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 */
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 #ifdef illumos
62 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
63 #endif
64
65 static int
dbuf_compare(const void * x1,const void * x2)66 dbuf_compare(const void *x1, const void *x2)
67 {
68 const dmu_buf_impl_t *d1 = x1;
69 const dmu_buf_impl_t *d2 = x2;
70
71 if (d1->db_level < d2->db_level) {
72 return (-1);
73 }
74 if (d1->db_level > d2->db_level) {
75 return (1);
76 }
77
78 if (d1->db_blkid < d2->db_blkid) {
79 return (-1);
80 }
81 if (d1->db_blkid > d2->db_blkid) {
82 return (1);
83 }
84
85 if (d1->db_state == DB_SEARCH) {
86 ASSERT3S(d2->db_state, !=, DB_SEARCH);
87 return (-1);
88 } else if (d2->db_state == DB_SEARCH) {
89 ASSERT3S(d1->db_state, !=, DB_SEARCH);
90 return (1);
91 }
92
93 if ((uintptr_t)d1 < (uintptr_t)d2) {
94 return (-1);
95 }
96 if ((uintptr_t)d1 > (uintptr_t)d2) {
97 return (1);
98 }
99 return (0);
100 }
101
102 /* ARGSUSED */
103 static int
dnode_cons(void * arg,void * unused,int kmflag)104 dnode_cons(void *arg, void *unused, int kmflag)
105 {
106 dnode_t *dn = arg;
107 int i;
108
109 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
110 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
111 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
112 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
113
114 /*
115 * Every dbuf has a reference, and dropping a tracked reference is
116 * O(number of references), so don't track dn_holds.
117 */
118 refcount_create_untracked(&dn->dn_holds);
119 refcount_create(&dn->dn_tx_holds);
120 list_link_init(&dn->dn_link);
121
122 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
123 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
124 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
125 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
126 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
127 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
128 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
129
130 for (i = 0; i < TXG_SIZE; i++) {
131 list_link_init(&dn->dn_dirty_link[i]);
132 dn->dn_free_ranges[i] = NULL;
133 list_create(&dn->dn_dirty_records[i],
134 sizeof (dbuf_dirty_record_t),
135 offsetof(dbuf_dirty_record_t, dr_dirty_node));
136 }
137
138 dn->dn_allocated_txg = 0;
139 dn->dn_free_txg = 0;
140 dn->dn_assigned_txg = 0;
141 dn->dn_dirtyctx = 0;
142 dn->dn_dirtyctx_firstset = NULL;
143 dn->dn_bonus = NULL;
144 dn->dn_have_spill = B_FALSE;
145 dn->dn_zio = NULL;
146 dn->dn_oldused = 0;
147 dn->dn_oldflags = 0;
148 dn->dn_olduid = 0;
149 dn->dn_oldgid = 0;
150 dn->dn_newuid = 0;
151 dn->dn_newgid = 0;
152 dn->dn_id_flags = 0;
153
154 dn->dn_dbufs_count = 0;
155 dn->dn_unlisted_l0_blkid = 0;
156 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
157 offsetof(dmu_buf_impl_t, db_link));
158
159 dn->dn_moved = 0;
160 POINTER_INVALIDATE(&dn->dn_objset);
161 return (0);
162 }
163
164 /* ARGSUSED */
165 static void
dnode_dest(void * arg,void * unused)166 dnode_dest(void *arg, void *unused)
167 {
168 int i;
169 dnode_t *dn = arg;
170
171 rw_destroy(&dn->dn_struct_rwlock);
172 mutex_destroy(&dn->dn_mtx);
173 mutex_destroy(&dn->dn_dbufs_mtx);
174 cv_destroy(&dn->dn_notxholds);
175 refcount_destroy(&dn->dn_holds);
176 refcount_destroy(&dn->dn_tx_holds);
177 ASSERT(!list_link_active(&dn->dn_link));
178
179 for (i = 0; i < TXG_SIZE; i++) {
180 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
181 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
182 list_destroy(&dn->dn_dirty_records[i]);
183 ASSERT0(dn->dn_next_nblkptr[i]);
184 ASSERT0(dn->dn_next_nlevels[i]);
185 ASSERT0(dn->dn_next_indblkshift[i]);
186 ASSERT0(dn->dn_next_bonustype[i]);
187 ASSERT0(dn->dn_rm_spillblk[i]);
188 ASSERT0(dn->dn_next_bonuslen[i]);
189 ASSERT0(dn->dn_next_blksz[i]);
190 }
191
192 ASSERT0(dn->dn_allocated_txg);
193 ASSERT0(dn->dn_free_txg);
194 ASSERT0(dn->dn_assigned_txg);
195 ASSERT0(dn->dn_dirtyctx);
196 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
197 ASSERT3P(dn->dn_bonus, ==, NULL);
198 ASSERT(!dn->dn_have_spill);
199 ASSERT3P(dn->dn_zio, ==, NULL);
200 ASSERT0(dn->dn_oldused);
201 ASSERT0(dn->dn_oldflags);
202 ASSERT0(dn->dn_olduid);
203 ASSERT0(dn->dn_oldgid);
204 ASSERT0(dn->dn_newuid);
205 ASSERT0(dn->dn_newgid);
206 ASSERT0(dn->dn_id_flags);
207
208 ASSERT0(dn->dn_dbufs_count);
209 ASSERT0(dn->dn_unlisted_l0_blkid);
210 avl_destroy(&dn->dn_dbufs);
211 }
212
213 void
dnode_init(void)214 dnode_init(void)
215 {
216 ASSERT(dnode_cache == NULL);
217 dnode_cache = kmem_cache_create("dnode_t",
218 sizeof (dnode_t),
219 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
220 kmem_cache_set_move(dnode_cache, dnode_move);
221 }
222
223 void
dnode_fini(void)224 dnode_fini(void)
225 {
226 kmem_cache_destroy(dnode_cache);
227 dnode_cache = NULL;
228 }
229
230
231 #ifdef ZFS_DEBUG
232 void
dnode_verify(dnode_t * dn)233 dnode_verify(dnode_t *dn)
234 {
235 int drop_struct_lock = FALSE;
236
237 ASSERT(dn->dn_phys);
238 ASSERT(dn->dn_objset);
239 ASSERT(dn->dn_handle->dnh_dnode == dn);
240
241 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
242
243 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
244 return;
245
246 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
247 rw_enter(&dn->dn_struct_rwlock, RW_READER);
248 drop_struct_lock = TRUE;
249 }
250 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
251 int i;
252 ASSERT3U(dn->dn_indblkshift, >=, 0);
253 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
254 if (dn->dn_datablkshift) {
255 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
256 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
257 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
258 }
259 ASSERT3U(dn->dn_nlevels, <=, 30);
260 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
261 ASSERT3U(dn->dn_nblkptr, >=, 1);
262 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
263 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
264 ASSERT3U(dn->dn_datablksz, ==,
265 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
266 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
267 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
268 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
269 for (i = 0; i < TXG_SIZE; i++) {
270 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
271 }
272 }
273 if (dn->dn_phys->dn_type != DMU_OT_NONE)
274 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
275 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
276 if (dn->dn_dbuf != NULL) {
277 ASSERT3P(dn->dn_phys, ==,
278 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
279 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
280 }
281 if (drop_struct_lock)
282 rw_exit(&dn->dn_struct_rwlock);
283 }
284 #endif
285
286 void
dnode_byteswap(dnode_phys_t * dnp)287 dnode_byteswap(dnode_phys_t *dnp)
288 {
289 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
290 int i;
291
292 if (dnp->dn_type == DMU_OT_NONE) {
293 bzero(dnp, sizeof (dnode_phys_t));
294 return;
295 }
296
297 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
298 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
299 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
300 dnp->dn_used = BSWAP_64(dnp->dn_used);
301
302 /*
303 * dn_nblkptr is only one byte, so it's OK to read it in either
304 * byte order. We can't read dn_bouslen.
305 */
306 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
307 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
308 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
309 buf64[i] = BSWAP_64(buf64[i]);
310
311 /*
312 * OK to check dn_bonuslen for zero, because it won't matter if
313 * we have the wrong byte order. This is necessary because the
314 * dnode dnode is smaller than a regular dnode.
315 */
316 if (dnp->dn_bonuslen != 0) {
317 /*
318 * Note that the bonus length calculated here may be
319 * longer than the actual bonus buffer. This is because
320 * we always put the bonus buffer after the last block
321 * pointer (instead of packing it against the end of the
322 * dnode buffer).
323 */
324 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
325 size_t len = DN_MAX_BONUSLEN - off;
326 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
327 dmu_object_byteswap_t byteswap =
328 DMU_OT_BYTESWAP(dnp->dn_bonustype);
329 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
330 }
331
332 /* Swap SPILL block if we have one */
333 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
334 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
335
336 }
337
338 void
dnode_buf_byteswap(void * vbuf,size_t size)339 dnode_buf_byteswap(void *vbuf, size_t size)
340 {
341 dnode_phys_t *buf = vbuf;
342 int i;
343
344 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
345 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
346
347 size >>= DNODE_SHIFT;
348 for (i = 0; i < size; i++) {
349 dnode_byteswap(buf);
350 buf++;
351 }
352 }
353
354 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)355 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
356 {
357 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
358
359 dnode_setdirty(dn, tx);
360 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
361 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
362 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
363 dn->dn_bonuslen = newsize;
364 if (newsize == 0)
365 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
366 else
367 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
368 rw_exit(&dn->dn_struct_rwlock);
369 }
370
371 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)372 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
373 {
374 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
375 dnode_setdirty(dn, tx);
376 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
377 dn->dn_bonustype = newtype;
378 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
379 rw_exit(&dn->dn_struct_rwlock);
380 }
381
382 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)383 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
384 {
385 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
386 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
387 dnode_setdirty(dn, tx);
388 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
389 dn->dn_have_spill = B_FALSE;
390 }
391
392 static void
dnode_setdblksz(dnode_t * dn,int size)393 dnode_setdblksz(dnode_t *dn, int size)
394 {
395 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
396 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
397 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
398 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
399 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
400 dn->dn_datablksz = size;
401 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
402 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
403 }
404
405 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)406 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
407 uint64_t object, dnode_handle_t *dnh)
408 {
409 dnode_t *dn;
410
411 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
412 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
413 dn->dn_moved = 0;
414
415 /*
416 * Defer setting dn_objset until the dnode is ready to be a candidate
417 * for the dnode_move() callback.
418 */
419 dn->dn_object = object;
420 dn->dn_dbuf = db;
421 dn->dn_handle = dnh;
422 dn->dn_phys = dnp;
423
424 if (dnp->dn_datablkszsec) {
425 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
426 } else {
427 dn->dn_datablksz = 0;
428 dn->dn_datablkszsec = 0;
429 dn->dn_datablkshift = 0;
430 }
431 dn->dn_indblkshift = dnp->dn_indblkshift;
432 dn->dn_nlevels = dnp->dn_nlevels;
433 dn->dn_type = dnp->dn_type;
434 dn->dn_nblkptr = dnp->dn_nblkptr;
435 dn->dn_checksum = dnp->dn_checksum;
436 dn->dn_compress = dnp->dn_compress;
437 dn->dn_bonustype = dnp->dn_bonustype;
438 dn->dn_bonuslen = dnp->dn_bonuslen;
439 dn->dn_maxblkid = dnp->dn_maxblkid;
440 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
441 dn->dn_id_flags = 0;
442
443 dmu_zfetch_init(&dn->dn_zfetch, dn);
444
445 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
446
447 mutex_enter(&os->os_lock);
448 if (dnh->dnh_dnode != NULL) {
449 /* Lost the allocation race. */
450 mutex_exit(&os->os_lock);
451 kmem_cache_free(dnode_cache, dn);
452 return (dnh->dnh_dnode);
453 }
454
455 /*
456 * Exclude special dnodes from os_dnodes so an empty os_dnodes
457 * signifies that the special dnodes have no references from
458 * their children (the entries in os_dnodes). This allows
459 * dnode_destroy() to easily determine if the last child has
460 * been removed and then complete eviction of the objset.
461 */
462 if (!DMU_OBJECT_IS_SPECIAL(object))
463 list_insert_head(&os->os_dnodes, dn);
464 membar_producer();
465
466 /*
467 * Everything else must be valid before assigning dn_objset
468 * makes the dnode eligible for dnode_move().
469 */
470 dn->dn_objset = os;
471
472 dnh->dnh_dnode = dn;
473 mutex_exit(&os->os_lock);
474
475 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
476 return (dn);
477 }
478
479 /*
480 * Caller must be holding the dnode handle, which is released upon return.
481 */
482 static void
dnode_destroy(dnode_t * dn)483 dnode_destroy(dnode_t *dn)
484 {
485 objset_t *os = dn->dn_objset;
486 boolean_t complete_os_eviction = B_FALSE;
487
488 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
489
490 mutex_enter(&os->os_lock);
491 POINTER_INVALIDATE(&dn->dn_objset);
492 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
493 list_remove(&os->os_dnodes, dn);
494 complete_os_eviction =
495 list_is_empty(&os->os_dnodes) &&
496 list_link_active(&os->os_evicting_node);
497 }
498 mutex_exit(&os->os_lock);
499
500 /* the dnode can no longer move, so we can release the handle */
501 zrl_remove(&dn->dn_handle->dnh_zrlock);
502
503 dn->dn_allocated_txg = 0;
504 dn->dn_free_txg = 0;
505 dn->dn_assigned_txg = 0;
506
507 dn->dn_dirtyctx = 0;
508 if (dn->dn_dirtyctx_firstset != NULL) {
509 kmem_free(dn->dn_dirtyctx_firstset, 1);
510 dn->dn_dirtyctx_firstset = NULL;
511 }
512 if (dn->dn_bonus != NULL) {
513 mutex_enter(&dn->dn_bonus->db_mtx);
514 dbuf_evict(dn->dn_bonus);
515 dn->dn_bonus = NULL;
516 }
517 dn->dn_zio = NULL;
518
519 dn->dn_have_spill = B_FALSE;
520 dn->dn_oldused = 0;
521 dn->dn_oldflags = 0;
522 dn->dn_olduid = 0;
523 dn->dn_oldgid = 0;
524 dn->dn_newuid = 0;
525 dn->dn_newgid = 0;
526 dn->dn_id_flags = 0;
527 dn->dn_unlisted_l0_blkid = 0;
528
529 dmu_zfetch_fini(&dn->dn_zfetch);
530 kmem_cache_free(dnode_cache, dn);
531 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
532
533 if (complete_os_eviction)
534 dmu_objset_evict_done(os);
535 }
536
537 void
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)538 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
539 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
540 {
541 int i;
542
543 ASSERT3U(blocksize, <=,
544 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
545 if (blocksize == 0)
546 blocksize = 1 << zfs_default_bs;
547 else
548 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
549
550 if (ibs == 0)
551 ibs = zfs_default_ibs;
552
553 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
554
555 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
556 dn->dn_object, tx->tx_txg, blocksize, ibs);
557
558 ASSERT(dn->dn_type == DMU_OT_NONE);
559 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
560 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
561 ASSERT(ot != DMU_OT_NONE);
562 ASSERT(DMU_OT_IS_VALID(ot));
563 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
564 (bonustype == DMU_OT_SA && bonuslen == 0) ||
565 (bonustype != DMU_OT_NONE && bonuslen != 0));
566 ASSERT(DMU_OT_IS_VALID(bonustype));
567 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
568 ASSERT(dn->dn_type == DMU_OT_NONE);
569 ASSERT0(dn->dn_maxblkid);
570 ASSERT0(dn->dn_allocated_txg);
571 ASSERT0(dn->dn_assigned_txg);
572 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
573 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
574 ASSERT(avl_is_empty(&dn->dn_dbufs));
575
576 for (i = 0; i < TXG_SIZE; i++) {
577 ASSERT0(dn->dn_next_nblkptr[i]);
578 ASSERT0(dn->dn_next_nlevels[i]);
579 ASSERT0(dn->dn_next_indblkshift[i]);
580 ASSERT0(dn->dn_next_bonuslen[i]);
581 ASSERT0(dn->dn_next_bonustype[i]);
582 ASSERT0(dn->dn_rm_spillblk[i]);
583 ASSERT0(dn->dn_next_blksz[i]);
584 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
585 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
586 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
587 }
588
589 dn->dn_type = ot;
590 dnode_setdblksz(dn, blocksize);
591 dn->dn_indblkshift = ibs;
592 dn->dn_nlevels = 1;
593 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
594 dn->dn_nblkptr = 1;
595 else
596 dn->dn_nblkptr = 1 +
597 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
598 dn->dn_bonustype = bonustype;
599 dn->dn_bonuslen = bonuslen;
600 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
601 dn->dn_compress = ZIO_COMPRESS_INHERIT;
602 dn->dn_dirtyctx = 0;
603
604 dn->dn_free_txg = 0;
605 if (dn->dn_dirtyctx_firstset) {
606 kmem_free(dn->dn_dirtyctx_firstset, 1);
607 dn->dn_dirtyctx_firstset = NULL;
608 }
609
610 dn->dn_allocated_txg = tx->tx_txg;
611 dn->dn_id_flags = 0;
612
613 dnode_setdirty(dn, tx);
614 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
615 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
616 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
617 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
618 }
619
620 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)621 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
622 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
623 {
624 int nblkptr;
625
626 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
627 ASSERT3U(blocksize, <=,
628 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
629 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
630 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
631 ASSERT(tx->tx_txg != 0);
632 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
633 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
634 (bonustype == DMU_OT_SA && bonuslen == 0));
635 ASSERT(DMU_OT_IS_VALID(bonustype));
636 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
637
638 /* clean up any unreferenced dbufs */
639 dnode_evict_dbufs(dn);
640
641 dn->dn_id_flags = 0;
642
643 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
644 dnode_setdirty(dn, tx);
645 if (dn->dn_datablksz != blocksize) {
646 /* change blocksize */
647 ASSERT(dn->dn_maxblkid == 0 &&
648 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
649 dnode_block_freed(dn, 0)));
650 dnode_setdblksz(dn, blocksize);
651 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
652 }
653 if (dn->dn_bonuslen != bonuslen)
654 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
655
656 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
657 nblkptr = 1;
658 else
659 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
660 if (dn->dn_bonustype != bonustype)
661 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
662 if (dn->dn_nblkptr != nblkptr)
663 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
664 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
665 dbuf_rm_spill(dn, tx);
666 dnode_rm_spill(dn, tx);
667 }
668 rw_exit(&dn->dn_struct_rwlock);
669
670 /* change type */
671 dn->dn_type = ot;
672
673 /* change bonus size and type */
674 mutex_enter(&dn->dn_mtx);
675 dn->dn_bonustype = bonustype;
676 dn->dn_bonuslen = bonuslen;
677 dn->dn_nblkptr = nblkptr;
678 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
679 dn->dn_compress = ZIO_COMPRESS_INHERIT;
680 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
681
682 /* fix up the bonus db_size */
683 if (dn->dn_bonus) {
684 dn->dn_bonus->db.db_size =
685 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
686 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
687 }
688
689 dn->dn_allocated_txg = tx->tx_txg;
690 mutex_exit(&dn->dn_mtx);
691 }
692
693 #ifdef DNODE_STATS
694 static struct {
695 uint64_t dms_dnode_invalid;
696 uint64_t dms_dnode_recheck1;
697 uint64_t dms_dnode_recheck2;
698 uint64_t dms_dnode_special;
699 uint64_t dms_dnode_handle;
700 uint64_t dms_dnode_rwlock;
701 uint64_t dms_dnode_active;
702 } dnode_move_stats;
703 #endif /* DNODE_STATS */
704
705 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)706 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
707 {
708 int i;
709
710 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
711 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
712 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
713 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
714
715 /* Copy fields. */
716 ndn->dn_objset = odn->dn_objset;
717 ndn->dn_object = odn->dn_object;
718 ndn->dn_dbuf = odn->dn_dbuf;
719 ndn->dn_handle = odn->dn_handle;
720 ndn->dn_phys = odn->dn_phys;
721 ndn->dn_type = odn->dn_type;
722 ndn->dn_bonuslen = odn->dn_bonuslen;
723 ndn->dn_bonustype = odn->dn_bonustype;
724 ndn->dn_nblkptr = odn->dn_nblkptr;
725 ndn->dn_checksum = odn->dn_checksum;
726 ndn->dn_compress = odn->dn_compress;
727 ndn->dn_nlevels = odn->dn_nlevels;
728 ndn->dn_indblkshift = odn->dn_indblkshift;
729 ndn->dn_datablkshift = odn->dn_datablkshift;
730 ndn->dn_datablkszsec = odn->dn_datablkszsec;
731 ndn->dn_datablksz = odn->dn_datablksz;
732 ndn->dn_maxblkid = odn->dn_maxblkid;
733 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
734 sizeof (odn->dn_next_nblkptr));
735 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
736 sizeof (odn->dn_next_nlevels));
737 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
738 sizeof (odn->dn_next_indblkshift));
739 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
740 sizeof (odn->dn_next_bonustype));
741 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
742 sizeof (odn->dn_rm_spillblk));
743 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
744 sizeof (odn->dn_next_bonuslen));
745 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
746 sizeof (odn->dn_next_blksz));
747 for (i = 0; i < TXG_SIZE; i++) {
748 list_move_tail(&ndn->dn_dirty_records[i],
749 &odn->dn_dirty_records[i]);
750 }
751 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
752 sizeof (odn->dn_free_ranges));
753 ndn->dn_allocated_txg = odn->dn_allocated_txg;
754 ndn->dn_free_txg = odn->dn_free_txg;
755 ndn->dn_assigned_txg = odn->dn_assigned_txg;
756 ndn->dn_dirtyctx = odn->dn_dirtyctx;
757 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
758 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
759 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
760 ASSERT(avl_is_empty(&ndn->dn_dbufs));
761 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
762 ndn->dn_dbufs_count = odn->dn_dbufs_count;
763 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
764 ndn->dn_bonus = odn->dn_bonus;
765 ndn->dn_have_spill = odn->dn_have_spill;
766 ndn->dn_zio = odn->dn_zio;
767 ndn->dn_oldused = odn->dn_oldused;
768 ndn->dn_oldflags = odn->dn_oldflags;
769 ndn->dn_olduid = odn->dn_olduid;
770 ndn->dn_oldgid = odn->dn_oldgid;
771 ndn->dn_newuid = odn->dn_newuid;
772 ndn->dn_newgid = odn->dn_newgid;
773 ndn->dn_id_flags = odn->dn_id_flags;
774 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
775 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
776 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
777
778 /*
779 * Update back pointers. Updating the handle fixes the back pointer of
780 * every descendant dbuf as well as the bonus dbuf.
781 */
782 ASSERT(ndn->dn_handle->dnh_dnode == odn);
783 ndn->dn_handle->dnh_dnode = ndn;
784 if (ndn->dn_zfetch.zf_dnode == odn) {
785 ndn->dn_zfetch.zf_dnode = ndn;
786 }
787
788 /*
789 * Invalidate the original dnode by clearing all of its back pointers.
790 */
791 odn->dn_dbuf = NULL;
792 odn->dn_handle = NULL;
793 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
794 offsetof(dmu_buf_impl_t, db_link));
795 odn->dn_dbufs_count = 0;
796 odn->dn_unlisted_l0_blkid = 0;
797 odn->dn_bonus = NULL;
798 odn->dn_zfetch.zf_dnode = NULL;
799
800 /*
801 * Set the low bit of the objset pointer to ensure that dnode_move()
802 * recognizes the dnode as invalid in any subsequent callback.
803 */
804 POINTER_INVALIDATE(&odn->dn_objset);
805
806 /*
807 * Satisfy the destructor.
808 */
809 for (i = 0; i < TXG_SIZE; i++) {
810 list_create(&odn->dn_dirty_records[i],
811 sizeof (dbuf_dirty_record_t),
812 offsetof(dbuf_dirty_record_t, dr_dirty_node));
813 odn->dn_free_ranges[i] = NULL;
814 odn->dn_next_nlevels[i] = 0;
815 odn->dn_next_indblkshift[i] = 0;
816 odn->dn_next_bonustype[i] = 0;
817 odn->dn_rm_spillblk[i] = 0;
818 odn->dn_next_bonuslen[i] = 0;
819 odn->dn_next_blksz[i] = 0;
820 }
821 odn->dn_allocated_txg = 0;
822 odn->dn_free_txg = 0;
823 odn->dn_assigned_txg = 0;
824 odn->dn_dirtyctx = 0;
825 odn->dn_dirtyctx_firstset = NULL;
826 odn->dn_have_spill = B_FALSE;
827 odn->dn_zio = NULL;
828 odn->dn_oldused = 0;
829 odn->dn_oldflags = 0;
830 odn->dn_olduid = 0;
831 odn->dn_oldgid = 0;
832 odn->dn_newuid = 0;
833 odn->dn_newgid = 0;
834 odn->dn_id_flags = 0;
835
836 /*
837 * Mark the dnode.
838 */
839 ndn->dn_moved = 1;
840 odn->dn_moved = (uint8_t)-1;
841 }
842
843 #ifdef illumos
844 #ifdef _KERNEL
845 /*ARGSUSED*/
846 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)847 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
848 {
849 dnode_t *odn = buf, *ndn = newbuf;
850 objset_t *os;
851 int64_t refcount;
852 uint32_t dbufs;
853
854 /*
855 * The dnode is on the objset's list of known dnodes if the objset
856 * pointer is valid. We set the low bit of the objset pointer when
857 * freeing the dnode to invalidate it, and the memory patterns written
858 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
859 * A newly created dnode sets the objset pointer last of all to indicate
860 * that the dnode is known and in a valid state to be moved by this
861 * function.
862 */
863 os = odn->dn_objset;
864 if (!POINTER_IS_VALID(os)) {
865 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
866 return (KMEM_CBRC_DONT_KNOW);
867 }
868
869 /*
870 * Ensure that the objset does not go away during the move.
871 */
872 rw_enter(&os_lock, RW_WRITER);
873 if (os != odn->dn_objset) {
874 rw_exit(&os_lock);
875 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
876 return (KMEM_CBRC_DONT_KNOW);
877 }
878
879 /*
880 * If the dnode is still valid, then so is the objset. We know that no
881 * valid objset can be freed while we hold os_lock, so we can safely
882 * ensure that the objset remains in use.
883 */
884 mutex_enter(&os->os_lock);
885
886 /*
887 * Recheck the objset pointer in case the dnode was removed just before
888 * acquiring the lock.
889 */
890 if (os != odn->dn_objset) {
891 mutex_exit(&os->os_lock);
892 rw_exit(&os_lock);
893 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
894 return (KMEM_CBRC_DONT_KNOW);
895 }
896
897 /*
898 * At this point we know that as long as we hold os->os_lock, the dnode
899 * cannot be freed and fields within the dnode can be safely accessed.
900 * The objset listing this dnode cannot go away as long as this dnode is
901 * on its list.
902 */
903 rw_exit(&os_lock);
904 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
905 mutex_exit(&os->os_lock);
906 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
907 return (KMEM_CBRC_NO);
908 }
909 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
910
911 /*
912 * Lock the dnode handle to prevent the dnode from obtaining any new
913 * holds. This also prevents the descendant dbufs and the bonus dbuf
914 * from accessing the dnode, so that we can discount their holds. The
915 * handle is safe to access because we know that while the dnode cannot
916 * go away, neither can its handle. Once we hold dnh_zrlock, we can
917 * safely move any dnode referenced only by dbufs.
918 */
919 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
920 mutex_exit(&os->os_lock);
921 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
922 return (KMEM_CBRC_LATER);
923 }
924
925 /*
926 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
927 * We need to guarantee that there is a hold for every dbuf in order to
928 * determine whether the dnode is actively referenced. Falsely matching
929 * a dbuf to an active hold would lead to an unsafe move. It's possible
930 * that a thread already having an active dnode hold is about to add a
931 * dbuf, and we can't compare hold and dbuf counts while the add is in
932 * progress.
933 */
934 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
935 zrl_exit(&odn->dn_handle->dnh_zrlock);
936 mutex_exit(&os->os_lock);
937 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
938 return (KMEM_CBRC_LATER);
939 }
940
941 /*
942 * A dbuf may be removed (evicted) without an active dnode hold. In that
943 * case, the dbuf count is decremented under the handle lock before the
944 * dbuf's hold is released. This order ensures that if we count the hold
945 * after the dbuf is removed but before its hold is released, we will
946 * treat the unmatched hold as active and exit safely. If we count the
947 * hold before the dbuf is removed, the hold is discounted, and the
948 * removal is blocked until the move completes.
949 */
950 refcount = refcount_count(&odn->dn_holds);
951 ASSERT(refcount >= 0);
952 dbufs = odn->dn_dbufs_count;
953
954 /* We can't have more dbufs than dnode holds. */
955 ASSERT3U(dbufs, <=, refcount);
956 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
957 uint32_t, dbufs);
958
959 if (refcount > dbufs) {
960 rw_exit(&odn->dn_struct_rwlock);
961 zrl_exit(&odn->dn_handle->dnh_zrlock);
962 mutex_exit(&os->os_lock);
963 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
964 return (KMEM_CBRC_LATER);
965 }
966
967 rw_exit(&odn->dn_struct_rwlock);
968
969 /*
970 * At this point we know that anyone with a hold on the dnode is not
971 * actively referencing it. The dnode is known and in a valid state to
972 * move. We're holding the locks needed to execute the critical section.
973 */
974 dnode_move_impl(odn, ndn);
975
976 list_link_replace(&odn->dn_link, &ndn->dn_link);
977 /* If the dnode was safe to move, the refcount cannot have changed. */
978 ASSERT(refcount == refcount_count(&ndn->dn_holds));
979 ASSERT(dbufs == ndn->dn_dbufs_count);
980 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
981 mutex_exit(&os->os_lock);
982
983 return (KMEM_CBRC_YES);
984 }
985 #endif /* _KERNEL */
986 #endif /* illumos */
987
988 void
dnode_special_close(dnode_handle_t * dnh)989 dnode_special_close(dnode_handle_t *dnh)
990 {
991 dnode_t *dn = dnh->dnh_dnode;
992
993 /*
994 * Wait for final references to the dnode to clear. This can
995 * only happen if the arc is asyncronously evicting state that
996 * has a hold on this dnode while we are trying to evict this
997 * dnode.
998 */
999 while (refcount_count(&dn->dn_holds) > 0)
1000 delay(1);
1001 ASSERT(dn->dn_dbuf == NULL ||
1002 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1003 zrl_add(&dnh->dnh_zrlock);
1004 dnode_destroy(dn); /* implicit zrl_remove() */
1005 zrl_destroy(&dnh->dnh_zrlock);
1006 dnh->dnh_dnode = NULL;
1007 }
1008
1009 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1010 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1011 dnode_handle_t *dnh)
1012 {
1013 dnode_t *dn;
1014
1015 dn = dnode_create(os, dnp, NULL, object, dnh);
1016 zrl_init(&dnh->dnh_zrlock);
1017 DNODE_VERIFY(dn);
1018 }
1019
1020 static void
dnode_buf_pageout(void * dbu)1021 dnode_buf_pageout(void *dbu)
1022 {
1023 dnode_children_t *children_dnodes = dbu;
1024 int i;
1025
1026 for (i = 0; i < children_dnodes->dnc_count; i++) {
1027 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1028 dnode_t *dn;
1029
1030 /*
1031 * The dnode handle lock guards against the dnode moving to
1032 * another valid address, so there is no need here to guard
1033 * against changes to or from NULL.
1034 */
1035 if (dnh->dnh_dnode == NULL) {
1036 zrl_destroy(&dnh->dnh_zrlock);
1037 continue;
1038 }
1039
1040 zrl_add(&dnh->dnh_zrlock);
1041 dn = dnh->dnh_dnode;
1042 /*
1043 * If there are holds on this dnode, then there should
1044 * be holds on the dnode's containing dbuf as well; thus
1045 * it wouldn't be eligible for eviction and this function
1046 * would not have been called.
1047 */
1048 ASSERT(refcount_is_zero(&dn->dn_holds));
1049 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1050
1051 dnode_destroy(dn); /* implicit zrl_remove() */
1052 zrl_destroy(&dnh->dnh_zrlock);
1053 dnh->dnh_dnode = NULL;
1054 }
1055 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1056 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1057 }
1058
1059 /*
1060 * errors:
1061 * EINVAL - invalid object number.
1062 * EIO - i/o error.
1063 * succeeds even for free dnodes.
1064 */
1065 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1066 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1067 void *tag, dnode_t **dnp)
1068 {
1069 int epb, idx, err;
1070 int drop_struct_lock = FALSE;
1071 int type;
1072 uint64_t blk;
1073 dnode_t *mdn, *dn;
1074 dmu_buf_impl_t *db;
1075 dnode_children_t *children_dnodes;
1076 dnode_handle_t *dnh;
1077
1078 /*
1079 * If you are holding the spa config lock as writer, you shouldn't
1080 * be asking the DMU to do *anything* unless it's the root pool
1081 * which may require us to read from the root filesystem while
1082 * holding some (not all) of the locks as writer.
1083 */
1084 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1085 (spa_is_root(os->os_spa) &&
1086 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1087
1088 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1089 dn = (object == DMU_USERUSED_OBJECT) ?
1090 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1091 if (dn == NULL)
1092 return (SET_ERROR(ENOENT));
1093 type = dn->dn_type;
1094 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1095 return (SET_ERROR(ENOENT));
1096 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1097 return (SET_ERROR(EEXIST));
1098 DNODE_VERIFY(dn);
1099 (void) refcount_add(&dn->dn_holds, tag);
1100 *dnp = dn;
1101 return (0);
1102 }
1103
1104 if (object == 0 || object >= DN_MAX_OBJECT)
1105 return (SET_ERROR(EINVAL));
1106
1107 mdn = DMU_META_DNODE(os);
1108 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1109
1110 DNODE_VERIFY(mdn);
1111
1112 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1113 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1114 drop_struct_lock = TRUE;
1115 }
1116
1117 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1118
1119 db = dbuf_hold(mdn, blk, FTAG);
1120 if (drop_struct_lock)
1121 rw_exit(&mdn->dn_struct_rwlock);
1122 if (db == NULL)
1123 return (SET_ERROR(EIO));
1124 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1125 if (err) {
1126 dbuf_rele(db, FTAG);
1127 return (err);
1128 }
1129
1130 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1131 epb = db->db.db_size >> DNODE_SHIFT;
1132
1133 idx = object & (epb-1);
1134
1135 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1136 children_dnodes = dmu_buf_get_user(&db->db);
1137 if (children_dnodes == NULL) {
1138 int i;
1139 dnode_children_t *winner;
1140 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1141 epb * sizeof (dnode_handle_t), KM_SLEEP);
1142 children_dnodes->dnc_count = epb;
1143 dnh = &children_dnodes->dnc_children[0];
1144 for (i = 0; i < epb; i++) {
1145 zrl_init(&dnh[i].dnh_zrlock);
1146 }
1147 dmu_buf_init_user(&children_dnodes->dnc_dbu,
1148 dnode_buf_pageout, NULL);
1149 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1150 if (winner != NULL) {
1151
1152 for (i = 0; i < epb; i++) {
1153 zrl_destroy(&dnh[i].dnh_zrlock);
1154 }
1155
1156 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1157 epb * sizeof (dnode_handle_t));
1158 children_dnodes = winner;
1159 }
1160 }
1161 ASSERT(children_dnodes->dnc_count == epb);
1162
1163 dnh = &children_dnodes->dnc_children[idx];
1164 zrl_add(&dnh->dnh_zrlock);
1165 dn = dnh->dnh_dnode;
1166 if (dn == NULL) {
1167 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1168
1169 dn = dnode_create(os, phys, db, object, dnh);
1170 }
1171
1172 mutex_enter(&dn->dn_mtx);
1173 type = dn->dn_type;
1174 if (dn->dn_free_txg ||
1175 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1176 ((flag & DNODE_MUST_BE_FREE) &&
1177 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1178 mutex_exit(&dn->dn_mtx);
1179 zrl_remove(&dnh->dnh_zrlock);
1180 dbuf_rele(db, FTAG);
1181 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1182 }
1183 if (refcount_add(&dn->dn_holds, tag) == 1)
1184 dbuf_add_ref(db, dnh);
1185 mutex_exit(&dn->dn_mtx);
1186
1187 /* Now we can rely on the hold to prevent the dnode from moving. */
1188 zrl_remove(&dnh->dnh_zrlock);
1189
1190 DNODE_VERIFY(dn);
1191 ASSERT3P(dn->dn_dbuf, ==, db);
1192 ASSERT3U(dn->dn_object, ==, object);
1193 dbuf_rele(db, FTAG);
1194
1195 *dnp = dn;
1196 return (0);
1197 }
1198
1199 /*
1200 * Return held dnode if the object is allocated, NULL if not.
1201 */
1202 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1203 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1204 {
1205 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1206 }
1207
1208 /*
1209 * Can only add a reference if there is already at least one
1210 * reference on the dnode. Returns FALSE if unable to add a
1211 * new reference.
1212 */
1213 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1214 dnode_add_ref(dnode_t *dn, void *tag)
1215 {
1216 mutex_enter(&dn->dn_mtx);
1217 if (refcount_is_zero(&dn->dn_holds)) {
1218 mutex_exit(&dn->dn_mtx);
1219 return (FALSE);
1220 }
1221 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1222 mutex_exit(&dn->dn_mtx);
1223 return (TRUE);
1224 }
1225
1226 void
dnode_rele(dnode_t * dn,void * tag)1227 dnode_rele(dnode_t *dn, void *tag)
1228 {
1229 mutex_enter(&dn->dn_mtx);
1230 dnode_rele_and_unlock(dn, tag);
1231 }
1232
1233 void
dnode_rele_and_unlock(dnode_t * dn,void * tag)1234 dnode_rele_and_unlock(dnode_t *dn, void *tag)
1235 {
1236 uint64_t refs;
1237 /* Get while the hold prevents the dnode from moving. */
1238 dmu_buf_impl_t *db = dn->dn_dbuf;
1239 dnode_handle_t *dnh = dn->dn_handle;
1240
1241 refs = refcount_remove(&dn->dn_holds, tag);
1242 mutex_exit(&dn->dn_mtx);
1243
1244 /*
1245 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1246 * indirectly by dbuf_rele() while relying on the dnode handle to
1247 * prevent the dnode from moving, since releasing the last hold could
1248 * result in the dnode's parent dbuf evicting its dnode handles. For
1249 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1250 * other direct or indirect hold on the dnode must first drop the dnode
1251 * handle.
1252 */
1253 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1254
1255 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1256 if (refs == 0 && db != NULL) {
1257 /*
1258 * Another thread could add a hold to the dnode handle in
1259 * dnode_hold_impl() while holding the parent dbuf. Since the
1260 * hold on the parent dbuf prevents the handle from being
1261 * destroyed, the hold on the handle is OK. We can't yet assert
1262 * that the handle has zero references, but that will be
1263 * asserted anyway when the handle gets destroyed.
1264 */
1265 dbuf_rele(db, dnh);
1266 }
1267 }
1268
1269 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1270 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1271 {
1272 objset_t *os = dn->dn_objset;
1273 uint64_t txg = tx->tx_txg;
1274
1275 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1276 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1277 return;
1278 }
1279
1280 DNODE_VERIFY(dn);
1281
1282 #ifdef ZFS_DEBUG
1283 mutex_enter(&dn->dn_mtx);
1284 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1285 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1286 mutex_exit(&dn->dn_mtx);
1287 #endif
1288
1289 /*
1290 * Determine old uid/gid when necessary
1291 */
1292 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1293
1294 mutex_enter(&os->os_lock);
1295
1296 /*
1297 * If we are already marked dirty, we're done.
1298 */
1299 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1300 mutex_exit(&os->os_lock);
1301 return;
1302 }
1303
1304 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1305 !avl_is_empty(&dn->dn_dbufs));
1306 ASSERT(dn->dn_datablksz != 0);
1307 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1308 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1309 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1310
1311 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1312 dn->dn_object, txg);
1313
1314 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1315 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1316 } else {
1317 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1318 }
1319
1320 mutex_exit(&os->os_lock);
1321
1322 /*
1323 * The dnode maintains a hold on its containing dbuf as
1324 * long as there are holds on it. Each instantiated child
1325 * dbuf maintains a hold on the dnode. When the last child
1326 * drops its hold, the dnode will drop its hold on the
1327 * containing dbuf. We add a "dirty hold" here so that the
1328 * dnode will hang around after we finish processing its
1329 * children.
1330 */
1331 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1332
1333 (void) dbuf_dirty(dn->dn_dbuf, tx);
1334
1335 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1336 }
1337
1338 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1339 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1340 {
1341 int txgoff = tx->tx_txg & TXG_MASK;
1342
1343 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1344
1345 /* we should be the only holder... hopefully */
1346 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1347
1348 mutex_enter(&dn->dn_mtx);
1349 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1350 mutex_exit(&dn->dn_mtx);
1351 return;
1352 }
1353 dn->dn_free_txg = tx->tx_txg;
1354 mutex_exit(&dn->dn_mtx);
1355
1356 /*
1357 * If the dnode is already dirty, it needs to be moved from
1358 * the dirty list to the free list.
1359 */
1360 mutex_enter(&dn->dn_objset->os_lock);
1361 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1362 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1363 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1364 mutex_exit(&dn->dn_objset->os_lock);
1365 } else {
1366 mutex_exit(&dn->dn_objset->os_lock);
1367 dnode_setdirty(dn, tx);
1368 }
1369 }
1370
1371 /*
1372 * Try to change the block size for the indicated dnode. This can only
1373 * succeed if there are no blocks allocated or dirty beyond first block
1374 */
1375 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1376 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1377 {
1378 dmu_buf_impl_t *db;
1379 int err;
1380
1381 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1382 if (size == 0)
1383 size = SPA_MINBLOCKSIZE;
1384 else
1385 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1386
1387 if (ibs == dn->dn_indblkshift)
1388 ibs = 0;
1389
1390 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1391 return (0);
1392
1393 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1394
1395 /* Check for any allocated blocks beyond the first */
1396 if (dn->dn_maxblkid != 0)
1397 goto fail;
1398
1399 mutex_enter(&dn->dn_dbufs_mtx);
1400 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1401 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1402 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1403 db->db_blkid != DMU_SPILL_BLKID) {
1404 mutex_exit(&dn->dn_dbufs_mtx);
1405 goto fail;
1406 }
1407 }
1408 mutex_exit(&dn->dn_dbufs_mtx);
1409
1410 if (ibs && dn->dn_nlevels != 1)
1411 goto fail;
1412
1413 /* resize the old block */
1414 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1415 if (err == 0)
1416 dbuf_new_size(db, size, tx);
1417 else if (err != ENOENT)
1418 goto fail;
1419
1420 dnode_setdblksz(dn, size);
1421 dnode_setdirty(dn, tx);
1422 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1423 if (ibs) {
1424 dn->dn_indblkshift = ibs;
1425 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1426 }
1427 /* rele after we have fixed the blocksize in the dnode */
1428 if (db)
1429 dbuf_rele(db, FTAG);
1430
1431 rw_exit(&dn->dn_struct_rwlock);
1432 return (0);
1433
1434 fail:
1435 rw_exit(&dn->dn_struct_rwlock);
1436 return (SET_ERROR(ENOTSUP));
1437 }
1438
1439 /* read-holding callers must not rely on the lock being continuously held */
1440 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1441 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1442 {
1443 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1444 int epbs, new_nlevels;
1445 uint64_t sz;
1446
1447 ASSERT(blkid != DMU_BONUS_BLKID);
1448
1449 ASSERT(have_read ?
1450 RW_READ_HELD(&dn->dn_struct_rwlock) :
1451 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1452
1453 /*
1454 * if we have a read-lock, check to see if we need to do any work
1455 * before upgrading to a write-lock.
1456 */
1457 if (have_read) {
1458 if (blkid <= dn->dn_maxblkid)
1459 return;
1460
1461 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1462 rw_exit(&dn->dn_struct_rwlock);
1463 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1464 }
1465 }
1466
1467 if (blkid <= dn->dn_maxblkid)
1468 goto out;
1469
1470 dn->dn_maxblkid = blkid;
1471
1472 /*
1473 * Compute the number of levels necessary to support the new maxblkid.
1474 */
1475 new_nlevels = 1;
1476 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1477 for (sz = dn->dn_nblkptr;
1478 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1479 new_nlevels++;
1480
1481 if (new_nlevels > dn->dn_nlevels) {
1482 int old_nlevels = dn->dn_nlevels;
1483 dmu_buf_impl_t *db;
1484 list_t *list;
1485 dbuf_dirty_record_t *new, *dr, *dr_next;
1486
1487 dn->dn_nlevels = new_nlevels;
1488
1489 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1490 dn->dn_next_nlevels[txgoff] = new_nlevels;
1491
1492 /* dirty the left indirects */
1493 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1494 ASSERT(db != NULL);
1495 new = dbuf_dirty(db, tx);
1496 dbuf_rele(db, FTAG);
1497
1498 /* transfer the dirty records to the new indirect */
1499 mutex_enter(&dn->dn_mtx);
1500 mutex_enter(&new->dt.di.dr_mtx);
1501 list = &dn->dn_dirty_records[txgoff];
1502 for (dr = list_head(list); dr; dr = dr_next) {
1503 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1504 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1505 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1506 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1507 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1508 list_remove(&dn->dn_dirty_records[txgoff], dr);
1509 list_insert_tail(&new->dt.di.dr_children, dr);
1510 dr->dr_parent = new;
1511 }
1512 }
1513 mutex_exit(&new->dt.di.dr_mtx);
1514 mutex_exit(&dn->dn_mtx);
1515 }
1516
1517 out:
1518 if (have_read)
1519 rw_downgrade(&dn->dn_struct_rwlock);
1520 }
1521
1522 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1523 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1524 {
1525 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1526 if (db != NULL) {
1527 dmu_buf_will_dirty(&db->db, tx);
1528 dbuf_rele(db, FTAG);
1529 }
1530 }
1531
1532 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1533 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1534 {
1535 dmu_buf_impl_t *db;
1536 uint64_t blkoff, blkid, nblks;
1537 int blksz, blkshift, head, tail;
1538 int trunc = FALSE;
1539 int epbs;
1540
1541 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1542 blksz = dn->dn_datablksz;
1543 blkshift = dn->dn_datablkshift;
1544 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1545
1546 if (len == DMU_OBJECT_END) {
1547 len = UINT64_MAX - off;
1548 trunc = TRUE;
1549 }
1550
1551 /*
1552 * First, block align the region to free:
1553 */
1554 if (ISP2(blksz)) {
1555 head = P2NPHASE(off, blksz);
1556 blkoff = P2PHASE(off, blksz);
1557 if ((off >> blkshift) > dn->dn_maxblkid)
1558 goto out;
1559 } else {
1560 ASSERT(dn->dn_maxblkid == 0);
1561 if (off == 0 && len >= blksz) {
1562 /*
1563 * Freeing the whole block; fast-track this request.
1564 * Note that we won't dirty any indirect blocks,
1565 * which is fine because we will be freeing the entire
1566 * file and thus all indirect blocks will be freed
1567 * by free_children().
1568 */
1569 blkid = 0;
1570 nblks = 1;
1571 goto done;
1572 } else if (off >= blksz) {
1573 /* Freeing past end-of-data */
1574 goto out;
1575 } else {
1576 /* Freeing part of the block. */
1577 head = blksz - off;
1578 ASSERT3U(head, >, 0);
1579 }
1580 blkoff = off;
1581 }
1582 /* zero out any partial block data at the start of the range */
1583 if (head) {
1584 ASSERT3U(blkoff + head, ==, blksz);
1585 if (len < head)
1586 head = len;
1587 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1588 TRUE, FALSE, FTAG, &db) == 0) {
1589 caddr_t data;
1590
1591 /* don't dirty if it isn't on disk and isn't dirty */
1592 if (db->db_last_dirty ||
1593 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1594 rw_exit(&dn->dn_struct_rwlock);
1595 dmu_buf_will_dirty(&db->db, tx);
1596 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1597 data = db->db.db_data;
1598 bzero(data + blkoff, head);
1599 }
1600 dbuf_rele(db, FTAG);
1601 }
1602 off += head;
1603 len -= head;
1604 }
1605
1606 /* If the range was less than one block, we're done */
1607 if (len == 0)
1608 goto out;
1609
1610 /* If the remaining range is past end of file, we're done */
1611 if ((off >> blkshift) > dn->dn_maxblkid)
1612 goto out;
1613
1614 ASSERT(ISP2(blksz));
1615 if (trunc)
1616 tail = 0;
1617 else
1618 tail = P2PHASE(len, blksz);
1619
1620 ASSERT0(P2PHASE(off, blksz));
1621 /* zero out any partial block data at the end of the range */
1622 if (tail) {
1623 if (len < tail)
1624 tail = len;
1625 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1626 TRUE, FALSE, FTAG, &db) == 0) {
1627 /* don't dirty if not on disk and not dirty */
1628 if (db->db_last_dirty ||
1629 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1630 rw_exit(&dn->dn_struct_rwlock);
1631 dmu_buf_will_dirty(&db->db, tx);
1632 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1633 bzero(db->db.db_data, tail);
1634 }
1635 dbuf_rele(db, FTAG);
1636 }
1637 len -= tail;
1638 }
1639
1640 /* If the range did not include a full block, we are done */
1641 if (len == 0)
1642 goto out;
1643
1644 ASSERT(IS_P2ALIGNED(off, blksz));
1645 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1646 blkid = off >> blkshift;
1647 nblks = len >> blkshift;
1648 if (trunc)
1649 nblks += 1;
1650
1651 /*
1652 * Dirty all the indirect blocks in this range. Note that only
1653 * the first and last indirect blocks can actually be written
1654 * (if they were partially freed) -- they must be dirtied, even if
1655 * they do not exist on disk yet. The interior blocks will
1656 * be freed by free_children(), so they will not actually be written.
1657 * Even though these interior blocks will not be written, we
1658 * dirty them for two reasons:
1659 *
1660 * - It ensures that the indirect blocks remain in memory until
1661 * syncing context. (They have already been prefetched by
1662 * dmu_tx_hold_free(), so we don't have to worry about reading
1663 * them serially here.)
1664 *
1665 * - The dirty space accounting will put pressure on the txg sync
1666 * mechanism to begin syncing, and to delay transactions if there
1667 * is a large amount of freeing. Even though these indirect
1668 * blocks will not be written, we could need to write the same
1669 * amount of space if we copy the freed BPs into deadlists.
1670 */
1671 if (dn->dn_nlevels > 1) {
1672 uint64_t first, last;
1673
1674 first = blkid >> epbs;
1675 dnode_dirty_l1(dn, first, tx);
1676 if (trunc)
1677 last = dn->dn_maxblkid >> epbs;
1678 else
1679 last = (blkid + nblks - 1) >> epbs;
1680 if (last != first)
1681 dnode_dirty_l1(dn, last, tx);
1682
1683 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1684 SPA_BLKPTRSHIFT;
1685 for (uint64_t i = first + 1; i < last; i++) {
1686 /*
1687 * Set i to the blockid of the next non-hole
1688 * level-1 indirect block at or after i. Note
1689 * that dnode_next_offset() operates in terms of
1690 * level-0-equivalent bytes.
1691 */
1692 uint64_t ibyte = i << shift;
1693 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1694 &ibyte, 2, 1, 0);
1695 i = ibyte >> shift;
1696 if (i >= last)
1697 break;
1698
1699 /*
1700 * Normally we should not see an error, either
1701 * from dnode_next_offset() or dbuf_hold_level()
1702 * (except for ESRCH from dnode_next_offset).
1703 * If there is an i/o error, then when we read
1704 * this block in syncing context, it will use
1705 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1706 * to the "failmode" property. dnode_next_offset()
1707 * doesn't have a flag to indicate MUSTSUCCEED.
1708 */
1709 if (err != 0)
1710 break;
1711
1712 dnode_dirty_l1(dn, i, tx);
1713 }
1714 }
1715
1716 done:
1717 /*
1718 * Add this range to the dnode range list.
1719 * We will finish up this free operation in the syncing phase.
1720 */
1721 mutex_enter(&dn->dn_mtx);
1722 int txgoff = tx->tx_txg & TXG_MASK;
1723 if (dn->dn_free_ranges[txgoff] == NULL) {
1724 dn->dn_free_ranges[txgoff] =
1725 range_tree_create(NULL, NULL, &dn->dn_mtx);
1726 }
1727 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1728 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1729 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1730 blkid, nblks, tx->tx_txg);
1731 mutex_exit(&dn->dn_mtx);
1732
1733 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1734 dnode_setdirty(dn, tx);
1735 out:
1736
1737 rw_exit(&dn->dn_struct_rwlock);
1738 }
1739
1740 static boolean_t
dnode_spill_freed(dnode_t * dn)1741 dnode_spill_freed(dnode_t *dn)
1742 {
1743 int i;
1744
1745 mutex_enter(&dn->dn_mtx);
1746 for (i = 0; i < TXG_SIZE; i++) {
1747 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1748 break;
1749 }
1750 mutex_exit(&dn->dn_mtx);
1751 return (i < TXG_SIZE);
1752 }
1753
1754 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1755 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1756 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1757 {
1758 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1759 int i;
1760
1761 if (blkid == DMU_BONUS_BLKID)
1762 return (FALSE);
1763
1764 /*
1765 * If we're in the process of opening the pool, dp will not be
1766 * set yet, but there shouldn't be anything dirty.
1767 */
1768 if (dp == NULL)
1769 return (FALSE);
1770
1771 if (dn->dn_free_txg)
1772 return (TRUE);
1773
1774 if (blkid == DMU_SPILL_BLKID)
1775 return (dnode_spill_freed(dn));
1776
1777 mutex_enter(&dn->dn_mtx);
1778 for (i = 0; i < TXG_SIZE; i++) {
1779 if (dn->dn_free_ranges[i] != NULL &&
1780 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1781 break;
1782 }
1783 mutex_exit(&dn->dn_mtx);
1784 return (i < TXG_SIZE);
1785 }
1786
1787 /* call from syncing context when we actually write/free space for this dnode */
1788 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1789 dnode_diduse_space(dnode_t *dn, int64_t delta)
1790 {
1791 uint64_t space;
1792 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1793 dn, dn->dn_phys,
1794 (u_longlong_t)dn->dn_phys->dn_used,
1795 (longlong_t)delta);
1796
1797 mutex_enter(&dn->dn_mtx);
1798 space = DN_USED_BYTES(dn->dn_phys);
1799 if (delta > 0) {
1800 ASSERT3U(space + delta, >=, space); /* no overflow */
1801 } else {
1802 ASSERT3U(space, >=, -delta); /* no underflow */
1803 }
1804 space += delta;
1805 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1806 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1807 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1808 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1809 } else {
1810 dn->dn_phys->dn_used = space;
1811 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1812 }
1813 mutex_exit(&dn->dn_mtx);
1814 }
1815
1816 /*
1817 * Call when we think we're going to write/free space in open context to track
1818 * the amount of memory in use by the currently open txg.
1819 */
1820 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1821 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1822 {
1823 objset_t *os = dn->dn_objset;
1824 dsl_dataset_t *ds = os->os_dsl_dataset;
1825 int64_t aspace = spa_get_asize(os->os_spa, space);
1826
1827 if (ds != NULL) {
1828 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1829 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1830 }
1831
1832 dmu_tx_willuse_space(tx, aspace);
1833 }
1834
1835 /*
1836 * Scans a block at the indicated "level" looking for a hole or data,
1837 * depending on 'flags'.
1838 *
1839 * If level > 0, then we are scanning an indirect block looking at its
1840 * pointers. If level == 0, then we are looking at a block of dnodes.
1841 *
1842 * If we don't find what we are looking for in the block, we return ESRCH.
1843 * Otherwise, return with *offset pointing to the beginning (if searching
1844 * forwards) or end (if searching backwards) of the range covered by the
1845 * block pointer we matched on (or dnode).
1846 *
1847 * The basic search algorithm used below by dnode_next_offset() is to
1848 * use this function to search up the block tree (widen the search) until
1849 * we find something (i.e., we don't return ESRCH) and then search back
1850 * down the tree (narrow the search) until we reach our original search
1851 * level.
1852 */
1853 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1854 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1855 int lvl, uint64_t blkfill, uint64_t txg)
1856 {
1857 dmu_buf_impl_t *db = NULL;
1858 void *data = NULL;
1859 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1860 uint64_t epb = 1ULL << epbs;
1861 uint64_t minfill, maxfill;
1862 boolean_t hole;
1863 int i, inc, error, span;
1864
1865 dprintf("probing object %llu offset %llx level %d of %u\n",
1866 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1867
1868 hole = ((flags & DNODE_FIND_HOLE) != 0);
1869 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1870 ASSERT(txg == 0 || !hole);
1871
1872 if (lvl == dn->dn_phys->dn_nlevels) {
1873 error = 0;
1874 epb = dn->dn_phys->dn_nblkptr;
1875 data = dn->dn_phys->dn_blkptr;
1876 } else {
1877 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1878 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1879 if (error) {
1880 if (error != ENOENT)
1881 return (error);
1882 if (hole)
1883 return (0);
1884 /*
1885 * This can only happen when we are searching up
1886 * the block tree for data. We don't really need to
1887 * adjust the offset, as we will just end up looking
1888 * at the pointer to this block in its parent, and its
1889 * going to be unallocated, so we will skip over it.
1890 */
1891 return (SET_ERROR(ESRCH));
1892 }
1893 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1894 if (error) {
1895 dbuf_rele(db, FTAG);
1896 return (error);
1897 }
1898 data = db->db.db_data;
1899 }
1900
1901
1902 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1903 db->db_blkptr->blk_birth <= txg ||
1904 BP_IS_HOLE(db->db_blkptr))) {
1905 /*
1906 * This can only happen when we are searching up the tree
1907 * and these conditions mean that we need to keep climbing.
1908 */
1909 error = SET_ERROR(ESRCH);
1910 } else if (lvl == 0) {
1911 dnode_phys_t *dnp = data;
1912 span = DNODE_SHIFT;
1913 ASSERT(dn->dn_type == DMU_OT_DNODE);
1914
1915 for (i = (*offset >> span) & (blkfill - 1);
1916 i >= 0 && i < blkfill; i += inc) {
1917 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1918 break;
1919 *offset += (1ULL << span) * inc;
1920 }
1921 if (i < 0 || i == blkfill)
1922 error = SET_ERROR(ESRCH);
1923 } else {
1924 blkptr_t *bp = data;
1925 uint64_t start = *offset;
1926 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1927 minfill = 0;
1928 maxfill = blkfill << ((lvl - 1) * epbs);
1929
1930 if (hole)
1931 maxfill--;
1932 else
1933 minfill++;
1934
1935 *offset = *offset >> span;
1936 for (i = BF64_GET(*offset, 0, epbs);
1937 i >= 0 && i < epb; i += inc) {
1938 if (BP_GET_FILL(&bp[i]) >= minfill &&
1939 BP_GET_FILL(&bp[i]) <= maxfill &&
1940 (hole || bp[i].blk_birth > txg))
1941 break;
1942 if (inc > 0 || *offset > 0)
1943 *offset += inc;
1944 }
1945 *offset = *offset << span;
1946 if (inc < 0) {
1947 /* traversing backwards; position offset at the end */
1948 ASSERT3U(*offset, <=, start);
1949 *offset = MIN(*offset + (1ULL << span) - 1, start);
1950 } else if (*offset < start) {
1951 *offset = start;
1952 }
1953 if (i < 0 || i >= epb)
1954 error = SET_ERROR(ESRCH);
1955 }
1956
1957 if (db)
1958 dbuf_rele(db, FTAG);
1959
1960 return (error);
1961 }
1962
1963 /*
1964 * Find the next hole, data, or sparse region at or after *offset.
1965 * The value 'blkfill' tells us how many items we expect to find
1966 * in an L0 data block; this value is 1 for normal objects,
1967 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1968 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1969 *
1970 * Examples:
1971 *
1972 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1973 * Finds the next/previous hole/data in a file.
1974 * Used in dmu_offset_next().
1975 *
1976 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1977 * Finds the next free/allocated dnode an objset's meta-dnode.
1978 * Only finds objects that have new contents since txg (ie.
1979 * bonus buffer changes and content removal are ignored).
1980 * Used in dmu_object_next().
1981 *
1982 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1983 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1984 * Used in dmu_object_alloc().
1985 */
1986 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1987 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1988 int minlvl, uint64_t blkfill, uint64_t txg)
1989 {
1990 uint64_t initial_offset = *offset;
1991 int lvl, maxlvl;
1992 int error = 0;
1993
1994 if (!(flags & DNODE_FIND_HAVELOCK))
1995 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1996
1997 if (dn->dn_phys->dn_nlevels == 0) {
1998 error = SET_ERROR(ESRCH);
1999 goto out;
2000 }
2001
2002 if (dn->dn_datablkshift == 0) {
2003 if (*offset < dn->dn_datablksz) {
2004 if (flags & DNODE_FIND_HOLE)
2005 *offset = dn->dn_datablksz;
2006 } else {
2007 error = SET_ERROR(ESRCH);
2008 }
2009 goto out;
2010 }
2011
2012 maxlvl = dn->dn_phys->dn_nlevels;
2013
2014 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2015 error = dnode_next_offset_level(dn,
2016 flags, offset, lvl, blkfill, txg);
2017 if (error != ESRCH)
2018 break;
2019 }
2020
2021 while (error == 0 && --lvl >= minlvl) {
2022 error = dnode_next_offset_level(dn,
2023 flags, offset, lvl, blkfill, txg);
2024 }
2025
2026 /*
2027 * There's always a "virtual hole" at the end of the object, even
2028 * if all BP's which physically exist are non-holes.
2029 */
2030 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2031 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2032 error = 0;
2033 }
2034
2035 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2036 initial_offset < *offset : initial_offset > *offset))
2037 error = SET_ERROR(ESRCH);
2038 out:
2039 if (!(flags & DNODE_FIND_HAVELOCK))
2040 rw_exit(&dn->dn_struct_rwlock);
2041
2042 return (error);
2043 }
2044