dn_bonus);
553 dn->dn_bonus = NULL;
554 }
555 dn->dn_zio = NULL;
556
557 dn->dn_have_spill = B_FALSE;
558 dn->dn_oldused = 0;
559 dn->dn_oldflags = 0;
560 dn->dn_olduid = 0;
561 dn->dn_oldgid = 0;
562 dn->dn_oldprojid = ZFS_DEFAULT_PROJID;
563 dn->dn_newuid = 0;
564 dn->dn_newgid = 0;
565 dn->dn_newprojid = ZFS_DEFAULT_PROJID;
566 dn->dn_id_flags = 0;
567
568 dmu_zfetch_fini(&dn->dn_zfetch);
569 kmem_cache_free(dnode_cache, dn);
570 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
571
572 if (complete_os_eviction)
573 dmu_objset_evict_done(os);
574 }
575
576 void
577 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
578 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
579 {
580 int i;
581
582 ASSERT3U(dn_slots, >, 0);
583 ASSERT3U(dn_slots << DNODE_SHIFT, <=,
584 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
585 ASSERT3U(blocksize, <=,
586 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
587 if (blocksize == 0)
588 blocksize = 1 << zfs_default_bs;
589 else
590 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
591
592 if (ibs == 0)
593 ibs = zfs_default_ibs;
594
595 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
596
597 dprintf("os=%p obj=%" PRIu64 " txg=%" PRIu64
598 " blocksize=%d ibs=%d dn_slots=%d\n",
599 dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
600 DNODE_STAT_BUMP(dnode_allocate);
601
602 ASSERT(dn->dn_type == DMU_OT_NONE);
603 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
604 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
605 ASSERT(ot != DMU_OT_NONE);
606 ASSERT(DMU_OT_IS_VALID(ot));
607 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
608 (bonustype == DMU_OT_SA && bonuslen == 0) ||
609 (bonustype != DMU_OT_NONE && bonuslen != 0));
610 ASSERT(DMU_OT_IS_VALID(bonustype));
611 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
612 ASSERT(dn->dn_type == DMU_OT_NONE);
613 ASSERT0(dn->dn_maxblkid);
614 ASSERT0(dn->dn_allocated_txg);
615 ASSERT0(dn->dn_dirty_txg);
616 ASSERT0(dn->dn_assigned_txg);
617 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
618 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1);
619 ASSERT(avl_is_empty(&dn->dn_dbufs));
620
621 for (i = 0; i < TXG_SIZE; i++) {
622 ASSERT0(dn->dn_next_nblkptr[i]);
623 ASSERT0(dn->dn_next_nlevels[i]);
624 ASSERT0(dn->dn_next_indblkshift[i]);
625 ASSERT0(dn->dn_next_bonuslen[i]);
626 ASSERT0(dn->dn_next_bonustype[i]);
627 ASSERT0(dn->dn_rm_spillblk[i]);
628 ASSERT0(dn->dn_next_blksz[i]);
629 ASSERT0(dn->dn_next_maxblkid[i]);
630 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i]));
631 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
632 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
633 }
634
635 dn->dn_type = ot;
636 dnode_setdblksz(dn, blocksize);
637 dn->dn_indblkshift = ibs;
638 dn->dn_nlevels = 1;
639 dn->dn_num_slots = dn_slots;
640 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
641 dn->dn_nblkptr = 1;
642 else {
643 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
644 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
645 SPA_BLKPTRSHIFT));
646 }
647
648 dn->dn_bonustype = bonustype;
649 dn->dn_bonuslen = bonuslen;
650 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
651 dn->dn_compress = ZIO_COMPRESS_INHERIT;
652 dn->dn_dirtyctx = 0;
653
654 dn->dn_free_txg = 0;
655 if (dn->dn_dirtyctx_firstset) {
656 kmem_free(dn->dn_dirtyctx_firstset, 1);
657 dn->dn_dirtyctx_firstset = NULL;
658 }
659
660 dn->dn_allocated_txg = tx->tx_txg;
661 dn->dn_id_flags = 0;
662
663 dnode_setdirty(dn, tx);
664 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
665 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
666 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
667 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
668 }
669
670 void
671 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
672 dmu_object_type_t bonustype, int bonuslen, int dn_slots,
673 boolean_t keep_spill, dmu_tx_t *tx)
674 {
675 int nblkptr;
676
677 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
678 ASSERT3U(blocksize, <=,
679 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
680 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
681 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
682 ASSERT(tx->tx_txg != 0);
683 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
684 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
685 (bonustype == DMU_OT_SA && bonuslen == 0));
686 ASSERT(DMU_OT_IS_VALID(bonustype));
687 ASSERT3U(bonuslen, <=,
688 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
689 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT));
690
691 dnode_free_interior_slots(dn);
692 DNODE_STAT_BUMP(dnode_reallocate);
693
694 /* clean up any unreferenced dbufs */
695 dnode_evict_dbufs(dn);
696
697 dn->dn_id_flags = 0;
698
699 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
700 dnode_setdirty(dn, tx);
701 if (dn->dn_datablksz != blocksize) {
702 /* change blocksize */
703 ASSERT(dn->dn_maxblkid == 0 &&
704 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
705 dnode_block_freed(dn, 0)));
706 dnode_setdblksz(dn, blocksize);
707 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
708 }
709 if (dn->dn_bonuslen != bonuslen)
710 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
711
712 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
713 nblkptr = 1;
714 else
715 nblkptr = MIN(DN_MAX_NBLKPTR,
716 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
717 SPA_BLKPTRSHIFT));
718 if (dn->dn_bonustype != bonustype)
719 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
720 if (dn->dn_nblkptr != nblkptr)
721 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
722 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) {
723 dbuf_rm_spill(dn, tx);
724 dnode_rm_spill(dn, tx);
725 }
726 rw_exit(&dn->dn_struct_rwlock);
727
728 /* change type */
729 dn->dn_type = ot;
730
731 /* change bonus size and type */
732 mutex_enter(&dn->dn_mtx);
733 dn->dn_bonustype = bonustype;
734 dn->dn_bonuslen = bonuslen;
735 dn->dn_num_slots = dn_slots;
736 dn->dn_nblkptr = nblkptr;
737 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
738 dn->dn_compress = ZIO_COMPRESS_INHERIT;
739 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
740
741 /* fix up the bonus db_size */
742 if (dn->dn_bonus) {
743 dn->dn_bonus->db.db_size =
744 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
745 (dn->dn_nblkptr - 1) * sizeof (blkptr_t);
746 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
747 }
748
749 dn->dn_allocated_txg = tx->tx_txg;
750 mutex_exit(&dn->dn_mtx);
751 }
752
753 #ifdef _KERNEL
754 static void
755 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
756 {
757 int i;
758
759 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
760 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
761 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
762 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
763
764 /* Copy fields. */
765 ndn->dn_objset = odn->dn_objset;
766 ndn->dn_object = odn->dn_object;
767 ndn->dn_dbuf = odn->dn_dbuf;
768 ndn->dn_handle = odn->dn_handle;
769 ndn->dn_phys = odn->dn_phys;
770 ndn->dn_type = odn->dn_type;
771 ndn->dn_bonuslen = odn->dn_bonuslen;
772 ndn->dn_bonustype = odn->dn_bonustype;
773 ndn->dn_nblkptr = odn->dn_nblkptr;
774 ndn->dn_checksum = odn->dn_checksum;
775 ndn->dn_compress = odn->dn_compress;
776 ndn->dn_nlevels = odn->dn_nlevels;
777 ndn->dn_indblkshift = odn->dn_indblkshift;
778 ndn->dn_datablkshift = odn->dn_datablkshift;
779 ndn->dn_datablkszsec = odn->dn_datablkszsec;
780 ndn->dn_datablksz = odn->dn_datablksz;
781 ndn->dn_maxblkid = odn->dn_maxblkid;
782 ndn->dn_num_slots = odn->dn_num_slots;
783 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
784 sizeof (odn->dn_next_type));
785 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
786 sizeof (odn->dn_next_nblkptr));
787 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
788 sizeof (odn->dn_next_nlevels));
789 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
790 sizeof (odn->dn_next_indblkshift));
791 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
792 sizeof (odn->dn_next_bonustype));
793 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
794 sizeof (odn->dn_rm_spillblk));
795 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
796 sizeof (odn->dn_next_bonuslen));
797 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
798 sizeof (odn->dn_next_blksz));
799 bcopy(&odn->dn_next_maxblkid[0], &ndn->dn_next_maxblkid[0],
800 sizeof (odn->dn_next_maxblkid));
801 for (i = 0; i < TXG_SIZE; i++) {
802 list_move_tail(&ndn->dn_dirty_records[i],
803 &odn->dn_dirty_records[i]);
804 }
805 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
806 sizeof (odn->dn_free_ranges));
807 ndn->dn_allocated_txg = odn->dn_allocated_txg;
808 ndn->dn_free_txg = odn->dn_free_txg;
809 ndn->dn_assigned_txg = odn->dn_assigned_txg;
810 ndn->dn_dirty_txg = odn->dn_dirty_txg;
811 ndn->dn_dirtyctx = odn->dn_dirtyctx;
812 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
813 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0);
814 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
815 ASSERT(avl_is_empty(&ndn->dn_dbufs));
816 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
817 ndn->dn_dbufs_count = odn->dn_dbufs_count;
818 ndn->dn_bonus = odn->dn_bonus;
819 ndn->dn_have_spill = odn->dn_have_spill;
820 ndn->dn_zio = odn->dn_zio;
821 ndn->dn_oldused = odn->dn_oldused;
822 ndn->dn_oldflags = odn->dn_oldflags;
823 ndn->dn_olduid = odn->dn_olduid;
824 ndn->dn_oldgid = odn->dn_oldgid;
825 ndn->dn_oldprojid = odn->dn_oldprojid;
826 ndn->dn_newuid = odn->dn_newuid;
827 ndn->dn_newgid = odn->dn_newgid;
828 ndn->dn_newprojid = odn->dn_newprojid;
829 ndn->dn_id_flags = odn->dn_id_flags;
830 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
831 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
832 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
833
834 /*
835 * Update back pointers. Updating the handle fixes the back pointer of
836 * every descendant dbuf as well as the bonus dbuf.
837 */
838 ASSERT(ndn->dn_handle->dnh_dnode == odn);
839 ndn->dn_handle->dnh_dnode = ndn;
840 if (ndn->dn_zfetch.zf_dnode == odn) {
841 ndn->dn_zfetch.zf_dnode = ndn;
842 }
843
844 /*
845 * Invalidate the original dnode by clearing all of its back pointers.
846 */
847 odn->dn_dbuf = NULL;
848 odn->dn_handle = NULL;
849 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
850 offsetof(dmu_buf_impl_t, db_link));
851 odn->dn_dbufs_count = 0;
852 odn->dn_bonus = NULL;
853 odn->dn_zfetch.zf_dnode = NULL;
854
855 /*
856 * Set the low bit of the objset pointer to ensure that dnode_move()
857 * recognizes the dnode as invalid in any subsequent callback.
858 */
859 POINTER_INVALIDATE(&odn->dn_objset);
860
861 /*
862 * Satisfy the destructor.
863 */
864 for (i = 0; i < TXG_SIZE; i++) {
865 list_create(&odn->dn_dirty_records[i],
866 sizeof (dbuf_dirty_record_t),
867 offsetof(dbuf_dirty_record_t, dr_dirty_node));
868 odn->dn_free_ranges[i] = NULL;
869 odn->dn_next_nlevels[i] = 0;
870 odn->dn_next_indblkshift[i] = 0;
871 odn->dn_next_bonustype[i] = 0;
872 odn->dn_rm_spillblk[i] = 0;
873 odn->dn_next_bonuslen[i] = 0;
874 odn->dn_next_blksz[i] = 0;
875 }
876 odn->dn_allocated_txg = 0;
877 odn->dn_free_txg = 0;
878 odn->dn_assigned_txg = 0;
879 odn->dn_dirty_txg = 0;
880 odn->dn_dirtyctx = 0;
881 odn->dn_dirtyctx_firstset = NULL;
882 odn->dn_have_spill = B_FALSE;
883 odn->dn_zio = NULL;
884 odn->dn_oldused = 0;
885 odn->dn_oldflags = 0;
886 odn->dn_olduid = 0;
887 odn->dn_oldgid = 0;
888 odn->dn_oldprojid = ZFS_DEFAULT_PROJID;
889 odn->dn_newuid = 0;
890 odn->dn_newgid = 0;
891 odn->dn_newprojid = ZFS_DEFAULT_PROJID;
892 odn->dn_id_flags = 0;
893
894 /*
895 * Mark the dnode.
896 */
897 ndn->dn_moved = 1;
898 odn->dn_moved = (uint8_t)-1;
899 }
900
901 /*ARGSUSED*/
902 static kmem_cbrc_t
903 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
904 {
905 dnode_t *odn = buf, *ndn = newbuf;
906 objset_t *os;
907 int64_t refcount;
908 uint32_t dbufs;
909
910 /*
911 * The dnode is on the objset's list of known dnodes if the objset
912 * pointer is valid. We set the low bit of the objset pointer when
913 * freeing the dnode to invalidate it, and the memory patterns written
914 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
915 * A newly created dnode sets the objset pointer last of all to indicate
916 * that the dnode is known and in a valid state to be moved by this
917 * function.
918 */
919 os = odn->dn_objset;
920 if (!POINTER_IS_VALID(os)) {
921 DNODE_STAT_BUMP(dnode_move_invalid);
922 return (KMEM_CBRC_DONT_KNOW);
923 }
924
925 /*
926 * Ensure that the objset does not go away during the move.
927 */
928 rw_enter(&os_lock, RW_WRITER);
929 if (os != odn->dn_objset) {
930 rw_exit(&os_lock);
931 DNODE_STAT_BUMP(dnode_move_recheck1);
932 return (KMEM_CBRC_DONT_KNOW);
933 }
934
935 /*
936 * If the dnode is still valid, then so is the objset. We know that no
937 * valid objset can be freed while we hold os_lock, so we can safely
938 * ensure that the objset remains in use.
939 */
940 mutex_enter(&os->os_lock);
941
942 /*
943 * Recheck the objset pointer in case the dnode was removed just before
944 * acquiring the lock.
945 */
946 if (os != odn->dn_objset) {
947 mutex_exit(&os->os_lock);
948 rw_exit(&os_lock);
949 DNODE_STAT_BUMP(dnode_move_recheck2);
950 return (KMEM_CBRC_DONT_KNOW);
951 }
952
953 /*
954 * At this point we know that as long as we hold os->os_lock, the dnode
955 * cannot be freed and fields within the dnode can be safely accessed.
956 * The objset listing this dnode cannot go away as long as this dnode is
957 * on its list.
958 */
959 rw_exit(&os_lock);
960 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
961 mutex_exit(&os->os_lock);
962 DNODE_STAT_BUMP(dnode_move_special);
963 return (KMEM_CBRC_NO);
964 }
965 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
966
967 /*
968 * Lock the dnode handle to prevent the dnode from obtaining any new
969 * holds. This also prevents the descendant dbufs and the bonus dbuf
970 * from accessing the dnode, so that we can discount their holds. The
971 * handle is safe to access because we know that while the dnode cannot
972 * go away, neither can its handle. Once we hold dnh_zrlock, we can
973 * safely move any dnode referenced only by dbufs.
974 */
975 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
976 mutex_exit(&os->os_lock);
977 DNODE_STAT_BUMP(dnode_move_handle);
978 return (KMEM_CBRC_LATER);
979 }
980
981 /*
982 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
983 * We need to guarantee that there is a hold for every dbuf in order to
984 * determine whether the dnode is actively referenced. Falsely matching
985 * a dbuf to an active hold would lead to an unsafe move. It's possible
986 * that a thread already having an active dnode hold is about to add a
987 * dbuf, and we can't compare hold and dbuf counts while the add is in
988 * progress.
989 */
990 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
991 zrl_exit(&odn->dn_handle->dnh_zrlock);
992 mutex_exit(&os->os_lock);
993 DNODE_STAT_BUMP(dnode_move_rwlock);
994 return (KMEM_CBRC_LATER);
995 }
996
997 /*
998 * A dbuf may be removed (evicted) without an active dnode hold. In that
999 * case, the dbuf count is decremented under the handle lock before the
1000 * dbuf's hold is released. This order ensures that if we count the hold
1001 * after the dbuf is removed but before its hold is released, we will
1002 * treat the unmatched hold as active and exit safely. If we count the
1003 * hold before the dbuf is removed, the hold is discounted, and the
1004 * removal is blocked until the move completes.
1005 */
1006 refcount = zfs_refcount_count(&odn->dn_holds);
1007 ASSERT(refcount >= 0);
1008 dbufs = odn->dn_dbufs_count;
1009
1010 /* We can't have more dbufs than dnode holds. */
1011 ASSERT3U(dbufs, <=, refcount);
1012 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
1013 uint32_t, dbufs);
1014
1015 if (refcount > dbufs) {
1016 rw_exit(&odn->dn_struct_rwlock);
1017 zrl_exit(&odn->dn_handle->dnh_zrlock);
1018 mutex_exit(&os->os_lock);
1019 DNODE_STAT_BUMP(dnode_move_active);
1020 return (KMEM_CBRC_LATER);
1021 }
1022
1023 rw_exit(&odn->dn_struct_rwlock);
1024
1025 /*
1026 * At this point we know that anyone with a hold on the dnode is not
1027 * actively referencing it. The dnode is known and in a valid state to
1028 * move. We're holding the locks needed to execute the critical section.
1029 */
1030 dnode_move_impl(odn, ndn);
1031
1032 list_link_replace(&odn->dn_link, &ndn->dn_link);
1033 /* If the dnode was safe to move, the refcount cannot have changed. */
1034 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds));
1035 ASSERT(dbufs == ndn->dn_dbufs_count);
1036 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
1037 mutex_exit(&os->os_lock);
1038
1039 return (KMEM_CBRC_YES);
1040 }
1041 #endif /* _KERNEL */
1042
1043 static void
1044 dnode_slots_hold(dnode_children_t *children, int idx, int slots)
1045 {
1046 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1047
1048 for (int i = idx; i < idx + slots; i++) {
1049 dnode_handle_t *dnh = &children->dnc_children[i];
1050 zrl_add(&dnh->dnh_zrlock);
1051 }
1052 }
1053
1054 static void
1055 dnode_slots_rele(dnode_children_t *children, int idx, int slots)
1056 {
1057 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1058
1059 for (int i = idx; i < idx + slots; i++) {
1060 dnode_handle_t *dnh = &children->dnc_children[i];
1061
1062 if (zrl_is_locked(&dnh->dnh_zrlock))
1063 zrl_exit(&dnh->dnh_zrlock);
1064 else
1065 zrl_remove(&dnh->dnh_zrlock);
1066 }
1067 }
1068
1069 static int
1070 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots)
1071 {
1072 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1073
1074 for (int i = idx; i < idx + slots; i++) {
1075 dnode_handle_t *dnh = &children->dnc_children[i];
1076
1077 if (!zrl_tryenter(&dnh->dnh_zrlock)) {
1078 for (int j = idx; j < i; j++) {
1079 dnh = &children->dnc_children[j];
1080 zrl_exit(&dnh->dnh_zrlock);
1081 }
1082
1083 return (0);
1084 }
1085 }
1086
1087 return (1);
1088 }
1089
1090 static void
1091 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr)
1092 {
1093 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1094
1095 for (int i = idx; i < idx + slots; i++) {
1096 dnode_handle_t *dnh = &children->dnc_children[i];
1097 dnh->dnh_dnode = ptr;
1098 }
1099 }
1100
1101 static boolean_t
1102 dnode_check_slots_free(dnode_children_t *children, int idx, int slots)
1103 {
1104 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1105
1106 /*
1107 * If all dnode slots are either already free or
1108 * evictable return B_TRUE.
1109 */
1110 for (int i = idx; i < idx + slots; i++) {
1111 dnode_handle_t *dnh = &children->dnc_children[i];
1112 dnode_t *dn = dnh->dnh_dnode;
1113
1114 if (dn == DN_SLOT_FREE) {
1115 continue;
1116 } else if (DN_SLOT_IS_PTR(dn)) {
1117 mutex_enter(&dn->dn_mtx);
1118 boolean_t can_free = (dn->dn_type == DMU_OT_NONE &&
1119 zfs_refcount_is_zero(&dn->dn_holds) &&
1120 !DNODE_IS_DIRTY(dn));
1121 mutex_exit(&dn->dn_mtx);
1122
1123 if (!can_free)
1124 return (B_FALSE);
1125 else
1126 continue;
1127 } else {
1128 return (B_FALSE);
1129 }
1130 }
1131
1132 return (B_TRUE);
1133 }
1134
1135 static void
1136 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots)
1137 {
1138 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1139
1140 for (int i = idx; i < idx + slots; i++) {
1141 dnode_handle_t *dnh = &children->dnc_children[i];
1142
1143 ASSERT(zrl_is_locked(&dnh->dnh_zrlock));
1144
1145 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1146 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE);
1147 dnode_destroy(dnh->dnh_dnode);
1148 dnh->dnh_dnode = DN_SLOT_FREE;
1149 }
1150 }
1151 }
1152
1153 void
1154 dnode_free_interior_slots(dnode_t *dn)
1155 {
1156 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db);
1157 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT;
1158 int idx = (dn->dn_object & (epb - 1)) + 1;
1159 int slots = dn->dn_num_slots - 1;
1160
1161 if (slots == 0)
1162 return;
1163
1164 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK);
1165
1166 while (!dnode_slots_tryenter(children, idx, slots))
1167 DNODE_STAT_BUMP(dnode_free_interior_lock_retry);
1168
1169 dnode_set_slots(children, idx, slots, DN_SLOT_FREE);
1170 dnode_slots_rele(children, idx, slots);
1171 }
1172
1173 void
1174 dnode_special_close(dnode_handle_t *dnh)
1175 {
1176 dnode_t *dn = dnh->dnh_dnode;
1177
1178 /*
1179 * Wait for final references to the dnode to clear. This can
1180 * only happen if the arc is asynchronously evicting state that
1181 * has a hold on this dnode while we are trying to evict this
1182 * dnode.
1183 */
1184 while (zfs_refcount_count(&dn->dn_holds) > 0)
1185 delay(1);
1186 ASSERT(dn->dn_dbuf == NULL ||
1187 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1188 zrl_add(&dnh->dnh_zrlock);
1189 dnode_destroy(dn); /* implicit zrl_remove() */
1190 zrl_destroy(&dnh->dnh_zrlock);
1191 dnh->dnh_dnode = NULL;
1192 }
1193
1194 void
1195 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1196 dnode_handle_t *dnh)
1197 {
1198 dnode_t *dn;
1199
1200 zrl_init(&dnh->dnh_zrlock);
1201 zrl_tryenter(&dnh->dnh_zrlock);
1202
1203 dn = dnode_create(os, dnp, NULL, object, dnh);
1204 DNODE_VERIFY(dn);
1205
1206 zrl_exit(&dnh->dnh_zrlock);
1207 }
1208
1209 static void
1210 dnode_buf_evict_async(void *dbu)
1211 {
1212 dnode_children_t *dnc = dbu;
1213
1214 DNODE_STAT_BUMP(dnode_buf_evict);
1215
1216 for (int i = 0; i < dnc->dnc_count; i++) {
1217 dnode_handle_t *dnh = &dnc->dnc_children[i];
1218 dnode_t *dn;
1219
1220 /*
1221 * The dnode handle lock guards against the dnode moving to
1222 * another valid address, so there is no need here to guard
1223 * against changes to or from NULL.
1224 */
1225 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1226 zrl_destroy(&dnh->dnh_zrlock);
1227 dnh->dnh_dnode = DN_SLOT_UNINIT;
1228 continue;
1229 }
1230
1231 zrl_add(&dnh->dnh_zrlock);
1232 dn = dnh->dnh_dnode;
1233 /*
1234 * If there are holds on this dnode, then there should
1235 * be holds on the dnode's containing dbuf as well; thus
1236 * it wouldn't be eligible for eviction and this function
1237 * would not have been called.
1238 */
1239 ASSERT(zfs_refcount_is_zero(&dn->dn_holds));
1240 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds));
1241
1242 dnode_destroy(dn); /* implicit zrl_remove() for first slot */
1243 zrl_destroy(&dnh->dnh_zrlock);
1244 dnh->dnh_dnode = DN_SLOT_UNINIT;
1245 }
1246 kmem_free(dnc, sizeof (dnode_children_t) +
1247 dnc->dnc_count * sizeof (dnode_handle_t));
1248 }
1249
1250 /*
1251 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used
1252 * to ensure the hole at the specified object offset is large enough to
1253 * hold the dnode being created. The slots parameter is also used to ensure
1254 * a dnode does not span multiple dnode blocks. In both of these cases, if
1255 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases
1256 * are only possible when using DNODE_MUST_BE_FREE.
1257 *
1258 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
1259 * dnode_hold_impl() will check if the requested dnode is already consumed
1260 * as an extra dnode slot by an large dnode, in which case it returns
1261 * ENOENT.
1262 *
1263 * errors:
1264 * EINVAL - invalid object number or flags.
1265 * ENOSPC - hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE)
1266 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE)
1267 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE)
1268 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED)
1269 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED)
1270 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED)
1271 * EIO - i/o error error when reading the meta dnode dbuf.
1272 * succeeds even for free dnodes.
1273 */
1274 int
1275 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
1276 void *tag, dnode_t **dnp)
1277 {
1278 int epb, idx, err;
1279 int drop_struct_lock = FALSE;
1280 int type;
1281 uint64_t blk;
1282 dnode_t *mdn, *dn;
1283 dmu_buf_impl_t *db;
1284 dnode_children_t *dnc;
1285 dnode_phys_t *dn_block;
1286 dnode_handle_t *dnh;
1287
1288 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
1289 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
1290
1291 /*
1292 * If you are holding the spa config lock as writer, you shouldn't
1293 * be asking the DMU to do *anything* unless it's the root pool
1294 * which may require us to read from the root filesystem while
1295 * holding some (not all) of the locks as writer.
1296 */
1297 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1298 (spa_is_root(os->os_spa) &&
1299 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1300
1301 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1302
1303 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT ||
1304 object == DMU_PROJECTUSED_OBJECT) {
1305 if (object == DMU_USERUSED_OBJECT)
1306 dn = DMU_USERUSED_DNODE(os);
1307 else if (object == DMU_GROUPUSED_OBJECT)
1308 dn = DMU_GROUPUSED_DNODE(os);
1309 else
1310 dn = DMU_PROJECTUSED_DNODE(os);
1311 if (dn == NULL)
1312 return (SET_ERROR(ENOENT));
1313 type = dn->dn_type;
1314 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1315 return (SET_ERROR(ENOENT));
1316 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1317 return (SET_ERROR(EEXIST));
1318 DNODE_VERIFY(dn);
1319 (void) zfs_refcount_add(&dn->dn_holds, tag);
1320 *dnp = dn;
1321 return (0);
1322 }
1323
1324 if (object == 0 || object >= DN_MAX_OBJECT)
1325 return (SET_ERROR(EINVAL));
1326
1327 mdn = DMU_META_DNODE(os);
1328 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1329
1330 DNODE_VERIFY(mdn);
1331
1332 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1333 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1334 drop_struct_lock = TRUE;
1335 }
1336
1337 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1338
1339 db = dbuf_hold(mdn, blk, FTAG);
1340 if (drop_struct_lock)
1341 rw_exit(&mdn->dn_struct_rwlock);
1342 if (db == NULL) {
1343 DNODE_STAT_BUMP(dnode_hold_dbuf_hold);
1344 return (SET_ERROR(EIO));
1345 }
1346 /*
1347 * We do not need to decrypt to read the dnode so it doesn't matter
1348 * if we get the encrypted or decrypted version.
1349 */
1350 err = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_NO_DECRYPT);
1351 if (err) {
1352 DNODE_STAT_BUMP(dnode_hold_dbuf_read);
1353 dbuf_rele(db, FTAG);
1354 return (err);
1355 }
1356
1357 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1358 epb = db->db.db_size >> DNODE_SHIFT;
1359
1360 idx = object & (epb - 1);
1361 dn_block = (dnode_phys_t *)db->db.db_data;
1362
1363 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1364 dnc = dmu_buf_get_user(&db->db);
1365 dnh = NULL;
1366 if (dnc == NULL) {
1367 dnode_children_t *winner;
1368 int skip = 0;
1369
1370 dnc = kmem_zalloc(sizeof (dnode_children_t) +
1371 epb * sizeof (dnode_handle_t), KM_SLEEP);
1372 dnc->dnc_count = epb;
1373 dnh = &dnc->dnc_children[0];
1374
1375 /* Initialize dnode slot status from dnode_phys_t */
1376 for (int i = 0; i < epb; i++) {
1377 zrl_init(&dnh[i].dnh_zrlock);
1378
1379 if (skip) {
1380 skip--;
1381 continue;
1382 }
1383
1384 if (dn_block[i].dn_type != DMU_OT_NONE) {
1385 int interior = dn_block[i].dn_extra_slots;
1386
1387 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED);
1388 dnode_set_slots(dnc, i + 1, interior,
1389 DN_SLOT_INTERIOR);
1390 skip = interior;
1391 } else {
1392 dnh[i].dnh_dnode = DN_SLOT_FREE;
1393 skip = 0;
1394 }
1395 }
1396
1397 dmu_buf_init_user(&dnc->dnc_dbu, NULL,
1398 dnode_buf_evict_async, NULL);
1399 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu);
1400 if (winner != NULL) {
1401
1402 for (int i = 0; i < epb; i++)
1403 zrl_destroy(&dnh[i].dnh_zrlock);
1404
1405 kmem_free(dnc, sizeof (dnode_children_t) +
1406 epb * sizeof (dnode_handle_t));
1407 dnc = winner;
1408 }
1409 }
1410
1411 ASSERT(dnc->dnc_count == epb);
1412 dn = DN_SLOT_UNINIT;
1413
1414 if (flag & DNODE_MUST_BE_ALLOCATED) {
1415 slots = 1;
1416
1417 while (dn == DN_SLOT_UNINIT) {
1418 dnode_slots_hold(dnc, idx, slots);
1419 dnh = &dnc->dnc_children[idx];
1420
1421 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1422 dn = dnh->dnh_dnode;
1423 break;
1424 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) {
1425 DNODE_STAT_BUMP(dnode_hold_alloc_interior);
1426 dnode_slots_rele(dnc, idx, slots);
1427 dbuf_rele(db, FTAG);
1428 return (SET_ERROR(EEXIST));
1429 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) {
1430 DNODE_STAT_BUMP(dnode_hold_alloc_misses);
1431 dnode_slots_rele(dnc, idx, slots);
1432 dbuf_rele(db, FTAG);
1433 return (SET_ERROR(ENOENT));
1434 }
1435
1436 dnode_slots_rele(dnc, idx, slots);
1437 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1438 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry);
1439 continue;
1440 }
1441
1442 /*
1443 * Someone else won the race and called dnode_create()
1444 * after we checked DN_SLOT_IS_PTR() above but before
1445 * we acquired the lock.
1446 */
1447 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1448 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses);
1449 dn = dnh->dnh_dnode;
1450 } else {
1451 dn = dnode_create(os, dn_block + idx, db,
1452 object, dnh);
1453 }
1454 }
1455
1456 mutex_enter(&dn->dn_mtx);
1457 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) {
1458 DNODE_STAT_BUMP(dnode_hold_alloc_type_none);
1459 mutex_exit(&dn->dn_mtx);
1460 dnode_slots_rele(dnc, idx, slots);
1461 dbuf_rele(db, FTAG);
1462 return (SET_ERROR(ENOENT));
1463 }
1464
1465 DNODE_STAT_BUMP(dnode_hold_alloc_hits);
1466 } else if (flag & DNODE_MUST_BE_FREE) {
1467
1468 if (idx + slots - 1 >= DNODES_PER_BLOCK) {
1469 DNODE_STAT_BUMP(dnode_hold_free_overflow);
1470 dbuf_rele(db, FTAG);
1471 return (SET_ERROR(ENOSPC));
1472 }
1473
1474 while (dn == DN_SLOT_UNINIT) {
1475 dnode_slots_hold(dnc, idx, slots);
1476
1477 if (!dnode_check_slots_free(dnc, idx, slots)) {
1478 DNODE_STAT_BUMP(dnode_hold_free_misses);
1479 dnode_slots_rele(dnc, idx, slots);
1480 dbuf_rele(db, FTAG);
1481 return (SET_ERROR(ENOSPC));
1482 }
1483
1484 dnode_slots_rele(dnc, idx, slots);
1485 if (!dnode_slots_tryenter(dnc, idx, slots)) {
1486 DNODE_STAT_BUMP(dnode_hold_free_lock_retry);
1487 continue;
1488 }
1489
1490 if (!dnode_check_slots_free(dnc, idx, slots)) {
1491 DNODE_STAT_BUMP(dnode_hold_free_lock_misses);
1492 dnode_slots_rele(dnc, idx, slots);
1493 dbuf_rele(db, FTAG);
1494 return (SET_ERROR(ENOSPC));
1495 }
1496
1497 /*
1498 * Allocated but otherwise free dnodes which would
1499 * be in the interior of a multi-slot dnodes need
1500 * to be freed. Single slot dnodes can be safely
1501 * re-purposed as a performance optimization.
1502 */
1503 if (slots > 1)
1504 dnode_reclaim_slots(dnc, idx + 1, slots - 1);
1505
1506 dnh = &dnc->dnc_children[idx];
1507 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) {
1508 dn = dnh->dnh_dnode;
1509 } else {
1510 dn = dnode_create(os, dn_block + idx, db,
1511 object, dnh);
1512 }
1513 }
1514
1515 mutex_enter(&dn->dn_mtx);
1516 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) {
1517 DNODE_STAT_BUMP(dnode_hold_free_refcount);
1518 mutex_exit(&dn->dn_mtx);
1519 dnode_slots_rele(dnc, idx, slots);
1520 dbuf_rele(db, FTAG);
1521 return (SET_ERROR(EEXIST));
1522 }
1523
1524 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR);
1525 DNODE_STAT_BUMP(dnode_hold_free_hits);
1526 } else {
1527 dbuf_rele(db, FTAG);
1528 return (SET_ERROR(EINVAL));
1529 }
1530
1531 if (dn->dn_free_txg) {
1532 DNODE_STAT_BUMP(dnode_hold_free_txg);
1533 type = dn->dn_type;
1534 mutex_exit(&dn->dn_mtx);
1535 dnode_slots_rele(dnc, idx, slots);
1536 dbuf_rele(db, FTAG);
1537 return (SET_ERROR((flag & DNODE_MUST_BE_ALLOCATED) ?
1538 ENOENT : EEXIST));
1539 }
1540
1541 if (zfs_refcount_add(&dn->dn_holds, tag) == 1)
1542 dbuf_add_ref(db, dnh);
1543
1544 mutex_exit(&dn->dn_mtx);
1545
1546 /* Now we can rely on the hold to prevent the dnode from moving. */
1547 dnode_slots_rele(dnc, idx, slots);
1548
1549 DNODE_VERIFY(dn);
1550 ASSERT3P(dn->dn_dbuf, ==, db);
1551 ASSERT3U(dn->dn_object, ==, object);
1552 dbuf_rele(db, FTAG);
1553
1554 *dnp = dn;
1555 return (0);
1556 }
1557
1558 /*
1559 * Return held dnode if the object is allocated, NULL if not.
1560 */
1561 int
1562 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1563 {
1564 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
1565 dnp));
1566 }
1567
1568 /*
1569 * Can only add a reference if there is already at least one
1570 * reference on the dnode. Returns FALSE if unable to add a
1571 * new reference.
1572 */
1573 boolean_t
1574 dnode_add_ref(dnode_t *dn, void *tag)
1575 {
1576 mutex_enter(&dn->dn_mtx);
1577 if (zfs_refcount_is_zero(&dn->dn_holds)) {
1578 mutex_exit(&dn->dn_mtx);
1579 return (FALSE);
1580 }
1581 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag));
1582 mutex_exit(&dn->dn_mtx);
1583 return (TRUE);
1584 }
1585
1586 void
1587 dnode_rele(dnode_t *dn, void *tag)
1588 {
1589 mutex_enter(&dn->dn_mtx);
1590 dnode_rele_and_unlock(dn, tag, B_FALSE);
1591 }
1592
1593 void
1594 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1595 {
1596 uint64_t refs;
1597 /* Get while the hold prevents the dnode from moving. */
1598 dmu_buf_impl_t *db = dn->dn_dbuf;
1599 dnode_handle_t *dnh = dn->dn_handle;
1600
1601 refs = zfs_refcount_remove(&dn->dn_holds, tag);
1602 mutex_exit(&dn->dn_mtx);
1603
1604 /*
1605 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1606 * indirectly by dbuf_rele() while relying on the dnode handle to
1607 * prevent the dnode from moving, since releasing the last hold could
1608 * result in the dnode's parent dbuf evicting its dnode handles. For
1609 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1610 * other direct or indirect hold on the dnode must first drop the dnode
1611 * handle.
1612 */
1613 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1614
1615 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1616 if (refs == 0 && db != NULL) {
1617 /*
1618 * Another thread could add a hold to the dnode handle in
1619 * dnode_hold_impl() while holding the parent dbuf. Since the
1620 * hold on the parent dbuf prevents the handle from being
1621 * destroyed, the hold on the handle is OK. We can't yet assert
1622 * that the handle has zero references, but that will be
1623 * asserted anyway when the handle gets destroyed.
1624 */
1625 mutex_enter(&db->db_mtx);
1626 dbuf_rele_and_unlock(db, dnh, evicting);
1627 }
1628 }
1629
1630 void
1631 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1632 {
1633 objset_t *os = dn->dn_objset;
1634 uint64_t txg = tx->tx_txg;
1635
1636 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1637 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1638 return;
1639 }
1640
1641 DNODE_VERIFY(dn);
1642
1643 #ifdef ZFS_DEBUG
1644 mutex_enter(&dn->dn_mtx);
1645 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1646 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1647 mutex_exit(&dn->dn_mtx);
1648 #endif
1649
1650 /*
1651 * Determine old uid/gid when necessary
1652 */
1653 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1654
1655 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1656 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1657
1658 /*
1659 * If we are already marked dirty, we're done.
1660 */
1661 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1662 multilist_sublist_unlock(mls);
1663 return;
1664 }
1665
1666 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) ||
1667 !avl_is_empty(&dn->dn_dbufs));
1668 ASSERT(dn->dn_datablksz != 0);
1669 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1670 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1671 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1672
1673 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1674 dn->dn_object, txg);
1675
1676 multilist_sublist_insert_head(mls, dn);
1677
1678 multilist_sublist_unlock(mls);
1679
1680 /*
1681 * The dnode maintains a hold on its containing dbuf as
1682 * long as there are holds on it. Each instantiated child
1683 * dbuf maintains a hold on the dnode. When the last child
1684 * drops its hold, the dnode will drop its hold on the
1685 * containing dbuf. We add a "dirty hold" here so that the
1686 * dnode will hang around after we finish processing its
1687 * children.
1688 */
1689 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1690
1691 (void) dbuf_dirty(dn->dn_dbuf, tx);
1692
1693 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1694 }
1695
1696 void
1697 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1698 {
1699 mutex_enter(&dn->dn_mtx);
1700 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1701 mutex_exit(&dn->dn_mtx);
1702 return;
1703 }
1704 dn->dn_free_txg = tx->tx_txg;
1705 mutex_exit(&dn->dn_mtx);
1706
1707 dnode_setdirty(dn, tx);
1708 }
1709
1710 /*
1711 * Try to change the block size for the indicated dnode. This can only
1712 * succeed if there are no blocks allocated or dirty beyond first block
1713 */
1714 int
1715 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1716 {
1717 dmu_buf_impl_t *db;
1718 int err;
1719
1720 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1721 if (size == 0)
1722 size = SPA_MINBLOCKSIZE;
1723 else
1724 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1725
1726 if (ibs == dn->dn_indblkshift)
1727 ibs = 0;
1728
1729 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1730 return (0);
1731
1732 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1733
1734 /* Check for any allocated blocks beyond the first */
1735 if (dn->dn_maxblkid != 0)
1736 goto fail;
1737
1738 mutex_enter(&dn->dn_dbufs_mtx);
1739 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1740 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1741 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1742 db->db_blkid != DMU_SPILL_BLKID) {
1743 mutex_exit(&dn->dn_dbufs_mtx);
1744 goto fail;
1745 }
1746 }
1747 mutex_exit(&dn->dn_dbufs_mtx);
1748
1749 if (ibs && dn->dn_nlevels != 1)
1750 goto fail;
1751
1752 /* resize the old block */
1753 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1754 if (err == 0)
1755 dbuf_new_size(db, size, tx);
1756 else if (err != ENOENT)
1757 goto fail;
1758
1759 dnode_setdblksz(dn, size);
1760 dnode_setdirty(dn, tx);
1761 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1762 if (ibs) {
1763 dn->dn_indblkshift = ibs;
1764 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1765 }
1766 /* rele after we have fixed the blocksize in the dnode */
1767 if (db)
1768 dbuf_rele(db, FTAG);
1769
1770 rw_exit(&dn->dn_struct_rwlock);
1771 return (0);
1772
1773 fail:
1774 rw_exit(&dn->dn_struct_rwlock);
1775 return (SET_ERROR(ENOTSUP));
1776 }
1777
1778 static void
1779 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx)
1780 {
1781 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1782 int old_nlevels = dn->dn_nlevels;
1783 dmu_buf_impl_t *db;
1784 list_t *list;
1785 dbuf_dirty_record_t *new, *dr, *dr_next;
1786
1787 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1788
1789 dn->dn_nlevels = new_nlevels;
1790
1791 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1792 dn->dn_next_nlevels[txgoff] = new_nlevels;
1793
1794 /* dirty the left indirects */
1795 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1796 ASSERT(db != NULL);
1797 new = dbuf_dirty(db, tx);
1798 dbuf_rele(db, FTAG);
1799
1800 /* transfer the dirty records to the new indirect */
1801 mutex_enter(&dn->dn_mtx);
1802 mutex_enter(&new->dt.di.dr_mtx);
1803 list = &dn->dn_dirty_records[txgoff];
1804 for (dr = list_head(list); dr; dr = dr_next) {
1805 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1806 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1807 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1808 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1809 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1810 list_remove(&dn->dn_dirty_records[txgoff], dr);
1811 list_insert_tail(&new->dt.di.dr_children, dr);
1812 dr->dr_parent = new;
1813 }
1814 }
1815 mutex_exit(&new->dt.di.dr_mtx);
1816 mutex_exit(&dn->dn_mtx);
1817 }
1818
1819 int
1820 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx)
1821 {
1822 int ret = 0;
1823
1824 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1825
1826 if (dn->dn_nlevels == nlevels) {
1827 ret = 0;
1828 goto out;
1829 } else if (nlevels < dn->dn_nlevels) {
1830 ret = SET_ERROR(EINVAL);
1831 goto out;
1832 }
1833
1834 dnode_set_nlevels_impl(dn, nlevels, tx);
1835
1836 out:
1837 rw_exit(&dn->dn_struct_rwlock);
1838 return (ret);
1839 }
1840
1841 /* read-holding callers must not rely on the lock being continuously held */
1842 void
1843 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read,
1844 boolean_t force)
1845 {
1846 int epbs, new_nlevels;
1847 uint64_t sz;
1848
1849 ASSERT(blkid != DMU_BONUS_BLKID);
1850
1851 ASSERT(have_read ?
1852 RW_READ_HELD(&dn->dn_struct_rwlock) :
1853 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1854
1855 /*
1856 * if we have a read-lock, check to see if we need to do any work
1857 * before upgrading to a write-lock.
1858 */
1859 if (have_read) {
1860 if (blkid <= dn->dn_maxblkid)
1861 return;
1862
1863 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1864 rw_exit(&dn->dn_struct_rwlock);
1865 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1866 }
1867 }
1868
1869 /*
1870 * Raw sends (indicated by the force flag) require that we take the
1871 * given blkid even if the value is lower than the current value.
1872 */
1873 if (!force && blkid <= dn->dn_maxblkid)
1874 goto out;
1875
1876 /*
1877 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff]
1878 * to indicate that this field is set. This allows us to set the
1879 * maxblkid to 0 on an existing object in dnode_sync().
1880 */
1881 dn->dn_maxblkid = blkid;
1882 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] =
1883 blkid | DMU_NEXT_MAXBLKID_SET;
1884
1885 /*
1886 * Compute the number of levels necessary to support the new maxblkid.
1887 * Raw sends will ensure nlevels is set correctly for us.
1888 */
1889 new_nlevels = 1;
1890 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1891 for (sz = dn->dn_nblkptr;
1892 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1893 new_nlevels++;
1894
1895 if (!force) {
1896 if (new_nlevels > dn->dn_nlevels)
1897 dnode_set_nlevels_impl(dn, new_nlevels, tx);
1898 } else {
1899 ASSERT3U(dn->dn_nlevels, >=, new_nlevels);
1900 }
1901
1902 out:
1903 if (have_read)
1904 rw_downgrade(&dn->dn_struct_rwlock);
1905 }
1906
1907 static void
1908 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1909 {
1910 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1911 if (db != NULL) {
1912 dmu_buf_will_dirty(&db->db, tx);
1913 dbuf_rele(db, FTAG);
1914 }
1915 }
1916
1917 /*
1918 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1919 * and end_blkid.
1920 */
1921 static void
1922 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1923 dmu_tx_t *tx)
1924 {
1925 dmu_buf_impl_t db_search;
1926 dmu_buf_impl_t *db;
1927 avl_index_t where;
1928
1929 mutex_enter(&dn->dn_dbufs_mtx);
1930
1931 db_search.db_level = 1;
1932 db_search.db_blkid = start_blkid + 1;
1933 db_search.db_state = DB_SEARCH;
1934 for (;;) {
1935
1936 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1937 if (db == NULL)
1938 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1939
1940 if (db == NULL || db->db_level != 1 ||
1941 db->db_blkid >= end_blkid) {
1942 break;
1943 }
1944
1945 /*
1946 * Setup the next blkid we want to search for.
1947 */
1948 db_search.db_blkid = db->db_blkid + 1;
1949 ASSERT3U(db->db_blkid, >=, start_blkid);
1950
1951 /*
1952 * If the dbuf transitions to DB_EVICTING while we're trying
1953 * to dirty it, then we will be unable to discover it in
1954 * the dbuf hash table. This will result in a call to
1955 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1956 * lock. To avoid a deadlock, we drop the lock before
1957 * dirtying the level-1 dbuf.
1958 */
1959 mutex_exit(&dn->dn_dbufs_mtx);
1960 dnode_dirty_l1(dn, db->db_blkid, tx);
1961 mutex_enter(&dn->dn_dbufs_mtx);
1962 }
1963
1964 #ifdef ZFS_DEBUG
1965 /*
1966 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1967 */
1968 db_search.db_level = 1;
1969 db_search.db_blkid = start_blkid + 1;
1970 db_search.db_state = DB_SEARCH;
1971 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1972 if (db == NULL)
1973 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1974 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
1975 if (db->db_level != 1 || db->db_blkid >= end_blkid)
1976 break;
1977 ASSERT(db->db_dirtycnt > 0);
1978 }
1979 #endif
1980 mutex_exit(&dn->dn_dbufs_mtx);
1981 }
1982
1983 void
1984 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1985 {
1986 dmu_buf_impl_t *db;
1987 uint64_t blkoff, blkid, nblks;
1988 int blksz, blkshift, head, tail;
1989 int trunc = FALSE;
1990 int epbs;
1991
1992 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1993 blksz = dn->dn_datablksz;
1994 blkshift = dn->dn_datablkshift;
1995 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1996
1997 if (len == DMU_OBJECT_END) {
1998 len = UINT64_MAX - off;
1999 trunc = TRUE;
2000 }
2001
2002 /*
2003 * First, block align the region to free:
2004 */
2005 if (ISP2(blksz)) {
2006 head = P2NPHASE(off, blksz);
2007 blkoff = P2PHASE(off, blksz);
2008 if ((off >> blkshift) > dn->dn_maxblkid)
2009 goto out;
2010 } else {
2011 ASSERT(dn->dn_maxblkid == 0);
2012 if (off == 0 && len >= blksz) {
2013 /*
2014 * Freeing the whole block; fast-track this request.
2015 */
2016 blkid = 0;
2017 nblks = 1;
2018 if (dn->dn_nlevels > 1)
2019 dnode_dirty_l1(dn, 0, tx);
2020 goto done;
2021 } else if (off >= blksz) {
2022 /* Freeing past end-of-data */
2023 goto out;
2024 } else {
2025 /* Freeing part of the block. */
2026 head = blksz - off;
2027 ASSERT3U(head, >, 0);
2028 }
2029 blkoff = off;
2030 }
2031 /* zero out any partial block data at the start of the range */
2032 if (head) {
2033 ASSERT3U(blkoff + head, ==, blksz);
2034 if (len < head)
2035 head = len;
2036 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
2037 TRUE, FALSE, FTAG, &db) == 0) {
2038 caddr_t data;
2039
2040 /* don't dirty if it isn't on disk and isn't dirty */
2041 if (db->db_last_dirty ||
2042 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2043 rw_exit(&dn->dn_struct_rwlock);
2044 dmu_buf_will_dirty(&db->db, tx);
2045 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2046 data = db->db.db_data;
2047 bzero(data + blkoff, head);
2048 }
2049 dbuf_rele(db, FTAG);
2050 }
2051 off += head;
2052 len -= head;
2053 }
2054
2055 /* If the range was less than one block, we're done */
2056 if (len == 0)
2057 goto out;
2058
2059 /* If the remaining range is past end of file, we're done */
2060 if ((off >> blkshift) > dn->dn_maxblkid)
2061 goto out;
2062
2063 ASSERT(ISP2(blksz));
2064 if (trunc)
2065 tail = 0;
2066 else
2067 tail = P2PHASE(len, blksz);
2068
2069 ASSERT0(P2PHASE(off, blksz));
2070 /* zero out any partial block data at the end of the range */
2071 if (tail) {
2072 if (len < tail)
2073 tail = len;
2074 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
2075 TRUE, FALSE, FTAG, &db) == 0) {
2076 /* don't dirty if not on disk and not dirty */
2077 if (db->db_last_dirty ||
2078 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
2079 rw_exit(&dn->dn_struct_rwlock);
2080 dmu_buf_will_dirty(&db->db, tx);
2081 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2082 bzero(db->db.db_data, tail);
2083 }
2084 dbuf_rele(db, FTAG);
2085 }
2086 len -= tail;
2087 }
2088
2089 /* If the range did not include a full block, we are done */
2090 if (len == 0)
2091 goto out;
2092
2093 ASSERT(IS_P2ALIGNED(off, blksz));
2094 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
2095 blkid = off >> blkshift;
2096 nblks = len >> blkshift;
2097 if (trunc)
2098 nblks += 1;
2099
2100 /*
2101 * Dirty all the indirect blocks in this range. Note that only
2102 * the first and last indirect blocks can actually be written
2103 * (if they were partially freed) -- they must be dirtied, even if
2104 * they do not exist on disk yet. The interior blocks will
2105 * be freed by free_children(), so they will not actually be written.
2106 * Even though these interior blocks will not be written, we
2107 * dirty them for two reasons:
2108 *
2109 * - It ensures that the indirect blocks remain in memory until
2110 * syncing context. (They have already been prefetched by
2111 * dmu_tx_hold_free(), so we don't have to worry about reading
2112 * them serially here.)
2113 *
2114 * - The dirty space accounting will put pressure on the txg sync
2115 * mechanism to begin syncing, and to delay transactions if there
2116 * is a large amount of freeing. Even though these indirect
2117 * blocks will not be written, we could need to write the same
2118 * amount of space if we copy the freed BPs into deadlists.
2119 */
2120 if (dn->dn_nlevels > 1) {
2121 uint64_t first, last;
2122
2123 first = blkid >> epbs;
2124 dnode_dirty_l1(dn, first, tx);
2125 if (trunc)
2126 last = dn->dn_maxblkid >> epbs;
2127 else
2128 last = (blkid + nblks - 1) >> epbs;
2129 if (last != first)
2130 dnode_dirty_l1(dn, last, tx);
2131
2132 dnode_dirty_l1range(dn, first, last, tx);
2133
2134 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
2135 SPA_BLKPTRSHIFT;
2136 for (uint64_t i = first + 1; i < last; i++) {
2137 /*
2138 * Set i to the blockid of the next non-hole
2139 * level-1 indirect block at or after i. Note
2140 * that dnode_next_offset() operates in terms of
2141 * level-0-equivalent bytes.
2142 */
2143 uint64_t ibyte = i << shift;
2144 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
2145 &ibyte, 2, 1, 0);
2146 i = ibyte >> shift;
2147 if (i >= last)
2148 break;
2149
2150 /*
2151 * Normally we should not see an error, either
2152 * from dnode_next_offset() or dbuf_hold_level()
2153 * (except for ESRCH from dnode_next_offset).
2154 * If there is an i/o error, then when we read
2155 * this block in syncing context, it will use
2156 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
2157 * to the "failmode" property. dnode_next_offset()
2158 * doesn't have a flag to indicate MUSTSUCCEED.
2159 */
2160 if (err != 0)
2161 break;
2162
2163 dnode_dirty_l1(dn, i, tx);
2164 }
2165 }
2166
2167 done:
2168 /*
2169 * Add this range to the dnode range list.
2170 * We will finish up this free operation in the syncing phase.
2171 */
2172 mutex_enter(&dn->dn_mtx);
2173 int txgoff = tx->tx_txg & TXG_MASK;
2174 if (dn->dn_free_ranges[txgoff] == NULL) {
2175 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
2176 }
2177 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
2178 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
2179 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
2180 blkid, nblks, tx->tx_txg);
2181 mutex_exit(&dn->dn_mtx);
2182
2183 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
2184 dnode_setdirty(dn, tx);
2185 out:
2186
2187 rw_exit(&dn->dn_struct_rwlock);
2188 }
2189
2190 static boolean_t
2191 dnode_spill_freed(dnode_t *dn)
2192 {
2193 int i;
2194
2195 mutex_enter(&dn->dn_mtx);
2196 for (i = 0; i < TXG_SIZE; i++) {
2197 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
2198 break;
2199 }
2200 mutex_exit(&dn->dn_mtx);
2201 return (i < TXG_SIZE);
2202 }
2203
2204 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
2205 uint64_t
2206 dnode_block_freed(dnode_t *dn, uint64_t blkid)
2207 {
2208 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
2209 int i;
2210
2211 if (blkid == DMU_BONUS_BLKID)
2212 return (FALSE);
2213
2214 /*
2215 * If we're in the process of opening the pool, dp will not be
2216 * set yet, but there shouldn't be anything dirty.
2217 */
2218 if (dp == NULL)
2219 return (FALSE);
2220
2221 if (dn->dn_free_txg)
2222 return (TRUE);
2223
2224 if (blkid == DMU_SPILL_BLKID)
2225 return (dnode_spill_freed(dn));
2226
2227 mutex_enter(&dn->dn_mtx);
2228 for (i = 0; i < TXG_SIZE; i++) {
2229 if (dn->dn_free_ranges[i] != NULL &&
2230 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
2231 break;
2232 }
2233 mutex_exit(&dn->dn_mtx);
2234 return (i < TXG_SIZE);
2235 }
2236
2237 /* call from syncing context when we actually write/free space for this dnode */
2238 void
2239 dnode_diduse_space(dnode_t *dn, int64_t delta)
2240 {
2241 uint64_t space;
2242 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
2243 dn, dn->dn_phys,
2244 (u_longlong_t)dn->dn_phys->dn_used,
2245 (longlong_t)delta);
2246
2247 mutex_enter(&dn->dn_mtx);
2248 space = DN_USED_BYTES(dn->dn_phys);
2249 if (delta > 0) {
2250 ASSERT3U(space + delta, >=, space); /* no overflow */
2251 } else {
2252 ASSERT3U(space, >=, -delta); /* no underflow */
2253 }
2254 space += delta;
2255 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
2256 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
2257 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
2258 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
2259 } else {
2260 dn->dn_phys->dn_used = space;
2261 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
2262 }
2263 mutex_exit(&dn->dn_mtx);
2264 }
2265
2266 /*
2267 * Scans a block at the indicated "level" looking for a hole or data,
2268 * depending on 'flags'.
2269 *
2270 * If level > 0, then we are scanning an indirect block looking at its
2271 * pointers. If level == 0, then we are looking at a block of dnodes.
2272 *
2273 * If we don't find what we are looking for in the block, we return ESRCH.
2274 * Otherwise, return with *offset pointing to the beginning (if searching
2275 * forwards) or end (if searching backwards) of the range covered by the
2276 * block pointer we matched on (or dnode).
2277 *
2278 * The basic search algorithm used below by dnode_next_offset() is to
2279 * use this function to search up the block tree (widen the search) until
2280 * we find something (i.e., we don't return ESRCH) and then search back
2281 * down the tree (narrow the search) until we reach our original search
2282 * level.
2283 */
2284 static int
2285 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
2286 int lvl, uint64_t blkfill, uint64_t txg)
2287 {
2288 dmu_buf_impl_t *db = NULL;
2289 void *data = NULL;
2290 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2291 uint64_t epb = 1ULL << epbs;
2292 uint64_t minfill, maxfill;
2293 boolean_t hole;
2294 int i, inc, error, span;
2295
2296 dprintf("probing object %llu offset %llx level %d of %u\n",
2297 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
2298
2299 hole = ((flags & DNODE_FIND_HOLE) != 0);
2300 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
2301 ASSERT(txg == 0 || !hole);
2302
2303 if (lvl == dn->dn_phys->dn_nlevels) {
2304 error = 0;
2305 epb = dn->dn_phys->dn_nblkptr;
2306 data = dn->dn_phys->dn_blkptr;
2307 } else {
2308 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
2309 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
2310 if (error) {
2311 if (error != ENOENT)
2312 return (error);
2313 if (hole)
2314 return (0);
2315 /*
2316 * This can only happen when we are searching up
2317 * the block tree for data. We don't really need to
2318 * adjust the offset, as we will just end up looking
2319 * at the pointer to this block in its parent, and its
2320 * going to be unallocated, so we will skip over it.
2321 */
2322 return (SET_ERROR(ESRCH));
2323 }
2324 error = dbuf_read(db, NULL,
2325 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | DB_RF_NO_DECRYPT);
2326 if (error) {
2327 dbuf_rele(db, FTAG);
2328 return (error);
2329 }
2330 data = db->db.db_data;
2331 }
2332
2333
2334 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
2335 db->db_blkptr->blk_birth <= txg ||
2336 BP_IS_HOLE(db->db_blkptr))) {
2337 /*
2338 * This can only happen when we are searching up the tree
2339 * and these conditions mean that we need to keep climbing.
2340 */
2341 error = SET_ERROR(ESRCH);
2342 } else if (lvl == 0) {
2343 dnode_phys_t *dnp = data;
2344
2345 ASSERT(dn->dn_type == DMU_OT_DNODE);
2346 ASSERT(!(flags & DNODE_FIND_BACKWARDS));
2347
2348 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
2349 i < blkfill; i += dnp[i].dn_extra_slots + 1) {
2350 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
2351 break;
2352 }
2353
2354 if (i == blkfill)
2355 error = SET_ERROR(ESRCH);
2356
2357 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
2358 (i << DNODE_SHIFT);
2359 } else {
2360 blkptr_t *bp = data;
2361 uint64_t start = *offset;
2362 span = (lvl - 1) * epbs + dn->dn_datablkshift;
2363 minfill = 0;
2364 maxfill = blkfill << ((lvl - 1) * epbs);
2365
2366 if (hole)
2367 maxfill--;
2368 else
2369 minfill++;
2370
2371 *offset = *offset >> span;
2372 for (i = BF64_GET(*offset, 0, epbs);
2373 i >= 0 && i < epb; i += inc) {
2374 if (BP_GET_FILL(&bp[i]) >= minfill &&
2375 BP_GET_FILL(&bp[i]) <= maxfill &&
2376 (hole || bp[i].blk_birth > txg))
2377 break;
2378 if (inc > 0 || *offset > 0)
2379 *offset += inc;
2380 }
2381 *offset = *offset << span;
2382 if (inc < 0) {
2383 /* traversing backwards; position offset at the end */
2384 ASSERT3U(*offset, <=, start);
2385 *offset = MIN(*offset + (1ULL << span) - 1, start);
2386 } else if (*offset < start) {
2387 *offset = start;
2388 }
2389 if (i < 0 || i >= epb)
2390 error = SET_ERROR(ESRCH);
2391 }
2392
2393 if (db)
2394 dbuf_rele(db, FTAG);
2395
2396 return (error);
2397 }
2398
2399 /*
2400 * Find the next hole, data, or sparse region at or after *offset.
2401 * The value 'blkfill' tells us how many items we expect to find
2402 * in an L0 data block; this value is 1 for normal objects,
2403 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
2404 * DNODES_PER_BLOCK when searching for sparse regions thereof.
2405 *
2406 * Examples:
2407 *
2408 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
2409 * Finds the next/previous hole/data in a file.
2410 * Used in dmu_offset_next().
2411 *
2412 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2413 * Finds the next free/allocated dnode an objset's meta-dnode.
2414 * Only finds objects that have new contents since txg (ie.
2415 * bonus buffer changes and content removal are ignored).
2416 * Used in dmu_object_next().
2417 *
2418 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2419 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2420 * Used in dmu_object_alloc().
2421 */
2422 int
2423 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2424 int minlvl, uint64_t blkfill, uint64_t txg)
2425 {
2426 uint64_t initial_offset = *offset;
2427 int lvl, maxlvl;
2428 int error = 0;
2429
2430 if (!(flags & DNODE_FIND_HAVELOCK))
2431 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2432
2433 if (dn->dn_phys->dn_nlevels == 0) {
2434 error = SET_ERROR(ESRCH);
2435 goto out;
2436 }
2437
2438 if (dn->dn_datablkshift == 0) {
2439 if (*offset < dn->dn_datablksz) {
2440 if (flags & DNODE_FIND_HOLE)
2441 *offset = dn->dn_datablksz;
2442 } else {
2443 error = SET_ERROR(ESRCH);
2444 }
2445 goto out;
2446 }
2447
2448 maxlvl = dn->dn_phys->dn_nlevels;
2449
2450 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2451 error = dnode_next_offset_level(dn,
2452 flags, offset, lvl, blkfill, txg);
2453 if (error != ESRCH)
2454 break;
2455 }
2456
2457 while (error == 0 && --lvl >= minlvl) {
2458 error = dnode_next_offset_level(dn,
2459 flags, offset, lvl, blkfill, txg);
2460 }
2461
2462 /*
2463 * There's always a "virtual hole" at the end of the object, even
2464 * if all BP's which physically exist are non-holes.
2465 */
2466 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2467 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2468 error = 0;
2469 }
2470
2471 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2472 initial_offset < *offset : initial_offset > *offset))
2473 error = SET_ERROR(ESRCH);
2474 out:
2475 if (!(flags & DNODE_FIND_HAVELOCK))
2476 rw_exit(&dn->dn_struct_rwlock);
2477
2478 return (error);
2479 }