Commit | Line | Data |
---|---|---|
ccd979bd MF |
1 | /* -*- mode: c; c-basic-offset: 8; -*- |
2 | * vim: noexpandtab sw=8 ts=8 sts=0: | |
3 | * | |
4 | * dcache.c | |
5 | * | |
6 | * dentry cache handling code | |
7 | * | |
8 | * Copyright (C) 2002, 2004 Oracle. All rights reserved. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or | |
11 | * modify it under the terms of the GNU General Public | |
12 | * License as published by the Free Software Foundation; either | |
13 | * version 2 of the License, or (at your option) any later version. | |
14 | * | |
15 | * This program is distributed in the hope that it will be useful, | |
16 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
18 | * General Public License for more details. | |
19 | * | |
20 | * You should have received a copy of the GNU General Public | |
21 | * License along with this program; if not, write to the | |
22 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, | |
23 | * Boston, MA 021110-1307, USA. | |
24 | */ | |
25 | ||
26 | #include <linux/fs.h> | |
27 | #include <linux/types.h> | |
28 | #include <linux/slab.h> | |
29 | #include <linux/namei.h> | |
30 | ||
31 | #define MLOG_MASK_PREFIX ML_DCACHE | |
32 | #include <cluster/masklog.h> | |
33 | ||
34 | #include "ocfs2.h" | |
35 | ||
36 | #include "alloc.h" | |
37 | #include "dcache.h" | |
80c05846 | 38 | #include "dlmglue.h" |
ccd979bd MF |
39 | #include "file.h" |
40 | #include "inode.h" | |
41 | ||
80c05846 | 42 | |
ccd979bd MF |
43 | static int ocfs2_dentry_revalidate(struct dentry *dentry, |
44 | struct nameidata *nd) | |
45 | { | |
46 | struct inode *inode = dentry->d_inode; | |
47 | int ret = 0; /* if all else fails, just return false */ | |
80c05846 | 48 | struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb); |
ccd979bd MF |
49 | |
50 | mlog_entry("(0x%p, '%.*s')\n", dentry, | |
51 | dentry->d_name.len, dentry->d_name.name); | |
52 | ||
53 | /* Never trust a negative dentry - force a new lookup. */ | |
54 | if (inode == NULL) { | |
55 | mlog(0, "negative dentry: %.*s\n", dentry->d_name.len, | |
56 | dentry->d_name.name); | |
57 | goto bail; | |
58 | } | |
59 | ||
ccd979bd MF |
60 | BUG_ON(!osb); |
61 | ||
80c05846 MF |
62 | if (inode == osb->root_inode || is_bad_inode(inode)) |
63 | goto bail; | |
64 | ||
65 | spin_lock(&OCFS2_I(inode)->ip_lock); | |
66 | /* did we or someone else delete this inode? */ | |
67 | if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) { | |
ccd979bd | 68 | spin_unlock(&OCFS2_I(inode)->ip_lock); |
80c05846 MF |
69 | mlog(0, "inode (%llu) deleted, returning false\n", |
70 | (unsigned long long)OCFS2_I(inode)->ip_blkno); | |
71 | goto bail; | |
72 | } | |
73 | spin_unlock(&OCFS2_I(inode)->ip_lock); | |
ccd979bd | 74 | |
80c05846 MF |
75 | /* |
76 | * We don't need a cluster lock to test this because once an | |
77 | * inode nlink hits zero, it never goes back. | |
78 | */ | |
79 | if (inode->i_nlink == 0) { | |
80 | mlog(0, "Inode %llu orphaned, returning false " | |
81 | "dir = %d\n", | |
82 | (unsigned long long)OCFS2_I(inode)->ip_blkno, | |
83 | S_ISDIR(inode->i_mode)); | |
84 | goto bail; | |
ccd979bd MF |
85 | } |
86 | ||
87 | ret = 1; | |
88 | ||
89 | bail: | |
90 | mlog_exit(ret); | |
91 | ||
92 | return ret; | |
93 | } | |
94 | ||
80c05846 MF |
95 | static int ocfs2_match_dentry(struct dentry *dentry, |
96 | u64 parent_blkno, | |
97 | int skip_unhashed) | |
98 | { | |
99 | struct inode *parent; | |
100 | ||
101 | /* | |
102 | * ocfs2_lookup() does a d_splice_alias() _before_ attaching | |
103 | * to the lock data, so we skip those here, otherwise | |
104 | * ocfs2_dentry_attach_lock() will get its original dentry | |
105 | * back. | |
106 | */ | |
107 | if (!dentry->d_fsdata) | |
108 | return 0; | |
109 | ||
110 | if (!dentry->d_parent) | |
111 | return 0; | |
112 | ||
113 | if (skip_unhashed && d_unhashed(dentry)) | |
114 | return 0; | |
115 | ||
116 | parent = dentry->d_parent->d_inode; | |
117 | /* Negative parent dentry? */ | |
118 | if (!parent) | |
119 | return 0; | |
120 | ||
121 | /* Name is in a different directory. */ | |
122 | if (OCFS2_I(parent)->ip_blkno != parent_blkno) | |
123 | return 0; | |
124 | ||
125 | return 1; | |
126 | } | |
127 | ||
128 | /* | |
129 | * Walk the inode alias list, and find a dentry which has a given | |
130 | * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it | |
131 | * is looking for a dentry_lock reference. The vote thread is looking | |
132 | * to unhash aliases, so we allow it to skip any that already have | |
133 | * that property. | |
134 | */ | |
135 | struct dentry *ocfs2_find_local_alias(struct inode *inode, | |
136 | u64 parent_blkno, | |
137 | int skip_unhashed) | |
138 | { | |
139 | struct list_head *p; | |
140 | struct dentry *dentry = NULL; | |
141 | ||
142 | spin_lock(&dcache_lock); | |
143 | ||
144 | list_for_each(p, &inode->i_dentry) { | |
145 | dentry = list_entry(p, struct dentry, d_alias); | |
146 | ||
147 | if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) { | |
148 | mlog(0, "dentry found: %.*s\n", | |
149 | dentry->d_name.len, dentry->d_name.name); | |
150 | ||
151 | dget_locked(dentry); | |
152 | break; | |
153 | } | |
154 | ||
155 | dentry = NULL; | |
156 | } | |
157 | ||
158 | spin_unlock(&dcache_lock); | |
159 | ||
160 | return dentry; | |
161 | } | |
162 | ||
d680efe9 MF |
163 | DEFINE_SPINLOCK(dentry_attach_lock); |
164 | ||
80c05846 MF |
165 | /* |
166 | * Attach this dentry to a cluster lock. | |
167 | * | |
168 | * Dentry locks cover all links in a given directory to a particular | |
169 | * inode. We do this so that ocfs2 can build a lock name which all | |
170 | * nodes in the cluster can agree on at all times. Shoving full names | |
171 | * in the cluster lock won't work due to size restrictions. Covering | |
172 | * links inside of a directory is a good compromise because it still | |
173 | * allows us to use the parent directory lock to synchronize | |
174 | * operations. | |
175 | * | |
176 | * Call this function with the parent dir semaphore and the parent dir | |
177 | * cluster lock held. | |
178 | * | |
179 | * The dir semaphore will protect us from having to worry about | |
180 | * concurrent processes on our node trying to attach a lock at the | |
181 | * same time. | |
182 | * | |
183 | * The dir cluster lock (held at either PR or EX mode) protects us | |
184 | * from unlink and rename on other nodes. | |
185 | * | |
186 | * The 'create' flag tells us whether we're doing this as a result of | |
187 | * a file creation. | |
188 | * | |
189 | * A dput() can happen asynchronously due to pruning, so we cover | |
190 | * attaching and detaching the dentry lock with a | |
191 | * dentry_attach_lock. | |
192 | * | |
193 | * A node which has done lookup on a name retains a protected read | |
194 | * lock until final dput. If the user requests and unlink or rename, | |
195 | * the protected read is upgraded to an exclusive lock. Other nodes | |
196 | * who have seen the dentry will then be informed that they need to | |
197 | * downgrade their lock, which will involve d_delete on the | |
198 | * dentry. This happens in ocfs2_dentry_convert_worker(). | |
199 | */ | |
200 | int ocfs2_dentry_attach_lock(struct dentry *dentry, | |
201 | struct inode *inode, | |
202 | u64 parent_blkno, | |
203 | int create) | |
204 | { | |
205 | int ret; | |
206 | struct dentry *alias; | |
207 | struct ocfs2_dentry_lock *dl = dentry->d_fsdata; | |
208 | ||
209 | mlog(0, "Attach \"%.*s\", parent %llu, create %d, fsdata: %p\n", | |
210 | dentry->d_name.len, dentry->d_name.name, | |
211 | (unsigned long long)parent_blkno, create, dl); | |
212 | ||
213 | /* | |
214 | * Negative dentry. We ignore these for now. | |
215 | * | |
216 | * XXX: Could we can improve ocfs2_dentry_revalidate() by | |
217 | * tracking these? | |
218 | */ | |
219 | if (!inode) | |
220 | return 0; | |
221 | ||
222 | if (dl) { | |
223 | mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, | |
224 | " \"%.*s\": old parent: %llu, new: %llu\n", | |
225 | dentry->d_name.len, dentry->d_name.name, | |
226 | (unsigned long long)parent_blkno, | |
227 | (unsigned long long)dl->dl_parent_blkno); | |
228 | return 0; | |
229 | } | |
230 | ||
231 | alias = ocfs2_find_local_alias(inode, parent_blkno, 0); | |
232 | if (alias) { | |
233 | /* | |
234 | * Great, an alias exists, which means we must have a | |
235 | * dentry lock already. We can just grab the lock off | |
236 | * the alias and add it to the list. | |
237 | * | |
238 | * We're depending here on the fact that this dentry | |
239 | * was found and exists in the dcache and so must have | |
240 | * a reference to the dentry_lock because we can't | |
241 | * race creates. Final dput() cannot happen on it | |
242 | * since we have it pinned, so our reference is safe. | |
243 | */ | |
244 | dl = alias->d_fsdata; | |
245 | mlog_bug_on_msg(!dl, "parent %llu, ino %llu, create %d\n", | |
246 | (unsigned long long)parent_blkno, | |
247 | (unsigned long long)OCFS2_I(inode)->ip_blkno, | |
248 | create); | |
249 | ||
250 | mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, | |
251 | " \"%.*s\": old parent: %llu, new: %llu\n", | |
252 | dentry->d_name.len, dentry->d_name.name, | |
253 | (unsigned long long)parent_blkno, | |
254 | (unsigned long long)dl->dl_parent_blkno); | |
255 | ||
256 | mlog(0, "Found: %s\n", dl->dl_lockres.l_name); | |
257 | ||
258 | goto out_attach; | |
259 | } | |
260 | ||
261 | /* | |
262 | * There are no other aliases | |
263 | */ | |
264 | dl = kmalloc(sizeof(*dl), GFP_NOFS); | |
265 | if (!dl) { | |
266 | ret = -ENOMEM; | |
267 | mlog_errno(ret); | |
268 | return ret; | |
269 | } | |
270 | ||
271 | dl->dl_count = 0; | |
272 | /* | |
273 | * Does this have to happen below, for all attaches, in case | |
274 | * the struct inode gets blown away by votes? | |
275 | */ | |
276 | dl->dl_inode = igrab(inode); | |
277 | dl->dl_parent_blkno = parent_blkno; | |
278 | ocfs2_dentry_lock_res_init(dl, parent_blkno, inode); | |
279 | ||
280 | out_attach: | |
281 | spin_lock(&dentry_attach_lock); | |
282 | dentry->d_fsdata = dl; | |
283 | dl->dl_count++; | |
284 | spin_unlock(&dentry_attach_lock); | |
285 | ||
286 | /* | |
287 | * Creation of a new file means that nobody can possibly have | |
288 | * this name in the system, which means that acquiry of those | |
289 | * locks can easily be optimized. | |
290 | */ | |
291 | if (create) { | |
292 | ret = ocfs2_create_new_lock(OCFS2_SB(inode->i_sb), | |
293 | &dl->dl_lockres, 0); | |
294 | if (ret) | |
295 | mlog_errno(ret); | |
296 | goto out; | |
297 | } | |
298 | ||
299 | /* | |
300 | * This actually gets us our PRMODE level lock. From now on, | |
301 | * we'll have a notification if one of these names is | |
302 | * destroyed on another node. | |
303 | */ | |
304 | ret = ocfs2_dentry_lock(dentry, 0); | |
305 | if (ret) { | |
306 | mlog_errno(ret); | |
307 | goto out; | |
308 | } | |
309 | ocfs2_dentry_unlock(dentry, 0); | |
310 | ||
311 | out: | |
312 | dput(alias); | |
313 | ||
314 | return ret; | |
315 | } | |
316 | ||
317 | /* | |
318 | * ocfs2_dentry_iput() and friends. | |
319 | * | |
320 | * At this point, our particular dentry is detached from the inodes | |
321 | * alias list, so there's no way that the locking code can find it. | |
322 | * | |
323 | * The interesting stuff happens when we determine that our lock needs | |
324 | * to go away because this is the last subdir alias in the | |
325 | * system. This function needs to handle a couple things: | |
326 | * | |
327 | * 1) Synchronizing lock shutdown with the downconvert threads. This | |
328 | * is already handled for us via the lockres release drop function | |
329 | * called in ocfs2_release_dentry_lock() | |
330 | * | |
331 | * 2) A race may occur when we're doing our lock shutdown and | |
332 | * another process wants to create a new dentry lock. Right now we | |
333 | * let them race, which means that for a very short while, this | |
334 | * node might have two locks on a lock resource. This should be a | |
335 | * problem though because one of them is in the process of being | |
336 | * thrown out. | |
337 | */ | |
338 | static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb, | |
339 | struct ocfs2_dentry_lock *dl) | |
340 | { | |
341 | ocfs2_simple_drop_lockres(osb, &dl->dl_lockres); | |
342 | ocfs2_lock_res_free(&dl->dl_lockres); | |
343 | iput(dl->dl_inode); | |
344 | kfree(dl); | |
345 | } | |
346 | ||
347 | void ocfs2_dentry_lock_put(struct ocfs2_super *osb, | |
348 | struct ocfs2_dentry_lock *dl) | |
349 | { | |
350 | int unlock = 0; | |
351 | ||
352 | BUG_ON(dl->dl_count == 0); | |
353 | ||
354 | spin_lock(&dentry_attach_lock); | |
355 | dl->dl_count--; | |
356 | unlock = !dl->dl_count; | |
357 | spin_unlock(&dentry_attach_lock); | |
358 | ||
359 | if (unlock) | |
360 | ocfs2_drop_dentry_lock(osb, dl); | |
361 | } | |
362 | ||
363 | static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode) | |
364 | { | |
365 | struct ocfs2_dentry_lock *dl = dentry->d_fsdata; | |
366 | ||
367 | mlog_bug_on_msg(!dl && !(dentry->d_flags & DCACHE_DISCONNECTED), | |
368 | "dentry: %.*s\n", dentry->d_name.len, | |
369 | dentry->d_name.name); | |
370 | ||
371 | if (!dl) | |
372 | goto out; | |
373 | ||
374 | mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n", | |
375 | dentry->d_name.len, dentry->d_name.name, | |
376 | dl->dl_count); | |
377 | ||
378 | ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl); | |
379 | ||
380 | out: | |
381 | iput(inode); | |
382 | } | |
383 | ||
384 | /* | |
385 | * d_move(), but keep the locks in sync. | |
386 | * | |
387 | * When we are done, "dentry" will have the parent dir and name of | |
388 | * "target", which will be thrown away. | |
389 | * | |
390 | * We manually update the lock of "dentry" if need be. | |
391 | * | |
392 | * "target" doesn't have it's dentry lock touched - we allow the later | |
393 | * dput() to handle this for us. | |
394 | * | |
395 | * This is called during ocfs2_rename(), while holding parent | |
396 | * directory locks. The dentries have already been deleted on other | |
397 | * nodes via ocfs2_remote_dentry_delete(). | |
398 | * | |
399 | * Normally, the VFS handles the d_move() for the file sytem, after | |
400 | * the ->rename() callback. OCFS2 wants to handle this internally, so | |
401 | * the new lock can be created atomically with respect to the cluster. | |
402 | */ | |
403 | void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target, | |
404 | struct inode *old_dir, struct inode *new_dir) | |
405 | { | |
406 | int ret; | |
407 | struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb); | |
408 | struct inode *inode = dentry->d_inode; | |
409 | ||
410 | /* | |
411 | * Move within the same directory, so the actual lock info won't | |
412 | * change. | |
413 | * | |
414 | * XXX: Is there any advantage to dropping the lock here? | |
415 | */ | |
416 | if (old_dir == new_dir) | |
417 | return; | |
418 | ||
419 | ocfs2_dentry_lock_put(osb, dentry->d_fsdata); | |
420 | ||
421 | dentry->d_fsdata = NULL; | |
422 | ret = ocfs2_dentry_attach_lock(dentry, inode, | |
423 | OCFS2_I(new_dir)->ip_blkno, 0); | |
424 | if (ret) | |
425 | mlog_errno(ret); | |
426 | } | |
427 | ||
ccd979bd MF |
428 | struct dentry_operations ocfs2_dentry_ops = { |
429 | .d_revalidate = ocfs2_dentry_revalidate, | |
80c05846 | 430 | .d_iput = ocfs2_dentry_iput, |
ccd979bd | 431 | }; |