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, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/dbuf.h>
29 #include <sys/dnode.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/spa.h>
37 #include <sys/zio.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/range_tree.h>
40
41 static kmem_cache_t *dnode_cache;
42 /*
43 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
44 * turned on when DEBUG is also defined.
45 */
46 #ifdef DEBUG
47 #define DNODE_STATS
48 #endif /* DEBUG */
49
50 #ifdef DNODE_STATS
51 #define DNODE_STAT_ADD(stat) ((stat)++)
52 #else
53 #define DNODE_STAT_ADD(stat) /* nothing */
54 #endif /* DNODE_STATS */
55
56 static dnode_phys_t dnode_phys_zero;
57
58 int zfs_default_bs = SPA_MINBLOCKSHIFT;
59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
60
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
62
63 static int
64 dbuf_compare(const void *x1, const void *x2)
65 {
66 const dmu_buf_impl_t *d1 = x1;
67 const dmu_buf_impl_t *d2 = x2;
68
69 if (d1->db_level < d2->db_level) {
70 return (-1);
71 }
72 if (d1->db_level > d2->db_level) {
73 return (1);
74 }
75
76 if (d1->db_blkid < d2->db_blkid) {
77 return (-1);
78 }
79 if (d1->db_blkid > d2->db_blkid) {
80 return (1);
81 }
82
83 if (d1->db_state == DB_SEARCH) {
84 ASSERT3S(d2->db_state, !=, DB_SEARCH);
85 return (-1);
86 } else if (d2->db_state == DB_SEARCH) {
87 ASSERT3S(d1->db_state, !=, DB_SEARCH);
88 return (1);
89 }
90
91 if ((uintptr_t)d1 < (uintptr_t)d2) {
92 return (-1);
93 }
94 if ((uintptr_t)d1 > (uintptr_t)d2) {
95 return (1);
96 }
97 return (0);
98 }
99
100 /* ARGSUSED */
101 static int
102 dnode_cons(void *arg, void *unused, int kmflag)
103 {
104 dnode_t *dn = arg;
105 int i;
106
107 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
108 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
109 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
110 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
111
112 /*
113 * Every dbuf has a reference, and dropping a tracked reference is
114 * O(number of references), so don't track dn_holds.
115 */
116 refcount_create_untracked(&dn->dn_holds);
117 refcount_create(&dn->dn_tx_holds);
118 list_link_init(&dn->dn_link);
119
120 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
121 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
122 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
123 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
124 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
125 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
126 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
127
128 for (i = 0; i < TXG_SIZE; i++) {
129 list_link_init(&dn->dn_dirty_link[i]);
130 dn->dn_free_ranges[i] = NULL;
131 list_create(&dn->dn_dirty_records[i],
132 sizeof (dbuf_dirty_record_t),
133 offsetof(dbuf_dirty_record_t, dr_dirty_node));
134 }
135
136 dn->dn_allocated_txg = 0;
137 dn->dn_free_txg = 0;
138 dn->dn_assigned_txg = 0;
139 dn->dn_dirtyctx = 0;
140 dn->dn_dirtyctx_firstset = NULL;
141 dn->dn_bonus = NULL;
142 dn->dn_have_spill = B_FALSE;
143 dn->dn_zio = NULL;
144 dn->dn_oldused = 0;
145 dn->dn_oldflags = 0;
146 dn->dn_olduid = 0;
147 dn->dn_oldgid = 0;
148 dn->dn_newuid = 0;
149 dn->dn_newgid = 0;
150 dn->dn_id_flags = 0;
151
152 dn->dn_dbufs_count = 0;
153 dn->dn_unlisted_l0_blkid = 0;
154 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
155 offsetof(dmu_buf_impl_t, db_link));
156
157 dn->dn_moved = 0;
158 return (0);
159 }
160
161 /* ARGSUSED */
162 static void
163 dnode_dest(void *arg, void *unused)
164 {
165 int i;
166 dnode_t *dn = arg;
167
168 rw_destroy(&dn->dn_struct_rwlock);
169 mutex_destroy(&dn->dn_mtx);
170 mutex_destroy(&dn->dn_dbufs_mtx);
171 cv_destroy(&dn->dn_notxholds);
172 refcount_destroy(&dn->dn_holds);
173 refcount_destroy(&dn->dn_tx_holds);
174 ASSERT(!list_link_active(&dn->dn_link));
175
176 for (i = 0; i < TXG_SIZE; i++) {
177 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
178 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
179 list_destroy(&dn->dn_dirty_records[i]);
180 ASSERT0(dn->dn_next_nblkptr[i]);
181 ASSERT0(dn->dn_next_nlevels[i]);
182 ASSERT0(dn->dn_next_indblkshift[i]);
183 ASSERT0(dn->dn_next_bonustype[i]);
184 ASSERT0(dn->dn_rm_spillblk[i]);
185 ASSERT0(dn->dn_next_bonuslen[i]);
186 ASSERT0(dn->dn_next_blksz[i]);
187 }
188
189 ASSERT0(dn->dn_allocated_txg);
190 ASSERT0(dn->dn_free_txg);
191 ASSERT0(dn->dn_assigned_txg);
192 ASSERT0(dn->dn_dirtyctx);
193 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
194 ASSERT3P(dn->dn_bonus, ==, NULL);
195 ASSERT(!dn->dn_have_spill);
196 ASSERT3P(dn->dn_zio, ==, NULL);
197 ASSERT0(dn->dn_oldused);
198 ASSERT0(dn->dn_oldflags);
199 ASSERT0(dn->dn_olduid);
200 ASSERT0(dn->dn_oldgid);
201 ASSERT0(dn->dn_newuid);
202 ASSERT0(dn->dn_newgid);
203 ASSERT0(dn->dn_id_flags);
204
205 ASSERT0(dn->dn_dbufs_count);
206 ASSERT0(dn->dn_unlisted_l0_blkid);
207 avl_destroy(&dn->dn_dbufs);
208 }
209
210 void
211 dnode_init(void)
212 {
213 ASSERT(dnode_cache == NULL);
214 dnode_cache = kmem_cache_create("dnode_t",
215 sizeof (dnode_t),
216 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
217 kmem_cache_set_move(dnode_cache, dnode_move);
218 }
219
220 void
221 dnode_fini(void)
222 {
223 kmem_cache_destroy(dnode_cache);
224 dnode_cache = NULL;
225 }
226
227
228 #ifdef ZFS_DEBUG
229 void
230 dnode_verify(dnode_t *dn)
231 {
232 int drop_struct_lock = FALSE;
233
234 ASSERT(dn->dn_phys);
235 ASSERT(dn->dn_objset);
236 ASSERT(dn->dn_handle->dnh_dnode == dn);
237
238 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
239
240 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
241 return;
242
243 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
244 rw_enter(&dn->dn_struct_rwlock, RW_READER);
245 drop_struct_lock = TRUE;
246 }
247 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
248 int i;
249 ASSERT3U(dn->dn_indblkshift, >=, 0);
250 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
251 if (dn->dn_datablkshift) {
252 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
253 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
254 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
255 }
256 ASSERT3U(dn->dn_nlevels, <=, 30);
257 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
258 ASSERT3U(dn->dn_nblkptr, >=, 1);
259 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
260 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
261 ASSERT3U(dn->dn_datablksz, ==,
262 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
263 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
264 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
265 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
266 for (i = 0; i < TXG_SIZE; i++) {
267 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
268 }
269 }
270 if (dn->dn_phys->dn_type != DMU_OT_NONE)
271 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
272 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
273 if (dn->dn_dbuf != NULL) {
274 ASSERT3P(dn->dn_phys, ==,
275 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
276 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
277 }
278 if (drop_struct_lock)
279 rw_exit(&dn->dn_struct_rwlock);
280 }
281 #endif
282
283 void
284 dnode_byteswap(dnode_phys_t *dnp)
285 {
286 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
287 int i;
288
289 if (dnp->dn_type == DMU_OT_NONE) {
290 bzero(dnp, sizeof (dnode_phys_t));
291 return;
292 }
293
294 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
295 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
296 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
297 dnp->dn_used = BSWAP_64(dnp->dn_used);
298
299 /*
300 * dn_nblkptr is only one byte, so it's OK to read it in either
301 * byte order. We can't read dn_bouslen.
302 */
303 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
304 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
305 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
306 buf64[i] = BSWAP_64(buf64[i]);
307
308 /*
309 * OK to check dn_bonuslen for zero, because it won't matter if
310 * we have the wrong byte order. This is necessary because the
311 * dnode dnode is smaller than a regular dnode.
312 */
313 if (dnp->dn_bonuslen != 0) {
314 /*
315 * Note that the bonus length calculated here may be
316 * longer than the actual bonus buffer. This is because
317 * we always put the bonus buffer after the last block
318 * pointer (instead of packing it against the end of the
319 * dnode buffer).
320 */
321 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
322 size_t len = DN_MAX_BONUSLEN - off;
323 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
324 dmu_object_byteswap_t byteswap =
325 DMU_OT_BYTESWAP(dnp->dn_bonustype);
326 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
327 }
328
329 /* Swap SPILL block if we have one */
330 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
331 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
332
333 }
334
335 void
336 dnode_buf_byteswap(void *vbuf, size_t size)
337 {
338 dnode_phys_t *buf = vbuf;
339 int i;
340
341 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
342 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
343
344 size >>= DNODE_SHIFT;
345 for (i = 0; i < size; i++) {
346 dnode_byteswap(buf);
347 buf++;
348 }
349 }
350
351 void
352 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
353 {
354 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
355
356 dnode_setdirty(dn, tx);
357 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
358 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
359 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
360 dn->dn_bonuslen = newsize;
361 if (newsize == 0)
362 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
363 else
364 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
365 rw_exit(&dn->dn_struct_rwlock);
366 }
367
368 void
369 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
370 {
371 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
372 dnode_setdirty(dn, tx);
373 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
374 dn->dn_bonustype = newtype;
375 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
376 rw_exit(&dn->dn_struct_rwlock);
377 }
378
379 void
380 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
381 {
382 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
383 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
384 dnode_setdirty(dn, tx);
385 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
386 dn->dn_have_spill = B_FALSE;
387 }
388
389 static void
390 dnode_setdblksz(dnode_t *dn, int size)
391 {
392 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
393 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
394 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
395 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
396 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
397 dn->dn_datablksz = size;
398 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
399 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
400 }
401
402 static dnode_t *
403 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
404 uint64_t object, dnode_handle_t *dnh)
405 {
406 dnode_t *dn;
407
408 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
409 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
410 dn->dn_moved = 0;
411
412 /*
413 * Defer setting dn_objset until the dnode is ready to be a candidate
414 * for the dnode_move() callback.
415 */
416 dn->dn_object = object;
417 dn->dn_dbuf = db;
418 dn->dn_handle = dnh;
419 dn->dn_phys = dnp;
420
421 if (dnp->dn_datablkszsec) {
422 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
423 } else {
424 dn->dn_datablksz = 0;
425 dn->dn_datablkszsec = 0;
426 dn->dn_datablkshift = 0;
427 }
428 dn->dn_indblkshift = dnp->dn_indblkshift;
429 dn->dn_nlevels = dnp->dn_nlevels;
430 dn->dn_type = dnp->dn_type;
431 dn->dn_nblkptr = dnp->dn_nblkptr;
432 dn->dn_checksum = dnp->dn_checksum;
433 dn->dn_compress = dnp->dn_compress;
434 dn->dn_bonustype = dnp->dn_bonustype;
435 dn->dn_bonuslen = dnp->dn_bonuslen;
436 dn->dn_maxblkid = dnp->dn_maxblkid;
437 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
438 dn->dn_id_flags = 0;
439
440 dmu_zfetch_init(&dn->dn_zfetch, dn);
441
442 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
443
444 mutex_enter(&os->os_lock);
445 if (dnh->dnh_dnode != NULL) {
446 /* Lost the allocation race. */
447 mutex_exit(&os->os_lock);
448 kmem_cache_free(dnode_cache, dn);
449 return (dnh->dnh_dnode);
450 }
451
452 /*
453 * Exclude special dnodes from os_dnodes so an empty os_dnodes
454 * signifies that the special dnodes have no references from
455 * their children (the entries in os_dnodes). This allows
456 * dnode_destroy() to easily determine if the last child has
457 * been removed and then complete eviction of the objset.
458 */
459 if (!DMU_OBJECT_IS_SPECIAL(object))
460 list_insert_head(&os->os_dnodes, dn);
461 membar_producer();
462
463 /*
464 * Everything else must be valid before assigning dn_objset
465 * makes the dnode eligible for dnode_move().
466 */
467 dn->dn_objset = os;
468
469 dnh->dnh_dnode = dn;
470 mutex_exit(&os->os_lock);
471
472 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
473 return (dn);
474 }
475
476 /*
477 * Caller must be holding the dnode handle, which is released upon return.
478 */
479 static void
480 dnode_destroy(dnode_t *dn)
481 {
482 objset_t *os = dn->dn_objset;
483 boolean_t complete_os_eviction = B_FALSE;
484
485 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
486
487 mutex_enter(&os->os_lock);
488 POINTER_INVALIDATE(&dn->dn_objset);
489 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
490 list_remove(&os->os_dnodes, dn);
491 complete_os_eviction =
492 list_is_empty(&os->os_dnodes) &&
493 list_link_active(&os->os_evicting_node);
494 }
495 mutex_exit(&os->os_lock);
496
497 /* the dnode can no longer move, so we can release the handle */
498 zrl_remove(&dn->dn_handle->dnh_zrlock);
499
500 dn->dn_allocated_txg = 0;
501 dn->dn_free_txg = 0;
502 dn->dn_assigned_txg = 0;
503
504 dn->dn_dirtyctx = 0;
505 if (dn->dn_dirtyctx_firstset != NULL) {
506 kmem_free(dn->dn_dirtyctx_firstset, 1);
507 dn->dn_dirtyctx_firstset = NULL;
508 }
509 if (dn->dn_bonus != NULL) {
510 mutex_enter(&dn->dn_bonus->db_mtx);
511 dbuf_evict(dn->dn_bonus);
512 dn->dn_bonus = NULL;
513 }
514 dn->dn_zio = NULL;
515
516 dn->dn_have_spill = B_FALSE;
517 dn->dn_oldused = 0;
518 dn->dn_oldflags = 0;
519 dn->dn_olduid = 0;
520 dn->dn_oldgid = 0;
521 dn->dn_newuid = 0;
522 dn->dn_newgid = 0;
523 dn->dn_id_flags = 0;
524 dn->dn_unlisted_l0_blkid = 0;
525
526 dmu_zfetch_rele(&dn->dn_zfetch);
527 kmem_cache_free(dnode_cache, dn);
528 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
529
530 if (complete_os_eviction)
531 dmu_objset_evict_done(os);
532 }
533
534 void
535 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
536 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
537 {
538 int i;
539
540 ASSERT3U(blocksize, <=,
541 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
542 if (blocksize == 0)
543 blocksize = 1 << zfs_default_bs;
544 else
545 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
546
547 if (ibs == 0)
548 ibs = zfs_default_ibs;
549
550 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
551
552 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
553 dn->dn_object, tx->tx_txg, blocksize, ibs);
554
555 ASSERT(dn->dn_type == DMU_OT_NONE);
556 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
557 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
558 ASSERT(ot != DMU_OT_NONE);
559 ASSERT(DMU_OT_IS_VALID(ot));
560 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
561 (bonustype == DMU_OT_SA && bonuslen == 0) ||
562 (bonustype != DMU_OT_NONE && bonuslen != 0));
563 ASSERT(DMU_OT_IS_VALID(bonustype));
564 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
565 ASSERT(dn->dn_type == DMU_OT_NONE);
566 ASSERT0(dn->dn_maxblkid);
567 ASSERT0(dn->dn_allocated_txg);
568 ASSERT0(dn->dn_assigned_txg);
569 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
570 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
571 ASSERT(avl_is_empty(&dn->dn_dbufs));
572
573 for (i = 0; i < TXG_SIZE; i++) {
574 ASSERT0(dn->dn_next_nblkptr[i]);
575 ASSERT0(dn->dn_next_nlevels[i]);
576 ASSERT0(dn->dn_next_indblkshift[i]);
577 ASSERT0(dn->dn_next_bonuslen[i]);
578 ASSERT0(dn->dn_next_bonustype[i]);
579 ASSERT0(dn->dn_rm_spillblk[i]);
580 ASSERT0(dn->dn_next_blksz[i]);
581 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
582 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
583 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
584 }
585
586 dn->dn_type = ot;
587 dnode_setdblksz(dn, blocksize);
588 dn->dn_indblkshift = ibs;
589 dn->dn_nlevels = 1;
590 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
591 dn->dn_nblkptr = 1;
592 else
593 dn->dn_nblkptr = 1 +
594 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
595 dn->dn_bonustype = bonustype;
596 dn->dn_bonuslen = bonuslen;
597 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
598 dn->dn_compress = ZIO_COMPRESS_INHERIT;
599 dn->dn_dirtyctx = 0;
600
601 dn->dn_free_txg = 0;
602 if (dn->dn_dirtyctx_firstset) {
603 kmem_free(dn->dn_dirtyctx_firstset, 1);
604 dn->dn_dirtyctx_firstset = NULL;
605 }
606
607 dn->dn_allocated_txg = tx->tx_txg;
608 dn->dn_id_flags = 0;
609
610 dnode_setdirty(dn, tx);
611 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
612 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
613 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
614 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
615 }
616
617 void
618 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
619 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
620 {
621 int nblkptr;
622
623 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
624 ASSERT3U(blocksize, <=,
625 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
626 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
627 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
628 ASSERT(tx->tx_txg != 0);
629 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
630 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
631 (bonustype == DMU_OT_SA && bonuslen == 0));
632 ASSERT(DMU_OT_IS_VALID(bonustype));
633 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
634
635 /* clean up any unreferenced dbufs */
636 dnode_evict_dbufs(dn);
637
638 dn->dn_id_flags = 0;
639
640 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
641 dnode_setdirty(dn, tx);
642 if (dn->dn_datablksz != blocksize) {
643 /* change blocksize */
644 ASSERT(dn->dn_maxblkid == 0 &&
645 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
646 dnode_block_freed(dn, 0)));
647 dnode_setdblksz(dn, blocksize);
648 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
649 }
650 if (dn->dn_bonuslen != bonuslen)
651 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
652
653 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
654 nblkptr = 1;
655 else
656 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
657 if (dn->dn_bonustype != bonustype)
658 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
659 if (dn->dn_nblkptr != nblkptr)
660 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
661 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
662 dbuf_rm_spill(dn, tx);
663 dnode_rm_spill(dn, tx);
664 }
665 rw_exit(&dn->dn_struct_rwlock);
666
667 /* change type */
668 dn->dn_type = ot;
669
670 /* change bonus size and type */
671 mutex_enter(&dn->dn_mtx);
672 dn->dn_bonustype = bonustype;
673 dn->dn_bonuslen = bonuslen;
674 dn->dn_nblkptr = nblkptr;
675 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
676 dn->dn_compress = ZIO_COMPRESS_INHERIT;
677 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
678
679 /* fix up the bonus db_size */
680 if (dn->dn_bonus) {
681 dn->dn_bonus->db.db_size =
682 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
683 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
684 }
685
686 dn->dn_allocated_txg = tx->tx_txg;
687 mutex_exit(&dn->dn_mtx);
688 }
689
690 #ifdef DNODE_STATS
691 static struct {
692 uint64_t dms_dnode_invalid;
693 uint64_t dms_dnode_recheck1;
694 uint64_t dms_dnode_recheck2;
695 uint64_t dms_dnode_special;
696 uint64_t dms_dnode_handle;
697 uint64_t dms_dnode_rwlock;
698 uint64_t dms_dnode_active;
699 } dnode_move_stats;
700 #endif /* DNODE_STATS */
701
702 static void
703 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
704 {
705 int i;
706
707 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
708 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
709 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
710 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
711
712 /* Copy fields. */
713 ndn->dn_objset = odn->dn_objset;
714 ndn->dn_object = odn->dn_object;
715 ndn->dn_dbuf = odn->dn_dbuf;
716 ndn->dn_handle = odn->dn_handle;
717 ndn->dn_phys = odn->dn_phys;
718 ndn->dn_type = odn->dn_type;
719 ndn->dn_bonuslen = odn->dn_bonuslen;
720 ndn->dn_bonustype = odn->dn_bonustype;
721 ndn->dn_nblkptr = odn->dn_nblkptr;
722 ndn->dn_checksum = odn->dn_checksum;
723 ndn->dn_compress = odn->dn_compress;
724 ndn->dn_nlevels = odn->dn_nlevels;
725 ndn->dn_indblkshift = odn->dn_indblkshift;
726 ndn->dn_datablkshift = odn->dn_datablkshift;
727 ndn->dn_datablkszsec = odn->dn_datablkszsec;
728 ndn->dn_datablksz = odn->dn_datablksz;
729 ndn->dn_maxblkid = odn->dn_maxblkid;
730 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
731 sizeof (odn->dn_next_nblkptr));
732 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
733 sizeof (odn->dn_next_nlevels));
734 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
735 sizeof (odn->dn_next_indblkshift));
736 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
737 sizeof (odn->dn_next_bonustype));
738 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
739 sizeof (odn->dn_rm_spillblk));
740 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
741 sizeof (odn->dn_next_bonuslen));
742 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
743 sizeof (odn->dn_next_blksz));
744 for (i = 0; i < TXG_SIZE; i++) {
745 list_move_tail(&ndn->dn_dirty_records[i],
746 &odn->dn_dirty_records[i]);
747 }
748 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
749 sizeof (odn->dn_free_ranges));
750 ndn->dn_allocated_txg = odn->dn_allocated_txg;
751 ndn->dn_free_txg = odn->dn_free_txg;
752 ndn->dn_assigned_txg = odn->dn_assigned_txg;
753 ndn->dn_dirtyctx = odn->dn_dirtyctx;
754 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
755 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
756 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
757 ASSERT(avl_is_empty(&ndn->dn_dbufs));
758 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
759 ndn->dn_dbufs_count = odn->dn_dbufs_count;
760 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
761 ndn->dn_bonus = odn->dn_bonus;
762 ndn->dn_have_spill = odn->dn_have_spill;
763 ndn->dn_zio = odn->dn_zio;
764 ndn->dn_oldused = odn->dn_oldused;
765 ndn->dn_oldflags = odn->dn_oldflags;
766 ndn->dn_olduid = odn->dn_olduid;
767 ndn->dn_oldgid = odn->dn_oldgid;
768 ndn->dn_newuid = odn->dn_newuid;
769 ndn->dn_newgid = odn->dn_newgid;
770 ndn->dn_id_flags = odn->dn_id_flags;
771 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
772 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
773 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
774 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
775 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
776
777 /*
778 * Update back pointers. Updating the handle fixes the back pointer of
779 * every descendant dbuf as well as the bonus dbuf.
780 */
781 ASSERT(ndn->dn_handle->dnh_dnode == odn);
782 ndn->dn_handle->dnh_dnode = ndn;
783 if (ndn->dn_zfetch.zf_dnode == odn) {
784 ndn->dn_zfetch.zf_dnode = ndn;
785 }
786
787 /*
788 * Invalidate the original dnode by clearing all of its back pointers.
789 */
790 odn->dn_dbuf = NULL;
791 odn->dn_handle = NULL;
792 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
793 offsetof(dmu_buf_impl_t, db_link));
794 odn->dn_dbufs_count = 0;
795 odn->dn_unlisted_l0_blkid = 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 #ifdef _KERNEL
843 /*ARGSUSED*/
844 static kmem_cbrc_t
845 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
846 {
847 dnode_t *odn = buf, *ndn = newbuf;
848 objset_t *os;
849 int64_t refcount;
850 uint32_t dbufs;
851
852 /*
853 * The dnode is on the objset's list of known dnodes if the objset
854 * pointer is valid. We set the low bit of the objset pointer when
855 * freeing the dnode to invalidate it, and the memory patterns written
856 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
857 * A newly created dnode sets the objset pointer last of all to indicate
858 * that the dnode is known and in a valid state to be moved by this
859 * function.
860 */
861 os = odn->dn_objset;
862 if (!POINTER_IS_VALID(os)) {
863 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
864 return (KMEM_CBRC_DONT_KNOW);
865 }
866
867 /*
868 * Ensure that the objset does not go away during the move.
869 */
870 rw_enter(&os_lock, RW_WRITER);
871 if (os != odn->dn_objset) {
872 rw_exit(&os_lock);
873 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
874 return (KMEM_CBRC_DONT_KNOW);
875 }
876
877 /*
878 * If the dnode is still valid, then so is the objset. We know that no
879 * valid objset can be freed while we hold os_lock, so we can safely
880 * ensure that the objset remains in use.
881 */
882 mutex_enter(&os->os_lock);
883
884 /*
885 * Recheck the objset pointer in case the dnode was removed just before
886 * acquiring the lock.
887 */
888 if (os != odn->dn_objset) {
889 mutex_exit(&os->os_lock);
890 rw_exit(&os_lock);
891 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
892 return (KMEM_CBRC_DONT_KNOW);
893 }
894
895 /*
896 * At this point we know that as long as we hold os->os_lock, the dnode
897 * cannot be freed and fields within the dnode can be safely accessed.
898 * The objset listing this dnode cannot go away as long as this dnode is
899 * on its list.
900 */
901 rw_exit(&os_lock);
902 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
903 mutex_exit(&os->os_lock);
904 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
905 return (KMEM_CBRC_NO);
906 }
907 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
908
909 /*
910 * Lock the dnode handle to prevent the dnode from obtaining any new
911 * holds. This also prevents the descendant dbufs and the bonus dbuf
912 * from accessing the dnode, so that we can discount their holds. The
913 * handle is safe to access because we know that while the dnode cannot
914 * go away, neither can its handle. Once we hold dnh_zrlock, we can
915 * safely move any dnode referenced only by dbufs.
916 */
917 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
918 mutex_exit(&os->os_lock);
919 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
920 return (KMEM_CBRC_LATER);
921 }
922
923 /*
924 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
925 * We need to guarantee that there is a hold for every dbuf in order to
926 * determine whether the dnode is actively referenced. Falsely matching
927 * a dbuf to an active hold would lead to an unsafe move. It's possible
928 * that a thread already having an active dnode hold is about to add a
929 * dbuf, and we can't compare hold and dbuf counts while the add is in
930 * progress.
931 */
932 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
933 zrl_exit(&odn->dn_handle->dnh_zrlock);
934 mutex_exit(&os->os_lock);
935 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
936 return (KMEM_CBRC_LATER);
937 }
938
939 /*
940 * A dbuf may be removed (evicted) without an active dnode hold. In that
941 * case, the dbuf count is decremented under the handle lock before the
942 * dbuf's hold is released. This order ensures that if we count the hold
943 * after the dbuf is removed but before its hold is released, we will
944 * treat the unmatched hold as active and exit safely. If we count the
945 * hold before the dbuf is removed, the hold is discounted, and the
946 * removal is blocked until the move completes.
947 */
948 refcount = refcount_count(&odn->dn_holds);
949 ASSERT(refcount >= 0);
950 dbufs = odn->dn_dbufs_count;
951
952 /* We can't have more dbufs than dnode holds. */
953 ASSERT3U(dbufs, <=, refcount);
954 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
955 uint32_t, dbufs);
956
957 if (refcount > dbufs) {
958 rw_exit(&odn->dn_struct_rwlock);
959 zrl_exit(&odn->dn_handle->dnh_zrlock);
960 mutex_exit(&os->os_lock);
961 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
962 return (KMEM_CBRC_LATER);
963 }
964
965 rw_exit(&odn->dn_struct_rwlock);
966
967 /*
968 * At this point we know that anyone with a hold on the dnode is not
969 * actively referencing it. The dnode is known and in a valid state to
970 * move. We're holding the locks needed to execute the critical section.
971 */
972 dnode_move_impl(odn, ndn);
973
974 list_link_replace(&odn->dn_link, &ndn->dn_link);
975 /* If the dnode was safe to move, the refcount cannot have changed. */
976 ASSERT(refcount == refcount_count(&ndn->dn_holds));
977 ASSERT(dbufs == ndn->dn_dbufs_count);
978 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
979 mutex_exit(&os->os_lock);
980
981 return (KMEM_CBRC_YES);
982 }
983 #endif /* _KERNEL */
984
985 void
986 dnode_special_close(dnode_handle_t *dnh)
987 {
988 dnode_t *dn = dnh->dnh_dnode;
989
990 /*
991 * Wait for final references to the dnode to clear. This can
992 * only happen if the arc is asyncronously evicting state that
993 * has a hold on this dnode while we are trying to evict this
994 * dnode.
995 */
996 while (refcount_count(&dn->dn_holds) > 0)
997 delay(1);
998 ASSERT(dn->dn_dbuf == NULL ||
999 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1000 zrl_add(&dnh->dnh_zrlock);
1001 dnode_destroy(dn); /* implicit zrl_remove() */
1002 zrl_destroy(&dnh->dnh_zrlock);
1003 dnh->dnh_dnode = NULL;
1004 }
1005
1006 void
1007 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1008 dnode_handle_t *dnh)
1009 {
1010 dnode_t *dn;
1011
1012 dn = dnode_create(os, dnp, NULL, object, dnh);
1013 zrl_init(&dnh->dnh_zrlock);
1014 DNODE_VERIFY(dn);
1015 }
1016
1017 static void
1018 dnode_buf_pageout(void *dbu)
1019 {
1020 dnode_children_t *children_dnodes = dbu;
1021 int i;
1022
1023 for (i = 0; i < children_dnodes->dnc_count; i++) {
1024 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1025 dnode_t *dn;
1026
1027 /*
1028 * The dnode handle lock guards against the dnode moving to
1029 * another valid address, so there is no need here to guard
1030 * against changes to or from NULL.
1031 */
1032 if (dnh->dnh_dnode == NULL) {
1033 zrl_destroy(&dnh->dnh_zrlock);
1034 continue;
1035 }
1036
1037 zrl_add(&dnh->dnh_zrlock);
1038 dn = dnh->dnh_dnode;
1039 /*
1040 * If there are holds on this dnode, then there should
1041 * be holds on the dnode's containing dbuf as well; thus
1042 * it wouldn't be eligible for eviction and this function
1043 * would not have been called.
1044 */
1045 ASSERT(refcount_is_zero(&dn->dn_holds));
1046 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1047
1048 dnode_destroy(dn); /* implicit zrl_remove() */
1049 zrl_destroy(&dnh->dnh_zrlock);
1050 dnh->dnh_dnode = NULL;
1051 }
1052 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1053 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1054 }
1055
1056 /*
1057 * errors:
1058 * EINVAL - invalid object number.
1059 * EIO - i/o error.
1060 * succeeds even for free dnodes.
1061 */
1062 int
1063 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1064 void *tag, dnode_t **dnp)
1065 {
1066 int epb, idx, err;
1067 int drop_struct_lock = FALSE;
1068 int type;
1069 uint64_t blk;
1070 dnode_t *mdn, *dn;
1071 dmu_buf_impl_t *db;
1072 dnode_children_t *children_dnodes;
1073 dnode_handle_t *dnh;
1074
1075 /*
1076 * If you are holding the spa config lock as writer, you shouldn't
1077 * be asking the DMU to do *anything* unless it's the root pool
1078 * which may require us to read from the root filesystem while
1079 * holding some (not all) of the locks as writer.
1080 */
1081 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1082 (spa_is_root(os->os_spa) &&
1083 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1084
1085 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1086 dn = (object == DMU_USERUSED_OBJECT) ?
1087 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1088 if (dn == NULL)
1089 return (SET_ERROR(ENOENT));
1090 type = dn->dn_type;
1091 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1092 return (SET_ERROR(ENOENT));
1093 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1094 return (SET_ERROR(EEXIST));
1095 DNODE_VERIFY(dn);
1096 (void) refcount_add(&dn->dn_holds, tag);
1097 *dnp = dn;
1098 return (0);
1099 }
1100
1101 if (object == 0 || object >= DN_MAX_OBJECT)
1102 return (SET_ERROR(EINVAL));
1103
1104 mdn = DMU_META_DNODE(os);
1105 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1106
1107 DNODE_VERIFY(mdn);
1108
1109 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1110 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1111 drop_struct_lock = TRUE;
1112 }
1113
1114 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1115
1116 db = dbuf_hold(mdn, blk, FTAG);
1117 if (drop_struct_lock)
1118 rw_exit(&mdn->dn_struct_rwlock);
1119 if (db == NULL)
1120 return (SET_ERROR(EIO));
1121 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1122 if (err) {
1123 dbuf_rele(db, FTAG);
1124 return (err);
1125 }
1126
1127 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1128 epb = db->db.db_size >> DNODE_SHIFT;
1129
1130 idx = object & (epb-1);
1131
1132 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1133 children_dnodes = dmu_buf_get_user(&db->db);
1134 if (children_dnodes == NULL) {
1135 int i;
1136 dnode_children_t *winner;
1137 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1138 epb * sizeof (dnode_handle_t), KM_SLEEP);
1139 children_dnodes->dnc_count = epb;
1140 dnh = &children_dnodes->dnc_children[0];
1141 for (i = 0; i < epb; i++) {
1142 zrl_init(&dnh[i].dnh_zrlock);
1143 }
1144 dmu_buf_init_user(&children_dnodes->dnc_dbu,
1145 dnode_buf_pageout, NULL);
1146 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1147 if (winner != NULL) {
1148
1149 for (i = 0; i < epb; i++) {
1150 zrl_destroy(&dnh[i].dnh_zrlock);
1151 }
1152
1153 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1154 epb * sizeof (dnode_handle_t));
1155 children_dnodes = winner;
1156 }
1157 }
1158 ASSERT(children_dnodes->dnc_count == epb);
1159
1160 dnh = &children_dnodes->dnc_children[idx];
1161 zrl_add(&dnh->dnh_zrlock);
1162 dn = dnh->dnh_dnode;
1163 if (dn == NULL) {
1164 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1165
1166 dn = dnode_create(os, phys, db, object, dnh);
1167 }
1168
1169 mutex_enter(&dn->dn_mtx);
1170 type = dn->dn_type;
1171 if (dn->dn_free_txg ||
1172 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1173 ((flag & DNODE_MUST_BE_FREE) &&
1174 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1175 mutex_exit(&dn->dn_mtx);
1176 zrl_remove(&dnh->dnh_zrlock);
1177 dbuf_rele(db, FTAG);
1178 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1179 }
1180 if (refcount_add(&dn->dn_holds, tag) == 1)
1181 dbuf_add_ref(db, dnh);
1182 mutex_exit(&dn->dn_mtx);
1183
1184 /* Now we can rely on the hold to prevent the dnode from moving. */
1185 zrl_remove(&dnh->dnh_zrlock);
1186
1187 DNODE_VERIFY(dn);
1188 ASSERT3P(dn->dn_dbuf, ==, db);
1189 ASSERT3U(dn->dn_object, ==, object);
1190 dbuf_rele(db, FTAG);
1191
1192 *dnp = dn;
1193 return (0);
1194 }
1195
1196 /*
1197 * Return held dnode if the object is allocated, NULL if not.
1198 */
1199 int
1200 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1201 {
1202 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1203 }
1204
1205 /*
1206 * Can only add a reference if there is already at least one
1207 * reference on the dnode. Returns FALSE if unable to add a
1208 * new reference.
1209 */
1210 boolean_t
1211 dnode_add_ref(dnode_t *dn, void *tag)
1212 {
1213 mutex_enter(&dn->dn_mtx);
1214 if (refcount_is_zero(&dn->dn_holds)) {
1215 mutex_exit(&dn->dn_mtx);
1216 return (FALSE);
1217 }
1218 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1219 mutex_exit(&dn->dn_mtx);
1220 return (TRUE);
1221 }
1222
1223 void
1224 dnode_rele(dnode_t *dn, void *tag)
1225 {
1226 uint64_t refs;
1227 /* Get while the hold prevents the dnode from moving. */
1228 dmu_buf_impl_t *db = dn->dn_dbuf;
1229 dnode_handle_t *dnh = dn->dn_handle;
1230
1231 mutex_enter(&dn->dn_mtx);
1232 refs = refcount_remove(&dn->dn_holds, tag);
1233 mutex_exit(&dn->dn_mtx);
1234
1235 /*
1236 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1237 * indirectly by dbuf_rele() while relying on the dnode handle to
1238 * prevent the dnode from moving, since releasing the last hold could
1239 * result in the dnode's parent dbuf evicting its dnode handles. For
1240 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1241 * other direct or indirect hold on the dnode must first drop the dnode
1242 * handle.
1243 */
1244 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1245
1246 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1247 if (refs == 0 && db != NULL) {
1248 /*
1249 * Another thread could add a hold to the dnode handle in
1250 * dnode_hold_impl() while holding the parent dbuf. Since the
1251 * hold on the parent dbuf prevents the handle from being
1252 * destroyed, the hold on the handle is OK. We can't yet assert
1253 * that the handle has zero references, but that will be
1254 * asserted anyway when the handle gets destroyed.
1255 */
1256 dbuf_rele(db, dnh);
1257 }
1258 }
1259
1260 void
1261 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1262 {
1263 objset_t *os = dn->dn_objset;
1264 uint64_t txg = tx->tx_txg;
1265
1266 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1267 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1268 return;
1269 }
1270
1271 DNODE_VERIFY(dn);
1272
1273 #ifdef ZFS_DEBUG
1274 mutex_enter(&dn->dn_mtx);
1275 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1276 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1277 mutex_exit(&dn->dn_mtx);
1278 #endif
1279
1280 /*
1281 * Determine old uid/gid when necessary
1282 */
1283 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1284
1285 mutex_enter(&os->os_lock);
1286
1287 /*
1288 * If we are already marked dirty, we're done.
1289 */
1290 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1291 mutex_exit(&os->os_lock);
1292 return;
1293 }
1294
1295 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1296 !avl_is_empty(&dn->dn_dbufs));
1297 ASSERT(dn->dn_datablksz != 0);
1298 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1299 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1300 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1301
1302 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1303 dn->dn_object, txg);
1304
1305 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1306 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1307 } else {
1308 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1309 }
1310
1311 mutex_exit(&os->os_lock);
1312
1313 /*
1314 * The dnode maintains a hold on its containing dbuf as
1315 * long as there are holds on it. Each instantiated child
1316 * dbuf maintains a hold on the dnode. When the last child
1317 * drops its hold, the dnode will drop its hold on the
1318 * containing dbuf. We add a "dirty hold" here so that the
1319 * dnode will hang around after we finish processing its
1320 * children.
1321 */
1322 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1323
1324 (void) dbuf_dirty(dn->dn_dbuf, tx);
1325
1326 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1327 }
1328
1329 void
1330 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1331 {
1332 int txgoff = tx->tx_txg & TXG_MASK;
1333
1334 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1335
1336 /* we should be the only holder... hopefully */
1337 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1338
1339 mutex_enter(&dn->dn_mtx);
1340 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1341 mutex_exit(&dn->dn_mtx);
1342 return;
1343 }
1344 dn->dn_free_txg = tx->tx_txg;
1345 mutex_exit(&dn->dn_mtx);
1346
1347 /*
1348 * If the dnode is already dirty, it needs to be moved from
1349 * the dirty list to the free list.
1350 */
1351 mutex_enter(&dn->dn_objset->os_lock);
1352 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1353 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1354 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1355 mutex_exit(&dn->dn_objset->os_lock);
1356 } else {
1357 mutex_exit(&dn->dn_objset->os_lock);
1358 dnode_setdirty(dn, tx);
1359 }
1360 }
1361
1362 /*
1363 * Try to change the block size for the indicated dnode. This can only
1364 * succeed if there are no blocks allocated or dirty beyond first block
1365 */
1366 int
1367 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1368 {
1369 dmu_buf_impl_t *db;
1370 int err;
1371
1372 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1373 if (size == 0)
1374 size = SPA_MINBLOCKSIZE;
1375 else
1376 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1377
1378 if (ibs == dn->dn_indblkshift)
1379 ibs = 0;
1380
1381 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1382 return (0);
1383
1384 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1385
1386 /* Check for any allocated blocks beyond the first */
1387 if (dn->dn_maxblkid != 0)
1388 goto fail;
1389
1390 mutex_enter(&dn->dn_dbufs_mtx);
1391 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1392 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1393 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1394 db->db_blkid != DMU_SPILL_BLKID) {
1395 mutex_exit(&dn->dn_dbufs_mtx);
1396 goto fail;
1397 }
1398 }
1399 mutex_exit(&dn->dn_dbufs_mtx);
1400
1401 if (ibs && dn->dn_nlevels != 1)
1402 goto fail;
1403
1404 /* resize the old block */
1405 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1406 if (err == 0)
1407 dbuf_new_size(db, size, tx);
1408 else if (err != ENOENT)
1409 goto fail;
1410
1411 dnode_setdblksz(dn, size);
1412 dnode_setdirty(dn, tx);
1413 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1414 if (ibs) {
1415 dn->dn_indblkshift = ibs;
1416 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1417 }
1418 /* rele after we have fixed the blocksize in the dnode */
1419 if (db)
1420 dbuf_rele(db, FTAG);
1421
1422 rw_exit(&dn->dn_struct_rwlock);
1423 return (0);
1424
1425 fail:
1426 rw_exit(&dn->dn_struct_rwlock);
1427 return (SET_ERROR(ENOTSUP));
1428 }
1429
1430 /* read-holding callers must not rely on the lock being continuously held */
1431 void
1432 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1433 {
1434 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1435 int epbs, new_nlevels;
1436 uint64_t sz;
1437
1438 ASSERT(blkid != DMU_BONUS_BLKID);
1439
1440 ASSERT(have_read ?
1441 RW_READ_HELD(&dn->dn_struct_rwlock) :
1442 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1443
1444 /*
1445 * if we have a read-lock, check to see if we need to do any work
1446 * before upgrading to a write-lock.
1447 */
1448 if (have_read) {
1449 if (blkid <= dn->dn_maxblkid)
1450 return;
1451
1452 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1453 rw_exit(&dn->dn_struct_rwlock);
1454 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1455 }
1456 }
1457
1458 if (blkid <= dn->dn_maxblkid)
1459 goto out;
1460
1461 dn->dn_maxblkid = blkid;
1462
1463 /*
1464 * Compute the number of levels necessary to support the new maxblkid.
1465 */
1466 new_nlevels = 1;
1467 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1468 for (sz = dn->dn_nblkptr;
1469 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1470 new_nlevels++;
1471
1472 if (new_nlevels > dn->dn_nlevels) {
1473 int old_nlevels = dn->dn_nlevels;
1474 dmu_buf_impl_t *db;
1475 list_t *list;
1476 dbuf_dirty_record_t *new, *dr, *dr_next;
1477
1478 dn->dn_nlevels = new_nlevels;
1479
1480 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1481 dn->dn_next_nlevels[txgoff] = new_nlevels;
1482
1483 /* dirty the left indirects */
1484 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1485 ASSERT(db != NULL);
1486 new = dbuf_dirty(db, tx);
1487 dbuf_rele(db, FTAG);
1488
1489 /* transfer the dirty records to the new indirect */
1490 mutex_enter(&dn->dn_mtx);
1491 mutex_enter(&new->dt.di.dr_mtx);
1492 list = &dn->dn_dirty_records[txgoff];
1493 for (dr = list_head(list); dr; dr = dr_next) {
1494 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1495 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1496 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1497 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1498 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1499 list_remove(&dn->dn_dirty_records[txgoff], dr);
1500 list_insert_tail(&new->dt.di.dr_children, dr);
1501 dr->dr_parent = new;
1502 }
1503 }
1504 mutex_exit(&new->dt.di.dr_mtx);
1505 mutex_exit(&dn->dn_mtx);
1506 }
1507
1508 out:
1509 if (have_read)
1510 rw_downgrade(&dn->dn_struct_rwlock);
1511 }
1512
1513 void
1514 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1515 {
1516 dmu_buf_impl_t *db;
1517 uint64_t blkoff, blkid, nblks;
1518 int blksz, blkshift, head, tail;
1519 int trunc = FALSE;
1520 int epbs;
1521
1522 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1523 blksz = dn->dn_datablksz;
1524 blkshift = dn->dn_datablkshift;
1525 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1526
1527 if (len == DMU_OBJECT_END) {
1528 len = UINT64_MAX - off;
1529 trunc = TRUE;
1530 }
1531
1532 /*
1533 * First, block align the region to free:
1534 */
1535 if (ISP2(blksz)) {
1536 head = P2NPHASE(off, blksz);
1537 blkoff = P2PHASE(off, blksz);
1538 if ((off >> blkshift) > dn->dn_maxblkid)
1539 goto out;
1540 } else {
1541 ASSERT(dn->dn_maxblkid == 0);
1542 if (off == 0 && len >= blksz) {
1543 /*
1544 * Freeing the whole block; fast-track this request.
1545 * Note that we won't dirty any indirect blocks,
1546 * which is fine because we will be freeing the entire
1547 * file and thus all indirect blocks will be freed
1548 * by free_children().
1549 */
1550 blkid = 0;
1551 nblks = 1;
1552 goto done;
1553 } else if (off >= blksz) {
1554 /* Freeing past end-of-data */
1555 goto out;
1556 } else {
1557 /* Freeing part of the block. */
1558 head = blksz - off;
1559 ASSERT3U(head, >, 0);
1560 }
1561 blkoff = off;
1562 }
1563 /* zero out any partial block data at the start of the range */
1564 if (head) {
1565 ASSERT3U(blkoff + head, ==, blksz);
1566 if (len < head)
1567 head = len;
1568 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1569 FTAG, &db) == 0) {
1570 caddr_t data;
1571
1572 /* don't dirty if it isn't on disk and isn't dirty */
1573 if (db->db_last_dirty ||
1574 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1575 rw_exit(&dn->dn_struct_rwlock);
1576 dmu_buf_will_dirty(&db->db, tx);
1577 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1578 data = db->db.db_data;
1579 bzero(data + blkoff, head);
1580 }
1581 dbuf_rele(db, FTAG);
1582 }
1583 off += head;
1584 len -= head;
1585 }
1586
1587 /* If the range was less than one block, we're done */
1588 if (len == 0)
1589 goto out;
1590
1591 /* If the remaining range is past end of file, we're done */
1592 if ((off >> blkshift) > dn->dn_maxblkid)
1593 goto out;
1594
1595 ASSERT(ISP2(blksz));
1596 if (trunc)
1597 tail = 0;
1598 else
1599 tail = P2PHASE(len, blksz);
1600
1601 ASSERT0(P2PHASE(off, blksz));
1602 /* zero out any partial block data at the end of the range */
1603 if (tail) {
1604 if (len < tail)
1605 tail = len;
1606 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1607 TRUE, FTAG, &db) == 0) {
1608 /* don't dirty if not on disk and not dirty */
1609 if (db->db_last_dirty ||
1610 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1611 rw_exit(&dn->dn_struct_rwlock);
1612 dmu_buf_will_dirty(&db->db, tx);
1613 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1614 bzero(db->db.db_data, tail);
1615 }
1616 dbuf_rele(db, FTAG);
1617 }
1618 len -= tail;
1619 }
1620
1621 /* If the range did not include a full block, we are done */
1622 if (len == 0)
1623 goto out;
1624
1625 ASSERT(IS_P2ALIGNED(off, blksz));
1626 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1627 blkid = off >> blkshift;
1628 nblks = len >> blkshift;
1629 if (trunc)
1630 nblks += 1;
1631
1632 /*
1633 * Dirty the first and last indirect blocks, as they (and/or their
1634 * parents) will need to be written out if they were only
1635 * partially freed. Interior indirect blocks will be themselves freed,
1636 * by free_children(), so they need not be dirtied. Note that these
1637 * interior blocks have already been prefetched by dmu_tx_hold_free().
1638 */
1639 if (dn->dn_nlevels > 1) {
1640 uint64_t first, last;
1641
1642 first = blkid >> epbs;
1643 if (db = dbuf_hold_level(dn, 1, first, FTAG)) {
1644 dmu_buf_will_dirty(&db->db, tx);
1645 dbuf_rele(db, FTAG);
1646 }
1647 if (trunc)
1648 last = dn->dn_maxblkid >> epbs;
1649 else
1650 last = (blkid + nblks - 1) >> epbs;
1651 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) {
1652 dmu_buf_will_dirty(&db->db, tx);
1653 dbuf_rele(db, FTAG);
1654 }
1655 }
1656
1657 done:
1658 /*
1659 * Add this range to the dnode range list.
1660 * We will finish up this free operation in the syncing phase.
1661 */
1662 mutex_enter(&dn->dn_mtx);
1663 int txgoff = tx->tx_txg & TXG_MASK;
1664 if (dn->dn_free_ranges[txgoff] == NULL) {
1665 dn->dn_free_ranges[txgoff] =
1666 range_tree_create(NULL, NULL, &dn->dn_mtx);
1667 }
1668 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1669 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1670 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1671 blkid, nblks, tx->tx_txg);
1672 mutex_exit(&dn->dn_mtx);
1673
1674 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1675 dnode_setdirty(dn, tx);
1676 out:
1677
1678 rw_exit(&dn->dn_struct_rwlock);
1679 }
1680
1681 static boolean_t
1682 dnode_spill_freed(dnode_t *dn)
1683 {
1684 int i;
1685
1686 mutex_enter(&dn->dn_mtx);
1687 for (i = 0; i < TXG_SIZE; i++) {
1688 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1689 break;
1690 }
1691 mutex_exit(&dn->dn_mtx);
1692 return (i < TXG_SIZE);
1693 }
1694
1695 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1696 uint64_t
1697 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1698 {
1699 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1700 int i;
1701
1702 if (blkid == DMU_BONUS_BLKID)
1703 return (FALSE);
1704
1705 /*
1706 * If we're in the process of opening the pool, dp will not be
1707 * set yet, but there shouldn't be anything dirty.
1708 */
1709 if (dp == NULL)
1710 return (FALSE);
1711
1712 if (dn->dn_free_txg)
1713 return (TRUE);
1714
1715 if (blkid == DMU_SPILL_BLKID)
1716 return (dnode_spill_freed(dn));
1717
1718 mutex_enter(&dn->dn_mtx);
1719 for (i = 0; i < TXG_SIZE; i++) {
1720 if (dn->dn_free_ranges[i] != NULL &&
1721 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1722 break;
1723 }
1724 mutex_exit(&dn->dn_mtx);
1725 return (i < TXG_SIZE);
1726 }
1727
1728 /* call from syncing context when we actually write/free space for this dnode */
1729 void
1730 dnode_diduse_space(dnode_t *dn, int64_t delta)
1731 {
1732 uint64_t space;
1733 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1734 dn, dn->dn_phys,
1735 (u_longlong_t)dn->dn_phys->dn_used,
1736 (longlong_t)delta);
1737
1738 mutex_enter(&dn->dn_mtx);
1739 space = DN_USED_BYTES(dn->dn_phys);
1740 if (delta > 0) {
1741 ASSERT3U(space + delta, >=, space); /* no overflow */
1742 } else {
1743 ASSERT3U(space, >=, -delta); /* no underflow */
1744 }
1745 space += delta;
1746 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1747 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1748 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1749 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1750 } else {
1751 dn->dn_phys->dn_used = space;
1752 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1753 }
1754 mutex_exit(&dn->dn_mtx);
1755 }
1756
1757 /*
1758 * Call when we think we're going to write/free space in open context to track
1759 * the amount of memory in use by the currently open txg.
1760 */
1761 void
1762 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1763 {
1764 objset_t *os = dn->dn_objset;
1765 dsl_dataset_t *ds = os->os_dsl_dataset;
1766 int64_t aspace = spa_get_asize(os->os_spa, space);
1767
1768 if (ds != NULL) {
1769 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1770 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1771 }
1772
1773 dmu_tx_willuse_space(tx, aspace);
1774 }
1775
1776 /*
1777 * Scans a block at the indicated "level" looking for a hole or data,
1778 * depending on 'flags'.
1779 *
1780 * If level > 0, then we are scanning an indirect block looking at its
1781 * pointers. If level == 0, then we are looking at a block of dnodes.
1782 *
1783 * If we don't find what we are looking for in the block, we return ESRCH.
1784 * Otherwise, return with *offset pointing to the beginning (if searching
1785 * forwards) or end (if searching backwards) of the range covered by the
1786 * block pointer we matched on (or dnode).
1787 *
1788 * The basic search algorithm used below by dnode_next_offset() is to
1789 * use this function to search up the block tree (widen the search) until
1790 * we find something (i.e., we don't return ESRCH) and then search back
1791 * down the tree (narrow the search) until we reach our original search
1792 * level.
1793 */
1794 static int
1795 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1796 int lvl, uint64_t blkfill, uint64_t txg)
1797 {
1798 dmu_buf_impl_t *db = NULL;
1799 void *data = NULL;
1800 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1801 uint64_t epb = 1ULL << epbs;
1802 uint64_t minfill, maxfill;
1803 boolean_t hole;
1804 int i, inc, error, span;
1805
1806 dprintf("probing object %llu offset %llx level %d of %u\n",
1807 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1808
1809 hole = ((flags & DNODE_FIND_HOLE) != 0);
1810 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1811 ASSERT(txg == 0 || !hole);
1812
1813 if (lvl == dn->dn_phys->dn_nlevels) {
1814 error = 0;
1815 epb = dn->dn_phys->dn_nblkptr;
1816 data = dn->dn_phys->dn_blkptr;
1817 } else {
1818 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1819 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1820 if (error) {
1821 if (error != ENOENT)
1822 return (error);
1823 if (hole)
1824 return (0);
1825 /*
1826 * This can only happen when we are searching up
1827 * the block tree for data. We don't really need to
1828 * adjust the offset, as we will just end up looking
1829 * at the pointer to this block in its parent, and its
1830 * going to be unallocated, so we will skip over it.
1831 */
1832 return (SET_ERROR(ESRCH));
1833 }
1834 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1835 if (error) {
1836 dbuf_rele(db, FTAG);
1837 return (error);
1838 }
1839 data = db->db.db_data;
1840 }
1841
1842
1843 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1844 db->db_blkptr->blk_birth <= txg ||
1845 BP_IS_HOLE(db->db_blkptr))) {
1846 /*
1847 * This can only happen when we are searching up the tree
1848 * and these conditions mean that we need to keep climbing.
1849 */
1850 error = SET_ERROR(ESRCH);
1851 } else if (lvl == 0) {
1852 dnode_phys_t *dnp = data;
1853 span = DNODE_SHIFT;
1854 ASSERT(dn->dn_type == DMU_OT_DNODE);
1855
1856 for (i = (*offset >> span) & (blkfill - 1);
1857 i >= 0 && i < blkfill; i += inc) {
1858 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1859 break;
1860 *offset += (1ULL << span) * inc;
1861 }
1862 if (i < 0 || i == blkfill)
1863 error = SET_ERROR(ESRCH);
1864 } else {
1865 blkptr_t *bp = data;
1866 uint64_t start = *offset;
1867 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1868 minfill = 0;
1869 maxfill = blkfill << ((lvl - 1) * epbs);
1870
1871 if (hole)
1872 maxfill--;
1873 else
1874 minfill++;
1875
1876 *offset = *offset >> span;
1877 for (i = BF64_GET(*offset, 0, epbs);
1878 i >= 0 && i < epb; i += inc) {
1879 if (BP_GET_FILL(&bp[i]) >= minfill &&
1880 BP_GET_FILL(&bp[i]) <= maxfill &&
1881 (hole || bp[i].blk_birth > txg))
1882 break;
1883 if (inc > 0 || *offset > 0)
1884 *offset += inc;
1885 }
1886 *offset = *offset << span;
1887 if (inc < 0) {
1888 /* traversing backwards; position offset at the end */
1889 ASSERT3U(*offset, <=, start);
1890 *offset = MIN(*offset + (1ULL << span) - 1, start);
1891 } else if (*offset < start) {
1892 *offset = start;
1893 }
1894 if (i < 0 || i >= epb)
1895 error = SET_ERROR(ESRCH);
1896 }
1897
1898 if (db)
1899 dbuf_rele(db, FTAG);
1900
1901 return (error);
1902 }
1903
1904 /*
1905 * Find the next hole, data, or sparse region at or after *offset.
1906 * The value 'blkfill' tells us how many items we expect to find
1907 * in an L0 data block; this value is 1 for normal objects,
1908 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1909 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1910 *
1911 * Examples:
1912 *
1913 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1914 * Finds the next/previous hole/data in a file.
1915 * Used in dmu_offset_next().
1916 *
1917 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1918 * Finds the next free/allocated dnode an objset's meta-dnode.
1919 * Only finds objects that have new contents since txg (ie.
1920 * bonus buffer changes and content removal are ignored).
1921 * Used in dmu_object_next().
1922 *
1923 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1924 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1925 * Used in dmu_object_alloc().
1926 */
1927 int
1928 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1929 int minlvl, uint64_t blkfill, uint64_t txg)
1930 {
1931 uint64_t initial_offset = *offset;
1932 int lvl, maxlvl;
1933 int error = 0;
1934
1935 if (!(flags & DNODE_FIND_HAVELOCK))
1936 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1937
1938 if (dn->dn_phys->dn_nlevels == 0) {
1939 error = SET_ERROR(ESRCH);
1940 goto out;
1941 }
1942
1943 if (dn->dn_datablkshift == 0) {
1944 if (*offset < dn->dn_datablksz) {
1945 if (flags & DNODE_FIND_HOLE)
1946 *offset = dn->dn_datablksz;
1947 } else {
1948 error = SET_ERROR(ESRCH);
1949 }
1950 goto out;
1951 }
1952
1953 maxlvl = dn->dn_phys->dn_nlevels;
1954
1955 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1956 error = dnode_next_offset_level(dn,
1957 flags, offset, lvl, blkfill, txg);
1958 if (error != ESRCH)
1959 break;
1960 }
1961
1962 while (error == 0 && --lvl >= minlvl) {
1963 error = dnode_next_offset_level(dn,
1964 flags, offset, lvl, blkfill, txg);
1965 }
1966
1967 /*
1968 * There's always a "virtual hole" at the end of the object, even
1969 * if all BP's which physically exist are non-holes.
1970 */
1971 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
1972 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
1973 error = 0;
1974 }
1975
1976 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1977 initial_offset < *offset : initial_offset > *offset))
1978 error = SET_ERROR(ESRCH);
1979 out:
1980 if (!(flags & DNODE_FIND_HAVELOCK))
1981 rw_exit(&dn->dn_struct_rwlock);
1982
1983 return (error);
1984 }