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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28 #include <sys/zfs_context.h>
29 #include <sys/dbuf.h>
30 #include <sys/dnode.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/spa.h>
38 #include <sys/zio.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/range_tree.h>
41
42 static kmem_cache_t *dnode_cache;
43 /*
44 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
45 * turned on when DEBUG is also defined.
46 */
47 #ifdef DEBUG
48 #define DNODE_STATS
49 #endif /* DEBUG */
50
51 #ifdef DNODE_STATS
52 #define DNODE_STAT_ADD(stat) ((stat)++)
53 #else
54 #define DNODE_STAT_ADD(stat) /* nothing */
55 #endif /* DNODE_STATS */
56
57 static dnode_phys_t dnode_phys_zero;
58
59 int zfs_default_bs = SPA_MINBLOCKSHIFT;
60 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
61
62 #ifdef illumos
63 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
64 #endif
65
66 static int
dbuf_compare(const void * x1,const void * x2)67 dbuf_compare(const void *x1, const void *x2)
68 {
69 const dmu_buf_impl_t *d1 = x1;
70 const dmu_buf_impl_t *d2 = x2;
71
72 if (d1->db_level < d2->db_level) {
73 return (-1);
74 }
75 if (d1->db_level > d2->db_level) {
76 return (1);
77 }
78
79 if (d1->db_blkid < d2->db_blkid) {
80 return (-1);
81 }
82 if (d1->db_blkid > d2->db_blkid) {
83 return (1);
84 }
85
86 if (d1->db_state == DB_SEARCH) {
87 ASSERT3S(d2->db_state, !=, DB_SEARCH);
88 return (-1);
89 } else if (d2->db_state == DB_SEARCH) {
90 ASSERT3S(d1->db_state, !=, DB_SEARCH);
91 return (1);
92 }
93
94 if ((uintptr_t)d1 < (uintptr_t)d2) {
95 return (-1);
96 }
97 if ((uintptr_t)d1 > (uintptr_t)d2) {
98 return (1);
99 }
100 return (0);
101 }
102
103 /* ARGSUSED */
104 static int
dnode_cons(void * arg,void * unused,int kmflag)105 dnode_cons(void *arg, void *unused, int kmflag)
106 {
107 dnode_t *dn = arg;
108 int i;
109
110 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
111 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
112 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
113 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
114
115 /*
116 * Every dbuf has a reference, and dropping a tracked reference is
117 * O(number of references), so don't track dn_holds.
118 */
119 refcount_create_untracked(&dn->dn_holds);
120 refcount_create(&dn->dn_tx_holds);
121 list_link_init(&dn->dn_link);
122
123 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
124 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
125 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
126 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
127 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
128 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
129 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
130
131 for (i = 0; i < TXG_SIZE; i++) {
132 list_link_init(&dn->dn_dirty_link[i]);
133 dn->dn_free_ranges[i] = NULL;
134 list_create(&dn->dn_dirty_records[i],
135 sizeof (dbuf_dirty_record_t),
136 offsetof(dbuf_dirty_record_t, dr_dirty_node));
137 }
138
139 dn->dn_allocated_txg = 0;
140 dn->dn_free_txg = 0;
141 dn->dn_assigned_txg = 0;
142 dn->dn_dirtyctx = 0;
143 dn->dn_dirtyctx_firstset = NULL;
144 dn->dn_bonus = NULL;
145 dn->dn_have_spill = B_FALSE;
146 dn->dn_zio = NULL;
147 dn->dn_oldused = 0;
148 dn->dn_oldflags = 0;
149 dn->dn_olduid = 0;
150 dn->dn_oldgid = 0;
151 dn->dn_newuid = 0;
152 dn->dn_newgid = 0;
153 dn->dn_id_flags = 0;
154
155 dn->dn_dbufs_count = 0;
156 dn->dn_unlisted_l0_blkid = 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 POINTER_INVALIDATE(&dn->dn_objset);
162 return (0);
163 }
164
165 /* ARGSUSED */
166 static void
dnode_dest(void * arg,void * unused)167 dnode_dest(void *arg, void *unused)
168 {
169 int i;
170 dnode_t *dn = arg;
171
172 rw_destroy(&dn->dn_struct_rwlock);
173 mutex_destroy(&dn->dn_mtx);
174 mutex_destroy(&dn->dn_dbufs_mtx);
175 cv_destroy(&dn->dn_notxholds);
176 refcount_destroy(&dn->dn_holds);
177 refcount_destroy(&dn->dn_tx_holds);
178 ASSERT(!list_link_active(&dn->dn_link));
179
180 for (i = 0; i < TXG_SIZE; i++) {
181 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
182 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
183 list_destroy(&dn->dn_dirty_records[i]);
184 ASSERT0(dn->dn_next_nblkptr[i]);
185 ASSERT0(dn->dn_next_nlevels[i]);
186 ASSERT0(dn->dn_next_indblkshift[i]);
187 ASSERT0(dn->dn_next_bonustype[i]);
188 ASSERT0(dn->dn_rm_spillblk[i]);
189 ASSERT0(dn->dn_next_bonuslen[i]);
190 ASSERT0(dn->dn_next_blksz[i]);
191 }
192
193 ASSERT0(dn->dn_allocated_txg);
194 ASSERT0(dn->dn_free_txg);
195 ASSERT0(dn->dn_assigned_txg);
196 ASSERT0(dn->dn_dirtyctx);
197 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
198 ASSERT3P(dn->dn_bonus, ==, NULL);
199 ASSERT(!dn->dn_have_spill);
200 ASSERT3P(dn->dn_zio, ==, NULL);
201 ASSERT0(dn->dn_oldused);
202 ASSERT0(dn->dn_oldflags);
203 ASSERT0(dn->dn_olduid);
204 ASSERT0(dn->dn_oldgid);
205 ASSERT0(dn->dn_newuid);
206 ASSERT0(dn->dn_newgid);
207 ASSERT0(dn->dn_id_flags);
208
209 ASSERT0(dn->dn_dbufs_count);
210 ASSERT0(dn->dn_unlisted_l0_blkid);
211 avl_destroy(&dn->dn_dbufs);
212 }
213
214 void
dnode_init(void)215 dnode_init(void)
216 {
217 ASSERT(dnode_cache == NULL);
218 dnode_cache = kmem_cache_create("dnode_t",
219 sizeof (dnode_t),
220 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
221 kmem_cache_set_move(dnode_cache, dnode_move);
222 }
223
224 void
dnode_fini(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
dnode_verify(dnode_t * dn)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
dnode_byteswap(dnode_phys_t * dnp)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
dnode_buf_byteswap(void * vbuf,size_t size)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
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)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
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)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
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)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
dnode_setdblksz(dnode_t * dn,int size)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 *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)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
dnode_destroy(dnode_t * dn)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 dn->dn_unlisted_l0_blkid = 0;
529
530 dmu_zfetch_fini(&dn->dn_zfetch);
531 kmem_cache_free(dnode_cache, dn);
532 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
533
534 if (complete_os_eviction)
535 dmu_objset_evict_done(os);
536 }
537
538 void
dnode_allocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,int ibs,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)539 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
540 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
541 {
542 int i;
543
544 ASSERT3U(blocksize, <=,
545 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
546 if (blocksize == 0)
547 blocksize = 1 << zfs_default_bs;
548 else
549 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
550
551 if (ibs == 0)
552 ibs = zfs_default_ibs;
553
554 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
555
556 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
557 dn->dn_object, tx->tx_txg, blocksize, ibs);
558
559 ASSERT(dn->dn_type == DMU_OT_NONE);
560 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
561 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
562 ASSERT(ot != DMU_OT_NONE);
563 ASSERT(DMU_OT_IS_VALID(ot));
564 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
565 (bonustype == DMU_OT_SA && bonuslen == 0) ||
566 (bonustype != DMU_OT_NONE && bonuslen != 0));
567 ASSERT(DMU_OT_IS_VALID(bonustype));
568 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
569 ASSERT(dn->dn_type == DMU_OT_NONE);
570 ASSERT0(dn->dn_maxblkid);
571 ASSERT0(dn->dn_allocated_txg);
572 ASSERT0(dn->dn_assigned_txg);
573 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
574 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
575 ASSERT(avl_is_empty(&dn->dn_dbufs));
576
577 for (i = 0; i < TXG_SIZE; i++) {
578 ASSERT0(dn->dn_next_nblkptr[i]);
579 ASSERT0(dn->dn_next_nlevels[i]);
580 ASSERT0(dn->dn_next_indblkshift[i]);
581 ASSERT0(dn->dn_next_bonuslen[i]);
582 ASSERT0(dn->dn_next_bonustype[i]);
583 ASSERT0(dn->dn_rm_spillblk[i]);
584 ASSERT0(dn->dn_next_blksz[i]);
585 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
586 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
587 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
588 }
589
590 dn->dn_type = ot;
591 dnode_setdblksz(dn, blocksize);
592 dn->dn_indblkshift = ibs;
593 dn->dn_nlevels = 1;
594 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
595 dn->dn_nblkptr = 1;
596 else
597 dn->dn_nblkptr = 1 +
598 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
599 dn->dn_bonustype = bonustype;
600 dn->dn_bonuslen = bonuslen;
601 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
602 dn->dn_compress = ZIO_COMPRESS_INHERIT;
603 dn->dn_dirtyctx = 0;
604
605 dn->dn_free_txg = 0;
606 if (dn->dn_dirtyctx_firstset) {
607 kmem_free(dn->dn_dirtyctx_firstset, 1);
608 dn->dn_dirtyctx_firstset = NULL;
609 }
610
611 dn->dn_allocated_txg = tx->tx_txg;
612 dn->dn_id_flags = 0;
613
614 dnode_setdirty(dn, tx);
615 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
616 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
617 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
618 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
619 }
620
621 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)622 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
623 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
624 {
625 int nblkptr;
626
627 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
628 ASSERT3U(blocksize, <=,
629 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
630 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
631 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
632 ASSERT(tx->tx_txg != 0);
633 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
634 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
635 (bonustype == DMU_OT_SA && bonuslen == 0));
636 ASSERT(DMU_OT_IS_VALID(bonustype));
637 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
638
639 /* clean up any unreferenced dbufs */
640 dnode_evict_dbufs(dn);
641
642 dn->dn_id_flags = 0;
643
644 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
645 dnode_setdirty(dn, tx);
646 if (dn->dn_datablksz != blocksize) {
647 /* change blocksize */
648 ASSERT(dn->dn_maxblkid == 0 &&
649 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
650 dnode_block_freed(dn, 0)));
651 dnode_setdblksz(dn, blocksize);
652 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
653 }
654 if (dn->dn_bonuslen != bonuslen)
655 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
656
657 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
658 nblkptr = 1;
659 else
660 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
661 if (dn->dn_bonustype != bonustype)
662 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
663 if (dn->dn_nblkptr != nblkptr)
664 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
665 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
666 dbuf_rm_spill(dn, tx);
667 dnode_rm_spill(dn, tx);
668 }
669 rw_exit(&dn->dn_struct_rwlock);
670
671 /* change type */
672 dn->dn_type = ot;
673
674 /* change bonus size and type */
675 mutex_enter(&dn->dn_mtx);
676 dn->dn_bonustype = bonustype;
677 dn->dn_bonuslen = bonuslen;
678 dn->dn_nblkptr = nblkptr;
679 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
680 dn->dn_compress = ZIO_COMPRESS_INHERIT;
681 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
682
683 /* fix up the bonus db_size */
684 if (dn->dn_bonus) {
685 dn->dn_bonus->db.db_size =
686 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
687 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
688 }
689
690 dn->dn_allocated_txg = tx->tx_txg;
691 mutex_exit(&dn->dn_mtx);
692 }
693
694 #ifdef DNODE_STATS
695 static struct {
696 uint64_t dms_dnode_invalid;
697 uint64_t dms_dnode_recheck1;
698 uint64_t dms_dnode_recheck2;
699 uint64_t dms_dnode_special;
700 uint64_t dms_dnode_handle;
701 uint64_t dms_dnode_rwlock;
702 uint64_t dms_dnode_active;
703 } dnode_move_stats;
704 #endif /* DNODE_STATS */
705
706 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)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_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
765 ndn->dn_bonus = odn->dn_bonus;
766 ndn->dn_have_spill = odn->dn_have_spill;
767 ndn->dn_zio = odn->dn_zio;
768 ndn->dn_oldused = odn->dn_oldused;
769 ndn->dn_oldflags = odn->dn_oldflags;
770 ndn->dn_olduid = odn->dn_olduid;
771 ndn->dn_oldgid = odn->dn_oldgid;
772 ndn->dn_newuid = odn->dn_newuid;
773 ndn->dn_newgid = odn->dn_newgid;
774 ndn->dn_id_flags = odn->dn_id_flags;
775 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
776 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
777 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
778
779 /*
780 * Update back pointers. Updating the handle fixes the back pointer of
781 * every descendant dbuf as well as the bonus dbuf.
782 */
783 ASSERT(ndn->dn_handle->dnh_dnode == odn);
784 ndn->dn_handle->dnh_dnode = ndn;
785 if (ndn->dn_zfetch.zf_dnode == odn) {
786 ndn->dn_zfetch.zf_dnode = ndn;
787 }
788
789 /*
790 * Invalidate the original dnode by clearing all of its back pointers.
791 */
792 odn->dn_dbuf = NULL;
793 odn->dn_handle = NULL;
794 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
795 offsetof(dmu_buf_impl_t, db_link));
796 odn->dn_dbufs_count = 0;
797 odn->dn_unlisted_l0_blkid = 0;
798 odn->dn_bonus = NULL;
799 odn->dn_zfetch.zf_dnode = NULL;
800
801 /*
802 * Set the low bit of the objset pointer to ensure that dnode_move()
803 * recognizes the dnode as invalid in any subsequent callback.
804 */
805 POINTER_INVALIDATE(&odn->dn_objset);
806
807 /*
808 * Satisfy the destructor.
809 */
810 for (i = 0; i < TXG_SIZE; i++) {
811 list_create(&odn->dn_dirty_records[i],
812 sizeof (dbuf_dirty_record_t),
813 offsetof(dbuf_dirty_record_t, dr_dirty_node));
814 odn->dn_free_ranges[i] = NULL;
815 odn->dn_next_nlevels[i] = 0;
816 odn->dn_next_indblkshift[i] = 0;
817 odn->dn_next_bonustype[i] = 0;
818 odn->dn_rm_spillblk[i] = 0;
819 odn->dn_next_bonuslen[i] = 0;
820 odn->dn_next_blksz[i] = 0;
821 }
822 odn->dn_allocated_txg = 0;
823 odn->dn_free_txg = 0;
824 odn->dn_assigned_txg = 0;
825 odn->dn_dirtyctx = 0;
826 odn->dn_dirtyctx_firstset = NULL;
827 odn->dn_have_spill = B_FALSE;
828 odn->dn_zio = NULL;
829 odn->dn_oldused = 0;
830 odn->dn_oldflags = 0;
831 odn->dn_olduid = 0;
832 odn->dn_oldgid = 0;
833 odn->dn_newuid = 0;
834 odn->dn_newgid = 0;
835 odn->dn_id_flags = 0;
836
837 /*
838 * Mark the dnode.
839 */
840 ndn->dn_moved = 1;
841 odn->dn_moved = (uint8_t)-1;
842 }
843
844 #ifdef illumos
845 #ifdef _KERNEL
846 /*ARGSUSED*/
847 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)848 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
849 {
850 dnode_t *odn = buf, *ndn = newbuf;
851 objset_t *os;
852 int64_t refcount;
853 uint32_t dbufs;
854
855 /*
856 * The dnode is on the objset's list of known dnodes if the objset
857 * pointer is valid. We set the low bit of the objset pointer when
858 * freeing the dnode to invalidate it, and the memory patterns written
859 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
860 * A newly created dnode sets the objset pointer last of all to indicate
861 * that the dnode is known and in a valid state to be moved by this
862 * function.
863 */
864 os = odn->dn_objset;
865 if (!POINTER_IS_VALID(os)) {
866 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
867 return (KMEM_CBRC_DONT_KNOW);
868 }
869
870 /*
871 * Ensure that the objset does not go away during the move.
872 */
873 rw_enter(&os_lock, RW_WRITER);
874 if (os != odn->dn_objset) {
875 rw_exit(&os_lock);
876 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
877 return (KMEM_CBRC_DONT_KNOW);
878 }
879
880 /*
881 * If the dnode is still valid, then so is the objset. We know that no
882 * valid objset can be freed while we hold os_lock, so we can safely
883 * ensure that the objset remains in use.
884 */
885 mutex_enter(&os->os_lock);
886
887 /*
888 * Recheck the objset pointer in case the dnode was removed just before
889 * acquiring the lock.
890 */
891 if (os != odn->dn_objset) {
892 mutex_exit(&os->os_lock);
893 rw_exit(&os_lock);
894 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
895 return (KMEM_CBRC_DONT_KNOW);
896 }
897
898 /*
899 * At this point we know that as long as we hold os->os_lock, the dnode
900 * cannot be freed and fields within the dnode can be safely accessed.
901 * The objset listing this dnode cannot go away as long as this dnode is
902 * on its list.
903 */
904 rw_exit(&os_lock);
905 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
906 mutex_exit(&os->os_lock);
907 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
908 return (KMEM_CBRC_NO);
909 }
910 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
911
912 /*
913 * Lock the dnode handle to prevent the dnode from obtaining any new
914 * holds. This also prevents the descendant dbufs and the bonus dbuf
915 * from accessing the dnode, so that we can discount their holds. The
916 * handle is safe to access because we know that while the dnode cannot
917 * go away, neither can its handle. Once we hold dnh_zrlock, we can
918 * safely move any dnode referenced only by dbufs.
919 */
920 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
921 mutex_exit(&os->os_lock);
922 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
923 return (KMEM_CBRC_LATER);
924 }
925
926 /*
927 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
928 * We need to guarantee that there is a hold for every dbuf in order to
929 * determine whether the dnode is actively referenced. Falsely matching
930 * a dbuf to an active hold would lead to an unsafe move. It's possible
931 * that a thread already having an active dnode hold is about to add a
932 * dbuf, and we can't compare hold and dbuf counts while the add is in
933 * progress.
934 */
935 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
936 zrl_exit(&odn->dn_handle->dnh_zrlock);
937 mutex_exit(&os->os_lock);
938 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
939 return (KMEM_CBRC_LATER);
940 }
941
942 /*
943 * A dbuf may be removed (evicted) without an active dnode hold. In that
944 * case, the dbuf count is decremented under the handle lock before the
945 * dbuf's hold is released. This order ensures that if we count the hold
946 * after the dbuf is removed but before its hold is released, we will
947 * treat the unmatched hold as active and exit safely. If we count the
948 * hold before the dbuf is removed, the hold is discounted, and the
949 * removal is blocked until the move completes.
950 */
951 refcount = refcount_count(&odn->dn_holds);
952 ASSERT(refcount >= 0);
953 dbufs = odn->dn_dbufs_count;
954
955 /* We can't have more dbufs than dnode holds. */
956 ASSERT3U(dbufs, <=, refcount);
957 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
958 uint32_t, dbufs);
959
960 if (refcount > dbufs) {
961 rw_exit(&odn->dn_struct_rwlock);
962 zrl_exit(&odn->dn_handle->dnh_zrlock);
963 mutex_exit(&os->os_lock);
964 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
965 return (KMEM_CBRC_LATER);
966 }
967
968 rw_exit(&odn->dn_struct_rwlock);
969
970 /*
971 * At this point we know that anyone with a hold on the dnode is not
972 * actively referencing it. The dnode is known and in a valid state to
973 * move. We're holding the locks needed to execute the critical section.
974 */
975 dnode_move_impl(odn, ndn);
976
977 list_link_replace(&odn->dn_link, &ndn->dn_link);
978 /* If the dnode was safe to move, the refcount cannot have changed. */
979 ASSERT(refcount == refcount_count(&ndn->dn_holds));
980 ASSERT(dbufs == ndn->dn_dbufs_count);
981 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
982 mutex_exit(&os->os_lock);
983
984 return (KMEM_CBRC_YES);
985 }
986 #endif /* _KERNEL */
987 #endif /* illumos */
988
989 void
dnode_special_close(dnode_handle_t * dnh)990 dnode_special_close(dnode_handle_t *dnh)
991 {
992 dnode_t *dn = dnh->dnh_dnode;
993
994 /*
995 * Wait for final references to the dnode to clear. This can
996 * only happen if the arc is asyncronously evicting state that
997 * has a hold on this dnode while we are trying to evict this
998 * dnode.
999 */
1000 while (refcount_count(&dn->dn_holds) > 0)
1001 delay(1);
1002 ASSERT(dn->dn_dbuf == NULL ||
1003 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1004 zrl_add(&dnh->dnh_zrlock);
1005 dnode_destroy(dn); /* implicit zrl_remove() */
1006 zrl_destroy(&dnh->dnh_zrlock);
1007 dnh->dnh_dnode = NULL;
1008 }
1009
1010 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1011 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1012 dnode_handle_t *dnh)
1013 {
1014 dnode_t *dn;
1015
1016 dn = dnode_create(os, dnp, NULL, object, dnh);
1017 zrl_init(&dnh->dnh_zrlock);
1018 DNODE_VERIFY(dn);
1019 }
1020
1021 static void
dnode_buf_pageout(void * dbu)1022 dnode_buf_pageout(void *dbu)
1023 {
1024 dnode_children_t *children_dnodes = dbu;
1025 int i;
1026
1027 for (i = 0; i < children_dnodes->dnc_count; i++) {
1028 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1029 dnode_t *dn;
1030
1031 /*
1032 * The dnode handle lock guards against the dnode moving to
1033 * another valid address, so there is no need here to guard
1034 * against changes to or from NULL.
1035 */
1036 if (dnh->dnh_dnode == NULL) {
1037 zrl_destroy(&dnh->dnh_zrlock);
1038 continue;
1039 }
1040
1041 zrl_add(&dnh->dnh_zrlock);
1042 dn = dnh->dnh_dnode;
1043 /*
1044 * If there are holds on this dnode, then there should
1045 * be holds on the dnode's containing dbuf as well; thus
1046 * it wouldn't be eligible for eviction and this function
1047 * would not have been called.
1048 */
1049 ASSERT(refcount_is_zero(&dn->dn_holds));
1050 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1051
1052 dnode_destroy(dn); /* implicit zrl_remove() */
1053 zrl_destroy(&dnh->dnh_zrlock);
1054 dnh->dnh_dnode = NULL;
1055 }
1056 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1057 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1058 }
1059
1060 /*
1061 * errors:
1062 * EINVAL - invalid object number.
1063 * EIO - i/o error.
1064 * succeeds even for free dnodes.
1065 */
1066 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1067 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1068 void *tag, dnode_t **dnp)
1069 {
1070 int epb, idx, err;
1071 int drop_struct_lock = FALSE;
1072 int type;
1073 uint64_t blk;
1074 dnode_t *mdn, *dn;
1075 dmu_buf_impl_t *db;
1076 dnode_children_t *children_dnodes;
1077 dnode_handle_t *dnh;
1078
1079 /*
1080 * If you are holding the spa config lock as writer, you shouldn't
1081 * be asking the DMU to do *anything* unless it's the root pool
1082 * which may require us to read from the root filesystem while
1083 * holding some (not all) of the locks as writer.
1084 */
1085 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1086 (spa_is_root(os->os_spa) &&
1087 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1088
1089 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1090 dn = (object == DMU_USERUSED_OBJECT) ?
1091 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1092 if (dn == NULL)
1093 return (SET_ERROR(ENOENT));
1094 type = dn->dn_type;
1095 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1096 return (SET_ERROR(ENOENT));
1097 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1098 return (SET_ERROR(EEXIST));
1099 DNODE_VERIFY(dn);
1100 (void) refcount_add(&dn->dn_holds, tag);
1101 *dnp = dn;
1102 return (0);
1103 }
1104
1105 if (object == 0 || object >= DN_MAX_OBJECT)
1106 return (SET_ERROR(EINVAL));
1107
1108 mdn = DMU_META_DNODE(os);
1109 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1110
1111 DNODE_VERIFY(mdn);
1112
1113 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1114 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1115 drop_struct_lock = TRUE;
1116 }
1117
1118 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1119
1120 db = dbuf_hold(mdn, blk, FTAG);
1121 if (drop_struct_lock)
1122 rw_exit(&mdn->dn_struct_rwlock);
1123 if (db == NULL)
1124 return (SET_ERROR(EIO));
1125 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1126 if (err) {
1127 dbuf_rele(db, FTAG);
1128 return (err);
1129 }
1130
1131 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1132 epb = db->db.db_size >> DNODE_SHIFT;
1133
1134 idx = object & (epb-1);
1135
1136 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1137 children_dnodes = dmu_buf_get_user(&db->db);
1138 if (children_dnodes == NULL) {
1139 int i;
1140 dnode_children_t *winner;
1141 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1142 epb * sizeof (dnode_handle_t), KM_SLEEP);
1143 children_dnodes->dnc_count = epb;
1144 dnh = &children_dnodes->dnc_children[0];
1145 for (i = 0; i < epb; i++) {
1146 zrl_init(&dnh[i].dnh_zrlock);
1147 }
1148 dmu_buf_init_user(&children_dnodes->dnc_dbu,
1149 dnode_buf_pageout, NULL);
1150 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1151 if (winner != NULL) {
1152
1153 for (i = 0; i < epb; i++) {
1154 zrl_destroy(&dnh[i].dnh_zrlock);
1155 }
1156
1157 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1158 epb * sizeof (dnode_handle_t));
1159 children_dnodes = winner;
1160 }
1161 }
1162 ASSERT(children_dnodes->dnc_count == epb);
1163
1164 dnh = &children_dnodes->dnc_children[idx];
1165 zrl_add(&dnh->dnh_zrlock);
1166 dn = dnh->dnh_dnode;
1167 if (dn == NULL) {
1168 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1169
1170 dn = dnode_create(os, phys, db, object, dnh);
1171 }
1172
1173 mutex_enter(&dn->dn_mtx);
1174 type = dn->dn_type;
1175 if (dn->dn_free_txg ||
1176 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1177 ((flag & DNODE_MUST_BE_FREE) &&
1178 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1179 mutex_exit(&dn->dn_mtx);
1180 zrl_remove(&dnh->dnh_zrlock);
1181 dbuf_rele(db, FTAG);
1182 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1183 }
1184 if (refcount_add(&dn->dn_holds, tag) == 1)
1185 dbuf_add_ref(db, dnh);
1186 mutex_exit(&dn->dn_mtx);
1187
1188 /* Now we can rely on the hold to prevent the dnode from moving. */
1189 zrl_remove(&dnh->dnh_zrlock);
1190
1191 DNODE_VERIFY(dn);
1192 ASSERT3P(dn->dn_dbuf, ==, db);
1193 ASSERT3U(dn->dn_object, ==, object);
1194 dbuf_rele(db, FTAG);
1195
1196 *dnp = dn;
1197 return (0);
1198 }
1199
1200 /*
1201 * Return held dnode if the object is allocated, NULL if not.
1202 */
1203 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1204 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1205 {
1206 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1207 }
1208
1209 /*
1210 * Can only add a reference if there is already at least one
1211 * reference on the dnode. Returns FALSE if unable to add a
1212 * new reference.
1213 */
1214 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1215 dnode_add_ref(dnode_t *dn, void *tag)
1216 {
1217 mutex_enter(&dn->dn_mtx);
1218 if (refcount_is_zero(&dn->dn_holds)) {
1219 mutex_exit(&dn->dn_mtx);
1220 return (FALSE);
1221 }
1222 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1223 mutex_exit(&dn->dn_mtx);
1224 return (TRUE);
1225 }
1226
1227 void
dnode_rele(dnode_t * dn,void * tag)1228 dnode_rele(dnode_t *dn, void *tag)
1229 {
1230 mutex_enter(&dn->dn_mtx);
1231 dnode_rele_and_unlock(dn, tag);
1232 }
1233
1234 void
dnode_rele_and_unlock(dnode_t * dn,void * tag)1235 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1236 {
1237 uint64_t refs;
1238 /* Get while the hold prevents the dnode from moving. */
1239 dmu_buf_impl_t *db = dn->dn_dbuf;
1240 dnode_handle_t *dnh = dn->dn_handle;
1241
1242 refs = refcount_remove(&dn->dn_holds, tag);
1243 mutex_exit(&dn->dn_mtx);
1244
1245 /*
1246 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1247 * indirectly by dbuf_rele() while relying on the dnode handle to
1248 * prevent the dnode from moving, since releasing the last hold could
1249 * result in the dnode's parent dbuf evicting its dnode handles. For
1250 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1251 * other direct or indirect hold on the dnode must first drop the dnode
1252 * handle.
1253 */
1254 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1255
1256 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1257 if (refs == 0 && db != NULL) {
1258 /*
1259 * Another thread could add a hold to the dnode handle in
1260 * dnode_hold_impl() while holding the parent dbuf. Since the
1261 * hold on the parent dbuf prevents the handle from being
1262 * destroyed, the hold on the handle is OK. We can't yet assert
1263 * that the handle has zero references, but that will be
1264 * asserted anyway when the handle gets destroyed.
1265 */
1266 dbuf_rele(db, dnh);
1267 }
1268 }
1269
1270 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1271 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1272 {
1273 objset_t *os = dn->dn_objset;
1274 uint64_t txg = tx->tx_txg;
1275
1276 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1277 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1278 return;
1279 }
1280
1281 DNODE_VERIFY(dn);
1282
1283 #ifdef ZFS_DEBUG
1284 mutex_enter(&dn->dn_mtx);
1285 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1286 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1287 mutex_exit(&dn->dn_mtx);
1288 #endif
1289
1290 /*
1291 * Determine old uid/gid when necessary
1292 */
1293 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1294
1295 mutex_enter(&os->os_lock);
1296
1297 /*
1298 * If we are already marked dirty, we're done.
1299 */
1300 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1301 mutex_exit(&os->os_lock);
1302 return;
1303 }
1304
1305 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1306 !avl_is_empty(&dn->dn_dbufs));
1307 ASSERT(dn->dn_datablksz != 0);
1308 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1309 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1310 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1311
1312 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1313 dn->dn_object, txg);
1314
1315 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1316 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1317 } else {
1318 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1319 }
1320
1321 mutex_exit(&os->os_lock);
1322
1323 /*
1324 * The dnode maintains a hold on its containing dbuf as
1325 * long as there are holds on it. Each instantiated child
1326 * dbuf maintains a hold on the dnode. When the last child
1327 * drops its hold, the dnode will drop its hold on the
1328 * containing dbuf. We add a "dirty hold" here so that the
1329 * dnode will hang around after we finish processing its
1330 * children.
1331 */
1332 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1333
1334 (void) dbuf_dirty(dn->dn_dbuf, tx);
1335
1336 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1337 }
1338
1339 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1340 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1341 {
1342 int txgoff = tx->tx_txg & TXG_MASK;
1343
1344 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1345
1346 /* we should be the only holder... hopefully */
1347 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1348
1349 mutex_enter(&dn->dn_mtx);
1350 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1351 mutex_exit(&dn->dn_mtx);
1352 return;
1353 }
1354 dn->dn_free_txg = tx->tx_txg;
1355 mutex_exit(&dn->dn_mtx);
1356
1357 /*
1358 * If the dnode is already dirty, it needs to be moved from
1359 * the dirty list to the free list.
1360 */
1361 mutex_enter(&dn->dn_objset->os_lock);
1362 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1363 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1364 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1365 mutex_exit(&dn->dn_objset->os_lock);
1366 } else {
1367 mutex_exit(&dn->dn_objset->os_lock);
1368 dnode_setdirty(dn, tx);
1369 }
1370 }
1371
1372 /*
1373 * Try to change the block size for the indicated dnode. This can only
1374 * succeed if there are no blocks allocated or dirty beyond first block
1375 */
1376 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1377 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1378 {
1379 dmu_buf_impl_t *db;
1380 int err;
1381
1382 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1383 if (size == 0)
1384 size = SPA_MINBLOCKSIZE;
1385 else
1386 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1387
1388 if (ibs == dn->dn_indblkshift)
1389 ibs = 0;
1390
1391 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1392 return (0);
1393
1394 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1395
1396 /* Check for any allocated blocks beyond the first */
1397 if (dn->dn_maxblkid != 0)
1398 goto fail;
1399
1400 mutex_enter(&dn->dn_dbufs_mtx);
1401 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1402 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1403 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1404 db->db_blkid != DMU_SPILL_BLKID) {
1405 mutex_exit(&dn->dn_dbufs_mtx);
1406 goto fail;
1407 }
1408 }
1409 mutex_exit(&dn->dn_dbufs_mtx);
1410
1411 if (ibs && dn->dn_nlevels != 1)
1412 goto fail;
1413
1414 /* resize the old block */
1415 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1416 if (err == 0)
1417 dbuf_new_size(db, size, tx);
1418 else if (err != ENOENT)
1419 goto fail;
1420
1421 dnode_setdblksz(dn, size);
1422 dnode_setdirty(dn, tx);
1423 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1424 if (ibs) {
1425 dn->dn_indblkshift = ibs;
1426 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1427 }
1428 /* rele after we have fixed the blocksize in the dnode */
1429 if (db)
1430 dbuf_rele(db, FTAG);
1431
1432 rw_exit(&dn->dn_struct_rwlock);
1433 return (0);
1434
1435 fail:
1436 rw_exit(&dn->dn_struct_rwlock);
1437 return (SET_ERROR(ENOTSUP));
1438 }
1439
1440 /* read-holding callers must not rely on the lock being continuously held */
1441 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1442 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1443 {
1444 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1445 int epbs, new_nlevels;
1446 uint64_t sz;
1447
1448 ASSERT(blkid != DMU_BONUS_BLKID);
1449
1450 ASSERT(have_read ?
1451 RW_READ_HELD(&dn->dn_struct_rwlock) :
1452 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1453
1454 /*
1455 * if we have a read-lock, check to see if we need to do any work
1456 * before upgrading to a write-lock.
1457 */
1458 if (have_read) {
1459 if (blkid <= dn->dn_maxblkid)
1460 return;
1461
1462 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1463 rw_exit(&dn->dn_struct_rwlock);
1464 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1465 }
1466 }
1467
1468 if (blkid <= dn->dn_maxblkid)
1469 goto out;
1470
1471 dn->dn_maxblkid = blkid;
1472
1473 /*
1474 * Compute the number of levels necessary to support the new maxblkid.
1475 */
1476 new_nlevels = 1;
1477 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1478 for (sz = dn->dn_nblkptr;
1479 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1480 new_nlevels++;
1481
1482 if (new_nlevels > dn->dn_nlevels) {
1483 int old_nlevels = dn->dn_nlevels;
1484 dmu_buf_impl_t *db;
1485 list_t *list;
1486 dbuf_dirty_record_t *new, *dr, *dr_next;
1487
1488 dn->dn_nlevels = new_nlevels;
1489
1490 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1491 dn->dn_next_nlevels[txgoff] = new_nlevels;
1492
1493 /* dirty the left indirects */
1494 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1495 ASSERT(db != NULL);
1496 new = dbuf_dirty(db, tx);
1497 dbuf_rele(db, FTAG);
1498
1499 /* transfer the dirty records to the new indirect */
1500 mutex_enter(&dn->dn_mtx);
1501 mutex_enter(&new->dt.di.dr_mtx);
1502 list = &dn->dn_dirty_records[txgoff];
1503 for (dr = list_head(list); dr; dr = dr_next) {
1504 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1505 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1506 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1507 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1508 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1509 list_remove(&dn->dn_dirty_records[txgoff], dr);
1510 list_insert_tail(&new->dt.di.dr_children, dr);
1511 dr->dr_parent = new;
1512 }
1513 }
1514 mutex_exit(&new->dt.di.dr_mtx);
1515 mutex_exit(&dn->dn_mtx);
1516 }
1517
1518 out:
1519 if (have_read)
1520 rw_downgrade(&dn->dn_struct_rwlock);
1521 }
1522
1523 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1524 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1525 {
1526 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1527 if (db != NULL) {
1528 dmu_buf_will_dirty(&db->db, tx);
1529 dbuf_rele(db, FTAG);
1530 }
1531 }
1532
1533 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1534 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1535 {
1536 dmu_buf_impl_t *db;
1537 uint64_t blkoff, blkid, nblks;
1538 int blksz, blkshift, head, tail;
1539 int trunc = FALSE;
1540 int epbs;
1541
1542 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1543 blksz = dn->dn_datablksz;
1544 blkshift = dn->dn_datablkshift;
1545 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1546
1547 if (len == DMU_OBJECT_END) {
1548 len = UINT64_MAX - off;
1549 trunc = TRUE;
1550 }
1551
1552 /*
1553 * First, block align the region to free:
1554 */
1555 if (ISP2(blksz)) {
1556 head = P2NPHASE(off, blksz);
1557 blkoff = P2PHASE(off, blksz);
1558 if ((off >> blkshift) > dn->dn_maxblkid)
1559 goto out;
1560 } else {
1561 ASSERT(dn->dn_maxblkid == 0);
1562 if (off == 0 && len >= blksz) {
1563 /*
1564 * Freeing the whole block; fast-track this request.
1565 * Note that we won't dirty any indirect blocks,
1566 * which is fine because we will be freeing the entire
1567 * file and thus all indirect blocks will be freed
1568 * by free_children().
1569 */
1570 blkid = 0;
1571 nblks = 1;
1572 goto done;
1573 } else if (off >= blksz) {
1574 /* Freeing past end-of-data */
1575 goto out;
1576 } else {
1577 /* Freeing part of the block. */
1578 head = blksz - off;
1579 ASSERT3U(head, >, 0);
1580 }
1581 blkoff = off;
1582 }
1583 /* zero out any partial block data at the start of the range */
1584 if (head) {
1585 ASSERT3U(blkoff + head, ==, blksz);
1586 if (len < head)
1587 head = len;
1588 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1589 TRUE, FALSE, FTAG, &db) == 0) {
1590 caddr_t data;
1591
1592 /* don't dirty if it isn't on disk and isn't dirty */
1593 if (db->db_last_dirty ||
1594 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1595 rw_exit(&dn->dn_struct_rwlock);
1596 dmu_buf_will_dirty(&db->db, tx);
1597 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1598 data = db->db.db_data;
1599 bzero(data + blkoff, head);
1600 }
1601 dbuf_rele(db, FTAG);
1602 }
1603 off += head;
1604 len -= head;
1605 }
1606
1607 /* If the range was less than one block, we're done */
1608 if (len == 0)
1609 goto out;
1610
1611 /* If the remaining range is past end of file, we're done */
1612 if ((off >> blkshift) > dn->dn_maxblkid)
1613 goto out;
1614
1615 ASSERT(ISP2(blksz));
1616 if (trunc)
1617 tail = 0;
1618 else
1619 tail = P2PHASE(len, blksz);
1620
1621 ASSERT0(P2PHASE(off, blksz));
1622 /* zero out any partial block data at the end of the range */
1623 if (tail) {
1624 if (len < tail)
1625 tail = len;
1626 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1627 TRUE, FALSE, FTAG, &db) == 0) {
1628 /* don't dirty if not on disk and not dirty */
1629 if (db->db_last_dirty ||
1630 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1631 rw_exit(&dn->dn_struct_rwlock);
1632 dmu_buf_will_dirty(&db->db, tx);
1633 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1634 bzero(db->db.db_data, tail);
1635 }
1636 dbuf_rele(db, FTAG);
1637 }
1638 len -= tail;
1639 }
1640
1641 /* If the range did not include a full block, we are done */
1642 if (len == 0)
1643 goto out;
1644
1645 ASSERT(IS_P2ALIGNED(off, blksz));
1646 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1647 blkid = off >> blkshift;
1648 nblks = len >> blkshift;
1649 if (trunc)
1650 nblks += 1;
1651
1652 /*
1653 * Dirty all the indirect blocks in this range. Note that only
1654 * the first and last indirect blocks can actually be written
1655 * (if they were partially freed) -- they must be dirtied, even if
1656 * they do not exist on disk yet. The interior blocks will
1657 * be freed by free_children(), so they will not actually be written.
1658 * Even though these interior blocks will not be written, we
1659 * dirty them for two reasons:
1660 *
1661 * - It ensures that the indirect blocks remain in memory until
1662 * syncing context. (They have already been prefetched by
1663 * dmu_tx_hold_free(), so we don't have to worry about reading
1664 * them serially here.)
1665 *
1666 * - The dirty space accounting will put pressure on the txg sync
1667 * mechanism to begin syncing, and to delay transactions if there
1668 * is a large amount of freeing. Even though these indirect
1669 * blocks will not be written, we could need to write the same
1670 * amount of space if we copy the freed BPs into deadlists.
1671 */
1672 if (dn->dn_nlevels > 1) {
1673 uint64_t first, last;
1674
1675 first = blkid >> epbs;
1676 dnode_dirty_l1(dn, first, tx);
1677 if (trunc)
1678 last = dn->dn_maxblkid >> epbs;
1679 else
1680 last = (blkid + nblks - 1) >> epbs;
1681 if (last != first)
1682 dnode_dirty_l1(dn, last, tx);
1683
1684 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1685 SPA_BLKPTRSHIFT;
1686 for (uint64_t i = first + 1; i < last; i++) {
1687 /*
1688 * Set i to the blockid of the next non-hole
1689 * level-1 indirect block at or after i. Note
1690 * that dnode_next_offset() operates in terms of
1691 * level-0-equivalent bytes.
1692 */
1693 uint64_t ibyte = i << shift;
1694 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1695 &ibyte, 2, 1, 0);
1696 i = ibyte >> shift;
1697 if (i >= last)
1698 break;
1699
1700 /*
1701 * Normally we should not see an error, either
1702 * from dnode_next_offset() or dbuf_hold_level()
1703 * (except for ESRCH from dnode_next_offset).
1704 * If there is an i/o error, then when we read
1705 * this block in syncing context, it will use
1706 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1707 * to the "failmode" property. dnode_next_offset()
1708 * doesn't have a flag to indicate MUSTSUCCEED.
1709 */
1710 if (err != 0)
1711 break;
1712
1713 dnode_dirty_l1(dn, i, tx);
1714 }
1715 }
1716
1717 done:
1718 /*
1719 * Add this range to the dnode range list.
1720 * We will finish up this free operation in the syncing phase.
1721 */
1722 mutex_enter(&dn->dn_mtx);
1723 int txgoff = tx->tx_txg & TXG_MASK;
1724 if (dn->dn_free_ranges[txgoff] == NULL) {
1725 dn->dn_free_ranges[txgoff] =
1726 range_tree_create(NULL, NULL, &dn->dn_mtx);
1727 }
1728 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1729 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1730 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1731 blkid, nblks, tx->tx_txg);
1732 mutex_exit(&dn->dn_mtx);
1733
1734 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1735 dnode_setdirty(dn, tx);
1736 out:
1737
1738 rw_exit(&dn->dn_struct_rwlock);
1739 }
1740
1741 static boolean_t
dnode_spill_freed(dnode_t * dn)1742 dnode_spill_freed(dnode_t *dn)
1743 {
1744 int i;
1745
1746 mutex_enter(&dn->dn_mtx);
1747 for (i = 0; i < TXG_SIZE; i++) {
1748 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1749 break;
1750 }
1751 mutex_exit(&dn->dn_mtx);
1752 return (i < TXG_SIZE);
1753 }
1754
1755 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1756 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1757 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1758 {
1759 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1760 int i;
1761
1762 if (blkid == DMU_BONUS_BLKID)
1763 return (FALSE);
1764
1765 /*
1766 * If we're in the process of opening the pool, dp will not be
1767 * set yet, but there shouldn't be anything dirty.
1768 */
1769 if (dp == NULL)
1770 return (FALSE);
1771
1772 if (dn->dn_free_txg)
1773 return (TRUE);
1774
1775 if (blkid == DMU_SPILL_BLKID)
1776 return (dnode_spill_freed(dn));
1777
1778 mutex_enter(&dn->dn_mtx);
1779 for (i = 0; i < TXG_SIZE; i++) {
1780 if (dn->dn_free_ranges[i] != NULL &&
1781 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1782 break;
1783 }
1784 mutex_exit(&dn->dn_mtx);
1785 return (i < TXG_SIZE);
1786 }
1787
1788 /* call from syncing context when we actually write/free space for this dnode */
1789 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1790 dnode_diduse_space(dnode_t *dn, int64_t delta)
1791 {
1792 uint64_t space;
1793 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1794 dn, dn->dn_phys,
1795 (u_longlong_t)dn->dn_phys->dn_used,
1796 (longlong_t)delta);
1797
1798 mutex_enter(&dn->dn_mtx);
1799 space = DN_USED_BYTES(dn->dn_phys);
1800 if (delta > 0) {
1801 ASSERT3U(space + delta, >=, space); /* no overflow */
1802 } else {
1803 ASSERT3U(space, >=, -delta); /* no underflow */
1804 }
1805 space += delta;
1806 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1807 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1808 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1809 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1810 } else {
1811 dn->dn_phys->dn_used = space;
1812 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1813 }
1814 mutex_exit(&dn->dn_mtx);
1815 }
1816
1817 /*
1818 * Call when we think we're going to write/free space in open context to track
1819 * the amount of memory in use by the currently open txg.
1820 */
1821 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1822 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1823 {
1824 objset_t *os = dn->dn_objset;
1825 dsl_dataset_t *ds = os->os_dsl_dataset;
1826 int64_t aspace = spa_get_asize(os->os_spa, space);
1827
1828 if (ds != NULL) {
1829 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1830 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1831 }
1832
1833 dmu_tx_willuse_space(tx, aspace);
1834 }
1835
1836 /*
1837 * Scans a block at the indicated "level" looking for a hole or data,
1838 * depending on 'flags'.
1839 *
1840 * If level > 0, then we are scanning an indirect block looking at its
1841 * pointers. If level == 0, then we are looking at a block of dnodes.
1842 *
1843 * If we don't find what we are looking for in the block, we return ESRCH.
1844 * Otherwise, return with *offset pointing to the beginning (if searching
1845 * forwards) or end (if searching backwards) of the range covered by the
1846 * block pointer we matched on (or dnode).
1847 *
1848 * The basic search algorithm used below by dnode_next_offset() is to
1849 * use this function to search up the block tree (widen the search) until
1850 * we find something (i.e., we don't return ESRCH) and then search back
1851 * down the tree (narrow the search) until we reach our original search
1852 * level.
1853 */
1854 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1855 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1856 int lvl, uint64_t blkfill, uint64_t txg)
1857 {
1858 dmu_buf_impl_t *db = NULL;
1859 void *data = NULL;
1860 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1861 uint64_t epb = 1ULL << epbs;
1862 uint64_t minfill, maxfill;
1863 boolean_t hole;
1864 int i, inc, error, span;
1865
1866 dprintf("probing object %llu offset %llx level %d of %u\n",
1867 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1868
1869 hole = ((flags & DNODE_FIND_HOLE) != 0);
1870 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1871 ASSERT(txg == 0 || !hole);
1872
1873 if (lvl == dn->dn_phys->dn_nlevels) {
1874 error = 0;
1875 epb = dn->dn_phys->dn_nblkptr;
1876 data = dn->dn_phys->dn_blkptr;
1877 } else {
1878 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1879 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1880 if (error) {
1881 if (error != ENOENT)
1882 return (error);
1883 if (hole)
1884 return (0);
1885 /*
1886 * This can only happen when we are searching up
1887 * the block tree for data. We don't really need to
1888 * adjust the offset, as we will just end up looking
1889 * at the pointer to this block in its parent, and its
1890 * going to be unallocated, so we will skip over it.
1891 */
1892 return (SET_ERROR(ESRCH));
1893 }
1894 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1895 if (error) {
1896 dbuf_rele(db, FTAG);
1897 return (error);
1898 }
1899 data = db->db.db_data;
1900 }
1901
1902
1903 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1904 db->db_blkptr->blk_birth <= txg ||
1905 BP_IS_HOLE(db->db_blkptr))) {
1906 /*
1907 * This can only happen when we are searching up the tree
1908 * and these conditions mean that we need to keep climbing.
1909 */
1910 error = SET_ERROR(ESRCH);
1911 } else if (lvl == 0) {
1912 dnode_phys_t *dnp = data;
1913 span = DNODE_SHIFT;
1914 ASSERT(dn->dn_type == DMU_OT_DNODE);
1915
1916 for (i = (*offset >> span) & (blkfill - 1);
1917 i >= 0 && i < blkfill; i += inc) {
1918 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1919 break;
1920 *offset += (1ULL << span) * inc;
1921 }
1922 if (i < 0 || i == blkfill)
1923 error = SET_ERROR(ESRCH);
1924 } else {
1925 blkptr_t *bp = data;
1926 uint64_t start = *offset;
1927 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1928 minfill = 0;
1929 maxfill = blkfill << ((lvl - 1) * epbs);
1930
1931 if (hole)
1932 maxfill--;
1933 else
1934 minfill++;
1935
1936 *offset = *offset >> span;
1937 for (i = BF64_GET(*offset, 0, epbs);
1938 i >= 0 && i < epb; i += inc) {
1939 if (BP_GET_FILL(&bp[i]) >= minfill &&
1940 BP_GET_FILL(&bp[i]) <= maxfill &&
1941 (hole || bp[i].blk_birth > txg))
1942 break;
1943 if (inc > 0 || *offset > 0)
1944 *offset += inc;
1945 }
1946 *offset = *offset << span;
1947 if (inc < 0) {
1948 /* traversing backwards; position offset at the end */
1949 ASSERT3U(*offset, <=, start);
1950 *offset = MIN(*offset + (1ULL << span) - 1, start);
1951 } else if (*offset < start) {
1952 *offset = start;
1953 }
1954 if (i < 0 || i >= epb)
1955 error = SET_ERROR(ESRCH);
1956 }
1957
1958 if (db)
1959 dbuf_rele(db, FTAG);
1960
1961 return (error);
1962 }
1963
1964 /*
1965 * Find the next hole, data, or sparse region at or after *offset.
1966 * The value 'blkfill' tells us how many items we expect to find
1967 * in an L0 data block; this value is 1 for normal objects,
1968 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1969 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1970 *
1971 * Examples:
1972 *
1973 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1974 * Finds the next/previous hole/data in a file.
1975 * Used in dmu_offset_next().
1976 *
1977 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1978 * Finds the next free/allocated dnode an objset's meta-dnode.
1979 * Only finds objects that have new contents since txg (ie.
1980 * bonus buffer changes and content removal are ignored).
1981 * Used in dmu_object_next().
1982 *
1983 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1984 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1985 * Used in dmu_object_alloc().
1986 */
1987 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1988 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1989 int minlvl, uint64_t blkfill, uint64_t txg)
1990 {
1991 uint64_t initial_offset = *offset;
1992 int lvl, maxlvl;
1993 int error = 0;
1994
1995 if (!(flags & DNODE_FIND_HAVELOCK))
1996 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1997
1998 if (dn->dn_phys->dn_nlevels == 0) {
1999 error = SET_ERROR(ESRCH);
2000 goto out;
2001 }
2002
2003 if (dn->dn_datablkshift == 0) {
2004 if (*offset < dn->dn_datablksz) {
2005 if (flags & DNODE_FIND_HOLE)
2006 *offset = dn->dn_datablksz;
2007 } else {
2008 error = SET_ERROR(ESRCH);
2009 }
2010 goto out;
2011 }
2012
2013 maxlvl = dn->dn_phys->dn_nlevels;
2014
2015 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2016 error = dnode_next_offset_level(dn,
2017 flags, offset, lvl, blkfill, txg);
2018 if (error != ESRCH)
2019 break;
2020 }
2021
2022 while (error == 0 && --lvl >= minlvl) {
2023 error = dnode_next_offset_level(dn,
2024 flags, offset, lvl, blkfill, txg);
2025 }
2026
2027 /*
2028 * There's always a "virtual hole" at the end of the object, even
2029 * if all BP's which physically exist are non-holes.
2030 */
2031 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2032 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2033 error = 0;
2034 }
2035
2036 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2037 initial_offset < *offset : initial_offset > *offset))
2038 error = SET_ERROR(ESRCH);
2039 out:
2040 if (!(flags & DNODE_FIND_HAVELOCK))
2041 rw_exit(&dn->dn_struct_rwlock);
2042
2043 return (error);
2044 }
2045