1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* -*- mode: c; c-basic-offset: 8; -*-
3 * vim: noexpandtab sw=8 ts=8 sts=0:
4 *
5 * dcache.c
6 *
7 * dentry cache handling code
8 *
9 * Copyright (C) 2002, 2004 Oracle. All rights reserved.
10 */
11
12 #include <linux/fs.h>
13 #include <linux/types.h>
14 #include <linux/slab.h>
15 #include <linux/namei.h>
16
17 #include <cluster/masklog.h>
18
19 #include "ocfs2.h"
20
21 #include "alloc.h"
22 #include "dcache.h"
23 #include "dlmglue.h"
24 #include "file.h"
25 #include "inode.h"
26 #include "ocfs2_trace.h"
27
28 void ocfs2_dentry_attach_gen(struct dentry *dentry)
29 {
30 unsigned long gen =
31 OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
32 BUG_ON(d_inode(dentry));
33 dentry->d_fsdata = (void *)gen;
34 }
35
36
37 static int ocfs2_dentry_revalidate(struct dentry *dentry, unsigned int flags)
38 {
39 struct inode *inode;
40 int ret = 0; /* if all else fails, just return false */
41 struct ocfs2_super *osb;
42
43 if (flags & LOOKUP_RCU)
44 return -ECHILD;
45
46 inode = d_inode(dentry);
47 osb = OCFS2_SB(dentry->d_sb);
48
49 trace_ocfs2_dentry_revalidate(dentry, dentry->d_name.len,
50 dentry->d_name.name);
51
52 /* For a negative dentry -
53 * check the generation number of the parent and compare with the
54 * one stored in the inode.
55 */
56 if (inode == NULL) {
57 unsigned long gen = (unsigned long) dentry->d_fsdata;
58 unsigned long pgen;
59 spin_lock(&dentry->d_lock);
60 pgen = OCFS2_I(d_inode(dentry->d_parent))->ip_dir_lock_gen;
61 spin_unlock(&dentry->d_lock);
62 trace_ocfs2_dentry_revalidate_negative(dentry->d_name.len,
63 dentry->d_name.name,
64 pgen, gen);
65 if (gen != pgen)
66 goto bail;
67 goto valid;
68 }
69
70 BUG_ON(!osb);
71
72 if (inode == osb->root_inode || is_bad_inode(inode))
73 goto bail;
74
75 spin_lock(&OCFS2_I(inode)->ip_lock);
76 /* did we or someone else delete this inode? */
77 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
78 spin_unlock(&OCFS2_I(inode)->ip_lock);
79 trace_ocfs2_dentry_revalidate_delete(
80 (unsigned long long)OCFS2_I(inode)->ip_blkno);
81 goto bail;
82 }
83 spin_unlock(&OCFS2_I(inode)->ip_lock);
84
85 /*
86 * We don't need a cluster lock to test this because once an
87 * inode nlink hits zero, it never goes back.
88 */
89 if (inode->i_nlink == 0) {
90 trace_ocfs2_dentry_revalidate_orphaned(
91 (unsigned long long)OCFS2_I(inode)->ip_blkno,
92 S_ISDIR(inode->i_mode));
93 goto bail;
94 }
95
96 /*
97 * If the last lookup failed to create dentry lock, let us
98 * redo it.
99 */
100 if (!dentry->d_fsdata) {
101 trace_ocfs2_dentry_revalidate_nofsdata(
102 (unsigned long long)OCFS2_I(inode)->ip_blkno);
103 goto bail;
104 }
105
106 valid:
107 ret = 1;
108
109 bail:
110 trace_ocfs2_dentry_revalidate_ret(ret);
111 return ret;
112 }
113
114 static int ocfs2_match_dentry(struct dentry *dentry,
115 u64 parent_blkno,
116 int skip_unhashed)
117 {
118 struct inode *parent;
119
120 /*
121 * ocfs2_lookup() does a d_splice_alias() _before_ attaching
122 * to the lock data, so we skip those here, otherwise
123 * ocfs2_dentry_attach_lock() will get its original dentry
124 * back.
125 */
126 if (!dentry->d_fsdata)
127 return 0;
128
129 if (!dentry->d_parent)
130 return 0;
131
132 if (skip_unhashed && d_unhashed(dentry))
133 return 0;
134
135 parent = d_inode(dentry->d_parent);
136 /* Negative parent dentry? */
137 if (!parent)
138 return 0;
139
140 /* Name is in a different directory. */
141 if (OCFS2_I(parent)->ip_blkno != parent_blkno)
142 return 0;
143
144 return 1;
145 }
146
147 /*
148 * Walk the inode alias list, and find a dentry which has a given
149 * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it
150 * is looking for a dentry_lock reference. The downconvert thread is
151 * looking to unhash aliases, so we allow it to skip any that already
152 * have that property.
153 */
154 struct dentry *ocfs2_find_local_alias(struct inode *inode,
155 u64 parent_blkno,
156 int skip_unhashed)
157 {
158 struct dentry *dentry;
159
160 spin_lock(&inode->i_lock);
161 hlist_for_each_entry(dentry, &inode->i_dentry, d_u.d_alias) {
162 spin_lock(&dentry->d_lock);
163 if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) {
164 trace_ocfs2_find_local_alias(dentry->d_name.len,
165 dentry->d_name.name);
166
167 dget_dlock(dentry);
168 spin_unlock(&dentry->d_lock);
169 spin_unlock(&inode->i_lock);
170 return dentry;
171 }
172 spin_unlock(&dentry->d_lock);
173 }
174 spin_unlock(&inode->i_lock);
175 return NULL;
176 }
177
178 DEFINE_SPINLOCK(dentry_attach_lock);
179
180 /*
181 * Attach this dentry to a cluster lock.
182 *
183 * Dentry locks cover all links in a given directory to a particular
184 * inode. We do this so that ocfs2 can build a lock name which all
185 * nodes in the cluster can agree on at all times. Shoving full names
186 * in the cluster lock won't work due to size restrictions. Covering
187 * links inside of a directory is a good compromise because it still
188 * allows us to use the parent directory lock to synchronize
189 * operations.
190 *
191 * Call this function with the parent dir semaphore and the parent dir
192 * cluster lock held.
193 *
194 * The dir semaphore will protect us from having to worry about
195 * concurrent processes on our node trying to attach a lock at the
196 * same time.
197 *
198 * The dir cluster lock (held at either PR or EX mode) protects us
199 * from unlink and rename on other nodes.
200 *
201 * A dput() can happen asynchronously due to pruning, so we cover
202 * attaching and detaching the dentry lock with a
203 * dentry_attach_lock.
204 *
205 * A node which has done lookup on a name retains a protected read
206 * lock until final dput. If the user requests and unlink or rename,
207 * the protected read is upgraded to an exclusive lock. Other nodes
208 * who have seen the dentry will then be informed that they need to
209 * downgrade their lock, which will involve d_delete on the
210 * dentry. This happens in ocfs2_dentry_convert_worker().
211 */
212 int ocfs2_dentry_attach_lock(struct dentry *dentry,
213 struct inode *inode,
214 u64 parent_blkno)
215 {
216 int ret;
217 struct dentry *alias;
218 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
219
220 trace_ocfs2_dentry_attach_lock(dentry->d_name.len, dentry->d_name.name,
221 (unsigned long long)parent_blkno, dl);
222
223 /*
224 * Negative dentry. We ignore these for now.
225 *
226 * XXX: Could we can improve ocfs2_dentry_revalidate() by
227 * tracking these?
228 */
229 if (!inode)
230 return 0;
231
232 if (d_really_is_negative(dentry) && dentry->d_fsdata) {
233 /* Converting a negative dentry to positive
234 Clear dentry->d_fsdata */
235 dentry->d_fsdata = dl = NULL;
236 }
237
238 if (dl) {
239 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
240 " \"%pd\": old parent: %llu, new: %llu\n",
241 dentry,
242 (unsigned long long)parent_blkno,
243 (unsigned long long)dl->dl_parent_blkno);
244 return 0;
245 }
246
247 alias = ocfs2_find_local_alias(inode, parent_blkno, 0);
248 if (alias) {
249 /*
250 * Great, an alias exists, which means we must have a
251 * dentry lock already. We can just grab the lock off
252 * the alias and add it to the list.
253 *
254 * We're depending here on the fact that this dentry
255 * was found and exists in the dcache and so must have
256 * a reference to the dentry_lock because we can't
257 * race creates. Final dput() cannot happen on it
258 * since we have it pinned, so our reference is safe.
259 */
260 dl = alias->d_fsdata;
261 mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n",
262 (unsigned long long)parent_blkno,
263 (unsigned long long)OCFS2_I(inode)->ip_blkno);
264
265 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
266 " \"%pd\": old parent: %llu, new: %llu\n",
267 dentry,
268 (unsigned long long)parent_blkno,
269 (unsigned long long)dl->dl_parent_blkno);
270
271 trace_ocfs2_dentry_attach_lock_found(dl->dl_lockres.l_name,
272 (unsigned long long)parent_blkno,
273 (unsigned long long)OCFS2_I(inode)->ip_blkno);
274
275 goto out_attach;
276 }
277
278 /*
279 * There are no other aliases
280 */
281 dl = kmalloc(sizeof(*dl), GFP_NOFS);
282 if (!dl) {
283 ret = -ENOMEM;
284 mlog_errno(ret);
285 return ret;
286 }
287
288 dl->dl_count = 0;
289 /*
290 * Does this have to happen below, for all attaches, in case
291 * the struct inode gets blown away by the downconvert thread?
292 */
293 dl->dl_inode = igrab(inode);
294 dl->dl_parent_blkno = parent_blkno;
295 ocfs2_dentry_lock_res_init(dl, parent_blkno, inode);
296
297 out_attach:
298 spin_lock(&dentry_attach_lock);
299 if (unlikely(dentry->d_fsdata && !alias)) {
300 /* d_fsdata is set by a racing thread which is doing
301 * the same thing as this thread is doing. Leave the racing
302 * thread going ahead and we return here.
303 */
304 spin_unlock(&dentry_attach_lock);
305 iput(dl->dl_inode);
306 ocfs2_lock_res_free(&dl->dl_lockres);
307 kfree(dl);
308 return 0;
309 }
310
311 dentry->d_fsdata = dl;
312 dl->dl_count++;
313 spin_unlock(&dentry_attach_lock);
314
315 /*
316 * This actually gets us our PRMODE level lock. From now on,
317 * we'll have a notification if one of these names is
318 * destroyed on another node.
319 */
320 ret = ocfs2_dentry_lock(dentry, 0);
321 if (!ret)
322 ocfs2_dentry_unlock(dentry, 0);
323 else
324 mlog_errno(ret);
325
326 /*
327 * In case of error, manually free the allocation and do the iput().
328 * We need to do this because error here means no d_instantiate(),
329 * which means iput() will not be called during dput(dentry).
330 */
331 if (ret < 0 && !alias) {
332 ocfs2_lock_res_free(&dl->dl_lockres);
333 BUG_ON(dl->dl_count != 1);
334 spin_lock(&dentry_attach_lock);
335 dentry->d_fsdata = NULL;
336 spin_unlock(&dentry_attach_lock);
337 kfree(dl);
338 iput(inode);
339 }
340
341 dput(alias);
342
343 return ret;
344 }
345
346 /*
347 * ocfs2_dentry_iput() and friends.
348 *
349 * At this point, our particular dentry is detached from the inodes
350 * alias list, so there's no way that the locking code can find it.
351 *
352 * The interesting stuff happens when we determine that our lock needs
353 * to go away because this is the last subdir alias in the
354 * system. This function needs to handle a couple things:
355 *
356 * 1) Synchronizing lock shutdown with the downconvert threads. This
357 * is already handled for us via the lockres release drop function
358 * called in ocfs2_release_dentry_lock()
359 *
360 * 2) A race may occur when we're doing our lock shutdown and
361 * another process wants to create a new dentry lock. Right now we
362 * let them race, which means that for a very short while, this
363 * node might have two locks on a lock resource. This should be a
364 * problem though because one of them is in the process of being
365 * thrown out.
366 */
367 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb,
368 struct ocfs2_dentry_lock *dl)
369 {
370 iput(dl->dl_inode);
371 ocfs2_simple_drop_lockres(osb, &dl->dl_lockres);
372 ocfs2_lock_res_free(&dl->dl_lockres);
373 kfree(dl);
374 }
375
376 void ocfs2_dentry_lock_put(struct ocfs2_super *osb,
377 struct ocfs2_dentry_lock *dl)
378 {
379 int unlock = 0;
380
381 BUG_ON(dl->dl_count == 0);
382
383 spin_lock(&dentry_attach_lock);
384 dl->dl_count--;
385 unlock = !dl->dl_count;
386 spin_unlock(&dentry_attach_lock);
387
388 if (unlock)
389 ocfs2_drop_dentry_lock(osb, dl);
390 }
391
392 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode)
393 {
394 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
395
396 if (!dl) {
397 /*
398 * No dentry lock is ok if we're disconnected or
399 * unhashed.
400 */
401 if (!(dentry->d_flags & DCACHE_DISCONNECTED) &&
402 !d_unhashed(dentry)) {
403 unsigned long long ino = 0ULL;
404 if (inode)
405 ino = (unsigned long long)OCFS2_I(inode)->ip_blkno;
406 mlog(ML_ERROR, "Dentry is missing cluster lock. "
407 "inode: %llu, d_flags: 0x%x, d_name: %pd\n",
408 ino, dentry->d_flags, dentry);
409 }
410
411 goto out;
412 }
413
414 mlog_bug_on_msg(dl->dl_count == 0, "dentry: %pd, count: %u\n",
415 dentry, dl->dl_count);
416
417 ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl);
418
419 out:
420 iput(inode);
421 }
422
423 /*
424 * d_move(), but keep the locks in sync.
425 *
426 * When we are done, "dentry" will have the parent dir and name of
427 * "target", which will be thrown away.
428 *
429 * We manually update the lock of "dentry" if need be.
430 *
431 * "target" doesn't have it's dentry lock touched - we allow the later
432 * dput() to handle this for us.
433 *
434 * This is called during ocfs2_rename(), while holding parent
435 * directory locks. The dentries have already been deleted on other
436 * nodes via ocfs2_remote_dentry_delete().
437 *
438 * Normally, the VFS handles the d_move() for the file system, after
439 * the ->rename() callback. OCFS2 wants to handle this internally, so
440 * the new lock can be created atomically with respect to the cluster.
441 */
442 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target,
443 struct inode *old_dir, struct inode *new_dir)
444 {
445 int ret;
446 struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb);
447 struct inode *inode = d_inode(dentry);
448
449 /*
450 * Move within the same directory, so the actual lock info won't
451 * change.
452 *
453 * XXX: Is there any advantage to dropping the lock here?
454 */
455 if (old_dir == new_dir)
456 goto out_move;
457
458 ocfs2_dentry_lock_put(osb, dentry->d_fsdata);
459
460 dentry->d_fsdata = NULL;
461 ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno);
462 if (ret)
463 mlog_errno(ret);
464
465 out_move:
466 d_move(dentry, target);
467 }
468
469 const struct dentry_operations ocfs2_dentry_ops = {
470 .d_revalidate = ocfs2_dentry_revalidate,
471 .d_iput = ocfs2_dentry_iput,
472 };