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
43 static kmem_cache_t *dnode_cache;
44 /*
45 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
46 * turned on when DEBUG is also defined.
47 */
48 #ifdef DEBUG
49 #define DNODE_STATS
50 #endif /* DEBUG */
51
52 #ifdef DNODE_STATS
53 #define DNODE_STAT_ADD(stat) ((stat)++)
54 #else
55 #define DNODE_STAT_ADD(stat) /* nothing */
56 #endif /* DNODE_STATS */
57
58 static dnode_phys_t dnode_phys_zero;
59
60 int zfs_default_bs = SPA_MINBLOCKSHIFT;
61 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
62
63 #ifdef _KERNEL
64 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
65 #endif /* _KERNEL */
66
67 static int
68 dbuf_compare(const void *x1, const void *x2)
69 {
70 const dmu_buf_impl_t *d1 = x1;
71 const dmu_buf_impl_t *d2 = x2;
72
73 if (d1->db_level < d2->db_level) {
74 return (-1);
75 }
76 if (d1->db_level > d2->db_level) {
77 return (1);
78 }
79
80 if (d1->db_blkid < d2->db_blkid) {
81 return (-1);
82 }
83 if (d1->db_blkid > d2->db_blkid) {
84 return (1);
85 }
86
87 if (d1->db_state == DB_SEARCH) {
88 ASSERT3S(d2->db_state, !=, DB_SEARCH);
89 return (-1);
90 } else if (d2->db_state == DB_SEARCH) {
91 ASSERT3S(d1->db_state, !=, DB_SEARCH);
92 return (1);
93 }
94
95 if ((uintptr_t)d1 < (uintptr_t)d2) {
96 return (-1);
97 }
98 if ((uintptr_t)d1 > (uintptr_t)d2) {
99 return (1);
100 }
101 return (0);
102 }
103
104 /* ARGSUSED */
105 static int
106 dnode_cons(void *arg, void *unused, int kmflag)
107 {
108 dnode_t *dn = arg;
109 int i;
110
111 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
112 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
113 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
114 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
115
116 /*
117 * Every dbuf has a reference, and dropping a tracked reference is
118 * O(number of references), so don't track dn_holds.
119 */
120 refcount_create_untracked(&dn->dn_holds);
121 refcount_create(&dn->dn_tx_holds);
122 list_link_init(&dn->dn_link);
123
124 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
125 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
126 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
127 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
128 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
129 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
130 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
131
132 for (i = 0; i < TXG_SIZE; i++) {
133 list_link_init(&dn->dn_dirty_link[i]);
134 dn->dn_free_ranges[i] = NULL;
135 list_create(&dn->dn_dirty_records[i],
136 sizeof (dbuf_dirty_record_t),
137 offsetof(dbuf_dirty_record_t, dr_dirty_node));
138 }
139
140 dn->dn_allocated_txg = 0;
141 dn->dn_free_txg = 0;
142 dn->dn_assigned_txg = 0;
143 dn->dn_dirtyctx = 0;
144 dn->dn_dirtyctx_firstset = NULL;
145 dn->dn_bonus = NULL;
146 dn->dn_have_spill = B_FALSE;
147 dn->dn_zio = NULL;
148 dn->dn_oldused = 0;
149 dn->dn_oldflags = 0;
150 dn->dn_olduid = 0;
151 dn->dn_oldgid = 0;
152 dn->dn_newuid = 0;
153 dn->dn_newgid = 0;
154 dn->dn_id_flags = 0;
155
156 dn->dn_dbufs_count = 0;
157 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
158 offsetof(dmu_buf_impl_t, db_link));
159
160 dn->dn_moved = 0;
161 return (0);
162 }
163
164 /* ARGSUSED */
165 static void
166 dnode_dest(void *arg, void *unused)
167 {
168 int i;
169 dnode_t *dn = arg;
170
171 rw_destroy(&dn->dn_struct_rwlock);
172 mutex_destroy(&dn->dn_mtx);
173 mutex_destroy(&dn->dn_dbufs_mtx);
174 cv_destroy(&dn->dn_notxholds);
175 refcount_destroy(&dn->dn_holds);
176 refcount_destroy(&dn->dn_tx_holds);
177 ASSERT(!list_link_active(&dn->dn_link));
178
179 for (i = 0; i < TXG_SIZE; i++) {
180 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
181 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
182 list_destroy(&dn->dn_dirty_records[i]);
183 ASSERT0(dn->dn_next_nblkptr[i]);
184 ASSERT0(dn->dn_next_nlevels[i]);
185 ASSERT0(dn->dn_next_indblkshift[i]);
186 ASSERT0(dn->dn_next_bonustype[i]);
187 ASSERT0(dn->dn_rm_spillblk[i]);
188 ASSERT0(dn->dn_next_bonuslen[i]);
189 ASSERT0(dn->dn_next_blksz[i]);
190 }
191
192 ASSERT0(dn->dn_allocated_txg);
193 ASSERT0(dn->dn_free_txg);
194 ASSERT0(dn->dn_assigned_txg);
195 ASSERT0(dn->dn_dirtyctx);
196 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
197 ASSERT3P(dn->dn_bonus, ==, NULL);
198 ASSERT(!dn->dn_have_spill);
199 ASSERT3P(dn->dn_zio, ==, NULL);
200 ASSERT0(dn->dn_oldused);
201 ASSERT0(dn->dn_oldflags);
202 ASSERT0(dn->dn_olduid);
203 ASSERT0(dn->dn_oldgid);
204 ASSERT0(dn->dn_newuid);
205 ASSERT0(dn->dn_newgid);
206 ASSERT0(dn->dn_id_flags);
207
208 ASSERT0(dn->dn_dbufs_count);
209 avl_destroy(&dn->dn_dbufs);
210 }
211
212 void
213 dnode_init(void)
214 {
215 ASSERT(dnode_cache == NULL);
216 dnode_cache = kmem_cache_create("dnode_t",
217 sizeof (dnode_t),
218 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
219 #ifdef _KERNEL
220 kmem_cache_set_move(dnode_cache, dnode_move);
221 #endif /* _KERNEL */
222 }
223
224 void
225 dnode_fini(void)
226 {
227 kmem_cache_destroy(dnode_cache);
228 dnode_cache = NULL;
229 }
230
231
232 #ifdef ZFS_DEBUG
233 void
234 dnode_verify(dnode_t *dn)
235 {
236 int drop_struct_lock = FALSE;
237
238 ASSERT(dn->dn_phys);
239 ASSERT(dn->dn_objset);
240 ASSERT(dn->dn_handle->dnh_dnode == dn);
241
242 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
243
244 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
245 return;
246
247 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
248 rw_enter(&dn->dn_struct_rwlock, RW_READER);
249 drop_struct_lock = TRUE;
250 }
251 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
252 int i;
253 ASSERT3U(dn->dn_indblkshift, >=, 0);
254 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
255 if (dn->dn_datablkshift) {
256 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
257 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
258 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
259 }
260 ASSERT3U(dn->dn_nlevels, <=, 30);
261 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
262 ASSERT3U(dn->dn_nblkptr, >=, 1);
263 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
264 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
265 ASSERT3U(dn->dn_datablksz, ==,
266 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
267 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
268 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
269 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
270 for (i = 0; i < TXG_SIZE; i++) {
271 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
272 }
273 }
274 if (dn->dn_phys->dn_type != DMU_OT_NONE)
275 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
276 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
277 if (dn->dn_dbuf != NULL) {
278 ASSERT3P(dn->dn_phys, ==,
279 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
280 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
281 }
282 if (drop_struct_lock)
283 rw_exit(&dn->dn_struct_rwlock);
284 }
285 #endif
286
287 void
288 dnode_byteswap(dnode_phys_t *dnp)
289 {
290 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
291 int i;
292
293 if (dnp->dn_type == DMU_OT_NONE) {
294 bzero(dnp, sizeof (dnode_phys_t));
295 return;
296 }
297
298 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
299 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
300 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
301 dnp->dn_used = BSWAP_64(dnp->dn_used);
302
303 /*
304 * dn_nblkptr is only one byte, so it's OK to read it in either
305 * byte order. We can't read dn_bouslen.
306 */
307 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
308 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
309 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
310 buf64[i] = BSWAP_64(buf64[i]);
311
312 /*
313 * OK to check dn_bonuslen for zero, because it won't matter if
314 * we have the wrong byte order. This is necessary because the
315 * dnode dnode is smaller than a regular dnode.
316 */
317 if (dnp->dn_bonuslen != 0) {
318 /*
319 * Note that the bonus length calculated here may be
320 * longer than the actual bonus buffer. This is because
321 * we always put the bonus buffer after the last block
322 * pointer (instead of packing it against the end of the
323 * dnode buffer).
324 */
325 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
326 size_t len = DN_MAX_BONUSLEN - off;
327 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
328 dmu_object_byteswap_t byteswap =
329 DMU_OT_BYTESWAP(dnp->dn_bonustype);
330 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
331 }
332
333 /* Swap SPILL block if we have one */
334 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
335 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
336
337 }
338
339 void
340 dnode_buf_byteswap(void *vbuf, size_t size)
341 {
342 dnode_phys_t *buf = vbuf;
343 int i;
344
345 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
346 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
347
348 size >>= DNODE_SHIFT;
349 for (i = 0; i < size; i++) {
350 dnode_byteswap(buf);
351 buf++;
352 }
353 }
354
355 void
356 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
357 {
358 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
359
360 dnode_setdirty(dn, tx);
361 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
362 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
363 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
364 dn->dn_bonuslen = newsize;
365 if (newsize == 0)
366 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
367 else
368 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
369 rw_exit(&dn->dn_struct_rwlock);
370 }
371
372 void
373 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
374 {
375 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
376 dnode_setdirty(dn, tx);
377 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
378 dn->dn_bonustype = newtype;
379 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
380 rw_exit(&dn->dn_struct_rwlock);
381 }
382
383 void
384 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
385 {
386 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
387 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
388 dnode_setdirty(dn, tx);
389 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
390 dn->dn_have_spill = B_FALSE;
391 }
392
393 static void
394 dnode_setdblksz(dnode_t *dn, int size)
395 {
396 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
397 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
398 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
399 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
400 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
401 dn->dn_datablksz = size;
402 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
403 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
404 }
405
406 static dnode_t *
407 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
408 uint64_t object, dnode_handle_t *dnh)
409 {
410 dnode_t *dn;
411
412 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
413 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
414 dn->dn_moved = 0;
415
416 /*
417 * Defer setting dn_objset until the dnode is ready to be a candidate
418 * for the dnode_move() callback.
419 */
420 dn->dn_object = object;
421 dn->dn_dbuf = db;
422 dn->dn_handle = dnh;
423 dn->dn_phys = dnp;
424
425 if (dnp->dn_datablkszsec) {
426 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
427 } else {
428 dn->dn_datablksz = 0;
429 dn->dn_datablkszsec = 0;
430 dn->dn_datablkshift = 0;
431 }
432 dn->dn_indblkshift = dnp->dn_indblkshift;
433 dn->dn_nlevels = dnp->dn_nlevels;
434 dn->dn_type = dnp->dn_type;
435 dn->dn_nblkptr = dnp->dn_nblkptr;
436 dn->dn_checksum = dnp->dn_checksum;
437 dn->dn_compress = dnp->dn_compress;
438 dn->dn_bonustype = dnp->dn_bonustype;
439 dn->dn_bonuslen = dnp->dn_bonuslen;
440 dn->dn_maxblkid = dnp->dn_maxblkid;
441 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
442 dn->dn_id_flags = 0;
443
444 dmu_zfetch_init(&dn->dn_zfetch, dn);
445
446 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
447
448 mutex_enter(&os->os_lock);
449 if (dnh->dnh_dnode != NULL) {
450 /* Lost the allocation race. */
451 mutex_exit(&os->os_lock);
452 kmem_cache_free(dnode_cache, dn);
453 return (dnh->dnh_dnode);
454 }
455
456 /*
457 * Exclude special dnodes from os_dnodes so an empty os_dnodes
458 * signifies that the special dnodes have no references from
459 * their children (the entries in os_dnodes). This allows
460 * dnode_destroy() to easily determine if the last child has
461 * been removed and then complete eviction of the objset.
462 */
463 if (!DMU_OBJECT_IS_SPECIAL(object))
464 list_insert_head(&os->os_dnodes, dn);
465 membar_producer();
466
467 /*
468 * Everything else must be valid before assigning dn_objset
469 * makes the dnode eligible for dnode_move().
470 */
471 dn->dn_objset = os;
472
473 dnh->dnh_dnode = dn;
474 mutex_exit(&os->os_lock);
475
476 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
477 return (dn);
478 }
479
480 /*
481 * Caller must be holding the dnode handle, which is released upon return.
482 */
483 static void
484 dnode_destroy(dnode_t *dn)
485 {
486 objset_t *os = dn->dn_objset;
487 boolean_t complete_os_eviction = B_FALSE;
488
489 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
490
491 mutex_enter(&os->os_lock);
492 POINTER_INVALIDATE(&dn->dn_objset);
493 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
494 list_remove(&os->os_dnodes, dn);
495 complete_os_eviction =
496 list_is_empty(&os->os_dnodes) &&
497 list_link_active(&os->os_evicting_node);
498 }
499 mutex_exit(&os->os_lock);
500
501 /* the dnode can no longer move, so we can release the handle */
502 zrl_remove(&dn->dn_handle->dnh_zrlock);
503
504 dn->dn_allocated_txg = 0;
505 dn->dn_free_txg = 0;
506 dn->dn_assigned_txg = 0;
507
508 dn->dn_dirtyctx = 0;
509 if (dn->dn_dirtyctx_firstset != NULL) {
510 kmem_free(dn->dn_dirtyctx_firstset, 1);
511 dn->dn_dirtyctx_firstset = NULL;
512 }
513 if (dn->dn_bonus != NULL) {
514 mutex_enter(&dn->dn_bonus->db_mtx);
515 dbuf_destroy(dn->dn_bonus);
516 dn->dn_bonus = NULL;
517 }
518 dn->dn_zio = NULL;
519
520 dn->dn_have_spill = B_FALSE;
521 dn->dn_oldused = 0;
522 dn->dn_oldflags = 0;
523 dn->dn_olduid = 0;
524 dn->dn_oldgid = 0;
525 dn->dn_newuid = 0;
526 dn->dn_newgid = 0;
527 dn->dn_id_flags = 0;
528
529 dmu_zfetch_fini(&dn->dn_zfetch);
530 kmem_cache_free(dnode_cache, dn);
531 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
532
533 if (complete_os_eviction)
534 dmu_objset_evict_done(os);
535 }
536
537 void
538 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
539 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
540 {
541 int i;
542
543 ASSERT3U(blocksize, <=,
544 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
545 if (blocksize == 0)
546 blocksize = 1 << zfs_default_bs;
547 else
548 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
549
550 if (ibs == 0)
551 ibs = zfs_default_ibs;
552
553 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
554
555 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
556 dn->dn_object, tx->tx_txg, blocksize, ibs);
557
558 ASSERT(dn->dn_type == DMU_OT_NONE);
559 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
560 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
561 ASSERT(ot != DMU_OT_NONE);
562 ASSERT(DMU_OT_IS_VALID(ot));
563 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
564 (bonustype == DMU_OT_SA && bonuslen == 0) ||
565 (bonustype != DMU_OT_NONE && bonuslen != 0));
566 ASSERT(DMU_OT_IS_VALID(bonustype));
567 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
568 ASSERT(dn->dn_type == DMU_OT_NONE);
569 ASSERT0(dn->dn_maxblkid);
570 ASSERT0(dn->dn_allocated_txg);
571 ASSERT0(dn->dn_assigned_txg);
572 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
573 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
574 ASSERT(avl_is_empty(&dn->dn_dbufs));
575
576 for (i = 0; i < TXG_SIZE; i++) {
577 ASSERT0(dn->dn_next_nblkptr[i]);
578 ASSERT0(dn->dn_next_nlevels[i]);
579 ASSERT0(dn->dn_next_indblkshift[i]);
580 ASSERT0(dn->dn_next_bonuslen[i]);
581 ASSERT0(dn->dn_next_bonustype[i]);
582 ASSERT0(dn->dn_rm_spillblk[i]);
583 ASSERT0(dn->dn_next_blksz[i]);
584 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
585 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
586 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
587 }
588
589 dn->dn_type = ot;
590 dnode_setdblksz(dn, blocksize);
591 dn->dn_indblkshift = ibs;
592 dn->dn_nlevels = 1;
593 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
594 dn->dn_nblkptr = 1;
595 else
596 dn->dn_nblkptr = 1 +
597 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
598 dn->dn_bonustype = bonustype;
599 dn->dn_bonuslen = bonuslen;
600 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
601 dn->dn_compress = ZIO_COMPRESS_INHERIT;
602 dn->dn_dirtyctx = 0;
603
604 dn->dn_free_txg = 0;
605 if (dn->dn_dirtyctx_firstset) {
606 kmem_free(dn->dn_dirtyctx_firstset, 1);
607 dn->dn_dirtyctx_firstset = NULL;
608 }
609
610 dn->dn_allocated_txg = tx->tx_txg;
611 dn->dn_id_flags = 0;
612
613 dnode_setdirty(dn, tx);
614 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
615 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
616 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
617 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
618 }
619
620 void
621 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
622 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
623 {
624 int nblkptr;
625
626 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
627 ASSERT3U(blocksize, <=,
628 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
629 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
630 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
631 ASSERT(tx->tx_txg != 0);
632 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
633 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
634 (bonustype == DMU_OT_SA && bonuslen == 0));
635 ASSERT(DMU_OT_IS_VALID(bonustype));
636 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
637
638 /* clean up any unreferenced dbufs */
639 dnode_evict_dbufs(dn);
640
641 dn->dn_id_flags = 0;
642
643 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
644 dnode_setdirty(dn, tx);
645 if (dn->dn_datablksz != blocksize) {
646 /* change blocksize */
647 ASSERT(dn->dn_maxblkid == 0 &&
648 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
649 dnode_block_freed(dn, 0)));
650 dnode_setdblksz(dn, blocksize);
651 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
652 }
653 if (dn->dn_bonuslen != bonuslen)
654 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
655
656 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
657 nblkptr = 1;
658 else
659 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
660 if (dn->dn_bonustype != bonustype)
661 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
662 if (dn->dn_nblkptr != nblkptr)
663 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
664 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
665 dbuf_rm_spill(dn, tx);
666 dnode_rm_spill(dn, tx);
667 }
668 rw_exit(&dn->dn_struct_rwlock);
669
670 /* change type */
671 dn->dn_type = ot;
672
673 /* change bonus size and type */
674 mutex_enter(&dn->dn_mtx);
675 dn->dn_bonustype = bonustype;
676 dn->dn_bonuslen = bonuslen;
677 dn->dn_nblkptr = nblkptr;
678 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
679 dn->dn_compress = ZIO_COMPRESS_INHERIT;
680 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
681
682 /* fix up the bonus db_size */
683 if (dn->dn_bonus) {
684 dn->dn_bonus->db.db_size =
685 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
686 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
687 }
688
689 dn->dn_allocated_txg = tx->tx_txg;
690 mutex_exit(&dn->dn_mtx);
691 }
692
693 #ifdef DNODE_STATS
694 static struct {
695 uint64_t dms_dnode_invalid;
696 uint64_t dms_dnode_recheck1;
697 uint64_t dms_dnode_recheck2;
698 uint64_t dms_dnode_special;
699 uint64_t dms_dnode_handle;
700 uint64_t dms_dnode_rwlock;
701 uint64_t dms_dnode_active;
702 } dnode_move_stats;
703 #endif /* DNODE_STATS */
704
705 #ifdef _KERNEL
706 static void
707 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
708 {
709 int i;
710
711 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
712 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
713 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
714 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
715
716 /* Copy fields. */
717 ndn->dn_objset = odn->dn_objset;
718 ndn->dn_object = odn->dn_object;
719 ndn->dn_dbuf = odn->dn_dbuf;
720 ndn->dn_handle = odn->dn_handle;
721 ndn->dn_phys = odn->dn_phys;
722 ndn->dn_type = odn->dn_type;
723 ndn->dn_bonuslen = odn->dn_bonuslen;
724 ndn->dn_bonustype = odn->dn_bonustype;
725 ndn->dn_nblkptr = odn->dn_nblkptr;
726 ndn->dn_checksum = odn->dn_checksum;
727 ndn->dn_compress = odn->dn_compress;
728 ndn->dn_nlevels = odn->dn_nlevels;
729 ndn->dn_indblkshift = odn->dn_indblkshift;
730 ndn->dn_datablkshift = odn->dn_datablkshift;
731 ndn->dn_datablkszsec = odn->dn_datablkszsec;
732 ndn->dn_datablksz = odn->dn_datablksz;
733 ndn->dn_maxblkid = odn->dn_maxblkid;
734 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
735 sizeof (odn->dn_next_nblkptr));
736 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
737 sizeof (odn->dn_next_nlevels));
738 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
739 sizeof (odn->dn_next_indblkshift));
740 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
741 sizeof (odn->dn_next_bonustype));
742 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
743 sizeof (odn->dn_rm_spillblk));
744 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
745 sizeof (odn->dn_next_bonuslen));
746 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
747 sizeof (odn->dn_next_blksz));
748 for (i = 0; i < TXG_SIZE; i++) {
749 list_move_tail(&ndn->dn_dirty_records[i],
750 &odn->dn_dirty_records[i]);
751 }
752 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
753 sizeof (odn->dn_free_ranges));
754 ndn->dn_allocated_txg = odn->dn_allocated_txg;
755 ndn->dn_free_txg = odn->dn_free_txg;
756 ndn->dn_assigned_txg = odn->dn_assigned_txg;
757 ndn->dn_dirtyctx = odn->dn_dirtyctx;
758 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
759 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
760 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
761 ASSERT(avl_is_empty(&ndn->dn_dbufs));
762 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
763 ndn->dn_dbufs_count = odn->dn_dbufs_count;
764 ndn->dn_bonus = odn->dn_bonus;
765 ndn->dn_have_spill = odn->dn_have_spill;
766 ndn->dn_zio = odn->dn_zio;
767 ndn->dn_oldused = odn->dn_oldused;
768 ndn->dn_oldflags = odn->dn_oldflags;
769 ndn->dn_olduid = odn->dn_olduid;
770 ndn->dn_oldgid = odn->dn_oldgid;
771 ndn->dn_newuid = odn->dn_newuid;
772 ndn->dn_newgid = odn->dn_newgid;
773 ndn->dn_id_flags = odn->dn_id_flags;
774 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
775 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
776 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
777
778 /*
779 * Update back pointers. Updating the handle fixes the back pointer of
780 * every descendant dbuf as well as the bonus dbuf.
781 */
782 ASSERT(ndn->dn_handle->dnh_dnode == odn);
783 ndn->dn_handle->dnh_dnode = ndn;
784 if (ndn->dn_zfetch.zf_dnode == odn) {
785 ndn->dn_zfetch.zf_dnode = ndn;
786 }
787
788 /*
789 * Invalidate the original dnode by clearing all of its back pointers.
790 */
791 odn->dn_dbuf = NULL;
792 odn->dn_handle = NULL;
793 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
794 offsetof(dmu_buf_impl_t, db_link));
795 odn->dn_dbufs_count = 0;
796 odn->dn_bonus = NULL;
797 odn->dn_zfetch.zf_dnode = NULL;
798
799 /*
800 * Set the low bit of the objset pointer to ensure that dnode_move()
801 * recognizes the dnode as invalid in any subsequent callback.
802 */
803 POINTER_INVALIDATE(&odn->dn_objset);
804
805 /*
806 * Satisfy the destructor.
807 */
808 for (i = 0; i < TXG_SIZE; i++) {
809 list_create(&odn->dn_dirty_records[i],
810 sizeof (dbuf_dirty_record_t),
811 offsetof(dbuf_dirty_record_t, dr_dirty_node));
812 odn->dn_free_ranges[i] = NULL;
813 odn->dn_next_nlevels[i] = 0;
814 odn->dn_next_indblkshift[i] = 0;
815 odn->dn_next_bonustype[i] = 0;
816 odn->dn_rm_spillblk[i] = 0;
817 odn->dn_next_bonuslen[i] = 0;
818 odn->dn_next_blksz[i] = 0;
819 }
820 odn->dn_allocated_txg = 0;
821 odn->dn_free_txg = 0;
822 odn->dn_assigned_txg = 0;
823 odn->dn_dirtyctx = 0;
824 odn->dn_dirtyctx_firstset = NULL;
825 odn->dn_have_spill = B_FALSE;
826 odn->dn_zio = NULL;
827 odn->dn_oldused = 0;
828 odn->dn_oldflags = 0;
829 odn->dn_olduid = 0;
830 odn->dn_oldgid = 0;
831 odn->dn_newuid = 0;
832 odn->dn_newgid = 0;
833 odn->dn_id_flags = 0;
834
835 /*
836 * Mark the dnode.
837 */
838 ndn->dn_moved = 1;
839 odn->dn_moved = (uint8_t)-1;
840 }
841
842 /*ARGSUSED*/
843 static kmem_cbrc_t
844 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
845 {
846 dnode_t *odn = buf, *ndn = newbuf;
847 objset_t *os;
848 int64_t refcount;
849 uint32_t dbufs;
850
851 /*
852 * The dnode is on the objset's list of known dnodes if the objset
853 * pointer is valid. We set the low bit of the objset pointer when
854 * freeing the dnode to invalidate it, and the memory patterns written
855 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
856 * A newly created dnode sets the objset pointer last of all to indicate
857 * that the dnode is known and in a valid state to be moved by this
858 * function.
859 */
860 os = odn->dn_objset;
861 if (!POINTER_IS_VALID(os)) {
862 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
863 return (KMEM_CBRC_DONT_KNOW);
864 }
865
866 /*
867 * Ensure that the objset does not go away during the move.
868 */
869 rw_enter(&os_lock, RW_WRITER);
870 if (os != odn->dn_objset) {
871 rw_exit(&os_lock);
872 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
873 return (KMEM_CBRC_DONT_KNOW);
874 }
875
876 /*
877 * If the dnode is still valid, then so is the objset. We know that no
878 * valid objset can be freed while we hold os_lock, so we can safely
879 * ensure that the objset remains in use.
880 */
881 mutex_enter(&os->os_lock);
882
883 /*
884 * Recheck the objset pointer in case the dnode was removed just before
885 * acquiring the lock.
886 */
887 if (os != odn->dn_objset) {
888 mutex_exit(&os->os_lock);
889 rw_exit(&os_lock);
890 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
891 return (KMEM_CBRC_DONT_KNOW);
892 }
893
894 /*
895 * At this point we know that as long as we hold os->os_lock, the dnode
896 * cannot be freed and fields within the dnode can be safely accessed.
897 * The objset listing this dnode cannot go away as long as this dnode is
898 * on its list.
899 */
900 rw_exit(&os_lock);
901 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
902 mutex_exit(&os->os_lock);
903 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
904 return (KMEM_CBRC_NO);
905 }
906 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
907
908 /*
909 * Lock the dnode handle to prevent the dnode from obtaining any new
910 * holds. This also prevents the descendant dbufs and the bonus dbuf
911 * from accessing the dnode, so that we can discount their holds. The
912 * handle is safe to access because we know that while the dnode cannot
913 * go away, neither can its handle. Once we hold dnh_zrlock, we can
914 * safely move any dnode referenced only by dbufs.
915 */
916 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
917 mutex_exit(&os->os_lock);
918 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
919 return (KMEM_CBRC_LATER);
920 }
921
922 /*
923 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
924 * We need to guarantee that there is a hold for every dbuf in order to
925 * determine whether the dnode is actively referenced. Falsely matching
926 * a dbuf to an active hold would lead to an unsafe move. It's possible
927 * that a thread already having an active dnode hold is about to add a
928 * dbuf, and we can't compare hold and dbuf counts while the add is in
929 * progress.
930 */
931 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
932 zrl_exit(&odn->dn_handle->dnh_zrlock);
933 mutex_exit(&os->os_lock);
934 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
935 return (KMEM_CBRC_LATER);
936 }
937
938 /*
939 * A dbuf may be removed (evicted) without an active dnode hold. In that
940 * case, the dbuf count is decremented under the handle lock before the
941 * dbuf's hold is released. This order ensures that if we count the hold
942 * after the dbuf is removed but before its hold is released, we will
943 * treat the unmatched hold as active and exit safely. If we count the
944 * hold before the dbuf is removed, the hold is discounted, and the
945 * removal is blocked until the move completes.
946 */
947 refcount = refcount_count(&odn->dn_holds);
948 ASSERT(refcount >= 0);
949 dbufs = odn->dn_dbufs_count;
950
951 /* We can't have more dbufs than dnode holds. */
952 ASSERT3U(dbufs, <=, refcount);
953 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
954 uint32_t, dbufs);
955
956 if (refcount > dbufs) {
957 rw_exit(&odn->dn_struct_rwlock);
958 zrl_exit(&odn->dn_handle->dnh_zrlock);
959 mutex_exit(&os->os_lock);
960 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
961 return (KMEM_CBRC_LATER);
962 }
963
964 rw_exit(&odn->dn_struct_rwlock);
965
966 /*
967 * At this point we know that anyone with a hold on the dnode is not
968 * actively referencing it. The dnode is known and in a valid state to
969 * move. We're holding the locks needed to execute the critical section.
970 */
971 dnode_move_impl(odn, ndn);
972
973 list_link_replace(&odn->dn_link, &ndn->dn_link);
974 /* If the dnode was safe to move, the refcount cannot have changed. */
975 ASSERT(refcount == refcount_count(&ndn->dn_holds));
976 ASSERT(dbufs == ndn->dn_dbufs_count);
977 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
978 mutex_exit(&os->os_lock);
979
980 return (KMEM_CBRC_YES);
981 }
982 #endif /* _KERNEL */
983
984 void
985 dnode_special_close(dnode_handle_t *dnh)
986 {
987 dnode_t *dn = dnh->dnh_dnode;
988
989 /*
990 * Wait for final references to the dnode to clear. This can
991 * only happen if the arc is asyncronously evicting state that
992 * has a hold on this dnode while we are trying to evict this
993 * dnode.
994 */
995 while (refcount_count(&dn->dn_holds) > 0)
996 delay(1);
997 ASSERT(dn->dn_dbuf == NULL ||
998 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
999 zrl_add(&dnh->dnh_zrlock);
1000 dnode_destroy(dn); /* implicit zrl_remove() */
1001 zrl_destroy(&dnh->dnh_zrlock);
1002 dnh->dnh_dnode = NULL;
1003 }
1004
1005 void
1006 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1007 dnode_handle_t *dnh)
1008 {
1009 dnode_t *dn;
1010
1011 dn = dnode_create(os, dnp, NULL, object, dnh);
1012 zrl_init(&dnh->dnh_zrlock);
1013 DNODE_VERIFY(dn);
1014 }
1015
1016 static void
1017 dnode_buf_evict_async(void *dbu)
1018 {
1019 dnode_children_t *children_dnodes = dbu;
1020 int i;
1021
1022 for (i = 0; i < children_dnodes->dnc_count; i++) {
1023 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1024 dnode_t *dn;
1025
1026 /*
1027 * The dnode handle lock guards against the dnode moving to
1028 * another valid address, so there is no need here to guard
1029 * against changes to or from NULL.
1030 */
1031 if (dnh->dnh_dnode == NULL) {
1032 zrl_destroy(&dnh->dnh_zrlock);
1033 continue;
1034 }
1035
1036 zrl_add(&dnh->dnh_zrlock);
1037 dn = dnh->dnh_dnode;
1038 /*
1039 * If there are holds on this dnode, then there should
1040 * be holds on the dnode's containing dbuf as well; thus
1041 * it wouldn't be eligible for eviction and this function
1042 * would not have been called.
1043 */
1044 ASSERT(refcount_is_zero(&dn->dn_holds));
1045 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1046
1047 dnode_destroy(dn); /* implicit zrl_remove() */
1048 zrl_destroy(&dnh->dnh_zrlock);
1049 dnh->dnh_dnode = NULL;
1050 }
1051 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1052 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1053 }
1054
1055 /*
1056 * errors:
1057 * EINVAL - invalid object number.
1058 * EIO - i/o error.
1059 * succeeds even for free dnodes.
1060 */
1061 int
1062 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1063 void *tag, dnode_t **dnp)
1064 {
1065 int epb, idx, err;
1066 int drop_struct_lock = FALSE;
1067 int type;
1068 uint64_t blk;
1069 dnode_t *mdn, *dn;
1070 dmu_buf_impl_t *db;
1071 dnode_children_t *children_dnodes;
1072 dnode_handle_t *dnh;
1073
1074 /*
1075 * If you are holding the spa config lock as writer, you shouldn't
1076 * be asking the DMU to do *anything* unless it's the root pool
1077 * which may require us to read from the root filesystem while
1078 * holding some (not all) of the locks as writer.
1079 */
1080 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1081 (spa_is_root(os->os_spa) &&
1082 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1083
1084 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1085 dn = (object == DMU_USERUSED_OBJECT) ?
1086 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1087 if (dn == NULL)
1088 return (SET_ERROR(ENOENT));
1089 type = dn->dn_type;
1090 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1091 return (SET_ERROR(ENOENT));
1092 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1093 return (SET_ERROR(EEXIST));
1094 DNODE_VERIFY(dn);
1095 (void) refcount_add(&dn->dn_holds, tag);
1096 *dnp = dn;
1097 return (0);
1098 }
1099
1100 if (object == 0 || object >= DN_MAX_OBJECT)
1101 return (SET_ERROR(EINVAL));
1102
1103 mdn = DMU_META_DNODE(os);
1104 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1105
1106 DNODE_VERIFY(mdn);
1107
1108 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1109 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1110 drop_struct_lock = TRUE;
1111 }
1112
1113 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1114
1115 db = dbuf_hold(mdn, blk, FTAG);
1116 if (drop_struct_lock)
1117 rw_exit(&mdn->dn_struct_rwlock);
1118 if (db == NULL)
1119 return (SET_ERROR(EIO));
1120 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1121 if (err) {
1122 dbuf_rele(db, FTAG);
1123 return (err);
1124 }
1125
1126 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1127 epb = db->db.db_size >> DNODE_SHIFT;
1128
1129 idx = object & (epb-1);
1130
1131 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1132 children_dnodes = dmu_buf_get_user(&db->db);
1133 if (children_dnodes == NULL) {
1134 int i;
1135 dnode_children_t *winner;
1136 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1137 epb * sizeof (dnode_handle_t), KM_SLEEP);
1138 children_dnodes->dnc_count = epb;
1139 dnh = &children_dnodes->dnc_children[0];
1140 for (i = 0; i < epb; i++) {
1141 zrl_init(&dnh[i].dnh_zrlock);
1142 }
1143 dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1144 dnode_buf_evict_async, NULL);
1145 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1146 if (winner != NULL) {
1147
1148 for (i = 0; i < epb; i++) {
1149 zrl_destroy(&dnh[i].dnh_zrlock);
1150 }
1151
1152 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1153 epb * sizeof (dnode_handle_t));
1154 children_dnodes = winner;
1155 }
1156 }
1157 ASSERT(children_dnodes->dnc_count == epb);
1158
1159 dnh = &children_dnodes->dnc_children[idx];
1160 zrl_add(&dnh->dnh_zrlock);
1161 dn = dnh->dnh_dnode;
1162 if (dn == NULL) {
1163 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1164
1165 dn = dnode_create(os, phys, db, object, dnh);
1166 }
1167
1168 mutex_enter(&dn->dn_mtx);
1169 type = dn->dn_type;
1170 if (dn->dn_free_txg ||
1171 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1172 ((flag & DNODE_MUST_BE_FREE) &&
1173 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1174 mutex_exit(&dn->dn_mtx);
1175 zrl_remove(&dnh->dnh_zrlock);
1176 dbuf_rele(db, FTAG);
1177 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1178 }
1179 if (refcount_add(&dn->dn_holds, tag) == 1)
1180 dbuf_add_ref(db, dnh);
1181 mutex_exit(&dn->dn_mtx);
1182
1183 /* Now we can rely on the hold to prevent the dnode from moving. */
1184 zrl_remove(&dnh->dnh_zrlock);
1185
1186 DNODE_VERIFY(dn);
1187 ASSERT3P(dn->dn_dbuf, ==, db);
1188 ASSERT3U(dn->dn_object, ==, object);
1189 dbuf_rele(db, FTAG);
1190
1191 *dnp = dn;
1192 return (0);
1193 }
1194
1195 /*
1196 * Return held dnode if the object is allocated, NULL if not.
1197 */
1198 int
1199 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1200 {
1201 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1202 }
1203
1204 /*
1205 * Can only add a reference if there is already at least one
1206 * reference on the dnode. Returns FALSE if unable to add a
1207 * new reference.
1208 */
1209 boolean_t
1210 dnode_add_ref(dnode_t *dn, void *tag)
1211 {
1212 mutex_enter(&dn->dn_mtx);
1213 if (refcount_is_zero(&dn->dn_holds)) {
1214 mutex_exit(&dn->dn_mtx);
1215 return (FALSE);
1216 }
1217 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1218 mutex_exit(&dn->dn_mtx);
1219 return (TRUE);
1220 }
1221
1222 void
1223 dnode_rele(dnode_t *dn, void *tag)
1224 {
1225 mutex_enter(&dn->dn_mtx);
1226 dnode_rele_and_unlock(dn, tag);
1227 }
1228
1229 void
1230 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1231 {
1232 uint64_t refs;
1233 /* Get while the hold prevents the dnode from moving. */
1234 dmu_buf_impl_t *db = dn->dn_dbuf;
1235 dnode_handle_t *dnh = dn->dn_handle;
1236
1237 refs = refcount_remove(&dn->dn_holds, tag);
1238 mutex_exit(&dn->dn_mtx);
1239
1240 /*
1241 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1242 * indirectly by dbuf_rele() while relying on the dnode handle to
1243 * prevent the dnode from moving, since releasing the last hold could
1244 * result in the dnode's parent dbuf evicting its dnode handles. For
1245 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1246 * other direct or indirect hold on the dnode must first drop the dnode
1247 * handle.
1248 */
1249 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1250
1251 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1252 if (refs == 0 && db != NULL) {
1253 /*
1254 * Another thread could add a hold to the dnode handle in
1255 * dnode_hold_impl() while holding the parent dbuf. Since the
1256 * hold on the parent dbuf prevents the handle from being
1257 * destroyed, the hold on the handle is OK. We can't yet assert
1258 * that the handle has zero references, but that will be
1259 * asserted anyway when the handle gets destroyed.
1260 */
1261 dbuf_rele(db, dnh);
1262 }
1263 }
1264
1265 void
1266 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1267 {
1268 objset_t *os = dn->dn_objset;
1269 uint64_t txg = tx->tx_txg;
1270
1271 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1272 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1273 return;
1274 }
1275
1276 DNODE_VERIFY(dn);
1277
1278 #ifdef ZFS_DEBUG
1279 mutex_enter(&dn->dn_mtx);
1280 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1281 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1282 mutex_exit(&dn->dn_mtx);
1283 #endif
1284
1285 /*
1286 * Determine old uid/gid when necessary
1287 */
1288 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1289
1290 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1291 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1292
1293 /*
1294 * If we are already marked dirty, we're done.
1295 */
1296 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1297 multilist_sublist_unlock(mls);
1298 return;
1299 }
1300
1301 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1302 !avl_is_empty(&dn->dn_dbufs));
1303 ASSERT(dn->dn_datablksz != 0);
1304 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1305 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1306 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1307
1308 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1309 dn->dn_object, txg);
1310
1311 multilist_sublist_insert_head(mls, dn);
1312
1313 multilist_sublist_unlock(mls);
1314
1315 /*
1316 * The dnode maintains a hold on its containing dbuf as
1317 * long as there are holds on it. Each instantiated child
1318 * dbuf maintains a hold on the dnode. When the last child
1319 * drops its hold, the dnode will drop its hold on the
1320 * containing dbuf. We add a "dirty hold" here so that the
1321 * dnode will hang around after we finish processing its
1322 * children.
1323 */
1324 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1325
1326 (void) dbuf_dirty(dn->dn_dbuf, tx);
1327
1328 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1329 }
1330
1331 void
1332 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1333 {
1334 mutex_enter(&dn->dn_mtx);
1335 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1336 mutex_exit(&dn->dn_mtx);
1337 return;
1338 }
1339 dn->dn_free_txg = tx->tx_txg;
1340 mutex_exit(&dn->dn_mtx);
1341
1342 dnode_setdirty(dn, tx);
1343 }
1344
1345 /*
1346 * Try to change the block size for the indicated dnode. This can only
1347 * succeed if there are no blocks allocated or dirty beyond first block
1348 */
1349 int
1350 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1351 {
1352 dmu_buf_impl_t *db;
1353 int err;
1354
1355 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1356 if (size == 0)
1357 size = SPA_MINBLOCKSIZE;
1358 else
1359 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1360
1361 if (ibs == dn->dn_indblkshift)
1362 ibs = 0;
1363
1364 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1365 return (0);
1366
1367 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1368
1369 /* Check for any allocated blocks beyond the first */
1370 if (dn->dn_maxblkid != 0)
1371 goto fail;
1372
1373 mutex_enter(&dn->dn_dbufs_mtx);
1374 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1375 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1376 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1377 db->db_blkid != DMU_SPILL_BLKID) {
1378 mutex_exit(&dn->dn_dbufs_mtx);
1379 goto fail;
1380 }
1381 }
1382 mutex_exit(&dn->dn_dbufs_mtx);
1383
1384 if (ibs && dn->dn_nlevels != 1)
1385 goto fail;
1386
1387 /* resize the old block */
1388 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1389 if (err == 0)
1390 dbuf_new_size(db, size, tx);
1391 else if (err != ENOENT)
1392 goto fail;
1393
1394 dnode_setdblksz(dn, size);
1395 dnode_setdirty(dn, tx);
1396 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1397 if (ibs) {
1398 dn->dn_indblkshift = ibs;
1399 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1400 }
1401 /* rele after we have fixed the blocksize in the dnode */
1402 if (db)
1403 dbuf_rele(db, FTAG);
1404
1405 rw_exit(&dn->dn_struct_rwlock);
1406 return (0);
1407
1408 fail:
1409 rw_exit(&dn->dn_struct_rwlock);
1410 return (SET_ERROR(ENOTSUP));
1411 }
1412
1413 /* read-holding callers must not rely on the lock being continuously held */
1414 void
1415 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1416 {
1417 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1418 int epbs, new_nlevels;
1419 uint64_t sz;
1420
1421 ASSERT(blkid != DMU_BONUS_BLKID);
1422
1423 ASSERT(have_read ?
1424 RW_READ_HELD(&dn->dn_struct_rwlock) :
1425 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1426
1427 /*
1428 * if we have a read-lock, check to see if we need to do any work
1429 * before upgrading to a write-lock.
1430 */
1431 if (have_read) {
1432 if (blkid <= dn->dn_maxblkid)
1433 return;
1434
1435 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1436 rw_exit(&dn->dn_struct_rwlock);
1437 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1438 }
1439 }
1440
1441 if (blkid <= dn->dn_maxblkid)
1442 goto out;
1443
1444 dn->dn_maxblkid = blkid;
1445
1446 /*
1447 * Compute the number of levels necessary to support the new maxblkid.
1448 */
1449 new_nlevels = 1;
1450 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1451 for (sz = dn->dn_nblkptr;
1452 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1453 new_nlevels++;
1454
1455 if (new_nlevels > dn->dn_nlevels) {
1456 int old_nlevels = dn->dn_nlevels;
1457 dmu_buf_impl_t *db;
1458 list_t *list;
1459 dbuf_dirty_record_t *new, *dr, *dr_next;
1460
1461 dn->dn_nlevels = new_nlevels;
1462
1463 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1464 dn->dn_next_nlevels[txgoff] = new_nlevels;
1465
1466 /* dirty the left indirects */
1467 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1468 ASSERT(db != NULL);
1469 new = dbuf_dirty(db, tx);
1470 dbuf_rele(db, FTAG);
1471
1472 /* transfer the dirty records to the new indirect */
1473 mutex_enter(&dn->dn_mtx);
1474 mutex_enter(&new->dt.di.dr_mtx);
1475 list = &dn->dn_dirty_records[txgoff];
1476 for (dr = list_head(list); dr; dr = dr_next) {
1477 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1478 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1479 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1480 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1481 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1482 list_remove(&dn->dn_dirty_records[txgoff], dr);
1483 list_insert_tail(&new->dt.di.dr_children, dr);
1484 dr->dr_parent = new;
1485 }
1486 }
1487 mutex_exit(&new->dt.di.dr_mtx);
1488 mutex_exit(&dn->dn_mtx);
1489 }
1490
1491 out:
1492 if (have_read)
1493 rw_downgrade(&dn->dn_struct_rwlock);
1494 }
1495
1496 static void
1497 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1498 {
1499 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1500 if (db != NULL) {
1501 dmu_buf_will_dirty(&db->db, tx);
1502 dbuf_rele(db, FTAG);
1503 }
1504 }
1505
1506 void
1507 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1508 {
1509 dmu_buf_impl_t *db;
1510 uint64_t blkoff, blkid, nblks;
1511 int blksz, blkshift, head, tail;
1512 int trunc = FALSE;
1513 int epbs;
1514
1515 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1516 blksz = dn->dn_datablksz;
1517 blkshift = dn->dn_datablkshift;
1518 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1519
1520 if (len == DMU_OBJECT_END) {
1521 len = UINT64_MAX - off;
1522 trunc = TRUE;
1523 }
1524
1525 /*
1526 * First, block align the region to free:
1527 */
1528 if (ISP2(blksz)) {
1529 head = P2NPHASE(off, blksz);
1530 blkoff = P2PHASE(off, blksz);
1531 if ((off >> blkshift) > dn->dn_maxblkid)
1532 goto out;
1533 } else {
1534 ASSERT(dn->dn_maxblkid == 0);
1535 if (off == 0 && len >= blksz) {
1536 /*
1537 * Freeing the whole block; fast-track this request.
1538 * Note that we won't dirty any indirect blocks,
1539 * which is fine because we will be freeing the entire
1540 * file and thus all indirect blocks will be freed
1541 * by free_children().
1542 */
1543 blkid = 0;
1544 nblks = 1;
1545 goto done;
1546 } else if (off >= blksz) {
1547 /* Freeing past end-of-data */
1548 goto out;
1549 } else {
1550 /* Freeing part of the block. */
1551 head = blksz - off;
1552 ASSERT3U(head, >, 0);
1553 }
1554 blkoff = off;
1555 }
1556 /* zero out any partial block data at the start of the range */
1557 if (head) {
1558 ASSERT3U(blkoff + head, ==, blksz);
1559 if (len < head)
1560 head = len;
1561 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1562 TRUE, FALSE, FTAG, &db) == 0) {
1563 caddr_t data;
1564
1565 /* don't dirty if it isn't on disk and isn't dirty */
1566 if (db->db_last_dirty ||
1567 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1568 rw_exit(&dn->dn_struct_rwlock);
1569 dmu_buf_will_dirty(&db->db, tx);
1570 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1571 data = db->db.db_data;
1572 bzero(data + blkoff, head);
1573 }
1574 dbuf_rele(db, FTAG);
1575 }
1576 off += head;
1577 len -= head;
1578 }
1579
1580 /* If the range was less than one block, we're done */
1581 if (len == 0)
1582 goto out;
1583
1584 /* If the remaining range is past end of file, we're done */
1585 if ((off >> blkshift) > dn->dn_maxblkid)
1586 goto out;
1587
1588 ASSERT(ISP2(blksz));
1589 if (trunc)
1590 tail = 0;
1591 else
1592 tail = P2PHASE(len, blksz);
1593
1594 ASSERT0(P2PHASE(off, blksz));
1595 /* zero out any partial block data at the end of the range */
1596 if (tail) {
1597 if (len < tail)
1598 tail = len;
1599 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1600 TRUE, FALSE, FTAG, &db) == 0) {
1601 /* don't dirty if not on disk and not dirty */
1602 if (db->db_last_dirty ||
1603 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1604 rw_exit(&dn->dn_struct_rwlock);
1605 dmu_buf_will_dirty(&db->db, tx);
1606 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1607 bzero(db->db.db_data, tail);
1608 }
1609 dbuf_rele(db, FTAG);
1610 }
1611 len -= tail;
1612 }
1613
1614 /* If the range did not include a full block, we are done */
1615 if (len == 0)
1616 goto out;
1617
1618 ASSERT(IS_P2ALIGNED(off, blksz));
1619 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1620 blkid = off >> blkshift;
1621 nblks = len >> blkshift;
1622 if (trunc)
1623 nblks += 1;
1624
1625 /*
1626 * Dirty all the indirect blocks in this range. Note that only
1627 * the first and last indirect blocks can actually be written
1628 * (if they were partially freed) -- they must be dirtied, even if
1629 * they do not exist on disk yet. The interior blocks will
1630 * be freed by free_children(), so they will not actually be written.
1631 * Even though these interior blocks will not be written, we
1632 * dirty them for two reasons:
1633 *
1634 * - It ensures that the indirect blocks remain in memory until
1635 * syncing context. (They have already been prefetched by
1636 * dmu_tx_hold_free(), so we don't have to worry about reading
1637 * them serially here.)
1638 *
1639 * - The dirty space accounting will put pressure on the txg sync
1640 * mechanism to begin syncing, and to delay transactions if there
1641 * is a large amount of freeing. Even though these indirect
1642 * blocks will not be written, we could need to write the same
1643 * amount of space if we copy the freed BPs into deadlists.
1644 */
1645 if (dn->dn_nlevels > 1) {
1646 uint64_t first, last;
1647
1648 first = blkid >> epbs;
1649 dnode_dirty_l1(dn, first, tx);
1650 if (trunc)
1651 last = dn->dn_maxblkid >> epbs;
1652 else
1653 last = (blkid + nblks - 1) >> epbs;
1654 if (last != first)
1655 dnode_dirty_l1(dn, last, tx);
1656
1657 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1658 SPA_BLKPTRSHIFT;
1659 for (uint64_t i = first + 1; i < last; i++) {
1660 /*
1661 * Set i to the blockid of the next non-hole
1662 * level-1 indirect block at or after i. Note
1663 * that dnode_next_offset() operates in terms of
1664 * level-0-equivalent bytes.
1665 */
1666 uint64_t ibyte = i << shift;
1667 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1668 &ibyte, 2, 1, 0);
1669 i = ibyte >> shift;
1670 if (i >= last)
1671 break;
1672
1673 /*
1674 * Normally we should not see an error, either
1675 * from dnode_next_offset() or dbuf_hold_level()
1676 * (except for ESRCH from dnode_next_offset).
1677 * If there is an i/o error, then when we read
1678 * this block in syncing context, it will use
1679 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1680 * to the "failmode" property. dnode_next_offset()
1681 * doesn't have a flag to indicate MUSTSUCCEED.
1682 */
1683 if (err != 0)
1684 break;
1685
1686 dnode_dirty_l1(dn, i, tx);
1687 }
1688 }
1689
1690 done:
1691 /*
1692 * Add this range to the dnode range list.
1693 * We will finish up this free operation in the syncing phase.
1694 */
1695 mutex_enter(&dn->dn_mtx);
1696 int txgoff = tx->tx_txg & TXG_MASK;
1697 if (dn->dn_free_ranges[txgoff] == NULL) {
1698 dn->dn_free_ranges[txgoff] =
1699 range_tree_create(NULL, NULL, &dn->dn_mtx);
1700 }
1701 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1702 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1703 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1704 blkid, nblks, tx->tx_txg);
1705 mutex_exit(&dn->dn_mtx);
1706
1707 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1708 dnode_setdirty(dn, tx);
1709 out:
1710
1711 rw_exit(&dn->dn_struct_rwlock);
1712 }
1713
1714 static boolean_t
1715 dnode_spill_freed(dnode_t *dn)
1716 {
1717 int i;
1718
1719 mutex_enter(&dn->dn_mtx);
1720 for (i = 0; i < TXG_SIZE; i++) {
1721 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1722 break;
1723 }
1724 mutex_exit(&dn->dn_mtx);
1725 return (i < TXG_SIZE);
1726 }
1727
1728 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1729 uint64_t
1730 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1731 {
1732 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1733 int i;
1734
1735 if (blkid == DMU_BONUS_BLKID)
1736 return (FALSE);
1737
1738 /*
1739 * If we're in the process of opening the pool, dp will not be
1740 * set yet, but there shouldn't be anything dirty.
1741 */
1742 if (dp == NULL)
1743 return (FALSE);
1744
1745 if (dn->dn_free_txg)
1746 return (TRUE);
1747
1748 if (blkid == DMU_SPILL_BLKID)
1749 return (dnode_spill_freed(dn));
1750
1751 mutex_enter(&dn->dn_mtx);
1752 for (i = 0; i < TXG_SIZE; i++) {
1753 if (dn->dn_free_ranges[i] != NULL &&
1754 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1755 break;
1756 }
1757 mutex_exit(&dn->dn_mtx);
1758 return (i < TXG_SIZE);
1759 }
1760
1761 /* call from syncing context when we actually write/free space for this dnode */
1762 void
1763 dnode_diduse_space(dnode_t *dn, int64_t delta)
1764 {
1765 uint64_t space;
1766 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1767 dn, dn->dn_phys,
1768 (u_longlong_t)dn->dn_phys->dn_used,
1769 (longlong_t)delta);
1770
1771 mutex_enter(&dn->dn_mtx);
1772 space = DN_USED_BYTES(dn->dn_phys);
1773 if (delta > 0) {
1774 ASSERT3U(space + delta, >=, space); /* no overflow */
1775 } else {
1776 ASSERT3U(space, >=, -delta); /* no underflow */
1777 }
1778 space += delta;
1779 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1780 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1781 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1782 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1783 } else {
1784 dn->dn_phys->dn_used = space;
1785 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1786 }
1787 mutex_exit(&dn->dn_mtx);
1788 }
1789
1790 /*
1791 * Scans a block at the indicated "level" looking for a hole or data,
1792 * depending on 'flags'.
1793 *
1794 * If level > 0, then we are scanning an indirect block looking at its
1795 * pointers. If level == 0, then we are looking at a block of dnodes.
1796 *
1797 * If we don't find what we are looking for in the block, we return ESRCH.
1798 * Otherwise, return with *offset pointing to the beginning (if searching
1799 * forwards) or end (if searching backwards) of the range covered by the
1800 * block pointer we matched on (or dnode).
1801 *
1802 * The basic search algorithm used below by dnode_next_offset() is to
1803 * use this function to search up the block tree (widen the search) until
1804 * we find something (i.e., we don't return ESRCH) and then search back
1805 * down the tree (narrow the search) until we reach our original search
1806 * level.
1807 */
1808 static int
1809 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1810 int lvl, uint64_t blkfill, uint64_t txg)
1811 {
1812 dmu_buf_impl_t *db = NULL;
1813 void *data = NULL;
1814 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1815 uint64_t epb = 1ULL << epbs;
1816 uint64_t minfill, maxfill;
1817 boolean_t hole;
1818 int i, inc, error, span;
1819
1820 dprintf("probing object %llu offset %llx level %d of %u\n",
1821 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1822
1823 hole = ((flags & DNODE_FIND_HOLE) != 0);
1824 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1825 ASSERT(txg == 0 || !hole);
1826
1827 if (lvl == dn->dn_phys->dn_nlevels) {
1828 error = 0;
1829 epb = dn->dn_phys->dn_nblkptr;
1830 data = dn->dn_phys->dn_blkptr;
1831 } else {
1832 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1833 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1834 if (error) {
1835 if (error != ENOENT)
1836 return (error);
1837 if (hole)
1838 return (0);
1839 /*
1840 * This can only happen when we are searching up
1841 * the block tree for data. We don't really need to
1842 * adjust the offset, as we will just end up looking
1843 * at the pointer to this block in its parent, and its
1844 * going to be unallocated, so we will skip over it.
1845 */
1846 return (SET_ERROR(ESRCH));
1847 }
1848 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1849 if (error) {
1850 dbuf_rele(db, FTAG);
1851 return (error);
1852 }
1853 data = db->db.db_data;
1854 }
1855
1856
1857 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1858 db->db_blkptr->blk_birth <= txg ||
1859 BP_IS_HOLE(db->db_blkptr))) {
1860 /*
1861 * This can only happen when we are searching up the tree
1862 * and these conditions mean that we need to keep climbing.
1863 */
1864 error = SET_ERROR(ESRCH);
1865 } else if (lvl == 0) {
1866 dnode_phys_t *dnp = data;
1867 span = DNODE_SHIFT;
1868 ASSERT(dn->dn_type == DMU_OT_DNODE);
1869
1870 for (i = (*offset >> span) & (blkfill - 1);
1871 i >= 0 && i < blkfill; i += inc) {
1872 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1873 break;
1874 *offset += (1ULL << span) * inc;
1875 }
1876 if (i < 0 || i == blkfill)
1877 error = SET_ERROR(ESRCH);
1878 } else {
1879 blkptr_t *bp = data;
1880 uint64_t start = *offset;
1881 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1882 minfill = 0;
1883 maxfill = blkfill << ((lvl - 1) * epbs);
1884
1885 if (hole)
1886 maxfill--;
1887 else
1888 minfill++;
1889
1890 *offset = *offset >> span;
1891 for (i = BF64_GET(*offset, 0, epbs);
1892 i >= 0 && i < epb; i += inc) {
1893 if (BP_GET_FILL(&bp[i]) >= minfill &&
1894 BP_GET_FILL(&bp[i]) <= maxfill &&
1895 (hole || bp[i].blk_birth > txg))
1896 break;
1897 if (inc > 0 || *offset > 0)
1898 *offset += inc;
1899 }
1900 *offset = *offset << span;
1901 if (inc < 0) {
1902 /* traversing backwards; position offset at the end */
1903 ASSERT3U(*offset, <=, start);
1904 *offset = MIN(*offset + (1ULL << span) - 1, start);
1905 } else if (*offset < start) {
1906 *offset = start;
1907 }
1908 if (i < 0 || i >= epb)
1909 error = SET_ERROR(ESRCH);
1910 }
1911
1912 if (db)
1913 dbuf_rele(db, FTAG);
1914
1915 return (error);
1916 }
1917
1918 /*
1919 * Find the next hole, data, or sparse region at or after *offset.
1920 * The value 'blkfill' tells us how many items we expect to find
1921 * in an L0 data block; this value is 1 for normal objects,
1922 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1923 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1924 *
1925 * Examples:
1926 *
1927 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1928 * Finds the next/previous hole/data in a file.
1929 * Used in dmu_offset_next().
1930 *
1931 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1932 * Finds the next free/allocated dnode an objset's meta-dnode.
1933 * Only finds objects that have new contents since txg (ie.
1934 * bonus buffer changes and content removal are ignored).
1935 * Used in dmu_object_next().
1936 *
1937 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1938 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1939 * Used in dmu_object_alloc().
1940 */
1941 int
1942 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1943 int minlvl, uint64_t blkfill, uint64_t txg)
1944 {
1945 uint64_t initial_offset = *offset;
1946 int lvl, maxlvl;
1947 int error = 0;
1948
1949 if (!(flags & DNODE_FIND_HAVELOCK))
1950 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1951
1952 if (dn->dn_phys->dn_nlevels == 0) {
1953 error = SET_ERROR(ESRCH);
1954 goto out;
1955 }
1956
1957 if (dn->dn_datablkshift == 0) {
1958 if (*offset < dn->dn_datablksz) {
1959 if (flags & DNODE_FIND_HOLE)
1960 *offset = dn->dn_datablksz;
1961 } else {
1962 error = SET_ERROR(ESRCH);
1963 }
1964 goto out;
1965 }
1966
1967 maxlvl = dn->dn_phys->dn_nlevels;
1968
1969 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1970 error = dnode_next_offset_level(dn,
1971 flags, offset, lvl, blkfill, txg);
1972 if (error != ESRCH)
1973 break;
1974 }
1975
1976 while (error == 0 && --lvl >= minlvl) {
1977 error = dnode_next_offset_level(dn,
1978 flags, offset, lvl, blkfill, txg);
1979 }
1980
1981 /*
1982 * There's always a "virtual hole" at the end of the object, even
1983 * if all BP's which physically exist are non-holes.
1984 */
1985 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1986 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1987 error = 0;
1988 }
1989
1990 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1991 initial_offset < *offset : initial_offset > *offset))
1992 error = SET_ERROR(ESRCH);
1993 out:
1994 if (!(flags & DNODE_FIND_HAVELOCK))
1995 rw_exit(&dn->dn_struct_rwlock);
1996
1997 return (error);
1998 }