CacheFiles: Fix occasional EIO on call to vfs_unlink()
[linux-2.6.git] / fs / cachefiles / namei.c
1 /* CacheFiles path walking and related routines
2  *
3  * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
4  * Written by David Howells (dhowells@redhat.com)
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public Licence
8  * as published by the Free Software Foundation; either version
9  * 2 of the Licence, or (at your option) any later version.
10  */
11
12 #include <linux/module.h>
13 #include <linux/sched.h>
14 #include <linux/file.h>
15 #include <linux/fs.h>
16 #include <linux/fsnotify.h>
17 #include <linux/quotaops.h>
18 #include <linux/xattr.h>
19 #include <linux/mount.h>
20 #include <linux/namei.h>
21 #include <linux/security.h>
22 #include <linux/slab.h>
23 #include "internal.h"
24
25 #define CACHEFILES_KEYBUF_SIZE 512
26
27 /*
28  * dump debugging info about an object
29  */
30 static noinline
31 void __cachefiles_printk_object(struct cachefiles_object *object,
32                                 const char *prefix,
33                                 u8 *keybuf)
34 {
35         struct fscache_cookie *cookie;
36         unsigned keylen, loop;
37
38         printk(KERN_ERR "%sobject: OBJ%x\n",
39                prefix, object->fscache.debug_id);
40         printk(KERN_ERR "%sobjstate=%s fl=%lx swfl=%lx ev=%lx[%lx]\n",
41                prefix, fscache_object_states[object->fscache.state],
42                object->fscache.flags, object->fscache.work.flags,
43                object->fscache.events,
44                object->fscache.event_mask & FSCACHE_OBJECT_EVENTS_MASK);
45         printk(KERN_ERR "%sops=%u inp=%u exc=%u\n",
46                prefix, object->fscache.n_ops, object->fscache.n_in_progress,
47                object->fscache.n_exclusive);
48         printk(KERN_ERR "%sparent=%p\n",
49                prefix, object->fscache.parent);
50
51         spin_lock(&object->fscache.lock);
52         cookie = object->fscache.cookie;
53         if (cookie) {
54                 printk(KERN_ERR "%scookie=%p [pr=%p nd=%p fl=%lx]\n",
55                        prefix,
56                        object->fscache.cookie,
57                        object->fscache.cookie->parent,
58                        object->fscache.cookie->netfs_data,
59                        object->fscache.cookie->flags);
60                 if (keybuf)
61                         keylen = cookie->def->get_key(cookie->netfs_data, keybuf,
62                                                       CACHEFILES_KEYBUF_SIZE);
63                 else
64                         keylen = 0;
65         } else {
66                 printk(KERN_ERR "%scookie=NULL\n", prefix);
67                 keylen = 0;
68         }
69         spin_unlock(&object->fscache.lock);
70
71         if (keylen) {
72                 printk(KERN_ERR "%skey=[%u] '", prefix, keylen);
73                 for (loop = 0; loop < keylen; loop++)
74                         printk("%02x", keybuf[loop]);
75                 printk("'\n");
76         }
77 }
78
79 /*
80  * dump debugging info about a pair of objects
81  */
82 static noinline void cachefiles_printk_object(struct cachefiles_object *object,
83                                               struct cachefiles_object *xobject)
84 {
85         u8 *keybuf;
86
87         keybuf = kmalloc(CACHEFILES_KEYBUF_SIZE, GFP_NOIO);
88         if (object)
89                 __cachefiles_printk_object(object, "", keybuf);
90         if (xobject)
91                 __cachefiles_printk_object(xobject, "x", keybuf);
92         kfree(keybuf);
93 }
94
95 /*
96  * mark the owner of a dentry, if there is one, to indicate that that dentry
97  * has been preemptively deleted
98  * - the caller must hold the i_mutex on the dentry's parent as required to
99  *   call vfs_unlink(), vfs_rmdir() or vfs_rename()
100  */
101 static void cachefiles_mark_object_buried(struct cachefiles_cache *cache,
102                                           struct dentry *dentry)
103 {
104         struct cachefiles_object *object;
105         struct rb_node *p;
106
107         _enter(",'%*.*s'",
108                dentry->d_name.len, dentry->d_name.len, dentry->d_name.name);
109
110         write_lock(&cache->active_lock);
111
112         p = cache->active_nodes.rb_node;
113         while (p) {
114                 object = rb_entry(p, struct cachefiles_object, active_node);
115                 if (object->dentry > dentry)
116                         p = p->rb_left;
117                 else if (object->dentry < dentry)
118                         p = p->rb_right;
119                 else
120                         goto found_dentry;
121         }
122
123         write_unlock(&cache->active_lock);
124         _leave(" [no owner]");
125         return;
126
127         /* found the dentry for  */
128 found_dentry:
129         kdebug("preemptive burial: OBJ%x [%s] %p",
130                object->fscache.debug_id,
131                fscache_object_states[object->fscache.state],
132                dentry);
133
134         if (object->fscache.state < FSCACHE_OBJECT_DYING) {
135                 printk(KERN_ERR "\n");
136                 printk(KERN_ERR "CacheFiles: Error:"
137                        " Can't preemptively bury live object\n");
138                 cachefiles_printk_object(object, NULL);
139         } else if (test_and_set_bit(CACHEFILES_OBJECT_BURIED, &object->flags)) {
140                 printk(KERN_ERR "CacheFiles: Error:"
141                        " Object already preemptively buried\n");
142         }
143
144         write_unlock(&cache->active_lock);
145         _leave(" [owner marked]");
146 }
147
148 /*
149  * record the fact that an object is now active
150  */
151 static int cachefiles_mark_object_active(struct cachefiles_cache *cache,
152                                          struct cachefiles_object *object)
153 {
154         struct cachefiles_object *xobject;
155         struct rb_node **_p, *_parent = NULL;
156         struct dentry *dentry;
157
158         _enter(",%p", object);
159
160 try_again:
161         write_lock(&cache->active_lock);
162
163         if (test_and_set_bit(CACHEFILES_OBJECT_ACTIVE, &object->flags)) {
164                 printk(KERN_ERR "CacheFiles: Error: Object already active\n");
165                 cachefiles_printk_object(object, NULL);
166                 BUG();
167         }
168
169         dentry = object->dentry;
170         _p = &cache->active_nodes.rb_node;
171         while (*_p) {
172                 _parent = *_p;
173                 xobject = rb_entry(_parent,
174                                    struct cachefiles_object, active_node);
175
176                 ASSERT(xobject != object);
177
178                 if (xobject->dentry > dentry)
179                         _p = &(*_p)->rb_left;
180                 else if (xobject->dentry < dentry)
181                         _p = &(*_p)->rb_right;
182                 else
183                         goto wait_for_old_object;
184         }
185
186         rb_link_node(&object->active_node, _parent, _p);
187         rb_insert_color(&object->active_node, &cache->active_nodes);
188
189         write_unlock(&cache->active_lock);
190         _leave(" = 0");
191         return 0;
192
193         /* an old object from a previous incarnation is hogging the slot - we
194          * need to wait for it to be destroyed */
195 wait_for_old_object:
196         if (xobject->fscache.state < FSCACHE_OBJECT_DYING) {
197                 printk(KERN_ERR "\n");
198                 printk(KERN_ERR "CacheFiles: Error:"
199                        " Unexpected object collision\n");
200                 cachefiles_printk_object(object, xobject);
201                 BUG();
202         }
203         atomic_inc(&xobject->usage);
204         write_unlock(&cache->active_lock);
205
206         if (test_bit(CACHEFILES_OBJECT_ACTIVE, &xobject->flags)) {
207                 wait_queue_head_t *wq;
208
209                 signed long timeout = 60 * HZ;
210                 wait_queue_t wait;
211                 bool requeue;
212
213                 /* if the object we're waiting for is queued for processing,
214                  * then just put ourselves on the queue behind it */
215                 if (slow_work_is_queued(&xobject->fscache.work)) {
216                         _debug("queue OBJ%x behind OBJ%x immediately",
217                                object->fscache.debug_id,
218                                xobject->fscache.debug_id);
219                         goto requeue;
220                 }
221
222                 /* otherwise we sleep until either the object we're waiting for
223                  * is done, or the slow-work facility wants the thread back to
224                  * do other work */
225                 wq = bit_waitqueue(&xobject->flags, CACHEFILES_OBJECT_ACTIVE);
226                 init_wait(&wait);
227                 requeue = false;
228                 do {
229                         prepare_to_wait(wq, &wait, TASK_UNINTERRUPTIBLE);
230                         if (!test_bit(CACHEFILES_OBJECT_ACTIVE, &xobject->flags))
231                                 break;
232                         requeue = slow_work_sleep_till_thread_needed(
233                                 &object->fscache.work, &timeout);
234                 } while (timeout > 0 && !requeue);
235                 finish_wait(wq, &wait);
236
237                 if (requeue &&
238                     test_bit(CACHEFILES_OBJECT_ACTIVE, &xobject->flags)) {
239                         _debug("queue OBJ%x behind OBJ%x after wait",
240                                object->fscache.debug_id,
241                                xobject->fscache.debug_id);
242                         goto requeue;
243                 }
244
245                 if (timeout <= 0) {
246                         printk(KERN_ERR "\n");
247                         printk(KERN_ERR "CacheFiles: Error: Overlong"
248                                " wait for old active object to go away\n");
249                         cachefiles_printk_object(object, xobject);
250                         goto requeue;
251                 }
252         }
253
254         ASSERT(!test_bit(CACHEFILES_OBJECT_ACTIVE, &xobject->flags));
255
256         cache->cache.ops->put_object(&xobject->fscache);
257         goto try_again;
258
259 requeue:
260         clear_bit(CACHEFILES_OBJECT_ACTIVE, &object->flags);
261         cache->cache.ops->put_object(&xobject->fscache);
262         _leave(" = -ETIMEDOUT");
263         return -ETIMEDOUT;
264 }
265
266 /*
267  * delete an object representation from the cache
268  * - file backed objects are unlinked
269  * - directory backed objects are stuffed into the graveyard for userspace to
270  *   delete
271  * - unlocks the directory mutex
272  */
273 static int cachefiles_bury_object(struct cachefiles_cache *cache,
274                                   struct dentry *dir,
275                                   struct dentry *rep,
276                                   bool preemptive)
277 {
278         struct dentry *grave, *trap;
279         char nbuffer[8 + 8 + 1];
280         int ret;
281
282         _enter(",'%*.*s','%*.*s'",
283                dir->d_name.len, dir->d_name.len, dir->d_name.name,
284                rep->d_name.len, rep->d_name.len, rep->d_name.name);
285
286         _debug("remove %p from %p", rep, dir);
287
288         /* non-directories can just be unlinked */
289         if (!S_ISDIR(rep->d_inode->i_mode)) {
290                 _debug("unlink stale object");
291                 ret = vfs_unlink(dir->d_inode, rep);
292
293                 if (preemptive)
294                         cachefiles_mark_object_buried(cache, rep);
295
296                 mutex_unlock(&dir->d_inode->i_mutex);
297
298                 if (ret == -EIO)
299                         cachefiles_io_error(cache, "Unlink failed");
300
301                 _leave(" = %d", ret);
302                 return ret;
303         }
304
305         /* directories have to be moved to the graveyard */
306         _debug("move stale object to graveyard");
307         mutex_unlock(&dir->d_inode->i_mutex);
308
309 try_again:
310         /* first step is to make up a grave dentry in the graveyard */
311         sprintf(nbuffer, "%08x%08x",
312                 (uint32_t) get_seconds(),
313                 (uint32_t) atomic_inc_return(&cache->gravecounter));
314
315         /* do the multiway lock magic */
316         trap = lock_rename(cache->graveyard, dir);
317
318         /* do some checks before getting the grave dentry */
319         if (rep->d_parent != dir) {
320                 /* the entry was probably culled when we dropped the parent dir
321                  * lock */
322                 unlock_rename(cache->graveyard, dir);
323                 _leave(" = 0 [culled?]");
324                 return 0;
325         }
326
327         if (!S_ISDIR(cache->graveyard->d_inode->i_mode)) {
328                 unlock_rename(cache->graveyard, dir);
329                 cachefiles_io_error(cache, "Graveyard no longer a directory");
330                 return -EIO;
331         }
332
333         if (trap == rep) {
334                 unlock_rename(cache->graveyard, dir);
335                 cachefiles_io_error(cache, "May not make directory loop");
336                 return -EIO;
337         }
338
339         if (d_mountpoint(rep)) {
340                 unlock_rename(cache->graveyard, dir);
341                 cachefiles_io_error(cache, "Mountpoint in cache");
342                 return -EIO;
343         }
344
345         grave = lookup_one_len(nbuffer, cache->graveyard, strlen(nbuffer));
346         if (IS_ERR(grave)) {
347                 unlock_rename(cache->graveyard, dir);
348
349                 if (PTR_ERR(grave) == -ENOMEM) {
350                         _leave(" = -ENOMEM");
351                         return -ENOMEM;
352                 }
353
354                 cachefiles_io_error(cache, "Lookup error %ld",
355                                     PTR_ERR(grave));
356                 return -EIO;
357         }
358
359         if (grave->d_inode) {
360                 unlock_rename(cache->graveyard, dir);
361                 dput(grave);
362                 grave = NULL;
363                 cond_resched();
364                 goto try_again;
365         }
366
367         if (d_mountpoint(grave)) {
368                 unlock_rename(cache->graveyard, dir);
369                 dput(grave);
370                 cachefiles_io_error(cache, "Mountpoint in graveyard");
371                 return -EIO;
372         }
373
374         /* target should not be an ancestor of source */
375         if (trap == grave) {
376                 unlock_rename(cache->graveyard, dir);
377                 dput(grave);
378                 cachefiles_io_error(cache, "May not make directory loop");
379                 return -EIO;
380         }
381
382         /* attempt the rename */
383         ret = vfs_rename(dir->d_inode, rep, cache->graveyard->d_inode, grave);
384         if (ret != 0 && ret != -ENOMEM)
385                 cachefiles_io_error(cache, "Rename failed with error %d", ret);
386
387         if (preemptive)
388                 cachefiles_mark_object_buried(cache, rep);
389
390         unlock_rename(cache->graveyard, dir);
391         dput(grave);
392         _leave(" = 0");
393         return 0;
394 }
395
396 /*
397  * delete an object representation from the cache
398  */
399 int cachefiles_delete_object(struct cachefiles_cache *cache,
400                              struct cachefiles_object *object)
401 {
402         struct dentry *dir;
403         int ret;
404
405         _enter(",OBJ%x{%p}", object->fscache.debug_id, object->dentry);
406
407         ASSERT(object->dentry);
408         ASSERT(object->dentry->d_inode);
409         ASSERT(object->dentry->d_parent);
410
411         dir = dget_parent(object->dentry);
412
413         mutex_lock_nested(&dir->d_inode->i_mutex, I_MUTEX_PARENT);
414
415         if (test_bit(CACHEFILES_OBJECT_BURIED, &object->flags)) {
416                 /* object allocation for the same key preemptively deleted this
417                  * object's file so that it could create its own file */
418                 _debug("object preemptively buried");
419                 mutex_unlock(&dir->d_inode->i_mutex);
420                 ret = 0;
421         } else {
422                 /* we need to check that our parent is _still_ our parent - it
423                  * may have been renamed */
424                 if (dir == object->dentry->d_parent) {
425                         ret = cachefiles_bury_object(cache, dir,
426                                                      object->dentry, false);
427                 } else {
428                         /* it got moved, presumably by cachefilesd culling it,
429                          * so it's no longer in the key path and we can ignore
430                          * it */
431                         mutex_unlock(&dir->d_inode->i_mutex);
432                         ret = 0;
433                 }
434         }
435
436         dput(dir);
437         _leave(" = %d", ret);
438         return ret;
439 }
440
441 /*
442  * walk from the parent object to the child object through the backing
443  * filesystem, creating directories as we go
444  */
445 int cachefiles_walk_to_object(struct cachefiles_object *parent,
446                               struct cachefiles_object *object,
447                               const char *key,
448                               struct cachefiles_xattr *auxdata)
449 {
450         struct cachefiles_cache *cache;
451         struct dentry *dir, *next = NULL;
452         unsigned long start;
453         const char *name;
454         int ret, nlen;
455
456         _enter("OBJ%x{%p},OBJ%x,%s,",
457                parent->fscache.debug_id, parent->dentry,
458                object->fscache.debug_id, key);
459
460         cache = container_of(parent->fscache.cache,
461                              struct cachefiles_cache, cache);
462
463         ASSERT(parent->dentry);
464         ASSERT(parent->dentry->d_inode);
465
466         if (!(S_ISDIR(parent->dentry->d_inode->i_mode))) {
467                 // TODO: convert file to dir
468                 _leave("looking up in none directory");
469                 return -ENOBUFS;
470         }
471
472         dir = dget(parent->dentry);
473
474 advance:
475         /* attempt to transit the first directory component */
476         name = key;
477         nlen = strlen(key);
478
479         /* key ends in a double NUL */
480         key = key + nlen + 1;
481         if (!*key)
482                 key = NULL;
483
484 lookup_again:
485         /* search the current directory for the element name */
486         _debug("lookup '%s'", name);
487
488         mutex_lock_nested(&dir->d_inode->i_mutex, I_MUTEX_PARENT);
489
490         start = jiffies;
491         next = lookup_one_len(name, dir, nlen);
492         cachefiles_hist(cachefiles_lookup_histogram, start);
493         if (IS_ERR(next))
494                 goto lookup_error;
495
496         _debug("next -> %p %s", next, next->d_inode ? "positive" : "negative");
497
498         if (!key)
499                 object->new = !next->d_inode;
500
501         /* if this element of the path doesn't exist, then the lookup phase
502          * failed, and we can release any readers in the certain knowledge that
503          * there's nothing for them to actually read */
504         if (!next->d_inode)
505                 fscache_object_lookup_negative(&object->fscache);
506
507         /* we need to create the object if it's negative */
508         if (key || object->type == FSCACHE_COOKIE_TYPE_INDEX) {
509                 /* index objects and intervening tree levels must be subdirs */
510                 if (!next->d_inode) {
511                         ret = cachefiles_has_space(cache, 1, 0);
512                         if (ret < 0)
513                                 goto create_error;
514
515                         start = jiffies;
516                         ret = vfs_mkdir(dir->d_inode, next, 0);
517                         cachefiles_hist(cachefiles_mkdir_histogram, start);
518                         if (ret < 0)
519                                 goto create_error;
520
521                         ASSERT(next->d_inode);
522
523                         _debug("mkdir -> %p{%p{ino=%lu}}",
524                                next, next->d_inode, next->d_inode->i_ino);
525
526                 } else if (!S_ISDIR(next->d_inode->i_mode)) {
527                         kerror("inode %lu is not a directory",
528                                next->d_inode->i_ino);
529                         ret = -ENOBUFS;
530                         goto error;
531                 }
532
533         } else {
534                 /* non-index objects start out life as files */
535                 if (!next->d_inode) {
536                         ret = cachefiles_has_space(cache, 1, 0);
537                         if (ret < 0)
538                                 goto create_error;
539
540                         start = jiffies;
541                         ret = vfs_create(dir->d_inode, next, S_IFREG, NULL);
542                         cachefiles_hist(cachefiles_create_histogram, start);
543                         if (ret < 0)
544                                 goto create_error;
545
546                         ASSERT(next->d_inode);
547
548                         _debug("create -> %p{%p{ino=%lu}}",
549                                next, next->d_inode, next->d_inode->i_ino);
550
551                 } else if (!S_ISDIR(next->d_inode->i_mode) &&
552                            !S_ISREG(next->d_inode->i_mode)
553                            ) {
554                         kerror("inode %lu is not a file or directory",
555                                next->d_inode->i_ino);
556                         ret = -ENOBUFS;
557                         goto error;
558                 }
559         }
560
561         /* process the next component */
562         if (key) {
563                 _debug("advance");
564                 mutex_unlock(&dir->d_inode->i_mutex);
565                 dput(dir);
566                 dir = next;
567                 next = NULL;
568                 goto advance;
569         }
570
571         /* we've found the object we were looking for */
572         object->dentry = next;
573
574         /* if we've found that the terminal object exists, then we need to
575          * check its attributes and delete it if it's out of date */
576         if (!object->new) {
577                 _debug("validate '%*.*s'",
578                        next->d_name.len, next->d_name.len, next->d_name.name);
579
580                 ret = cachefiles_check_object_xattr(object, auxdata);
581                 if (ret == -ESTALE) {
582                         /* delete the object (the deleter drops the directory
583                          * mutex) */
584                         object->dentry = NULL;
585
586                         ret = cachefiles_bury_object(cache, dir, next, true);
587                         dput(next);
588                         next = NULL;
589
590                         if (ret < 0)
591                                 goto delete_error;
592
593                         _debug("redo lookup");
594                         goto lookup_again;
595                 }
596         }
597
598         /* note that we're now using this object */
599         ret = cachefiles_mark_object_active(cache, object);
600
601         mutex_unlock(&dir->d_inode->i_mutex);
602         dput(dir);
603         dir = NULL;
604
605         if (ret == -ETIMEDOUT)
606                 goto mark_active_timed_out;
607
608         _debug("=== OBTAINED_OBJECT ===");
609
610         if (object->new) {
611                 /* attach data to a newly constructed terminal object */
612                 ret = cachefiles_set_object_xattr(object, auxdata);
613                 if (ret < 0)
614                         goto check_error;
615         } else {
616                 /* always update the atime on an object we've just looked up
617                  * (this is used to keep track of culling, and atimes are only
618                  * updated by read, write and readdir but not lookup or
619                  * open) */
620                 touch_atime(cache->mnt, next);
621         }
622
623         /* open a file interface onto a data file */
624         if (object->type != FSCACHE_COOKIE_TYPE_INDEX) {
625                 if (S_ISREG(object->dentry->d_inode->i_mode)) {
626                         const struct address_space_operations *aops;
627
628                         ret = -EPERM;
629                         aops = object->dentry->d_inode->i_mapping->a_ops;
630                         if (!aops->bmap)
631                                 goto check_error;
632
633                         object->backer = object->dentry;
634                 } else {
635                         BUG(); // TODO: open file in data-class subdir
636                 }
637         }
638
639         object->new = 0;
640         fscache_obtained_object(&object->fscache);
641
642         _leave(" = 0 [%lu]", object->dentry->d_inode->i_ino);
643         return 0;
644
645 create_error:
646         _debug("create error %d", ret);
647         if (ret == -EIO)
648                 cachefiles_io_error(cache, "Create/mkdir failed");
649         goto error;
650
651 mark_active_timed_out:
652         _debug("mark active timed out");
653         goto release_dentry;
654
655 check_error:
656         _debug("check error %d", ret);
657         write_lock(&cache->active_lock);
658         rb_erase(&object->active_node, &cache->active_nodes);
659         clear_bit(CACHEFILES_OBJECT_ACTIVE, &object->flags);
660         wake_up_bit(&object->flags, CACHEFILES_OBJECT_ACTIVE);
661         write_unlock(&cache->active_lock);
662 release_dentry:
663         dput(object->dentry);
664         object->dentry = NULL;
665         goto error_out;
666
667 delete_error:
668         _debug("delete error %d", ret);
669         goto error_out2;
670
671 lookup_error:
672         _debug("lookup error %ld", PTR_ERR(next));
673         ret = PTR_ERR(next);
674         if (ret == -EIO)
675                 cachefiles_io_error(cache, "Lookup failed");
676         next = NULL;
677 error:
678         mutex_unlock(&dir->d_inode->i_mutex);
679         dput(next);
680 error_out2:
681         dput(dir);
682 error_out:
683         _leave(" = error %d", -ret);
684         return ret;
685 }
686
687 /*
688  * get a subdirectory
689  */
690 struct dentry *cachefiles_get_directory(struct cachefiles_cache *cache,
691                                         struct dentry *dir,
692                                         const char *dirname)
693 {
694         struct dentry *subdir;
695         unsigned long start;
696         int ret;
697
698         _enter(",,%s", dirname);
699
700         /* search the current directory for the element name */
701         mutex_lock(&dir->d_inode->i_mutex);
702
703         start = jiffies;
704         subdir = lookup_one_len(dirname, dir, strlen(dirname));
705         cachefiles_hist(cachefiles_lookup_histogram, start);
706         if (IS_ERR(subdir)) {
707                 if (PTR_ERR(subdir) == -ENOMEM)
708                         goto nomem_d_alloc;
709                 goto lookup_error;
710         }
711
712         _debug("subdir -> %p %s",
713                subdir, subdir->d_inode ? "positive" : "negative");
714
715         /* we need to create the subdir if it doesn't exist yet */
716         if (!subdir->d_inode) {
717                 ret = cachefiles_has_space(cache, 1, 0);
718                 if (ret < 0)
719                         goto mkdir_error;
720
721                 _debug("attempt mkdir");
722
723                 ret = vfs_mkdir(dir->d_inode, subdir, 0700);
724                 if (ret < 0)
725                         goto mkdir_error;
726
727                 ASSERT(subdir->d_inode);
728
729                 _debug("mkdir -> %p{%p{ino=%lu}}",
730                        subdir,
731                        subdir->d_inode,
732                        subdir->d_inode->i_ino);
733         }
734
735         mutex_unlock(&dir->d_inode->i_mutex);
736
737         /* we need to make sure the subdir is a directory */
738         ASSERT(subdir->d_inode);
739
740         if (!S_ISDIR(subdir->d_inode->i_mode)) {
741                 kerror("%s is not a directory", dirname);
742                 ret = -EIO;
743                 goto check_error;
744         }
745
746         ret = -EPERM;
747         if (!subdir->d_inode->i_op ||
748             !subdir->d_inode->i_op->setxattr ||
749             !subdir->d_inode->i_op->getxattr ||
750             !subdir->d_inode->i_op->lookup ||
751             !subdir->d_inode->i_op->mkdir ||
752             !subdir->d_inode->i_op->create ||
753             !subdir->d_inode->i_op->rename ||
754             !subdir->d_inode->i_op->rmdir ||
755             !subdir->d_inode->i_op->unlink)
756                 goto check_error;
757
758         _leave(" = [%lu]", subdir->d_inode->i_ino);
759         return subdir;
760
761 check_error:
762         dput(subdir);
763         _leave(" = %d [check]", ret);
764         return ERR_PTR(ret);
765
766 mkdir_error:
767         mutex_unlock(&dir->d_inode->i_mutex);
768         dput(subdir);
769         kerror("mkdir %s failed with error %d", dirname, ret);
770         return ERR_PTR(ret);
771
772 lookup_error:
773         mutex_unlock(&dir->d_inode->i_mutex);
774         ret = PTR_ERR(subdir);
775         kerror("Lookup %s failed with error %d", dirname, ret);
776         return ERR_PTR(ret);
777
778 nomem_d_alloc:
779         mutex_unlock(&dir->d_inode->i_mutex);
780         _leave(" = -ENOMEM");
781         return ERR_PTR(-ENOMEM);
782 }
783
784 /*
785  * find out if an object is in use or not
786  * - if finds object and it's not in use:
787  *   - returns a pointer to the object and a reference on it
788  *   - returns with the directory locked
789  */
790 static struct dentry *cachefiles_check_active(struct cachefiles_cache *cache,
791                                               struct dentry *dir,
792                                               char *filename)
793 {
794         struct cachefiles_object *object;
795         struct rb_node *_n;
796         struct dentry *victim;
797         unsigned long start;
798         int ret;
799
800         //_enter(",%*.*s/,%s",
801         //       dir->d_name.len, dir->d_name.len, dir->d_name.name, filename);
802
803         /* look up the victim */
804         mutex_lock_nested(&dir->d_inode->i_mutex, 1);
805
806         start = jiffies;
807         victim = lookup_one_len(filename, dir, strlen(filename));
808         cachefiles_hist(cachefiles_lookup_histogram, start);
809         if (IS_ERR(victim))
810                 goto lookup_error;
811
812         //_debug("victim -> %p %s",
813         //       victim, victim->d_inode ? "positive" : "negative");
814
815         /* if the object is no longer there then we probably retired the object
816          * at the netfs's request whilst the cull was in progress
817          */
818         if (!victim->d_inode) {
819                 mutex_unlock(&dir->d_inode->i_mutex);
820                 dput(victim);
821                 _leave(" = -ENOENT [absent]");
822                 return ERR_PTR(-ENOENT);
823         }
824
825         /* check to see if we're using this object */
826         read_lock(&cache->active_lock);
827
828         _n = cache->active_nodes.rb_node;
829
830         while (_n) {
831                 object = rb_entry(_n, struct cachefiles_object, active_node);
832
833                 if (object->dentry > victim)
834                         _n = _n->rb_left;
835                 else if (object->dentry < victim)
836                         _n = _n->rb_right;
837                 else
838                         goto object_in_use;
839         }
840
841         read_unlock(&cache->active_lock);
842
843         //_leave(" = %p", victim);
844         return victim;
845
846 object_in_use:
847         read_unlock(&cache->active_lock);
848         mutex_unlock(&dir->d_inode->i_mutex);
849         dput(victim);
850         //_leave(" = -EBUSY [in use]");
851         return ERR_PTR(-EBUSY);
852
853 lookup_error:
854         mutex_unlock(&dir->d_inode->i_mutex);
855         ret = PTR_ERR(victim);
856         if (ret == -ENOENT) {
857                 /* file or dir now absent - probably retired by netfs */
858                 _leave(" = -ESTALE [absent]");
859                 return ERR_PTR(-ESTALE);
860         }
861
862         if (ret == -EIO) {
863                 cachefiles_io_error(cache, "Lookup failed");
864         } else if (ret != -ENOMEM) {
865                 kerror("Internal error: %d", ret);
866                 ret = -EIO;
867         }
868
869         _leave(" = %d", ret);
870         return ERR_PTR(ret);
871 }
872
873 /*
874  * cull an object if it's not in use
875  * - called only by cache manager daemon
876  */
877 int cachefiles_cull(struct cachefiles_cache *cache, struct dentry *dir,
878                     char *filename)
879 {
880         struct dentry *victim;
881         int ret;
882
883         _enter(",%*.*s/,%s",
884                dir->d_name.len, dir->d_name.len, dir->d_name.name, filename);
885
886         victim = cachefiles_check_active(cache, dir, filename);
887         if (IS_ERR(victim))
888                 return PTR_ERR(victim);
889
890         _debug("victim -> %p %s",
891                victim, victim->d_inode ? "positive" : "negative");
892
893         /* okay... the victim is not being used so we can cull it
894          * - start by marking it as stale
895          */
896         _debug("victim is cullable");
897
898         ret = cachefiles_remove_object_xattr(cache, victim);
899         if (ret < 0)
900                 goto error_unlock;
901
902         /*  actually remove the victim (drops the dir mutex) */
903         _debug("bury");
904
905         ret = cachefiles_bury_object(cache, dir, victim, false);
906         if (ret < 0)
907                 goto error;
908
909         dput(victim);
910         _leave(" = 0");
911         return 0;
912
913 error_unlock:
914         mutex_unlock(&dir->d_inode->i_mutex);
915 error:
916         dput(victim);
917         if (ret == -ENOENT) {
918                 /* file or dir now absent - probably retired by netfs */
919                 _leave(" = -ESTALE [absent]");
920                 return -ESTALE;
921         }
922
923         if (ret != -ENOMEM) {
924                 kerror("Internal error: %d", ret);
925                 ret = -EIO;
926         }
927
928         _leave(" = %d", ret);
929         return ret;
930 }
931
932 /*
933  * find out if an object is in use or not
934  * - called only by cache manager daemon
935  * - returns -EBUSY or 0 to indicate whether an object is in use or not
936  */
937 int cachefiles_check_in_use(struct cachefiles_cache *cache, struct dentry *dir,
938                             char *filename)
939 {
940         struct dentry *victim;
941
942         //_enter(",%*.*s/,%s",
943         //       dir->d_name.len, dir->d_name.len, dir->d_name.name, filename);
944
945         victim = cachefiles_check_active(cache, dir, filename);
946         if (IS_ERR(victim))
947                 return PTR_ERR(victim);
948
949         mutex_unlock(&dir->d_inode->i_mutex);
950         dput(victim);
951         //_leave(" = 0");
952         return 0;
953 }