[PATCH] knfsd: svcrpc: remove another silent drop from deferral code
[linux-2.6.git] / net / sunrpc / cache.c
1 /*
2  * net/sunrpc/cache.c
3  *
4  * Generic code for various authentication-related caches
5  * used by sunrpc clients and servers.
6  *
7  * Copyright (C) 2002 Neil Brown <neilb@cse.unsw.edu.au>
8  *
9  * Released under terms in GPL version 2.  See COPYING.
10  *
11  */
12
13 #include <linux/types.h>
14 #include <linux/fs.h>
15 #include <linux/file.h>
16 #include <linux/slab.h>
17 #include <linux/signal.h>
18 #include <linux/sched.h>
19 #include <linux/kmod.h>
20 #include <linux/list.h>
21 #include <linux/module.h>
22 #include <linux/ctype.h>
23 #include <asm/uaccess.h>
24 #include <linux/poll.h>
25 #include <linux/seq_file.h>
26 #include <linux/proc_fs.h>
27 #include <linux/net.h>
28 #include <linux/workqueue.h>
29 #include <linux/mutex.h>
30 #include <asm/ioctls.h>
31 #include <linux/sunrpc/types.h>
32 #include <linux/sunrpc/cache.h>
33 #include <linux/sunrpc/stats.h>
34
35 #define  RPCDBG_FACILITY RPCDBG_CACHE
36
37 static int cache_defer_req(struct cache_req *req, struct cache_head *item);
38 static void cache_revisit_request(struct cache_head *item);
39
40 static void cache_init(struct cache_head *h)
41 {
42         time_t now = get_seconds();
43         h->next = NULL;
44         h->flags = 0;
45         kref_init(&h->ref);
46         h->expiry_time = now + CACHE_NEW_EXPIRY;
47         h->last_refresh = now;
48 }
49
50 struct cache_head *sunrpc_cache_lookup(struct cache_detail *detail,
51                                        struct cache_head *key, int hash)
52 {
53         struct cache_head **head,  **hp;
54         struct cache_head *new = NULL;
55
56         head = &detail->hash_table[hash];
57
58         read_lock(&detail->hash_lock);
59
60         for (hp=head; *hp != NULL ; hp = &(*hp)->next) {
61                 struct cache_head *tmp = *hp;
62                 if (detail->match(tmp, key)) {
63                         cache_get(tmp);
64                         read_unlock(&detail->hash_lock);
65                         return tmp;
66                 }
67         }
68         read_unlock(&detail->hash_lock);
69         /* Didn't find anything, insert an empty entry */
70
71         new = detail->alloc();
72         if (!new)
73                 return NULL;
74         /* must fully initialise 'new', else
75          * we might get lose if we need to
76          * cache_put it soon.
77          */
78         cache_init(new);
79         detail->init(new, key);
80
81         write_lock(&detail->hash_lock);
82
83         /* check if entry appeared while we slept */
84         for (hp=head; *hp != NULL ; hp = &(*hp)->next) {
85                 struct cache_head *tmp = *hp;
86                 if (detail->match(tmp, key)) {
87                         cache_get(tmp);
88                         write_unlock(&detail->hash_lock);
89                         cache_put(new, detail);
90                         return tmp;
91                 }
92         }
93         new->next = *head;
94         *head = new;
95         detail->entries++;
96         cache_get(new);
97         write_unlock(&detail->hash_lock);
98
99         return new;
100 }
101 EXPORT_SYMBOL(sunrpc_cache_lookup);
102
103
104 static void queue_loose(struct cache_detail *detail, struct cache_head *ch);
105
106 static int cache_fresh_locked(struct cache_head *head, time_t expiry)
107 {
108         head->expiry_time = expiry;
109         head->last_refresh = get_seconds();
110         return !test_and_set_bit(CACHE_VALID, &head->flags);
111 }
112
113 static void cache_fresh_unlocked(struct cache_head *head,
114                         struct cache_detail *detail, int new)
115 {
116         if (new)
117                 cache_revisit_request(head);
118         if (test_and_clear_bit(CACHE_PENDING, &head->flags)) {
119                 cache_revisit_request(head);
120                 queue_loose(detail, head);
121         }
122 }
123
124 struct cache_head *sunrpc_cache_update(struct cache_detail *detail,
125                                        struct cache_head *new, struct cache_head *old, int hash)
126 {
127         /* The 'old' entry is to be replaced by 'new'.
128          * If 'old' is not VALID, we update it directly,
129          * otherwise we need to replace it
130          */
131         struct cache_head **head;
132         struct cache_head *tmp;
133         int is_new;
134
135         if (!test_bit(CACHE_VALID, &old->flags)) {
136                 write_lock(&detail->hash_lock);
137                 if (!test_bit(CACHE_VALID, &old->flags)) {
138                         if (test_bit(CACHE_NEGATIVE, &new->flags))
139                                 set_bit(CACHE_NEGATIVE, &old->flags);
140                         else
141                                 detail->update(old, new);
142                         is_new = cache_fresh_locked(old, new->expiry_time);
143                         write_unlock(&detail->hash_lock);
144                         cache_fresh_unlocked(old, detail, is_new);
145                         return old;
146                 }
147                 write_unlock(&detail->hash_lock);
148         }
149         /* We need to insert a new entry */
150         tmp = detail->alloc();
151         if (!tmp) {
152                 cache_put(old, detail);
153                 return NULL;
154         }
155         cache_init(tmp);
156         detail->init(tmp, old);
157         head = &detail->hash_table[hash];
158
159         write_lock(&detail->hash_lock);
160         if (test_bit(CACHE_NEGATIVE, &new->flags))
161                 set_bit(CACHE_NEGATIVE, &tmp->flags);
162         else
163                 detail->update(tmp, new);
164         tmp->next = *head;
165         *head = tmp;
166         detail->entries++;
167         cache_get(tmp);
168         is_new = cache_fresh_locked(tmp, new->expiry_time);
169         cache_fresh_locked(old, 0);
170         write_unlock(&detail->hash_lock);
171         cache_fresh_unlocked(tmp, detail, is_new);
172         cache_fresh_unlocked(old, detail, 0);
173         cache_put(old, detail);
174         return tmp;
175 }
176 EXPORT_SYMBOL(sunrpc_cache_update);
177
178 static int cache_make_upcall(struct cache_detail *detail, struct cache_head *h);
179 /*
180  * This is the generic cache management routine for all
181  * the authentication caches.
182  * It checks the currency of a cache item and will (later)
183  * initiate an upcall to fill it if needed.
184  *
185  *
186  * Returns 0 if the cache_head can be used, or cache_puts it and returns
187  * -EAGAIN if upcall is pending,
188  * -ETIMEDOUT if upcall failed and should be retried,
189  * -ENOENT if cache entry was negative
190  */
191 int cache_check(struct cache_detail *detail,
192                     struct cache_head *h, struct cache_req *rqstp)
193 {
194         int rv;
195         long refresh_age, age;
196
197         /* First decide return status as best we can */
198         if (!test_bit(CACHE_VALID, &h->flags) ||
199             h->expiry_time < get_seconds())
200                 rv = -EAGAIN;
201         else if (detail->flush_time > h->last_refresh)
202                 rv = -EAGAIN;
203         else {
204                 /* entry is valid */
205                 if (test_bit(CACHE_NEGATIVE, &h->flags))
206                         rv = -ENOENT;
207                 else rv = 0;
208         }
209
210         /* now see if we want to start an upcall */
211         refresh_age = (h->expiry_time - h->last_refresh);
212         age = get_seconds() - h->last_refresh;
213
214         if (rqstp == NULL) {
215                 if (rv == -EAGAIN)
216                         rv = -ENOENT;
217         } else if (rv == -EAGAIN || age > refresh_age/2) {
218                 dprintk("Want update, refage=%ld, age=%ld\n", refresh_age, age);
219                 if (!test_and_set_bit(CACHE_PENDING, &h->flags)) {
220                         switch (cache_make_upcall(detail, h)) {
221                         case -EINVAL:
222                                 clear_bit(CACHE_PENDING, &h->flags);
223                                 if (rv == -EAGAIN) {
224                                         set_bit(CACHE_NEGATIVE, &h->flags);
225                                         cache_fresh_unlocked(h, detail,
226                                              cache_fresh_locked(h, get_seconds()+CACHE_NEW_EXPIRY));
227                                         rv = -ENOENT;
228                                 }
229                                 break;
230
231                         case -EAGAIN:
232                                 clear_bit(CACHE_PENDING, &h->flags);
233                                 cache_revisit_request(h);
234                                 break;
235                         }
236                 }
237         }
238
239         if (rv == -EAGAIN)
240                 if (cache_defer_req(rqstp, h) != 0)
241                         rv = -ETIMEDOUT;
242
243         if (rv)
244                 cache_put(h, detail);
245         return rv;
246 }
247
248 /*
249  * caches need to be periodically cleaned.
250  * For this we maintain a list of cache_detail and
251  * a current pointer into that list and into the table
252  * for that entry.
253  *
254  * Each time clean_cache is called it finds the next non-empty entry
255  * in the current table and walks the list in that entry
256  * looking for entries that can be removed.
257  *
258  * An entry gets removed if:
259  * - The expiry is before current time
260  * - The last_refresh time is before the flush_time for that cache
261  *
262  * later we might drop old entries with non-NEVER expiry if that table
263  * is getting 'full' for some definition of 'full'
264  *
265  * The question of "how often to scan a table" is an interesting one
266  * and is answered in part by the use of the "nextcheck" field in the
267  * cache_detail.
268  * When a scan of a table begins, the nextcheck field is set to a time
269  * that is well into the future.
270  * While scanning, if an expiry time is found that is earlier than the
271  * current nextcheck time, nextcheck is set to that expiry time.
272  * If the flush_time is ever set to a time earlier than the nextcheck
273  * time, the nextcheck time is then set to that flush_time.
274  *
275  * A table is then only scanned if the current time is at least
276  * the nextcheck time.
277  * 
278  */
279
280 static LIST_HEAD(cache_list);
281 static DEFINE_SPINLOCK(cache_list_lock);
282 static struct cache_detail *current_detail;
283 static int current_index;
284
285 static struct file_operations cache_file_operations;
286 static struct file_operations content_file_operations;
287 static struct file_operations cache_flush_operations;
288
289 static void do_cache_clean(struct work_struct *work);
290 static DECLARE_DELAYED_WORK(cache_cleaner, do_cache_clean);
291
292 void cache_register(struct cache_detail *cd)
293 {
294         cd->proc_ent = proc_mkdir(cd->name, proc_net_rpc);
295         if (cd->proc_ent) {
296                 struct proc_dir_entry *p;
297                 cd->proc_ent->owner = cd->owner;
298                 cd->channel_ent = cd->content_ent = NULL;
299                 
300                 p = create_proc_entry("flush", S_IFREG|S_IRUSR|S_IWUSR,
301                                       cd->proc_ent);
302                 cd->flush_ent =  p;
303                 if (p) {
304                         p->proc_fops = &cache_flush_operations;
305                         p->owner = cd->owner;
306                         p->data = cd;
307                 }
308  
309                 if (cd->cache_request || cd->cache_parse) {
310                         p = create_proc_entry("channel", S_IFREG|S_IRUSR|S_IWUSR,
311                                               cd->proc_ent);
312                         cd->channel_ent = p;
313                         if (p) {
314                                 p->proc_fops = &cache_file_operations;
315                                 p->owner = cd->owner;
316                                 p->data = cd;
317                         }
318                 }
319                 if (cd->cache_show) {
320                         p = create_proc_entry("content", S_IFREG|S_IRUSR|S_IWUSR,
321                                               cd->proc_ent);
322                         cd->content_ent = p;
323                         if (p) {
324                                 p->proc_fops = &content_file_operations;
325                                 p->owner = cd->owner;
326                                 p->data = cd;
327                         }
328                 }
329         }
330         rwlock_init(&cd->hash_lock);
331         INIT_LIST_HEAD(&cd->queue);
332         spin_lock(&cache_list_lock);
333         cd->nextcheck = 0;
334         cd->entries = 0;
335         atomic_set(&cd->readers, 0);
336         cd->last_close = 0;
337         cd->last_warn = -1;
338         list_add(&cd->others, &cache_list);
339         spin_unlock(&cache_list_lock);
340
341         /* start the cleaning process */
342         schedule_delayed_work(&cache_cleaner, 0);
343 }
344
345 int cache_unregister(struct cache_detail *cd)
346 {
347         cache_purge(cd);
348         spin_lock(&cache_list_lock);
349         write_lock(&cd->hash_lock);
350         if (cd->entries || atomic_read(&cd->inuse)) {
351                 write_unlock(&cd->hash_lock);
352                 spin_unlock(&cache_list_lock);
353                 return -EBUSY;
354         }
355         if (current_detail == cd)
356                 current_detail = NULL;
357         list_del_init(&cd->others);
358         write_unlock(&cd->hash_lock);
359         spin_unlock(&cache_list_lock);
360         if (cd->proc_ent) {
361                 if (cd->flush_ent)
362                         remove_proc_entry("flush", cd->proc_ent);
363                 if (cd->channel_ent)
364                         remove_proc_entry("channel", cd->proc_ent);
365                 if (cd->content_ent)
366                         remove_proc_entry("content", cd->proc_ent);
367
368                 cd->proc_ent = NULL;
369                 remove_proc_entry(cd->name, proc_net_rpc);
370         }
371         if (list_empty(&cache_list)) {
372                 /* module must be being unloaded so its safe to kill the worker */
373                 cancel_delayed_work(&cache_cleaner);
374                 flush_scheduled_work();
375         }
376         return 0;
377 }
378
379 /* clean cache tries to find something to clean
380  * and cleans it.
381  * It returns 1 if it cleaned something,
382  *            0 if it didn't find anything this time
383  *           -1 if it fell off the end of the list.
384  */
385 static int cache_clean(void)
386 {
387         int rv = 0;
388         struct list_head *next;
389
390         spin_lock(&cache_list_lock);
391
392         /* find a suitable table if we don't already have one */
393         while (current_detail == NULL ||
394             current_index >= current_detail->hash_size) {
395                 if (current_detail)
396                         next = current_detail->others.next;
397                 else
398                         next = cache_list.next;
399                 if (next == &cache_list) {
400                         current_detail = NULL;
401                         spin_unlock(&cache_list_lock);
402                         return -1;
403                 }
404                 current_detail = list_entry(next, struct cache_detail, others);
405                 if (current_detail->nextcheck > get_seconds())
406                         current_index = current_detail->hash_size;
407                 else {
408                         current_index = 0;
409                         current_detail->nextcheck = get_seconds()+30*60;
410                 }
411         }
412
413         /* find a non-empty bucket in the table */
414         while (current_detail &&
415                current_index < current_detail->hash_size &&
416                current_detail->hash_table[current_index] == NULL)
417                 current_index++;
418
419         /* find a cleanable entry in the bucket and clean it, or set to next bucket */
420         
421         if (current_detail && current_index < current_detail->hash_size) {
422                 struct cache_head *ch, **cp;
423                 struct cache_detail *d;
424                 
425                 write_lock(&current_detail->hash_lock);
426
427                 /* Ok, now to clean this strand */
428                         
429                 cp = & current_detail->hash_table[current_index];
430                 ch = *cp;
431                 for (; ch; cp= & ch->next, ch= *cp) {
432                         if (current_detail->nextcheck > ch->expiry_time)
433                                 current_detail->nextcheck = ch->expiry_time+1;
434                         if (ch->expiry_time >= get_seconds()
435                             && ch->last_refresh >= current_detail->flush_time
436                                 )
437                                 continue;
438                         if (test_and_clear_bit(CACHE_PENDING, &ch->flags))
439                                 queue_loose(current_detail, ch);
440
441                         if (atomic_read(&ch->ref.refcount) == 1)
442                                 break;
443                 }
444                 if (ch) {
445                         *cp = ch->next;
446                         ch->next = NULL;
447                         current_detail->entries--;
448                         rv = 1;
449                 }
450                 write_unlock(&current_detail->hash_lock);
451                 d = current_detail;
452                 if (!ch)
453                         current_index ++;
454                 spin_unlock(&cache_list_lock);
455                 if (ch)
456                         cache_put(ch, d);
457         } else
458                 spin_unlock(&cache_list_lock);
459
460         return rv;
461 }
462
463 /*
464  * We want to regularly clean the cache, so we need to schedule some work ...
465  */
466 static void do_cache_clean(struct work_struct *work)
467 {
468         int delay = 5;
469         if (cache_clean() == -1)
470                 delay = 30*HZ;
471
472         if (list_empty(&cache_list))
473                 delay = 0;
474
475         if (delay)
476                 schedule_delayed_work(&cache_cleaner, delay);
477 }
478
479
480 /* 
481  * Clean all caches promptly.  This just calls cache_clean
482  * repeatedly until we are sure that every cache has had a chance to 
483  * be fully cleaned
484  */
485 void cache_flush(void)
486 {
487         while (cache_clean() != -1)
488                 cond_resched();
489         while (cache_clean() != -1)
490                 cond_resched();
491 }
492
493 void cache_purge(struct cache_detail *detail)
494 {
495         detail->flush_time = LONG_MAX;
496         detail->nextcheck = get_seconds();
497         cache_flush();
498         detail->flush_time = 1;
499 }
500
501
502
503 /*
504  * Deferral and Revisiting of Requests.
505  *
506  * If a cache lookup finds a pending entry, we
507  * need to defer the request and revisit it later.
508  * All deferred requests are stored in a hash table,
509  * indexed by "struct cache_head *".
510  * As it may be wasteful to store a whole request
511  * structure, we allow the request to provide a 
512  * deferred form, which must contain a
513  * 'struct cache_deferred_req'
514  * This cache_deferred_req contains a method to allow
515  * it to be revisited when cache info is available
516  */
517
518 #define DFR_HASHSIZE    (PAGE_SIZE/sizeof(struct list_head))
519 #define DFR_HASH(item)  ((((long)item)>>4 ^ (((long)item)>>13)) % DFR_HASHSIZE)
520
521 #define DFR_MAX 300     /* ??? */
522
523 static DEFINE_SPINLOCK(cache_defer_lock);
524 static LIST_HEAD(cache_defer_list);
525 static struct list_head cache_defer_hash[DFR_HASHSIZE];
526 static int cache_defer_cnt;
527
528 static int cache_defer_req(struct cache_req *req, struct cache_head *item)
529 {
530         struct cache_deferred_req *dreq;
531         int hash = DFR_HASH(item);
532
533         if (cache_defer_cnt >= DFR_MAX) {
534                 /* too much in the cache, randomly drop this one,
535                  * or continue and drop the oldest below
536                  */
537                 if (net_random()&1)
538                         return -ETIMEDOUT;
539         }
540         dreq = req->defer(req);
541         if (dreq == NULL)
542                 return -ETIMEDOUT;
543
544         dreq->item = item;
545         dreq->recv_time = get_seconds();
546
547         spin_lock(&cache_defer_lock);
548
549         list_add(&dreq->recent, &cache_defer_list);
550
551         if (cache_defer_hash[hash].next == NULL)
552                 INIT_LIST_HEAD(&cache_defer_hash[hash]);
553         list_add(&dreq->hash, &cache_defer_hash[hash]);
554
555         /* it is in, now maybe clean up */
556         dreq = NULL;
557         if (++cache_defer_cnt > DFR_MAX) {
558                 dreq = list_entry(cache_defer_list.prev,
559                                   struct cache_deferred_req, recent);
560                 list_del(&dreq->recent);
561                 list_del(&dreq->hash);
562                 cache_defer_cnt--;
563         }
564         spin_unlock(&cache_defer_lock);
565
566         if (dreq) {
567                 /* there was one too many */
568                 dreq->revisit(dreq, 1);
569         }
570         if (!test_bit(CACHE_PENDING, &item->flags)) {
571                 /* must have just been validated... */
572                 cache_revisit_request(item);
573         }
574         return 0;
575 }
576
577 static void cache_revisit_request(struct cache_head *item)
578 {
579         struct cache_deferred_req *dreq;
580         struct list_head pending;
581
582         struct list_head *lp;
583         int hash = DFR_HASH(item);
584
585         INIT_LIST_HEAD(&pending);
586         spin_lock(&cache_defer_lock);
587         
588         lp = cache_defer_hash[hash].next;
589         if (lp) {
590                 while (lp != &cache_defer_hash[hash]) {
591                         dreq = list_entry(lp, struct cache_deferred_req, hash);
592                         lp = lp->next;
593                         if (dreq->item == item) {
594                                 list_del(&dreq->hash);
595                                 list_move(&dreq->recent, &pending);
596                                 cache_defer_cnt--;
597                         }
598                 }
599         }
600         spin_unlock(&cache_defer_lock);
601
602         while (!list_empty(&pending)) {
603                 dreq = list_entry(pending.next, struct cache_deferred_req, recent);
604                 list_del_init(&dreq->recent);
605                 dreq->revisit(dreq, 0);
606         }
607 }
608
609 void cache_clean_deferred(void *owner)
610 {
611         struct cache_deferred_req *dreq, *tmp;
612         struct list_head pending;
613
614
615         INIT_LIST_HEAD(&pending);
616         spin_lock(&cache_defer_lock);
617         
618         list_for_each_entry_safe(dreq, tmp, &cache_defer_list, recent) {
619                 if (dreq->owner == owner) {
620                         list_del(&dreq->hash);
621                         list_move(&dreq->recent, &pending);
622                         cache_defer_cnt--;
623                 }
624         }
625         spin_unlock(&cache_defer_lock);
626
627         while (!list_empty(&pending)) {
628                 dreq = list_entry(pending.next, struct cache_deferred_req, recent);
629                 list_del_init(&dreq->recent);
630                 dreq->revisit(dreq, 1);
631         }
632 }
633
634 /*
635  * communicate with user-space
636  *
637  * We have a magic /proc file - /proc/sunrpc/cache
638  * On read, you get a full request, or block
639  * On write, an update request is processed
640  * Poll works if anything to read, and always allows write
641  *
642  * Implemented by linked list of requests.  Each open file has 
643  * a ->private that also exists in this list.  New request are added
644  * to the end and may wakeup and preceding readers.
645  * New readers are added to the head.  If, on read, an item is found with
646  * CACHE_UPCALLING clear, we free it from the list.
647  *
648  */
649
650 static DEFINE_SPINLOCK(queue_lock);
651 static DEFINE_MUTEX(queue_io_mutex);
652
653 struct cache_queue {
654         struct list_head        list;
655         int                     reader; /* if 0, then request */
656 };
657 struct cache_request {
658         struct cache_queue      q;
659         struct cache_head       *item;
660         char                    * buf;
661         int                     len;
662         int                     readers;
663 };
664 struct cache_reader {
665         struct cache_queue      q;
666         int                     offset; /* if non-0, we have a refcnt on next request */
667 };
668
669 static ssize_t
670 cache_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
671 {
672         struct cache_reader *rp = filp->private_data;
673         struct cache_request *rq;
674         struct cache_detail *cd = PDE(filp->f_path.dentry->d_inode)->data;
675         int err;
676
677         if (count == 0)
678                 return 0;
679
680         mutex_lock(&queue_io_mutex); /* protect against multiple concurrent
681                               * readers on this file */
682  again:
683         spin_lock(&queue_lock);
684         /* need to find next request */
685         while (rp->q.list.next != &cd->queue &&
686                list_entry(rp->q.list.next, struct cache_queue, list)
687                ->reader) {
688                 struct list_head *next = rp->q.list.next;
689                 list_move(&rp->q.list, next);
690         }
691         if (rp->q.list.next == &cd->queue) {
692                 spin_unlock(&queue_lock);
693                 mutex_unlock(&queue_io_mutex);
694                 BUG_ON(rp->offset);
695                 return 0;
696         }
697         rq = container_of(rp->q.list.next, struct cache_request, q.list);
698         BUG_ON(rq->q.reader);
699         if (rp->offset == 0)
700                 rq->readers++;
701         spin_unlock(&queue_lock);
702
703         if (rp->offset == 0 && !test_bit(CACHE_PENDING, &rq->item->flags)) {
704                 err = -EAGAIN;
705                 spin_lock(&queue_lock);
706                 list_move(&rp->q.list, &rq->q.list);
707                 spin_unlock(&queue_lock);
708         } else {
709                 if (rp->offset + count > rq->len)
710                         count = rq->len - rp->offset;
711                 err = -EFAULT;
712                 if (copy_to_user(buf, rq->buf + rp->offset, count))
713                         goto out;
714                 rp->offset += count;
715                 if (rp->offset >= rq->len) {
716                         rp->offset = 0;
717                         spin_lock(&queue_lock);
718                         list_move(&rp->q.list, &rq->q.list);
719                         spin_unlock(&queue_lock);
720                 }
721                 err = 0;
722         }
723  out:
724         if (rp->offset == 0) {
725                 /* need to release rq */
726                 spin_lock(&queue_lock);
727                 rq->readers--;
728                 if (rq->readers == 0 &&
729                     !test_bit(CACHE_PENDING, &rq->item->flags)) {
730                         list_del(&rq->q.list);
731                         spin_unlock(&queue_lock);
732                         cache_put(rq->item, cd);
733                         kfree(rq->buf);
734                         kfree(rq);
735                 } else
736                         spin_unlock(&queue_lock);
737         }
738         if (err == -EAGAIN)
739                 goto again;
740         mutex_unlock(&queue_io_mutex);
741         return err ? err :  count;
742 }
743
744 static char write_buf[8192]; /* protected by queue_io_mutex */
745
746 static ssize_t
747 cache_write(struct file *filp, const char __user *buf, size_t count,
748             loff_t *ppos)
749 {
750         int err;
751         struct cache_detail *cd = PDE(filp->f_path.dentry->d_inode)->data;
752
753         if (count == 0)
754                 return 0;
755         if (count >= sizeof(write_buf))
756                 return -EINVAL;
757
758         mutex_lock(&queue_io_mutex);
759
760         if (copy_from_user(write_buf, buf, count)) {
761                 mutex_unlock(&queue_io_mutex);
762                 return -EFAULT;
763         }
764         write_buf[count] = '\0';
765         if (cd->cache_parse)
766                 err = cd->cache_parse(cd, write_buf, count);
767         else
768                 err = -EINVAL;
769
770         mutex_unlock(&queue_io_mutex);
771         return err ? err : count;
772 }
773
774 static DECLARE_WAIT_QUEUE_HEAD(queue_wait);
775
776 static unsigned int
777 cache_poll(struct file *filp, poll_table *wait)
778 {
779         unsigned int mask;
780         struct cache_reader *rp = filp->private_data;
781         struct cache_queue *cq;
782         struct cache_detail *cd = PDE(filp->f_path.dentry->d_inode)->data;
783
784         poll_wait(filp, &queue_wait, wait);
785
786         /* alway allow write */
787         mask = POLL_OUT | POLLWRNORM;
788
789         if (!rp)
790                 return mask;
791
792         spin_lock(&queue_lock);
793
794         for (cq= &rp->q; &cq->list != &cd->queue;
795              cq = list_entry(cq->list.next, struct cache_queue, list))
796                 if (!cq->reader) {
797                         mask |= POLLIN | POLLRDNORM;
798                         break;
799                 }
800         spin_unlock(&queue_lock);
801         return mask;
802 }
803
804 static int
805 cache_ioctl(struct inode *ino, struct file *filp,
806             unsigned int cmd, unsigned long arg)
807 {
808         int len = 0;
809         struct cache_reader *rp = filp->private_data;
810         struct cache_queue *cq;
811         struct cache_detail *cd = PDE(ino)->data;
812
813         if (cmd != FIONREAD || !rp)
814                 return -EINVAL;
815
816         spin_lock(&queue_lock);
817
818         /* only find the length remaining in current request,
819          * or the length of the next request
820          */
821         for (cq= &rp->q; &cq->list != &cd->queue;
822              cq = list_entry(cq->list.next, struct cache_queue, list))
823                 if (!cq->reader) {
824                         struct cache_request *cr =
825                                 container_of(cq, struct cache_request, q);
826                         len = cr->len - rp->offset;
827                         break;
828                 }
829         spin_unlock(&queue_lock);
830
831         return put_user(len, (int __user *)arg);
832 }
833
834 static int
835 cache_open(struct inode *inode, struct file *filp)
836 {
837         struct cache_reader *rp = NULL;
838
839         nonseekable_open(inode, filp);
840         if (filp->f_mode & FMODE_READ) {
841                 struct cache_detail *cd = PDE(inode)->data;
842
843                 rp = kmalloc(sizeof(*rp), GFP_KERNEL);
844                 if (!rp)
845                         return -ENOMEM;
846                 rp->offset = 0;
847                 rp->q.reader = 1;
848                 atomic_inc(&cd->readers);
849                 spin_lock(&queue_lock);
850                 list_add(&rp->q.list, &cd->queue);
851                 spin_unlock(&queue_lock);
852         }
853         filp->private_data = rp;
854         return 0;
855 }
856
857 static int
858 cache_release(struct inode *inode, struct file *filp)
859 {
860         struct cache_reader *rp = filp->private_data;
861         struct cache_detail *cd = PDE(inode)->data;
862
863         if (rp) {
864                 spin_lock(&queue_lock);
865                 if (rp->offset) {
866                         struct cache_queue *cq;
867                         for (cq= &rp->q; &cq->list != &cd->queue;
868                              cq = list_entry(cq->list.next, struct cache_queue, list))
869                                 if (!cq->reader) {
870                                         container_of(cq, struct cache_request, q)
871                                                 ->readers--;
872                                         break;
873                                 }
874                         rp->offset = 0;
875                 }
876                 list_del(&rp->q.list);
877                 spin_unlock(&queue_lock);
878
879                 filp->private_data = NULL;
880                 kfree(rp);
881
882                 cd->last_close = get_seconds();
883                 atomic_dec(&cd->readers);
884         }
885         return 0;
886 }
887
888
889
890 static struct file_operations cache_file_operations = {
891         .owner          = THIS_MODULE,
892         .llseek         = no_llseek,
893         .read           = cache_read,
894         .write          = cache_write,
895         .poll           = cache_poll,
896         .ioctl          = cache_ioctl, /* for FIONREAD */
897         .open           = cache_open,
898         .release        = cache_release,
899 };
900
901
902 static void queue_loose(struct cache_detail *detail, struct cache_head *ch)
903 {
904         struct cache_queue *cq;
905         spin_lock(&queue_lock);
906         list_for_each_entry(cq, &detail->queue, list)
907                 if (!cq->reader) {
908                         struct cache_request *cr = container_of(cq, struct cache_request, q);
909                         if (cr->item != ch)
910                                 continue;
911                         if (cr->readers != 0)
912                                 continue;
913                         list_del(&cr->q.list);
914                         spin_unlock(&queue_lock);
915                         cache_put(cr->item, detail);
916                         kfree(cr->buf);
917                         kfree(cr);
918                         return;
919                 }
920         spin_unlock(&queue_lock);
921 }
922
923 /*
924  * Support routines for text-based upcalls.
925  * Fields are separated by spaces.
926  * Fields are either mangled to quote space tab newline slosh with slosh
927  * or a hexified with a leading \x
928  * Record is terminated with newline.
929  *
930  */
931
932 void qword_add(char **bpp, int *lp, char *str)
933 {
934         char *bp = *bpp;
935         int len = *lp;
936         char c;
937
938         if (len < 0) return;
939
940         while ((c=*str++) && len)
941                 switch(c) {
942                 case ' ':
943                 case '\t':
944                 case '\n':
945                 case '\\':
946                         if (len >= 4) {
947                                 *bp++ = '\\';
948                                 *bp++ = '0' + ((c & 0300)>>6);
949                                 *bp++ = '0' + ((c & 0070)>>3);
950                                 *bp++ = '0' + ((c & 0007)>>0);
951                         }
952                         len -= 4;
953                         break;
954                 default:
955                         *bp++ = c;
956                         len--;
957                 }
958         if (c || len <1) len = -1;
959         else {
960                 *bp++ = ' ';
961                 len--;
962         }
963         *bpp = bp;
964         *lp = len;
965 }
966
967 void qword_addhex(char **bpp, int *lp, char *buf, int blen)
968 {
969         char *bp = *bpp;
970         int len = *lp;
971
972         if (len < 0) return;
973
974         if (len > 2) {
975                 *bp++ = '\\';
976                 *bp++ = 'x';
977                 len -= 2;
978                 while (blen && len >= 2) {
979                         unsigned char c = *buf++;
980                         *bp++ = '0' + ((c&0xf0)>>4) + (c>=0xa0)*('a'-'9'-1);
981                         *bp++ = '0' + (c&0x0f) + ((c&0x0f)>=0x0a)*('a'-'9'-1);
982                         len -= 2;
983                         blen--;
984                 }
985         }
986         if (blen || len<1) len = -1;
987         else {
988                 *bp++ = ' ';
989                 len--;
990         }
991         *bpp = bp;
992         *lp = len;
993 }
994
995 static void warn_no_listener(struct cache_detail *detail)
996 {
997         if (detail->last_warn != detail->last_close) {
998                 detail->last_warn = detail->last_close;
999                 if (detail->warn_no_listener)
1000                         detail->warn_no_listener(detail);
1001         }
1002 }
1003
1004 /*
1005  * register an upcall request to user-space.
1006  * Each request is at most one page long.
1007  */
1008 static int cache_make_upcall(struct cache_detail *detail, struct cache_head *h)
1009 {
1010
1011         char *buf;
1012         struct cache_request *crq;
1013         char *bp;
1014         int len;
1015
1016         if (detail->cache_request == NULL)
1017                 return -EINVAL;
1018
1019         if (atomic_read(&detail->readers) == 0 &&
1020             detail->last_close < get_seconds() - 30) {
1021                         warn_no_listener(detail);
1022                         return -EINVAL;
1023         }
1024
1025         buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
1026         if (!buf)
1027                 return -EAGAIN;
1028
1029         crq = kmalloc(sizeof (*crq), GFP_KERNEL);
1030         if (!crq) {
1031                 kfree(buf);
1032                 return -EAGAIN;
1033         }
1034
1035         bp = buf; len = PAGE_SIZE;
1036
1037         detail->cache_request(detail, h, &bp, &len);
1038
1039         if (len < 0) {
1040                 kfree(buf);
1041                 kfree(crq);
1042                 return -EAGAIN;
1043         }
1044         crq->q.reader = 0;
1045         crq->item = cache_get(h);
1046         crq->buf = buf;
1047         crq->len = PAGE_SIZE - len;
1048         crq->readers = 0;
1049         spin_lock(&queue_lock);
1050         list_add_tail(&crq->q.list, &detail->queue);
1051         spin_unlock(&queue_lock);
1052         wake_up(&queue_wait);
1053         return 0;
1054 }
1055
1056 /*
1057  * parse a message from user-space and pass it
1058  * to an appropriate cache
1059  * Messages are, like requests, separated into fields by
1060  * spaces and dequotes as \xHEXSTRING or embedded \nnn octal
1061  *
1062  * Message is 
1063  *   reply cachename expiry key ... content....
1064  *
1065  * key and content are both parsed by cache 
1066  */
1067
1068 #define isodigit(c) (isdigit(c) && c <= '7')
1069 int qword_get(char **bpp, char *dest, int bufsize)
1070 {
1071         /* return bytes copied, or -1 on error */
1072         char *bp = *bpp;
1073         int len = 0;
1074
1075         while (*bp == ' ') bp++;
1076
1077         if (bp[0] == '\\' && bp[1] == 'x') {
1078                 /* HEX STRING */
1079                 bp += 2;
1080                 while (isxdigit(bp[0]) && isxdigit(bp[1]) && len < bufsize) {
1081                         int byte = isdigit(*bp) ? *bp-'0' : toupper(*bp)-'A'+10;
1082                         bp++;
1083                         byte <<= 4;
1084                         byte |= isdigit(*bp) ? *bp-'0' : toupper(*bp)-'A'+10;
1085                         *dest++ = byte;
1086                         bp++;
1087                         len++;
1088                 }
1089         } else {
1090                 /* text with \nnn octal quoting */
1091                 while (*bp != ' ' && *bp != '\n' && *bp && len < bufsize-1) {
1092                         if (*bp == '\\' &&
1093                             isodigit(bp[1]) && (bp[1] <= '3') &&
1094                             isodigit(bp[2]) &&
1095                             isodigit(bp[3])) {
1096                                 int byte = (*++bp -'0');
1097                                 bp++;
1098                                 byte = (byte << 3) | (*bp++ - '0');
1099                                 byte = (byte << 3) | (*bp++ - '0');
1100                                 *dest++ = byte;
1101                                 len++;
1102                         } else {
1103                                 *dest++ = *bp++;
1104                                 len++;
1105                         }
1106                 }
1107         }
1108
1109         if (*bp != ' ' && *bp != '\n' && *bp != '\0')
1110                 return -1;
1111         while (*bp == ' ') bp++;
1112         *bpp = bp;
1113         *dest = '\0';
1114         return len;
1115 }
1116
1117
1118 /*
1119  * support /proc/sunrpc/cache/$CACHENAME/content
1120  * as a seqfile.
1121  * We call ->cache_show passing NULL for the item to
1122  * get a header, then pass each real item in the cache
1123  */
1124
1125 struct handle {
1126         struct cache_detail *cd;
1127 };
1128
1129 static void *c_start(struct seq_file *m, loff_t *pos)
1130 {
1131         loff_t n = *pos;
1132         unsigned hash, entry;
1133         struct cache_head *ch;
1134         struct cache_detail *cd = ((struct handle*)m->private)->cd;
1135         
1136
1137         read_lock(&cd->hash_lock);
1138         if (!n--)
1139                 return SEQ_START_TOKEN;
1140         hash = n >> 32;
1141         entry = n & ((1LL<<32) - 1);
1142
1143         for (ch=cd->hash_table[hash]; ch; ch=ch->next)
1144                 if (!entry--)
1145                         return ch;
1146         n &= ~((1LL<<32) - 1);
1147         do {
1148                 hash++;
1149                 n += 1LL<<32;
1150         } while(hash < cd->hash_size && 
1151                 cd->hash_table[hash]==NULL);
1152         if (hash >= cd->hash_size)
1153                 return NULL;
1154         *pos = n+1;
1155         return cd->hash_table[hash];
1156 }
1157
1158 static void *c_next(struct seq_file *m, void *p, loff_t *pos)
1159 {
1160         struct cache_head *ch = p;
1161         int hash = (*pos >> 32);
1162         struct cache_detail *cd = ((struct handle*)m->private)->cd;
1163
1164         if (p == SEQ_START_TOKEN)
1165                 hash = 0;
1166         else if (ch->next == NULL) {
1167                 hash++;
1168                 *pos += 1LL<<32;
1169         } else {
1170                 ++*pos;
1171                 return ch->next;
1172         }
1173         *pos &= ~((1LL<<32) - 1);
1174         while (hash < cd->hash_size &&
1175                cd->hash_table[hash] == NULL) {
1176                 hash++;
1177                 *pos += 1LL<<32;
1178         }
1179         if (hash >= cd->hash_size)
1180                 return NULL;
1181         ++*pos;
1182         return cd->hash_table[hash];
1183 }
1184
1185 static void c_stop(struct seq_file *m, void *p)
1186 {
1187         struct cache_detail *cd = ((struct handle*)m->private)->cd;
1188         read_unlock(&cd->hash_lock);
1189 }
1190
1191 static int c_show(struct seq_file *m, void *p)
1192 {
1193         struct cache_head *cp = p;
1194         struct cache_detail *cd = ((struct handle*)m->private)->cd;
1195
1196         if (p == SEQ_START_TOKEN)
1197                 return cd->cache_show(m, cd, NULL);
1198
1199         ifdebug(CACHE)
1200                 seq_printf(m, "# expiry=%ld refcnt=%d flags=%lx\n",
1201                            cp->expiry_time, atomic_read(&cp->ref.refcount), cp->flags);
1202         cache_get(cp);
1203         if (cache_check(cd, cp, NULL))
1204                 /* cache_check does a cache_put on failure */
1205                 seq_printf(m, "# ");
1206         else
1207                 cache_put(cp, cd);
1208
1209         return cd->cache_show(m, cd, cp);
1210 }
1211
1212 static struct seq_operations cache_content_op = {
1213         .start  = c_start,
1214         .next   = c_next,
1215         .stop   = c_stop,
1216         .show   = c_show,
1217 };
1218
1219 static int content_open(struct inode *inode, struct file *file)
1220 {
1221         int res;
1222         struct handle *han;
1223         struct cache_detail *cd = PDE(inode)->data;
1224
1225         han = kmalloc(sizeof(*han), GFP_KERNEL);
1226         if (han == NULL)
1227                 return -ENOMEM;
1228
1229         han->cd = cd;
1230
1231         res = seq_open(file, &cache_content_op);
1232         if (res)
1233                 kfree(han);
1234         else
1235                 ((struct seq_file *)file->private_data)->private = han;
1236
1237         return res;
1238 }
1239 static int content_release(struct inode *inode, struct file *file)
1240 {
1241         struct seq_file *m = (struct seq_file *)file->private_data;
1242         struct handle *han = m->private;
1243         kfree(han);
1244         m->private = NULL;
1245         return seq_release(inode, file);
1246 }
1247
1248 static struct file_operations content_file_operations = {
1249         .open           = content_open,
1250         .read           = seq_read,
1251         .llseek         = seq_lseek,
1252         .release        = content_release,
1253 };
1254
1255 static ssize_t read_flush(struct file *file, char __user *buf,
1256                             size_t count, loff_t *ppos)
1257 {
1258         struct cache_detail *cd = PDE(file->f_path.dentry->d_inode)->data;
1259         char tbuf[20];
1260         unsigned long p = *ppos;
1261         int len;
1262
1263         sprintf(tbuf, "%lu\n", cd->flush_time);
1264         len = strlen(tbuf);
1265         if (p >= len)
1266                 return 0;
1267         len -= p;
1268         if (len > count) len = count;
1269         if (copy_to_user(buf, (void*)(tbuf+p), len))
1270                 len = -EFAULT;
1271         else
1272                 *ppos += len;
1273         return len;
1274 }
1275
1276 static ssize_t write_flush(struct file * file, const char __user * buf,
1277                              size_t count, loff_t *ppos)
1278 {
1279         struct cache_detail *cd = PDE(file->f_path.dentry->d_inode)->data;
1280         char tbuf[20];
1281         char *ep;
1282         long flushtime;
1283         if (*ppos || count > sizeof(tbuf)-1)
1284                 return -EINVAL;
1285         if (copy_from_user(tbuf, buf, count))
1286                 return -EFAULT;
1287         tbuf[count] = 0;
1288         flushtime = simple_strtoul(tbuf, &ep, 0);
1289         if (*ep && *ep != '\n')
1290                 return -EINVAL;
1291
1292         cd->flush_time = flushtime;
1293         cd->nextcheck = get_seconds();
1294         cache_flush();
1295
1296         *ppos += count;
1297         return count;
1298 }
1299
1300 static struct file_operations cache_flush_operations = {
1301         .open           = nonseekable_open,
1302         .read           = read_flush,
1303         .write          = write_flush,
1304 };