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" | |
ea455f8a | 41 | #include "super.h" |
ccd979bd | 42 | |
5e98d492 GR |
43 | void ocfs2_dentry_attach_gen(struct dentry *dentry) |
44 | { | |
45 | unsigned long gen = | |
46 | OCFS2_I(dentry->d_parent->d_inode)->ip_dir_lock_gen; | |
47 | BUG_ON(dentry->d_inode); | |
48 | dentry->d_fsdata = (void *)gen; | |
49 | } | |
50 | ||
80c05846 | 51 | |
ccd979bd MF |
52 | static int ocfs2_dentry_revalidate(struct dentry *dentry, |
53 | struct nameidata *nd) | |
54 | { | |
55 | struct inode *inode = dentry->d_inode; | |
56 | int ret = 0; /* if all else fails, just return false */ | |
80c05846 | 57 | struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb); |
ccd979bd MF |
58 | |
59 | mlog_entry("(0x%p, '%.*s')\n", dentry, | |
60 | dentry->d_name.len, dentry->d_name.name); | |
61 | ||
5e98d492 GR |
62 | /* For a negative dentry - |
63 | * check the generation number of the parent and compare with the | |
64 | * one stored in the inode. | |
65 | */ | |
ccd979bd | 66 | if (inode == NULL) { |
5e98d492 GR |
67 | unsigned long gen = (unsigned long) dentry->d_fsdata; |
68 | unsigned long pgen = | |
69 | OCFS2_I(dentry->d_parent->d_inode)->ip_dir_lock_gen; | |
70 | mlog(0, "negative dentry: %.*s parent gen: %lu " | |
71 | "dentry gen: %lu\n", | |
72 | dentry->d_name.len, dentry->d_name.name, pgen, gen); | |
73 | if (gen != pgen) | |
74 | goto bail; | |
75 | goto valid; | |
ccd979bd MF |
76 | } |
77 | ||
ccd979bd MF |
78 | BUG_ON(!osb); |
79 | ||
80c05846 MF |
80 | if (inode == osb->root_inode || is_bad_inode(inode)) |
81 | goto bail; | |
82 | ||
83 | spin_lock(&OCFS2_I(inode)->ip_lock); | |
84 | /* did we or someone else delete this inode? */ | |
85 | if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) { | |
ccd979bd | 86 | spin_unlock(&OCFS2_I(inode)->ip_lock); |
80c05846 MF |
87 | mlog(0, "inode (%llu) deleted, returning false\n", |
88 | (unsigned long long)OCFS2_I(inode)->ip_blkno); | |
89 | goto bail; | |
90 | } | |
91 | spin_unlock(&OCFS2_I(inode)->ip_lock); | |
ccd979bd | 92 | |
80c05846 MF |
93 | /* |
94 | * We don't need a cluster lock to test this because once an | |
95 | * inode nlink hits zero, it never goes back. | |
96 | */ | |
97 | if (inode->i_nlink == 0) { | |
98 | mlog(0, "Inode %llu orphaned, returning false " | |
99 | "dir = %d\n", | |
100 | (unsigned long long)OCFS2_I(inode)->ip_blkno, | |
101 | S_ISDIR(inode->i_mode)); | |
102 | goto bail; | |
ccd979bd MF |
103 | } |
104 | ||
a1b08e75 TM |
105 | /* |
106 | * If the last lookup failed to create dentry lock, let us | |
107 | * redo it. | |
108 | */ | |
109 | if (!dentry->d_fsdata) { | |
110 | mlog(0, "Inode %llu doesn't have dentry lock, " | |
111 | "returning false\n", | |
112 | (unsigned long long)OCFS2_I(inode)->ip_blkno); | |
113 | goto bail; | |
114 | } | |
115 | ||
5e98d492 | 116 | valid: |
ccd979bd MF |
117 | ret = 1; |
118 | ||
119 | bail: | |
120 | mlog_exit(ret); | |
121 | ||
122 | return ret; | |
123 | } | |
124 | ||
80c05846 MF |
125 | static int ocfs2_match_dentry(struct dentry *dentry, |
126 | u64 parent_blkno, | |
127 | int skip_unhashed) | |
128 | { | |
129 | struct inode *parent; | |
130 | ||
131 | /* | |
132 | * ocfs2_lookup() does a d_splice_alias() _before_ attaching | |
133 | * to the lock data, so we skip those here, otherwise | |
134 | * ocfs2_dentry_attach_lock() will get its original dentry | |
135 | * back. | |
136 | */ | |
137 | if (!dentry->d_fsdata) | |
138 | return 0; | |
139 | ||
140 | if (!dentry->d_parent) | |
141 | return 0; | |
142 | ||
143 | if (skip_unhashed && d_unhashed(dentry)) | |
144 | return 0; | |
145 | ||
146 | parent = dentry->d_parent->d_inode; | |
147 | /* Negative parent dentry? */ | |
148 | if (!parent) | |
149 | return 0; | |
150 | ||
151 | /* Name is in a different directory. */ | |
152 | if (OCFS2_I(parent)->ip_blkno != parent_blkno) | |
153 | return 0; | |
154 | ||
155 | return 1; | |
156 | } | |
157 | ||
158 | /* | |
159 | * Walk the inode alias list, and find a dentry which has a given | |
160 | * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it | |
34d024f8 MF |
161 | * is looking for a dentry_lock reference. The downconvert thread is |
162 | * looking to unhash aliases, so we allow it to skip any that already | |
163 | * have that property. | |
80c05846 MF |
164 | */ |
165 | struct dentry *ocfs2_find_local_alias(struct inode *inode, | |
166 | u64 parent_blkno, | |
167 | int skip_unhashed) | |
168 | { | |
169 | struct list_head *p; | |
170 | struct dentry *dentry = NULL; | |
171 | ||
172 | spin_lock(&dcache_lock); | |
173 | ||
174 | list_for_each(p, &inode->i_dentry) { | |
175 | dentry = list_entry(p, struct dentry, d_alias); | |
176 | ||
177 | if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) { | |
178 | mlog(0, "dentry found: %.*s\n", | |
179 | dentry->d_name.len, dentry->d_name.name); | |
180 | ||
181 | dget_locked(dentry); | |
182 | break; | |
183 | } | |
184 | ||
185 | dentry = NULL; | |
186 | } | |
187 | ||
188 | spin_unlock(&dcache_lock); | |
189 | ||
190 | return dentry; | |
191 | } | |
192 | ||
d680efe9 MF |
193 | DEFINE_SPINLOCK(dentry_attach_lock); |
194 | ||
80c05846 MF |
195 | /* |
196 | * Attach this dentry to a cluster lock. | |
197 | * | |
198 | * Dentry locks cover all links in a given directory to a particular | |
199 | * inode. We do this so that ocfs2 can build a lock name which all | |
200 | * nodes in the cluster can agree on at all times. Shoving full names | |
201 | * in the cluster lock won't work due to size restrictions. Covering | |
202 | * links inside of a directory is a good compromise because it still | |
203 | * allows us to use the parent directory lock to synchronize | |
204 | * operations. | |
205 | * | |
206 | * Call this function with the parent dir semaphore and the parent dir | |
207 | * cluster lock held. | |
208 | * | |
209 | * The dir semaphore will protect us from having to worry about | |
210 | * concurrent processes on our node trying to attach a lock at the | |
211 | * same time. | |
212 | * | |
213 | * The dir cluster lock (held at either PR or EX mode) protects us | |
214 | * from unlink and rename on other nodes. | |
215 | * | |
80c05846 MF |
216 | * A dput() can happen asynchronously due to pruning, so we cover |
217 | * attaching and detaching the dentry lock with a | |
218 | * dentry_attach_lock. | |
219 | * | |
220 | * A node which has done lookup on a name retains a protected read | |
221 | * lock until final dput. If the user requests and unlink or rename, | |
222 | * the protected read is upgraded to an exclusive lock. Other nodes | |
223 | * who have seen the dentry will then be informed that they need to | |
224 | * downgrade their lock, which will involve d_delete on the | |
225 | * dentry. This happens in ocfs2_dentry_convert_worker(). | |
226 | */ | |
227 | int ocfs2_dentry_attach_lock(struct dentry *dentry, | |
228 | struct inode *inode, | |
0027dd5b | 229 | u64 parent_blkno) |
80c05846 MF |
230 | { |
231 | int ret; | |
232 | struct dentry *alias; | |
233 | struct ocfs2_dentry_lock *dl = dentry->d_fsdata; | |
234 | ||
0027dd5b | 235 | mlog(0, "Attach \"%.*s\", parent %llu, fsdata: %p\n", |
80c05846 | 236 | dentry->d_name.len, dentry->d_name.name, |
0027dd5b | 237 | (unsigned long long)parent_blkno, dl); |
80c05846 MF |
238 | |
239 | /* | |
240 | * Negative dentry. We ignore these for now. | |
241 | * | |
242 | * XXX: Could we can improve ocfs2_dentry_revalidate() by | |
243 | * tracking these? | |
244 | */ | |
245 | if (!inode) | |
246 | return 0; | |
247 | ||
5e98d492 GR |
248 | if (!dentry->d_inode && dentry->d_fsdata) { |
249 | /* Converting a negative dentry to positive | |
250 | Clear dentry->d_fsdata */ | |
251 | dentry->d_fsdata = dl = NULL; | |
252 | } | |
253 | ||
80c05846 MF |
254 | if (dl) { |
255 | mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, | |
256 | " \"%.*s\": old parent: %llu, new: %llu\n", | |
257 | dentry->d_name.len, dentry->d_name.name, | |
258 | (unsigned long long)parent_blkno, | |
259 | (unsigned long long)dl->dl_parent_blkno); | |
260 | return 0; | |
261 | } | |
262 | ||
263 | alias = ocfs2_find_local_alias(inode, parent_blkno, 0); | |
264 | if (alias) { | |
265 | /* | |
266 | * Great, an alias exists, which means we must have a | |
267 | * dentry lock already. We can just grab the lock off | |
268 | * the alias and add it to the list. | |
269 | * | |
270 | * We're depending here on the fact that this dentry | |
271 | * was found and exists in the dcache and so must have | |
272 | * a reference to the dentry_lock because we can't | |
273 | * race creates. Final dput() cannot happen on it | |
274 | * since we have it pinned, so our reference is safe. | |
275 | */ | |
276 | dl = alias->d_fsdata; | |
0027dd5b | 277 | mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n", |
80c05846 | 278 | (unsigned long long)parent_blkno, |
0027dd5b | 279 | (unsigned long long)OCFS2_I(inode)->ip_blkno); |
80c05846 MF |
280 | |
281 | mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, | |
282 | " \"%.*s\": old parent: %llu, new: %llu\n", | |
283 | dentry->d_name.len, dentry->d_name.name, | |
284 | (unsigned long long)parent_blkno, | |
285 | (unsigned long long)dl->dl_parent_blkno); | |
286 | ||
287 | mlog(0, "Found: %s\n", dl->dl_lockres.l_name); | |
288 | ||
289 | goto out_attach; | |
290 | } | |
291 | ||
292 | /* | |
293 | * There are no other aliases | |
294 | */ | |
295 | dl = kmalloc(sizeof(*dl), GFP_NOFS); | |
296 | if (!dl) { | |
297 | ret = -ENOMEM; | |
298 | mlog_errno(ret); | |
299 | return ret; | |
300 | } | |
301 | ||
302 | dl->dl_count = 0; | |
303 | /* | |
304 | * Does this have to happen below, for all attaches, in case | |
34d024f8 | 305 | * the struct inode gets blown away by the downconvert thread? |
80c05846 MF |
306 | */ |
307 | dl->dl_inode = igrab(inode); | |
308 | dl->dl_parent_blkno = parent_blkno; | |
309 | ocfs2_dentry_lock_res_init(dl, parent_blkno, inode); | |
310 | ||
311 | out_attach: | |
312 | spin_lock(&dentry_attach_lock); | |
313 | dentry->d_fsdata = dl; | |
314 | dl->dl_count++; | |
315 | spin_unlock(&dentry_attach_lock); | |
316 | ||
80c05846 MF |
317 | /* |
318 | * This actually gets us our PRMODE level lock. From now on, | |
319 | * we'll have a notification if one of these names is | |
320 | * destroyed on another node. | |
321 | */ | |
322 | ret = ocfs2_dentry_lock(dentry, 0); | |
0027dd5b MF |
323 | if (!ret) |
324 | ocfs2_dentry_unlock(dentry, 0); | |
325 | else | |
80c05846 | 326 | mlog_errno(ret); |
80c05846 | 327 | |
a5a0a630 SM |
328 | /* |
329 | * In case of error, manually free the allocation and do the iput(). | |
330 | * We need to do this because error here means no d_instantiate(), | |
331 | * which means iput() will not be called during dput(dentry). | |
332 | */ | |
333 | if (ret < 0 && !alias) { | |
334 | ocfs2_lock_res_free(&dl->dl_lockres); | |
335 | BUG_ON(dl->dl_count != 1); | |
336 | spin_lock(&dentry_attach_lock); | |
337 | dentry->d_fsdata = NULL; | |
338 | spin_unlock(&dentry_attach_lock); | |
339 | kfree(dl); | |
340 | iput(inode); | |
341 | } | |
342 | ||
80c05846 MF |
343 | dput(alias); |
344 | ||
345 | return ret; | |
346 | } | |
347 | ||
f7b1aa69 | 348 | DEFINE_SPINLOCK(dentry_list_lock); |
ea455f8a JK |
349 | |
350 | /* We limit the number of dentry locks to drop in one go. We have | |
351 | * this limit so that we don't starve other users of ocfs2_wq. */ | |
352 | #define DL_INODE_DROP_COUNT 64 | |
353 | ||
354 | /* Drop inode references from dentry locks */ | |
f7b1aa69 | 355 | static void __ocfs2_drop_dl_inodes(struct ocfs2_super *osb, int drop_count) |
ea455f8a | 356 | { |
ea455f8a | 357 | struct ocfs2_dentry_lock *dl; |
ea455f8a JK |
358 | |
359 | spin_lock(&dentry_list_lock); | |
f7b1aa69 | 360 | while (osb->dentry_lock_list && (drop_count < 0 || drop_count--)) { |
ea455f8a JK |
361 | dl = osb->dentry_lock_list; |
362 | osb->dentry_lock_list = dl->dl_next; | |
363 | spin_unlock(&dentry_list_lock); | |
364 | iput(dl->dl_inode); | |
365 | kfree(dl); | |
366 | spin_lock(&dentry_list_lock); | |
367 | } | |
f7b1aa69 JK |
368 | spin_unlock(&dentry_list_lock); |
369 | } | |
370 | ||
371 | void ocfs2_drop_dl_inodes(struct work_struct *work) | |
372 | { | |
373 | struct ocfs2_super *osb = container_of(work, struct ocfs2_super, | |
374 | dentry_lock_work); | |
375 | ||
376 | __ocfs2_drop_dl_inodes(osb, DL_INODE_DROP_COUNT); | |
377 | /* | |
378 | * Don't queue dropping if umount is in progress. We flush the | |
379 | * list in ocfs2_dismount_volume | |
380 | */ | |
381 | spin_lock(&dentry_list_lock); | |
382 | if (osb->dentry_lock_list && | |
383 | !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) | |
ea455f8a JK |
384 | queue_work(ocfs2_wq, &osb->dentry_lock_work); |
385 | spin_unlock(&dentry_list_lock); | |
386 | } | |
387 | ||
f7b1aa69 JK |
388 | /* Flush the whole work queue */ |
389 | void ocfs2_drop_all_dl_inodes(struct ocfs2_super *osb) | |
390 | { | |
391 | __ocfs2_drop_dl_inodes(osb, -1); | |
392 | } | |
393 | ||
80c05846 MF |
394 | /* |
395 | * ocfs2_dentry_iput() and friends. | |
396 | * | |
397 | * At this point, our particular dentry is detached from the inodes | |
398 | * alias list, so there's no way that the locking code can find it. | |
399 | * | |
400 | * The interesting stuff happens when we determine that our lock needs | |
401 | * to go away because this is the last subdir alias in the | |
402 | * system. This function needs to handle a couple things: | |
403 | * | |
404 | * 1) Synchronizing lock shutdown with the downconvert threads. This | |
405 | * is already handled for us via the lockres release drop function | |
406 | * called in ocfs2_release_dentry_lock() | |
407 | * | |
408 | * 2) A race may occur when we're doing our lock shutdown and | |
409 | * another process wants to create a new dentry lock. Right now we | |
410 | * let them race, which means that for a very short while, this | |
411 | * node might have two locks on a lock resource. This should be a | |
412 | * problem though because one of them is in the process of being | |
413 | * thrown out. | |
414 | */ | |
415 | static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb, | |
416 | struct ocfs2_dentry_lock *dl) | |
417 | { | |
418 | ocfs2_simple_drop_lockres(osb, &dl->dl_lockres); | |
419 | ocfs2_lock_res_free(&dl->dl_lockres); | |
ea455f8a JK |
420 | |
421 | /* We leave dropping of inode reference to ocfs2_wq as that can | |
422 | * possibly lead to inode deletion which gets tricky */ | |
423 | spin_lock(&dentry_list_lock); | |
f7b1aa69 JK |
424 | if (!osb->dentry_lock_list && |
425 | !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) | |
ea455f8a JK |
426 | queue_work(ocfs2_wq, &osb->dentry_lock_work); |
427 | dl->dl_next = osb->dentry_lock_list; | |
428 | osb->dentry_lock_list = dl; | |
429 | spin_unlock(&dentry_list_lock); | |
80c05846 MF |
430 | } |
431 | ||
432 | void ocfs2_dentry_lock_put(struct ocfs2_super *osb, | |
433 | struct ocfs2_dentry_lock *dl) | |
434 | { | |
ea455f8a | 435 | int unlock; |
80c05846 MF |
436 | |
437 | BUG_ON(dl->dl_count == 0); | |
438 | ||
439 | spin_lock(&dentry_attach_lock); | |
440 | dl->dl_count--; | |
441 | unlock = !dl->dl_count; | |
442 | spin_unlock(&dentry_attach_lock); | |
443 | ||
444 | if (unlock) | |
445 | ocfs2_drop_dentry_lock(osb, dl); | |
446 | } | |
447 | ||
448 | static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode) | |
449 | { | |
450 | struct ocfs2_dentry_lock *dl = dentry->d_fsdata; | |
451 | ||
bccb9dad MF |
452 | if (!dl) { |
453 | /* | |
454 | * No dentry lock is ok if we're disconnected or | |
455 | * unhashed. | |
456 | */ | |
457 | if (!(dentry->d_flags & DCACHE_DISCONNECTED) && | |
458 | !d_unhashed(dentry)) { | |
459 | unsigned long long ino = 0ULL; | |
460 | if (inode) | |
461 | ino = (unsigned long long)OCFS2_I(inode)->ip_blkno; | |
462 | mlog(ML_ERROR, "Dentry is missing cluster lock. " | |
463 | "inode: %llu, d_flags: 0x%x, d_name: %.*s\n", | |
464 | ino, dentry->d_flags, dentry->d_name.len, | |
465 | dentry->d_name.name); | |
466 | } | |
80c05846 | 467 | |
80c05846 | 468 | goto out; |
bccb9dad | 469 | } |
80c05846 MF |
470 | |
471 | mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n", | |
472 | dentry->d_name.len, dentry->d_name.name, | |
473 | dl->dl_count); | |
474 | ||
475 | ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl); | |
476 | ||
477 | out: | |
478 | iput(inode); | |
5e98d492 | 479 | ocfs2_dentry_attach_gen(dentry); |
80c05846 MF |
480 | } |
481 | ||
482 | /* | |
483 | * d_move(), but keep the locks in sync. | |
484 | * | |
485 | * When we are done, "dentry" will have the parent dir and name of | |
486 | * "target", which will be thrown away. | |
487 | * | |
488 | * We manually update the lock of "dentry" if need be. | |
489 | * | |
490 | * "target" doesn't have it's dentry lock touched - we allow the later | |
491 | * dput() to handle this for us. | |
492 | * | |
493 | * This is called during ocfs2_rename(), while holding parent | |
494 | * directory locks. The dentries have already been deleted on other | |
495 | * nodes via ocfs2_remote_dentry_delete(). | |
496 | * | |
3a4fa0a2 | 497 | * Normally, the VFS handles the d_move() for the file system, after |
80c05846 MF |
498 | * the ->rename() callback. OCFS2 wants to handle this internally, so |
499 | * the new lock can be created atomically with respect to the cluster. | |
500 | */ | |
501 | void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target, | |
502 | struct inode *old_dir, struct inode *new_dir) | |
503 | { | |
504 | int ret; | |
505 | struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb); | |
506 | struct inode *inode = dentry->d_inode; | |
507 | ||
508 | /* | |
509 | * Move within the same directory, so the actual lock info won't | |
510 | * change. | |
511 | * | |
512 | * XXX: Is there any advantage to dropping the lock here? | |
513 | */ | |
514 | if (old_dir == new_dir) | |
1ba9da2f | 515 | goto out_move; |
80c05846 MF |
516 | |
517 | ocfs2_dentry_lock_put(osb, dentry->d_fsdata); | |
518 | ||
519 | dentry->d_fsdata = NULL; | |
0027dd5b | 520 | ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno); |
80c05846 MF |
521 | if (ret) |
522 | mlog_errno(ret); | |
1ba9da2f MF |
523 | |
524 | out_move: | |
525 | d_move(dentry, target); | |
80c05846 MF |
526 | } |
527 | ||
d8fba0ff | 528 | const struct dentry_operations ocfs2_dentry_ops = { |
ccd979bd | 529 | .d_revalidate = ocfs2_dentry_revalidate, |
80c05846 | 530 | .d_iput = ocfs2_dentry_iput, |
ccd979bd | 531 | }; |