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 */
25
26 #include <sys/zfs_context.h>
27 #include <sys/dbuf.h>
28 #include <sys/dnode.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dmu_objset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/spa.h>
36 #include <sys/zio.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/range_tree.h>
39
40 static kmem_cache_t *dnode_cache;
41 /*
42 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
43 * turned on when DEBUG is also defined.
44 */
45 #ifdef DEBUG
46 #define DNODE_STATS
47 #endif /* DEBUG */
48
49 #ifdef DNODE_STATS
50 #define DNODE_STAT_ADD(stat) ((stat)++)
51 #else
52 #define DNODE_STAT_ADD(stat) /* nothing */
53 #endif /* DNODE_STATS */
54
55 static dnode_phys_t dnode_phys_zero;
56
57 int zfs_default_bs = SPA_MINBLOCKSHIFT;
58 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
59
60 #ifdef sun
61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
62 #endif
63
64 /* ARGSUSED */
65 static int
dnode_cons(void * arg,void * unused,int kmflag)66 dnode_cons(void *arg, void *unused, int kmflag)
67 {
68 dnode_t *dn = arg;
69 int i;
70
71 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
72 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
73 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
74 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
75
76 /*
77 * Every dbuf has a reference, and dropping a tracked reference is
78 * O(number of references), so don't track dn_holds.
79 */
80 refcount_create_untracked(&dn->dn_holds);
81 refcount_create(&dn->dn_tx_holds);
82 list_link_init(&dn->dn_link);
83
84 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
85 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
86 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
87 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
88 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
89 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
90 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
91
92 for (i = 0; i < TXG_SIZE; i++) {
93 list_link_init(&dn->dn_dirty_link[i]);
94 dn->dn_free_ranges[i] = NULL;
95 list_create(&dn->dn_dirty_records[i],
96 sizeof (dbuf_dirty_record_t),
97 offsetof(dbuf_dirty_record_t, dr_dirty_node));
98 }
99
100 dn->dn_allocated_txg = 0;
101 dn->dn_free_txg = 0;
102 dn->dn_assigned_txg = 0;
103 dn->dn_dirtyctx = 0;
104 dn->dn_dirtyctx_firstset = NULL;
105 dn->dn_bonus = NULL;
106 dn->dn_have_spill = B_FALSE;
107 dn->dn_zio = NULL;
108 dn->dn_oldused = 0;
109 dn->dn_oldflags = 0;
110 dn->dn_olduid = 0;
111 dn->dn_oldgid = 0;
112 dn->dn_newuid = 0;
113 dn->dn_newgid = 0;
114 dn->dn_id_flags = 0;
115
116 dn->dn_dbufs_count = 0;
117 dn->dn_unlisted_l0_blkid = 0;
118 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t),
119 offsetof(dmu_buf_impl_t, db_link));
120
121 dn->dn_moved = 0;
122 POINTER_INVALIDATE(&dn->dn_objset);
123 return (0);
124 }
125
126 /* ARGSUSED */
127 static void
dnode_dest(void * arg,void * unused)128 dnode_dest(void *arg, void *unused)
129 {
130 int i;
131 dnode_t *dn = arg;
132
133 rw_destroy(&dn->dn_struct_rwlock);
134 mutex_destroy(&dn->dn_mtx);
135 mutex_destroy(&dn->dn_dbufs_mtx);
136 cv_destroy(&dn->dn_notxholds);
137 refcount_destroy(&dn->dn_holds);
138 refcount_destroy(&dn->dn_tx_holds);
139 ASSERT(!list_link_active(&dn->dn_link));
140
141 for (i = 0; i < TXG_SIZE; i++) {
142 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
143 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
144 list_destroy(&dn->dn_dirty_records[i]);
145 ASSERT0(dn->dn_next_nblkptr[i]);
146 ASSERT0(dn->dn_next_nlevels[i]);
147 ASSERT0(dn->dn_next_indblkshift[i]);
148 ASSERT0(dn->dn_next_bonustype[i]);
149 ASSERT0(dn->dn_rm_spillblk[i]);
150 ASSERT0(dn->dn_next_bonuslen[i]);
151 ASSERT0(dn->dn_next_blksz[i]);
152 }
153
154 ASSERT0(dn->dn_allocated_txg);
155 ASSERT0(dn->dn_free_txg);
156 ASSERT0(dn->dn_assigned_txg);
157 ASSERT0(dn->dn_dirtyctx);
158 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
159 ASSERT3P(dn->dn_bonus, ==, NULL);
160 ASSERT(!dn->dn_have_spill);
161 ASSERT3P(dn->dn_zio, ==, NULL);
162 ASSERT0(dn->dn_oldused);
163 ASSERT0(dn->dn_oldflags);
164 ASSERT0(dn->dn_olduid);
165 ASSERT0(dn->dn_oldgid);
166 ASSERT0(dn->dn_newuid);
167 ASSERT0(dn->dn_newgid);
168 ASSERT0(dn->dn_id_flags);
169
170 ASSERT0(dn->dn_dbufs_count);
171 ASSERT0(dn->dn_unlisted_l0_blkid);
172 list_destroy(&dn->dn_dbufs);
173 }
174
175 void
dnode_init(void)176 dnode_init(void)
177 {
178 ASSERT(dnode_cache == NULL);
179 dnode_cache = kmem_cache_create("dnode_t",
180 sizeof (dnode_t),
181 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
182 kmem_cache_set_move(dnode_cache, dnode_move);
183 }
184
185 void
dnode_fini(void)186 dnode_fini(void)
187 {
188 kmem_cache_destroy(dnode_cache);
189 dnode_cache = NULL;
190 }
191
192
193 #ifdef ZFS_DEBUG
194 void
dnode_verify(dnode_t * dn)195 dnode_verify(dnode_t *dn)
196 {
197 int drop_struct_lock = FALSE;
198
199 ASSERT(dn->dn_phys);
200 ASSERT(dn->dn_objset);
201 ASSERT(dn->dn_handle->dnh_dnode == dn);
202
203 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
204
205 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
206 return;
207
208 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
209 rw_enter(&dn->dn_struct_rwlock, RW_READER);
210 drop_struct_lock = TRUE;
211 }
212 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
213 int i;
214 ASSERT3U(dn->dn_indblkshift, >=, 0);
215 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
216 if (dn->dn_datablkshift) {
217 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
218 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
219 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
220 }
221 ASSERT3U(dn->dn_nlevels, <=, 30);
222 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
223 ASSERT3U(dn->dn_nblkptr, >=, 1);
224 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
225 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
226 ASSERT3U(dn->dn_datablksz, ==,
227 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
228 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
229 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
230 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
231 for (i = 0; i < TXG_SIZE; i++) {
232 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
233 }
234 }
235 if (dn->dn_phys->dn_type != DMU_OT_NONE)
236 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
237 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
238 if (dn->dn_dbuf != NULL) {
239 ASSERT3P(dn->dn_phys, ==,
240 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
241 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
242 }
243 if (drop_struct_lock)
244 rw_exit(&dn->dn_struct_rwlock);
245 }
246 #endif
247
248 void
dnode_byteswap(dnode_phys_t * dnp)249 dnode_byteswap(dnode_phys_t *dnp)
250 {
251 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
252 int i;
253
254 if (dnp->dn_type == DMU_OT_NONE) {
255 bzero(dnp, sizeof (dnode_phys_t));
256 return;
257 }
258
259 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
260 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
261 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
262 dnp->dn_used = BSWAP_64(dnp->dn_used);
263
264 /*
265 * dn_nblkptr is only one byte, so it's OK to read it in either
266 * byte order. We can't read dn_bouslen.
267 */
268 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
269 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
270 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
271 buf64[i] = BSWAP_64(buf64[i]);
272
273 /*
274 * OK to check dn_bonuslen for zero, because it won't matter if
275 * we have the wrong byte order. This is necessary because the
276 * dnode dnode is smaller than a regular dnode.
277 */
278 if (dnp->dn_bonuslen != 0) {
279 /*
280 * Note that the bonus length calculated here may be
281 * longer than the actual bonus buffer. This is because
282 * we always put the bonus buffer after the last block
283 * pointer (instead of packing it against the end of the
284 * dnode buffer).
285 */
286 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
287 size_t len = DN_MAX_BONUSLEN - off;
288 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
289 dmu_object_byteswap_t byteswap =
290 DMU_OT_BYTESWAP(dnp->dn_bonustype);
291 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
292 }
293
294 /* Swap SPILL block if we have one */
295 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
296 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
297
298 }
299
300 void
dnode_buf_byteswap(void * vbuf,size_t size)301 dnode_buf_byteswap(void *vbuf, size_t size)
302 {
303 dnode_phys_t *buf = vbuf;
304 int i;
305
306 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
307 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
308
309 size >>= DNODE_SHIFT;
310 for (i = 0; i < size; i++) {
311 dnode_byteswap(buf);
312 buf++;
313 }
314 }
315
316 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)317 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
318 {
319 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
320
321 dnode_setdirty(dn, tx);
322 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
323 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
324 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
325 dn->dn_bonuslen = newsize;
326 if (newsize == 0)
327 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
328 else
329 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
330 rw_exit(&dn->dn_struct_rwlock);
331 }
332
333 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)334 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
335 {
336 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
337 dnode_setdirty(dn, tx);
338 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
339 dn->dn_bonustype = newtype;
340 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
341 rw_exit(&dn->dn_struct_rwlock);
342 }
343
344 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)345 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
346 {
347 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
348 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
349 dnode_setdirty(dn, tx);
350 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
351 dn->dn_have_spill = B_FALSE;
352 }
353
354 static void
dnode_setdblksz(dnode_t * dn,int size)355 dnode_setdblksz(dnode_t *dn, int size)
356 {
357 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
358 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
359 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
360 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
361 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
362 dn->dn_datablksz = size;
363 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
364 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
365 }
366
367 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)368 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
369 uint64_t object, dnode_handle_t *dnh)
370 {
371 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
372
373 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
374 dn->dn_moved = 0;
375
376 /*
377 * Defer setting dn_objset until the dnode is ready to be a candidate
378 * for the dnode_move() callback.
379 */
380 dn->dn_object = object;
381 dn->dn_dbuf = db;
382 dn->dn_handle = dnh;
383 dn->dn_phys = dnp;
384
385 if (dnp->dn_datablkszsec) {
386 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
387 } else {
388 dn->dn_datablksz = 0;
389 dn->dn_datablkszsec = 0;
390 dn->dn_datablkshift = 0;
391 }
392 dn->dn_indblkshift = dnp->dn_indblkshift;
393 dn->dn_nlevels = dnp->dn_nlevels;
394 dn->dn_type = dnp->dn_type;
395 dn->dn_nblkptr = dnp->dn_nblkptr;
396 dn->dn_checksum = dnp->dn_checksum;
397 dn->dn_compress = dnp->dn_compress;
398 dn->dn_bonustype = dnp->dn_bonustype;
399 dn->dn_bonuslen = dnp->dn_bonuslen;
400 dn->dn_maxblkid = dnp->dn_maxblkid;
401 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
402 dn->dn_id_flags = 0;
403
404 dmu_zfetch_init(&dn->dn_zfetch, dn);
405
406 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
407
408 mutex_enter(&os->os_lock);
409 list_insert_head(&os->os_dnodes, dn);
410 membar_producer();
411 /*
412 * Everything else must be valid before assigning dn_objset makes the
413 * dnode eligible for dnode_move().
414 */
415 dn->dn_objset = os;
416 mutex_exit(&os->os_lock);
417
418 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
419 return (dn);
420 }
421
422 /*
423 * Caller must be holding the dnode handle, which is released upon return.
424 */
425 static void
dnode_destroy(dnode_t * dn)426 dnode_destroy(dnode_t *dn)
427 {
428 objset_t *os = dn->dn_objset;
429
430 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
431
432 mutex_enter(&os->os_lock);
433 POINTER_INVALIDATE(&dn->dn_objset);
434 list_remove(&os->os_dnodes, dn);
435 mutex_exit(&os->os_lock);
436
437 /* the dnode can no longer move, so we can release the handle */
438 zrl_remove(&dn->dn_handle->dnh_zrlock);
439
440 dn->dn_allocated_txg = 0;
441 dn->dn_free_txg = 0;
442 dn->dn_assigned_txg = 0;
443
444 dn->dn_dirtyctx = 0;
445 if (dn->dn_dirtyctx_firstset != NULL) {
446 kmem_free(dn->dn_dirtyctx_firstset, 1);
447 dn->dn_dirtyctx_firstset = NULL;
448 }
449 if (dn->dn_bonus != NULL) {
450 mutex_enter(&dn->dn_bonus->db_mtx);
451 dbuf_evict(dn->dn_bonus);
452 dn->dn_bonus = NULL;
453 }
454 dn->dn_zio = NULL;
455
456 dn->dn_have_spill = B_FALSE;
457 dn->dn_oldused = 0;
458 dn->dn_oldflags = 0;
459 dn->dn_olduid = 0;
460 dn->dn_oldgid = 0;
461 dn->dn_newuid = 0;
462 dn->dn_newgid = 0;
463 dn->dn_id_flags = 0;
464 dn->dn_unlisted_l0_blkid = 0;
465
466 dmu_zfetch_rele(&dn->dn_zfetch);
467 kmem_cache_free(dnode_cache, dn);
468 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
469 }
470
471 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)472 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
473 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
474 {
475 int i;
476
477 if (blocksize == 0)
478 blocksize = 1 << zfs_default_bs;
479 else if (blocksize > SPA_MAXBLOCKSIZE)
480 blocksize = SPA_MAXBLOCKSIZE;
481 else
482 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
483
484 if (ibs == 0)
485 ibs = zfs_default_ibs;
486
487 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
488
489 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
490 dn->dn_object, tx->tx_txg, blocksize, ibs);
491
492 ASSERT(dn->dn_type == DMU_OT_NONE);
493 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
494 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
495 ASSERT(ot != DMU_OT_NONE);
496 ASSERT(DMU_OT_IS_VALID(ot));
497 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
498 (bonustype == DMU_OT_SA && bonuslen == 0) ||
499 (bonustype != DMU_OT_NONE && bonuslen != 0));
500 ASSERT(DMU_OT_IS_VALID(bonustype));
501 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
502 ASSERT(dn->dn_type == DMU_OT_NONE);
503 ASSERT0(dn->dn_maxblkid);
504 ASSERT0(dn->dn_allocated_txg);
505 ASSERT0(dn->dn_assigned_txg);
506 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
507 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
508 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL);
509
510 for (i = 0; i < TXG_SIZE; i++) {
511 ASSERT0(dn->dn_next_nblkptr[i]);
512 ASSERT0(dn->dn_next_nlevels[i]);
513 ASSERT0(dn->dn_next_indblkshift[i]);
514 ASSERT0(dn->dn_next_bonuslen[i]);
515 ASSERT0(dn->dn_next_bonustype[i]);
516 ASSERT0(dn->dn_rm_spillblk[i]);
517 ASSERT0(dn->dn_next_blksz[i]);
518 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
519 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
520 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
521 }
522
523 dn->dn_type = ot;
524 dnode_setdblksz(dn, blocksize);
525 dn->dn_indblkshift = ibs;
526 dn->dn_nlevels = 1;
527 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
528 dn->dn_nblkptr = 1;
529 else
530 dn->dn_nblkptr = 1 +
531 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
532 dn->dn_bonustype = bonustype;
533 dn->dn_bonuslen = bonuslen;
534 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
535 dn->dn_compress = ZIO_COMPRESS_INHERIT;
536 dn->dn_dirtyctx = 0;
537
538 dn->dn_free_txg = 0;
539 if (dn->dn_dirtyctx_firstset) {
540 kmem_free(dn->dn_dirtyctx_firstset, 1);
541 dn->dn_dirtyctx_firstset = NULL;
542 }
543
544 dn->dn_allocated_txg = tx->tx_txg;
545 dn->dn_id_flags = 0;
546
547 dnode_setdirty(dn, tx);
548 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
549 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
550 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
551 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
552 }
553
554 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)555 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
556 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
557 {
558 int nblkptr;
559
560 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
561 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE);
562 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
563 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
564 ASSERT(tx->tx_txg != 0);
565 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
566 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
567 (bonustype == DMU_OT_SA && bonuslen == 0));
568 ASSERT(DMU_OT_IS_VALID(bonustype));
569 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
570
571 /* clean up any unreferenced dbufs */
572 dnode_evict_dbufs(dn);
573
574 dn->dn_id_flags = 0;
575
576 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
577 dnode_setdirty(dn, tx);
578 if (dn->dn_datablksz != blocksize) {
579 /* change blocksize */
580 ASSERT(dn->dn_maxblkid == 0 &&
581 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
582 dnode_block_freed(dn, 0)));
583 dnode_setdblksz(dn, blocksize);
584 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
585 }
586 if (dn->dn_bonuslen != bonuslen)
587 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
588
589 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
590 nblkptr = 1;
591 else
592 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
593 if (dn->dn_bonustype != bonustype)
594 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
595 if (dn->dn_nblkptr != nblkptr)
596 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
597 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
598 dbuf_rm_spill(dn, tx);
599 dnode_rm_spill(dn, tx);
600 }
601 rw_exit(&dn->dn_struct_rwlock);
602
603 /* change type */
604 dn->dn_type = ot;
605
606 /* change bonus size and type */
607 mutex_enter(&dn->dn_mtx);
608 dn->dn_bonustype = bonustype;
609 dn->dn_bonuslen = bonuslen;
610 dn->dn_nblkptr = nblkptr;
611 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
612 dn->dn_compress = ZIO_COMPRESS_INHERIT;
613 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
614
615 /* fix up the bonus db_size */
616 if (dn->dn_bonus) {
617 dn->dn_bonus->db.db_size =
618 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
619 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
620 }
621
622 dn->dn_allocated_txg = tx->tx_txg;
623 mutex_exit(&dn->dn_mtx);
624 }
625
626 #ifdef DNODE_STATS
627 static struct {
628 uint64_t dms_dnode_invalid;
629 uint64_t dms_dnode_recheck1;
630 uint64_t dms_dnode_recheck2;
631 uint64_t dms_dnode_special;
632 uint64_t dms_dnode_handle;
633 uint64_t dms_dnode_rwlock;
634 uint64_t dms_dnode_active;
635 } dnode_move_stats;
636 #endif /* DNODE_STATS */
637
638 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)639 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
640 {
641 int i;
642
643 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
644 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
645 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
646 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
647
648 /* Copy fields. */
649 ndn->dn_objset = odn->dn_objset;
650 ndn->dn_object = odn->dn_object;
651 ndn->dn_dbuf = odn->dn_dbuf;
652 ndn->dn_handle = odn->dn_handle;
653 ndn->dn_phys = odn->dn_phys;
654 ndn->dn_type = odn->dn_type;
655 ndn->dn_bonuslen = odn->dn_bonuslen;
656 ndn->dn_bonustype = odn->dn_bonustype;
657 ndn->dn_nblkptr = odn->dn_nblkptr;
658 ndn->dn_checksum = odn->dn_checksum;
659 ndn->dn_compress = odn->dn_compress;
660 ndn->dn_nlevels = odn->dn_nlevels;
661 ndn->dn_indblkshift = odn->dn_indblkshift;
662 ndn->dn_datablkshift = odn->dn_datablkshift;
663 ndn->dn_datablkszsec = odn->dn_datablkszsec;
664 ndn->dn_datablksz = odn->dn_datablksz;
665 ndn->dn_maxblkid = odn->dn_maxblkid;
666 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
667 sizeof (odn->dn_next_nblkptr));
668 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
669 sizeof (odn->dn_next_nlevels));
670 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
671 sizeof (odn->dn_next_indblkshift));
672 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
673 sizeof (odn->dn_next_bonustype));
674 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
675 sizeof (odn->dn_rm_spillblk));
676 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
677 sizeof (odn->dn_next_bonuslen));
678 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
679 sizeof (odn->dn_next_blksz));
680 for (i = 0; i < TXG_SIZE; i++) {
681 list_move_tail(&ndn->dn_dirty_records[i],
682 &odn->dn_dirty_records[i]);
683 }
684 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
685 sizeof (odn->dn_free_ranges));
686 ndn->dn_allocated_txg = odn->dn_allocated_txg;
687 ndn->dn_free_txg = odn->dn_free_txg;
688 ndn->dn_assigned_txg = odn->dn_assigned_txg;
689 ndn->dn_dirtyctx = odn->dn_dirtyctx;
690 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
691 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
692 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
693 ASSERT(list_is_empty(&ndn->dn_dbufs));
694 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs);
695 ndn->dn_dbufs_count = odn->dn_dbufs_count;
696 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
697 ndn->dn_bonus = odn->dn_bonus;
698 ndn->dn_have_spill = odn->dn_have_spill;
699 ndn->dn_zio = odn->dn_zio;
700 ndn->dn_oldused = odn->dn_oldused;
701 ndn->dn_oldflags = odn->dn_oldflags;
702 ndn->dn_olduid = odn->dn_olduid;
703 ndn->dn_oldgid = odn->dn_oldgid;
704 ndn->dn_newuid = odn->dn_newuid;
705 ndn->dn_newgid = odn->dn_newgid;
706 ndn->dn_id_flags = odn->dn_id_flags;
707 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
708 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
709 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
710 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt;
711 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail;
712
713 /*
714 * Update back pointers. Updating the handle fixes the back pointer of
715 * every descendant dbuf as well as the bonus dbuf.
716 */
717 ASSERT(ndn->dn_handle->dnh_dnode == odn);
718 ndn->dn_handle->dnh_dnode = ndn;
719 if (ndn->dn_zfetch.zf_dnode == odn) {
720 ndn->dn_zfetch.zf_dnode = ndn;
721 }
722
723 /*
724 * Invalidate the original dnode by clearing all of its back pointers.
725 */
726 odn->dn_dbuf = NULL;
727 odn->dn_handle = NULL;
728 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t),
729 offsetof(dmu_buf_impl_t, db_link));
730 odn->dn_dbufs_count = 0;
731 odn->dn_unlisted_l0_blkid = 0;
732 odn->dn_bonus = NULL;
733 odn->dn_zfetch.zf_dnode = NULL;
734
735 /*
736 * Set the low bit of the objset pointer to ensure that dnode_move()
737 * recognizes the dnode as invalid in any subsequent callback.
738 */
739 POINTER_INVALIDATE(&odn->dn_objset);
740
741 /*
742 * Satisfy the destructor.
743 */
744 for (i = 0; i < TXG_SIZE; i++) {
745 list_create(&odn->dn_dirty_records[i],
746 sizeof (dbuf_dirty_record_t),
747 offsetof(dbuf_dirty_record_t, dr_dirty_node));
748 odn->dn_free_ranges[i] = NULL;
749 odn->dn_next_nlevels[i] = 0;
750 odn->dn_next_indblkshift[i] = 0;
751 odn->dn_next_bonustype[i] = 0;
752 odn->dn_rm_spillblk[i] = 0;
753 odn->dn_next_bonuslen[i] = 0;
754 odn->dn_next_blksz[i] = 0;
755 }
756 odn->dn_allocated_txg = 0;
757 odn->dn_free_txg = 0;
758 odn->dn_assigned_txg = 0;
759 odn->dn_dirtyctx = 0;
760 odn->dn_dirtyctx_firstset = NULL;
761 odn->dn_have_spill = B_FALSE;
762 odn->dn_zio = NULL;
763 odn->dn_oldused = 0;
764 odn->dn_oldflags = 0;
765 odn->dn_olduid = 0;
766 odn->dn_oldgid = 0;
767 odn->dn_newuid = 0;
768 odn->dn_newgid = 0;
769 odn->dn_id_flags = 0;
770
771 /*
772 * Mark the dnode.
773 */
774 ndn->dn_moved = 1;
775 odn->dn_moved = (uint8_t)-1;
776 }
777
778 #ifdef sun
779 #ifdef _KERNEL
780 /*ARGSUSED*/
781 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)782 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
783 {
784 dnode_t *odn = buf, *ndn = newbuf;
785 objset_t *os;
786 int64_t refcount;
787 uint32_t dbufs;
788
789 /*
790 * The dnode is on the objset's list of known dnodes if the objset
791 * pointer is valid. We set the low bit of the objset pointer when
792 * freeing the dnode to invalidate it, and the memory patterns written
793 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
794 * A newly created dnode sets the objset pointer last of all to indicate
795 * that the dnode is known and in a valid state to be moved by this
796 * function.
797 */
798 os = odn->dn_objset;
799 if (!POINTER_IS_VALID(os)) {
800 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
801 return (KMEM_CBRC_DONT_KNOW);
802 }
803
804 /*
805 * Ensure that the objset does not go away during the move.
806 */
807 rw_enter(&os_lock, RW_WRITER);
808 if (os != odn->dn_objset) {
809 rw_exit(&os_lock);
810 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
811 return (KMEM_CBRC_DONT_KNOW);
812 }
813
814 /*
815 * If the dnode is still valid, then so is the objset. We know that no
816 * valid objset can be freed while we hold os_lock, so we can safely
817 * ensure that the objset remains in use.
818 */
819 mutex_enter(&os->os_lock);
820
821 /*
822 * Recheck the objset pointer in case the dnode was removed just before
823 * acquiring the lock.
824 */
825 if (os != odn->dn_objset) {
826 mutex_exit(&os->os_lock);
827 rw_exit(&os_lock);
828 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
829 return (KMEM_CBRC_DONT_KNOW);
830 }
831
832 /*
833 * At this point we know that as long as we hold os->os_lock, the dnode
834 * cannot be freed and fields within the dnode can be safely accessed.
835 * The objset listing this dnode cannot go away as long as this dnode is
836 * on its list.
837 */
838 rw_exit(&os_lock);
839 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
840 mutex_exit(&os->os_lock);
841 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
842 return (KMEM_CBRC_NO);
843 }
844 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
845
846 /*
847 * Lock the dnode handle to prevent the dnode from obtaining any new
848 * holds. This also prevents the descendant dbufs and the bonus dbuf
849 * from accessing the dnode, so that we can discount their holds. The
850 * handle is safe to access because we know that while the dnode cannot
851 * go away, neither can its handle. Once we hold dnh_zrlock, we can
852 * safely move any dnode referenced only by dbufs.
853 */
854 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
855 mutex_exit(&os->os_lock);
856 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
857 return (KMEM_CBRC_LATER);
858 }
859
860 /*
861 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
862 * We need to guarantee that there is a hold for every dbuf in order to
863 * determine whether the dnode is actively referenced. Falsely matching
864 * a dbuf to an active hold would lead to an unsafe move. It's possible
865 * that a thread already having an active dnode hold is about to add a
866 * dbuf, and we can't compare hold and dbuf counts while the add is in
867 * progress.
868 */
869 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
870 zrl_exit(&odn->dn_handle->dnh_zrlock);
871 mutex_exit(&os->os_lock);
872 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
873 return (KMEM_CBRC_LATER);
874 }
875
876 /*
877 * A dbuf may be removed (evicted) without an active dnode hold. In that
878 * case, the dbuf count is decremented under the handle lock before the
879 * dbuf's hold is released. This order ensures that if we count the hold
880 * after the dbuf is removed but before its hold is released, we will
881 * treat the unmatched hold as active and exit safely. If we count the
882 * hold before the dbuf is removed, the hold is discounted, and the
883 * removal is blocked until the move completes.
884 */
885 refcount = refcount_count(&odn->dn_holds);
886 ASSERT(refcount >= 0);
887 dbufs = odn->dn_dbufs_count;
888
889 /* We can't have more dbufs than dnode holds. */
890 ASSERT3U(dbufs, <=, refcount);
891 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
892 uint32_t, dbufs);
893
894 if (refcount > dbufs) {
895 rw_exit(&odn->dn_struct_rwlock);
896 zrl_exit(&odn->dn_handle->dnh_zrlock);
897 mutex_exit(&os->os_lock);
898 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
899 return (KMEM_CBRC_LATER);
900 }
901
902 rw_exit(&odn->dn_struct_rwlock);
903
904 /*
905 * At this point we know that anyone with a hold on the dnode is not
906 * actively referencing it. The dnode is known and in a valid state to
907 * move. We're holding the locks needed to execute the critical section.
908 */
909 dnode_move_impl(odn, ndn);
910
911 list_link_replace(&odn->dn_link, &ndn->dn_link);
912 /* If the dnode was safe to move, the refcount cannot have changed. */
913 ASSERT(refcount == refcount_count(&ndn->dn_holds));
914 ASSERT(dbufs == ndn->dn_dbufs_count);
915 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
916 mutex_exit(&os->os_lock);
917
918 return (KMEM_CBRC_YES);
919 }
920 #endif /* _KERNEL */
921 #endif /* sun */
922
923 void
dnode_special_close(dnode_handle_t * dnh)924 dnode_special_close(dnode_handle_t *dnh)
925 {
926 dnode_t *dn = dnh->dnh_dnode;
927
928 /*
929 * Wait for final references to the dnode to clear. This can
930 * only happen if the arc is asyncronously evicting state that
931 * has a hold on this dnode while we are trying to evict this
932 * dnode.
933 */
934 while (refcount_count(&dn->dn_holds) > 0)
935 delay(1);
936 zrl_add(&dnh->dnh_zrlock);
937 dnode_destroy(dn); /* implicit zrl_remove() */
938 zrl_destroy(&dnh->dnh_zrlock);
939 dnh->dnh_dnode = NULL;
940 }
941
942 dnode_t *
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)943 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
944 dnode_handle_t *dnh)
945 {
946 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh);
947 dnh->dnh_dnode = dn;
948 zrl_init(&dnh->dnh_zrlock);
949 DNODE_VERIFY(dn);
950 return (dn);
951 }
952
953 static void
dnode_buf_pageout(dmu_buf_t * db,void * arg)954 dnode_buf_pageout(dmu_buf_t *db, void *arg)
955 {
956 dnode_children_t *children_dnodes = arg;
957 int i;
958 int epb = db->db_size >> DNODE_SHIFT;
959
960 ASSERT(epb == children_dnodes->dnc_count);
961
962 for (i = 0; i < epb; i++) {
963 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
964 dnode_t *dn;
965
966 /*
967 * The dnode handle lock guards against the dnode moving to
968 * another valid address, so there is no need here to guard
969 * against changes to or from NULL.
970 */
971 if (dnh->dnh_dnode == NULL) {
972 zrl_destroy(&dnh->dnh_zrlock);
973 continue;
974 }
975
976 zrl_add(&dnh->dnh_zrlock);
977 dn = dnh->dnh_dnode;
978 /*
979 * If there are holds on this dnode, then there should
980 * be holds on the dnode's containing dbuf as well; thus
981 * it wouldn't be eligible for eviction and this function
982 * would not have been called.
983 */
984 ASSERT(refcount_is_zero(&dn->dn_holds));
985 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
986
987 dnode_destroy(dn); /* implicit zrl_remove() */
988 zrl_destroy(&dnh->dnh_zrlock);
989 dnh->dnh_dnode = NULL;
990 }
991 kmem_free(children_dnodes, sizeof (dnode_children_t) +
992 (epb - 1) * sizeof (dnode_handle_t));
993 }
994
995 /*
996 * errors:
997 * EINVAL - invalid object number.
998 * EIO - i/o error.
999 * succeeds even for free dnodes.
1000 */
1001 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1002 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1003 void *tag, dnode_t **dnp)
1004 {
1005 int epb, idx, err;
1006 int drop_struct_lock = FALSE;
1007 int type;
1008 uint64_t blk;
1009 dnode_t *mdn, *dn;
1010 dmu_buf_impl_t *db;
1011 dnode_children_t *children_dnodes;
1012 dnode_handle_t *dnh;
1013
1014 /*
1015 * If you are holding the spa config lock as writer, you shouldn't
1016 * be asking the DMU to do *anything* unless it's the root pool
1017 * which may require us to read from the root filesystem while
1018 * holding some (not all) of the locks as writer.
1019 */
1020 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1021 (spa_is_root(os->os_spa) &&
1022 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1023
1024 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1025 dn = (object == DMU_USERUSED_OBJECT) ?
1026 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1027 if (dn == NULL)
1028 return (SET_ERROR(ENOENT));
1029 type = dn->dn_type;
1030 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1031 return (SET_ERROR(ENOENT));
1032 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1033 return (SET_ERROR(EEXIST));
1034 DNODE_VERIFY(dn);
1035 (void) refcount_add(&dn->dn_holds, tag);
1036 *dnp = dn;
1037 return (0);
1038 }
1039
1040 if (object == 0 || object >= DN_MAX_OBJECT)
1041 return (SET_ERROR(EINVAL));
1042
1043 mdn = DMU_META_DNODE(os);
1044 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1045
1046 DNODE_VERIFY(mdn);
1047
1048 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1049 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1050 drop_struct_lock = TRUE;
1051 }
1052
1053 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t));
1054
1055 db = dbuf_hold(mdn, blk, FTAG);
1056 if (drop_struct_lock)
1057 rw_exit(&mdn->dn_struct_rwlock);
1058 if (db == NULL)
1059 return (SET_ERROR(EIO));
1060 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1061 if (err) {
1062 dbuf_rele(db, FTAG);
1063 return (err);
1064 }
1065
1066 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1067 epb = db->db.db_size >> DNODE_SHIFT;
1068
1069 idx = object & (epb-1);
1070
1071 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1072 children_dnodes = dmu_buf_get_user(&db->db);
1073 if (children_dnodes == NULL) {
1074 int i;
1075 dnode_children_t *winner;
1076 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1077 (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP);
1078 children_dnodes->dnc_count = epb;
1079 dnh = &children_dnodes->dnc_children[0];
1080 for (i = 0; i < epb; i++) {
1081 zrl_init(&dnh[i].dnh_zrlock);
1082 dnh[i].dnh_dnode = NULL;
1083 }
1084 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL,
1085 dnode_buf_pageout)) {
1086
1087 for (i = 0; i < epb; i++) {
1088 zrl_destroy(&dnh[i].dnh_zrlock);
1089 }
1090
1091 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1092 (epb - 1) * sizeof (dnode_handle_t));
1093 children_dnodes = winner;
1094 }
1095 }
1096 ASSERT(children_dnodes->dnc_count == epb);
1097
1098 dnh = &children_dnodes->dnc_children[idx];
1099 zrl_add(&dnh->dnh_zrlock);
1100 if ((dn = dnh->dnh_dnode) == NULL) {
1101 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1102 dnode_t *winner;
1103
1104 dn = dnode_create(os, phys, db, object, dnh);
1105 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn);
1106 if (winner != NULL) {
1107 zrl_add(&dnh->dnh_zrlock);
1108 dnode_destroy(dn); /* implicit zrl_remove() */
1109 dn = winner;
1110 }
1111 }
1112
1113 mutex_enter(&dn->dn_mtx);
1114 type = dn->dn_type;
1115 if (dn->dn_free_txg ||
1116 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1117 ((flag & DNODE_MUST_BE_FREE) &&
1118 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1119 mutex_exit(&dn->dn_mtx);
1120 zrl_remove(&dnh->dnh_zrlock);
1121 dbuf_rele(db, FTAG);
1122 return (type == DMU_OT_NONE ? ENOENT : EEXIST);
1123 }
1124 mutex_exit(&dn->dn_mtx);
1125
1126 if (refcount_add(&dn->dn_holds, tag) == 1)
1127 dbuf_add_ref(db, dnh);
1128 /* Now we can rely on the hold to prevent the dnode from moving. */
1129 zrl_remove(&dnh->dnh_zrlock);
1130
1131 DNODE_VERIFY(dn);
1132 ASSERT3P(dn->dn_dbuf, ==, db);
1133 ASSERT3U(dn->dn_object, ==, object);
1134 dbuf_rele(db, FTAG);
1135
1136 *dnp = dn;
1137 return (0);
1138 }
1139
1140 /*
1141 * Return held dnode if the object is allocated, NULL if not.
1142 */
1143 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1144 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1145 {
1146 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1147 }
1148
1149 /*
1150 * Can only add a reference if there is already at least one
1151 * reference on the dnode. Returns FALSE if unable to add a
1152 * new reference.
1153 */
1154 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1155 dnode_add_ref(dnode_t *dn, void *tag)
1156 {
1157 mutex_enter(&dn->dn_mtx);
1158 if (refcount_is_zero(&dn->dn_holds)) {
1159 mutex_exit(&dn->dn_mtx);
1160 return (FALSE);
1161 }
1162 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1163 mutex_exit(&dn->dn_mtx);
1164 return (TRUE);
1165 }
1166
1167 void
dnode_rele(dnode_t * dn,void * tag)1168 dnode_rele(dnode_t *dn, void *tag)
1169 {
1170 uint64_t refs;
1171 /* Get while the hold prevents the dnode from moving. */
1172 dmu_buf_impl_t *db = dn->dn_dbuf;
1173 dnode_handle_t *dnh = dn->dn_handle;
1174
1175 mutex_enter(&dn->dn_mtx);
1176 refs = refcount_remove(&dn->dn_holds, tag);
1177 mutex_exit(&dn->dn_mtx);
1178
1179 /*
1180 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1181 * indirectly by dbuf_rele() while relying on the dnode handle to
1182 * prevent the dnode from moving, since releasing the last hold could
1183 * result in the dnode's parent dbuf evicting its dnode handles. For
1184 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1185 * other direct or indirect hold on the dnode must first drop the dnode
1186 * handle.
1187 */
1188 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1189
1190 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1191 if (refs == 0 && db != NULL) {
1192 /*
1193 * Another thread could add a hold to the dnode handle in
1194 * dnode_hold_impl() while holding the parent dbuf. Since the
1195 * hold on the parent dbuf prevents the handle from being
1196 * destroyed, the hold on the handle is OK. We can't yet assert
1197 * that the handle has zero references, but that will be
1198 * asserted anyway when the handle gets destroyed.
1199 */
1200 dbuf_rele(db, dnh);
1201 }
1202 }
1203
1204 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1205 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1206 {
1207 objset_t *os = dn->dn_objset;
1208 uint64_t txg = tx->tx_txg;
1209
1210 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1211 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1212 return;
1213 }
1214
1215 DNODE_VERIFY(dn);
1216
1217 #ifdef ZFS_DEBUG
1218 mutex_enter(&dn->dn_mtx);
1219 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1220 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1221 mutex_exit(&dn->dn_mtx);
1222 #endif
1223
1224 /*
1225 * Determine old uid/gid when necessary
1226 */
1227 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1228
1229 mutex_enter(&os->os_lock);
1230
1231 /*
1232 * If we are already marked dirty, we're done.
1233 */
1234 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1235 mutex_exit(&os->os_lock);
1236 return;
1237 }
1238
1239 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs));
1240 ASSERT(dn->dn_datablksz != 0);
1241 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1242 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1243 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1244
1245 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1246 dn->dn_object, txg);
1247
1248 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
1249 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
1250 } else {
1251 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
1252 }
1253
1254 mutex_exit(&os->os_lock);
1255
1256 /*
1257 * The dnode maintains a hold on its containing dbuf as
1258 * long as there are holds on it. Each instantiated child
1259 * dbuf maintains a hold on the dnode. When the last child
1260 * drops its hold, the dnode will drop its hold on the
1261 * containing dbuf. We add a "dirty hold" here so that the
1262 * dnode will hang around after we finish processing its
1263 * children.
1264 */
1265 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1266
1267 (void) dbuf_dirty(dn->dn_dbuf, tx);
1268
1269 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1270 }
1271
1272 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1273 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1274 {
1275 int txgoff = tx->tx_txg & TXG_MASK;
1276
1277 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
1278
1279 /* we should be the only holder... hopefully */
1280 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
1281
1282 mutex_enter(&dn->dn_mtx);
1283 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1284 mutex_exit(&dn->dn_mtx);
1285 return;
1286 }
1287 dn->dn_free_txg = tx->tx_txg;
1288 mutex_exit(&dn->dn_mtx);
1289
1290 /*
1291 * If the dnode is already dirty, it needs to be moved from
1292 * the dirty list to the free list.
1293 */
1294 mutex_enter(&dn->dn_objset->os_lock);
1295 if (list_link_active(&dn->dn_dirty_link[txgoff])) {
1296 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
1297 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
1298 mutex_exit(&dn->dn_objset->os_lock);
1299 } else {
1300 mutex_exit(&dn->dn_objset->os_lock);
1301 dnode_setdirty(dn, tx);
1302 }
1303 }
1304
1305 /*
1306 * Try to change the block size for the indicated dnode. This can only
1307 * succeed if there are no blocks allocated or dirty beyond first block
1308 */
1309 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1310 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1311 {
1312 dmu_buf_impl_t *db, *db_next;
1313 int err;
1314
1315 if (size == 0)
1316 size = SPA_MINBLOCKSIZE;
1317 if (size > SPA_MAXBLOCKSIZE)
1318 size = SPA_MAXBLOCKSIZE;
1319 else
1320 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1321
1322 if (ibs == dn->dn_indblkshift)
1323 ibs = 0;
1324
1325 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1326 return (0);
1327
1328 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1329
1330 /* Check for any allocated blocks beyond the first */
1331 if (dn->dn_maxblkid != 0)
1332 goto fail;
1333
1334 mutex_enter(&dn->dn_dbufs_mtx);
1335 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
1336 db_next = list_next(&dn->dn_dbufs, db);
1337
1338 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1339 db->db_blkid != DMU_SPILL_BLKID) {
1340 mutex_exit(&dn->dn_dbufs_mtx);
1341 goto fail;
1342 }
1343 }
1344 mutex_exit(&dn->dn_dbufs_mtx);
1345
1346 if (ibs && dn->dn_nlevels != 1)
1347 goto fail;
1348
1349 /* resize the old block */
1350 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db);
1351 if (err == 0)
1352 dbuf_new_size(db, size, tx);
1353 else if (err != ENOENT)
1354 goto fail;
1355
1356 dnode_setdblksz(dn, size);
1357 dnode_setdirty(dn, tx);
1358 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1359 if (ibs) {
1360 dn->dn_indblkshift = ibs;
1361 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1362 }
1363 /* rele after we have fixed the blocksize in the dnode */
1364 if (db)
1365 dbuf_rele(db, FTAG);
1366
1367 rw_exit(&dn->dn_struct_rwlock);
1368 return (0);
1369
1370 fail:
1371 rw_exit(&dn->dn_struct_rwlock);
1372 return (SET_ERROR(ENOTSUP));
1373 }
1374
1375 /* read-holding callers must not rely on the lock being continuously held */
1376 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1377 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1378 {
1379 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1380 int epbs, new_nlevels;
1381 uint64_t sz;
1382
1383 ASSERT(blkid != DMU_BONUS_BLKID);
1384
1385 ASSERT(have_read ?
1386 RW_READ_HELD(&dn->dn_struct_rwlock) :
1387 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1388
1389 /*
1390 * if we have a read-lock, check to see if we need to do any work
1391 * before upgrading to a write-lock.
1392 */
1393 if (have_read) {
1394 if (blkid <= dn->dn_maxblkid)
1395 return;
1396
1397 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1398 rw_exit(&dn->dn_struct_rwlock);
1399 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1400 }
1401 }
1402
1403 if (blkid <= dn->dn_maxblkid)
1404 goto out;
1405
1406 dn->dn_maxblkid = blkid;
1407
1408 /*
1409 * Compute the number of levels necessary to support the new maxblkid.
1410 */
1411 new_nlevels = 1;
1412 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1413 for (sz = dn->dn_nblkptr;
1414 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1415 new_nlevels++;
1416
1417 if (new_nlevels > dn->dn_nlevels) {
1418 int old_nlevels = dn->dn_nlevels;
1419 dmu_buf_impl_t *db;
1420 list_t *list;
1421 dbuf_dirty_record_t *new, *dr, *dr_next;
1422
1423 dn->dn_nlevels = new_nlevels;
1424
1425 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1426 dn->dn_next_nlevels[txgoff] = new_nlevels;
1427
1428 /* dirty the left indirects */
1429 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1430 ASSERT(db != NULL);
1431 new = dbuf_dirty(db, tx);
1432 dbuf_rele(db, FTAG);
1433
1434 /* transfer the dirty records to the new indirect */
1435 mutex_enter(&dn->dn_mtx);
1436 mutex_enter(&new->dt.di.dr_mtx);
1437 list = &dn->dn_dirty_records[txgoff];
1438 for (dr = list_head(list); dr; dr = dr_next) {
1439 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1440 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1441 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1442 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1443 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1444 list_remove(&dn->dn_dirty_records[txgoff], dr);
1445 list_insert_tail(&new->dt.di.dr_children, dr);
1446 dr->dr_parent = new;
1447 }
1448 }
1449 mutex_exit(&new->dt.di.dr_mtx);
1450 mutex_exit(&dn->dn_mtx);
1451 }
1452
1453 out:
1454 if (have_read)
1455 rw_downgrade(&dn->dn_struct_rwlock);
1456 }
1457
1458 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1459 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1460 {
1461 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1462 if (db != NULL) {
1463 dmu_buf_will_dirty(&db->db, tx);
1464 dbuf_rele(db, FTAG);
1465 }
1466 }
1467
1468 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1469 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1470 {
1471 dmu_buf_impl_t *db;
1472 uint64_t blkoff, blkid, nblks;
1473 int blksz, blkshift, head, tail;
1474 int trunc = FALSE;
1475 int epbs;
1476
1477 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1478 blksz = dn->dn_datablksz;
1479 blkshift = dn->dn_datablkshift;
1480 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1481
1482 if (len == DMU_OBJECT_END) {
1483 len = UINT64_MAX - off;
1484 trunc = TRUE;
1485 }
1486
1487 /*
1488 * First, block align the region to free:
1489 */
1490 if (ISP2(blksz)) {
1491 head = P2NPHASE(off, blksz);
1492 blkoff = P2PHASE(off, blksz);
1493 if ((off >> blkshift) > dn->dn_maxblkid)
1494 goto out;
1495 } else {
1496 ASSERT(dn->dn_maxblkid == 0);
1497 if (off == 0 && len >= blksz) {
1498 /*
1499 * Freeing the whole block; fast-track this request.
1500 * Note that we won't dirty any indirect blocks,
1501 * which is fine because we will be freeing the entire
1502 * file and thus all indirect blocks will be freed
1503 * by free_children().
1504 */
1505 blkid = 0;
1506 nblks = 1;
1507 goto done;
1508 } else if (off >= blksz) {
1509 /* Freeing past end-of-data */
1510 goto out;
1511 } else {
1512 /* Freeing part of the block. */
1513 head = blksz - off;
1514 ASSERT3U(head, >, 0);
1515 }
1516 blkoff = off;
1517 }
1518 /* zero out any partial block data at the start of the range */
1519 if (head) {
1520 ASSERT3U(blkoff + head, ==, blksz);
1521 if (len < head)
1522 head = len;
1523 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE,
1524 FTAG, &db) == 0) {
1525 caddr_t data;
1526
1527 /* don't dirty if it isn't on disk and isn't dirty */
1528 if (db->db_last_dirty ||
1529 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1530 rw_exit(&dn->dn_struct_rwlock);
1531 dmu_buf_will_dirty(&db->db, tx);
1532 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1533 data = db->db.db_data;
1534 bzero(data + blkoff, head);
1535 }
1536 dbuf_rele(db, FTAG);
1537 }
1538 off += head;
1539 len -= head;
1540 }
1541
1542 /* If the range was less than one block, we're done */
1543 if (len == 0)
1544 goto out;
1545
1546 /* If the remaining range is past end of file, we're done */
1547 if ((off >> blkshift) > dn->dn_maxblkid)
1548 goto out;
1549
1550 ASSERT(ISP2(blksz));
1551 if (trunc)
1552 tail = 0;
1553 else
1554 tail = P2PHASE(len, blksz);
1555
1556 ASSERT0(P2PHASE(off, blksz));
1557 /* zero out any partial block data at the end of the range */
1558 if (tail) {
1559 if (len < tail)
1560 tail = len;
1561 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len),
1562 TRUE, FTAG, &db) == 0) {
1563 /* don't dirty if not on disk and not dirty */
1564 if (db->db_last_dirty ||
1565 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1566 rw_exit(&dn->dn_struct_rwlock);
1567 dmu_buf_will_dirty(&db->db, tx);
1568 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1569 bzero(db->db.db_data, tail);
1570 }
1571 dbuf_rele(db, FTAG);
1572 }
1573 len -= tail;
1574 }
1575
1576 /* If the range did not include a full block, we are done */
1577 if (len == 0)
1578 goto out;
1579
1580 ASSERT(IS_P2ALIGNED(off, blksz));
1581 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1582 blkid = off >> blkshift;
1583 nblks = len >> blkshift;
1584 if (trunc)
1585 nblks += 1;
1586
1587 /*
1588 * Dirty all the indirect blocks in this range. Note that only
1589 * the first and last indirect blocks can actually be written
1590 * (if they were partially freed) -- they must be dirtied, even if
1591 * they do not exist on disk yet. The interior blocks will
1592 * be freed by free_children(), so they will not actually be written.
1593 * Even though these interior blocks will not be written, we
1594 * dirty them for two reasons:
1595 *
1596 * - It ensures that the indirect blocks remain in memory until
1597 * syncing context. (They have already been prefetched by
1598 * dmu_tx_hold_free(), so we don't have to worry about reading
1599 * them serially here.)
1600 *
1601 * - The dirty space accounting will put pressure on the txg sync
1602 * mechanism to begin syncing, and to delay transactions if there
1603 * is a large amount of freeing. Even though these indirect
1604 * blocks will not be written, we could need to write the same
1605 * amount of space if we copy the freed BPs into deadlists.
1606 */
1607 if (dn->dn_nlevels > 1) {
1608 uint64_t first, last;
1609
1610 first = blkid >> epbs;
1611 dnode_dirty_l1(dn, first, tx);
1612 if (trunc)
1613 last = dn->dn_maxblkid >> epbs;
1614 else
1615 last = (blkid + nblks - 1) >> epbs;
1616 if (last != first)
1617 dnode_dirty_l1(dn, last, tx);
1618
1619 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1620 SPA_BLKPTRSHIFT;
1621 for (uint64_t i = first + 1; i < last; i++) {
1622 /*
1623 * Set i to the blockid of the next non-hole
1624 * level-1 indirect block at or after i. Note
1625 * that dnode_next_offset() operates in terms of
1626 * level-0-equivalent bytes.
1627 */
1628 uint64_t ibyte = i << shift;
1629 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1630 &ibyte, 2, 1, 0);
1631 i = ibyte >> shift;
1632 if (i >= last)
1633 break;
1634
1635 /*
1636 * Normally we should not see an error, either
1637 * from dnode_next_offset() or dbuf_hold_level()
1638 * (except for ESRCH from dnode_next_offset).
1639 * If there is an i/o error, then when we read
1640 * this block in syncing context, it will use
1641 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1642 * to the "failmode" property. dnode_next_offset()
1643 * doesn't have a flag to indicate MUSTSUCCEED.
1644 */
1645 if (err != 0)
1646 break;
1647
1648 dnode_dirty_l1(dn, i, tx);
1649 }
1650 }
1651
1652 done:
1653 /*
1654 * Add this range to the dnode range list.
1655 * We will finish up this free operation in the syncing phase.
1656 */
1657 mutex_enter(&dn->dn_mtx);
1658 int txgoff = tx->tx_txg & TXG_MASK;
1659 if (dn->dn_free_ranges[txgoff] == NULL) {
1660 dn->dn_free_ranges[txgoff] =
1661 range_tree_create(NULL, NULL, &dn->dn_mtx);
1662 }
1663 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1664 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1665 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1666 blkid, nblks, tx->tx_txg);
1667 mutex_exit(&dn->dn_mtx);
1668
1669 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1670 dnode_setdirty(dn, tx);
1671 out:
1672
1673 rw_exit(&dn->dn_struct_rwlock);
1674 }
1675
1676 static boolean_t
dnode_spill_freed(dnode_t * dn)1677 dnode_spill_freed(dnode_t *dn)
1678 {
1679 int i;
1680
1681 mutex_enter(&dn->dn_mtx);
1682 for (i = 0; i < TXG_SIZE; i++) {
1683 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1684 break;
1685 }
1686 mutex_exit(&dn->dn_mtx);
1687 return (i < TXG_SIZE);
1688 }
1689
1690 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1691 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1692 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1693 {
1694 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1695 int i;
1696
1697 if (blkid == DMU_BONUS_BLKID)
1698 return (FALSE);
1699
1700 /*
1701 * If we're in the process of opening the pool, dp will not be
1702 * set yet, but there shouldn't be anything dirty.
1703 */
1704 if (dp == NULL)
1705 return (FALSE);
1706
1707 if (dn->dn_free_txg)
1708 return (TRUE);
1709
1710 if (blkid == DMU_SPILL_BLKID)
1711 return (dnode_spill_freed(dn));
1712
1713 mutex_enter(&dn->dn_mtx);
1714 for (i = 0; i < TXG_SIZE; i++) {
1715 if (dn->dn_free_ranges[i] != NULL &&
1716 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1717 break;
1718 }
1719 mutex_exit(&dn->dn_mtx);
1720 return (i < TXG_SIZE);
1721 }
1722
1723 /* call from syncing context when we actually write/free space for this dnode */
1724 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1725 dnode_diduse_space(dnode_t *dn, int64_t delta)
1726 {
1727 uint64_t space;
1728 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1729 dn, dn->dn_phys,
1730 (u_longlong_t)dn->dn_phys->dn_used,
1731 (longlong_t)delta);
1732
1733 mutex_enter(&dn->dn_mtx);
1734 space = DN_USED_BYTES(dn->dn_phys);
1735 if (delta > 0) {
1736 ASSERT3U(space + delta, >=, space); /* no overflow */
1737 } else {
1738 ASSERT3U(space, >=, -delta); /* no underflow */
1739 }
1740 space += delta;
1741 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1742 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1743 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1744 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1745 } else {
1746 dn->dn_phys->dn_used = space;
1747 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1748 }
1749 mutex_exit(&dn->dn_mtx);
1750 }
1751
1752 /*
1753 * Call when we think we're going to write/free space in open context to track
1754 * the amount of memory in use by the currently open txg.
1755 */
1756 void
dnode_willuse_space(dnode_t * dn,int64_t space,dmu_tx_t * tx)1757 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
1758 {
1759 objset_t *os = dn->dn_objset;
1760 dsl_dataset_t *ds = os->os_dsl_dataset;
1761 int64_t aspace = spa_get_asize(os->os_spa, space);
1762
1763 if (ds != NULL) {
1764 dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
1765 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
1766 }
1767
1768 dmu_tx_willuse_space(tx, aspace);
1769 }
1770
1771 /*
1772 * Scans a block at the indicated "level" looking for a hole or data,
1773 * depending on 'flags'.
1774 *
1775 * If level > 0, then we are scanning an indirect block looking at its
1776 * pointers. If level == 0, then we are looking at a block of dnodes.
1777 *
1778 * If we don't find what we are looking for in the block, we return ESRCH.
1779 * Otherwise, return with *offset pointing to the beginning (if searching
1780 * forwards) or end (if searching backwards) of the range covered by the
1781 * block pointer we matched on (or dnode).
1782 *
1783 * The basic search algorithm used below by dnode_next_offset() is to
1784 * use this function to search up the block tree (widen the search) until
1785 * we find something (i.e., we don't return ESRCH) and then search back
1786 * down the tree (narrow the search) until we reach our original search
1787 * level.
1788 */
1789 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1790 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1791 int lvl, uint64_t blkfill, uint64_t txg)
1792 {
1793 dmu_buf_impl_t *db = NULL;
1794 void *data = NULL;
1795 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1796 uint64_t epb = 1ULL << epbs;
1797 uint64_t minfill, maxfill;
1798 boolean_t hole;
1799 int i, inc, error, span;
1800
1801 dprintf("probing object %llu offset %llx level %d of %u\n",
1802 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1803
1804 hole = ((flags & DNODE_FIND_HOLE) != 0);
1805 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1806 ASSERT(txg == 0 || !hole);
1807
1808 if (lvl == dn->dn_phys->dn_nlevels) {
1809 error = 0;
1810 epb = dn->dn_phys->dn_nblkptr;
1811 data = dn->dn_phys->dn_blkptr;
1812 } else {
1813 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl);
1814 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db);
1815 if (error) {
1816 if (error != ENOENT)
1817 return (error);
1818 if (hole)
1819 return (0);
1820 /*
1821 * This can only happen when we are searching up
1822 * the block tree for data. We don't really need to
1823 * adjust the offset, as we will just end up looking
1824 * at the pointer to this block in its parent, and its
1825 * going to be unallocated, so we will skip over it.
1826 */
1827 return (SET_ERROR(ESRCH));
1828 }
1829 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1830 if (error) {
1831 dbuf_rele(db, FTAG);
1832 return (error);
1833 }
1834 data = db->db.db_data;
1835 }
1836
1837
1838 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1839 db->db_blkptr->blk_birth <= txg ||
1840 BP_IS_HOLE(db->db_blkptr))) {
1841 /*
1842 * This can only happen when we are searching up the tree
1843 * and these conditions mean that we need to keep climbing.
1844 */
1845 error = SET_ERROR(ESRCH);
1846 } else if (lvl == 0) {
1847 dnode_phys_t *dnp = data;
1848 span = DNODE_SHIFT;
1849 ASSERT(dn->dn_type == DMU_OT_DNODE);
1850
1851 for (i = (*offset >> span) & (blkfill - 1);
1852 i >= 0 && i < blkfill; i += inc) {
1853 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1854 break;
1855 *offset += (1ULL << span) * inc;
1856 }
1857 if (i < 0 || i == blkfill)
1858 error = SET_ERROR(ESRCH);
1859 } else {
1860 blkptr_t *bp = data;
1861 uint64_t start = *offset;
1862 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1863 minfill = 0;
1864 maxfill = blkfill << ((lvl - 1) * epbs);
1865
1866 if (hole)
1867 maxfill--;
1868 else
1869 minfill++;
1870
1871 *offset = *offset >> span;
1872 for (i = BF64_GET(*offset, 0, epbs);
1873 i >= 0 && i < epb; i += inc) {
1874 if (bp[i].blk_fill >= minfill &&
1875 bp[i].blk_fill <= maxfill &&
1876 (hole || bp[i].blk_birth > txg))
1877 break;
1878 if (inc > 0 || *offset > 0)
1879 *offset += inc;
1880 }
1881 *offset = *offset << span;
1882 if (inc < 0) {
1883 /* traversing backwards; position offset at the end */
1884 ASSERT3U(*offset, <=, start);
1885 *offset = MIN(*offset + (1ULL << span) - 1, start);
1886 } else if (*offset < start) {
1887 *offset = start;
1888 }
1889 if (i < 0 || i >= epb)
1890 error = SET_ERROR(ESRCH);
1891 }
1892
1893 if (db)
1894 dbuf_rele(db, FTAG);
1895
1896 return (error);
1897 }
1898
1899 /*
1900 * Find the next hole, data, or sparse region at or after *offset.
1901 * The value 'blkfill' tells us how many items we expect to find
1902 * in an L0 data block; this value is 1 for normal objects,
1903 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1904 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1905 *
1906 * Examples:
1907 *
1908 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1909 * Finds the next/previous hole/data in a file.
1910 * Used in dmu_offset_next().
1911 *
1912 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
1913 * Finds the next free/allocated dnode an objset's meta-dnode.
1914 * Only finds objects that have new contents since txg (ie.
1915 * bonus buffer changes and content removal are ignored).
1916 * Used in dmu_object_next().
1917 *
1918 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
1919 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
1920 * Used in dmu_object_alloc().
1921 */
1922 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)1923 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
1924 int minlvl, uint64_t blkfill, uint64_t txg)
1925 {
1926 uint64_t initial_offset = *offset;
1927 int lvl, maxlvl;
1928 int error = 0;
1929
1930 if (!(flags & DNODE_FIND_HAVELOCK))
1931 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1932
1933 if (dn->dn_phys->dn_nlevels == 0) {
1934 error = SET_ERROR(ESRCH);
1935 goto out;
1936 }
1937
1938 if (dn->dn_datablkshift == 0) {
1939 if (*offset < dn->dn_datablksz) {
1940 if (flags & DNODE_FIND_HOLE)
1941 *offset = dn->dn_datablksz;
1942 } else {
1943 error = SET_ERROR(ESRCH);
1944 }
1945 goto out;
1946 }
1947
1948 maxlvl = dn->dn_phys->dn_nlevels;
1949
1950 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
1951 error = dnode_next_offset_level(dn,
1952 flags, offset, lvl, blkfill, txg);
1953 if (error != ESRCH)
1954 break;
1955 }
1956
1957 while (error == 0 && --lvl >= minlvl) {
1958 error = dnode_next_offset_level(dn,
1959 flags, offset, lvl, blkfill, txg);
1960 }
1961
1962 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
1963 initial_offset < *offset : initial_offset > *offset))
1964 error = SET_ERROR(ESRCH);
1965 out:
1966 if (!(flags & DNODE_FIND_HAVELOCK))
1967 rw_exit(&dn->dn_struct_rwlock);
1968
1969 return (error);
1970 }
1971