Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi...
[pandora-kernel.git] / fs / fuse / dev.c
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
4
5   This program can be distributed under the terms of the GNU GPL.
6   See the file COPYING.
7 */
8
9 #include "fuse_i.h"
10
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/poll.h>
14 #include <linux/uio.h>
15 #include <linux/miscdevice.h>
16 #include <linux/pagemap.h>
17 #include <linux/file.h>
18 #include <linux/slab.h>
19
20 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
21
22 static struct kmem_cache *fuse_req_cachep;
23
24 static struct fuse_conn *fuse_get_conn(struct file *file)
25 {
26         /*
27          * Lockless access is OK, because file->private data is set
28          * once during mount and is valid until the file is released.
29          */
30         return file->private_data;
31 }
32
33 static void fuse_request_init(struct fuse_req *req)
34 {
35         memset(req, 0, sizeof(*req));
36         INIT_LIST_HEAD(&req->list);
37         INIT_LIST_HEAD(&req->intr_entry);
38         init_waitqueue_head(&req->waitq);
39         atomic_set(&req->count, 1);
40 }
41
42 struct fuse_req *fuse_request_alloc(void)
43 {
44         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_KERNEL);
45         if (req)
46                 fuse_request_init(req);
47         return req;
48 }
49
50 struct fuse_req *fuse_request_alloc_nofs(void)
51 {
52         struct fuse_req *req = kmem_cache_alloc(fuse_req_cachep, GFP_NOFS);
53         if (req)
54                 fuse_request_init(req);
55         return req;
56 }
57
58 void fuse_request_free(struct fuse_req *req)
59 {
60         kmem_cache_free(fuse_req_cachep, req);
61 }
62
63 static void block_sigs(sigset_t *oldset)
64 {
65         sigset_t mask;
66
67         siginitsetinv(&mask, sigmask(SIGKILL));
68         sigprocmask(SIG_BLOCK, &mask, oldset);
69 }
70
71 static void restore_sigs(sigset_t *oldset)
72 {
73         sigprocmask(SIG_SETMASK, oldset, NULL);
74 }
75
76 static void __fuse_get_request(struct fuse_req *req)
77 {
78         atomic_inc(&req->count);
79 }
80
81 /* Must be called with > 1 refcount */
82 static void __fuse_put_request(struct fuse_req *req)
83 {
84         BUG_ON(atomic_read(&req->count) < 2);
85         atomic_dec(&req->count);
86 }
87
88 static void fuse_req_init_context(struct fuse_req *req)
89 {
90         req->in.h.uid = current_fsuid();
91         req->in.h.gid = current_fsgid();
92         req->in.h.pid = current->pid;
93 }
94
95 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
96 {
97         struct fuse_req *req;
98         sigset_t oldset;
99         int intr;
100         int err;
101
102         atomic_inc(&fc->num_waiting);
103         block_sigs(&oldset);
104         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
105         restore_sigs(&oldset);
106         err = -EINTR;
107         if (intr)
108                 goto out;
109
110         err = -ENOTCONN;
111         if (!fc->connected)
112                 goto out;
113
114         req = fuse_request_alloc();
115         err = -ENOMEM;
116         if (!req)
117                 goto out;
118
119         fuse_req_init_context(req);
120         req->waiting = 1;
121         return req;
122
123  out:
124         atomic_dec(&fc->num_waiting);
125         return ERR_PTR(err);
126 }
127
128 /*
129  * Return request in fuse_file->reserved_req.  However that may
130  * currently be in use.  If that is the case, wait for it to become
131  * available.
132  */
133 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
134                                          struct file *file)
135 {
136         struct fuse_req *req = NULL;
137         struct fuse_file *ff = file->private_data;
138
139         do {
140                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
141                 spin_lock(&fc->lock);
142                 if (ff->reserved_req) {
143                         req = ff->reserved_req;
144                         ff->reserved_req = NULL;
145                         get_file(file);
146                         req->stolen_file = file;
147                 }
148                 spin_unlock(&fc->lock);
149         } while (!req);
150
151         return req;
152 }
153
154 /*
155  * Put stolen request back into fuse_file->reserved_req
156  */
157 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
158 {
159         struct file *file = req->stolen_file;
160         struct fuse_file *ff = file->private_data;
161
162         spin_lock(&fc->lock);
163         fuse_request_init(req);
164         BUG_ON(ff->reserved_req);
165         ff->reserved_req = req;
166         wake_up_all(&fc->reserved_req_waitq);
167         spin_unlock(&fc->lock);
168         fput(file);
169 }
170
171 /*
172  * Gets a requests for a file operation, always succeeds
173  *
174  * This is used for sending the FLUSH request, which must get to
175  * userspace, due to POSIX locks which may need to be unlocked.
176  *
177  * If allocation fails due to OOM, use the reserved request in
178  * fuse_file.
179  *
180  * This is very unlikely to deadlock accidentally, since the
181  * filesystem should not have it's own file open.  If deadlock is
182  * intentional, it can still be broken by "aborting" the filesystem.
183  */
184 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
185 {
186         struct fuse_req *req;
187
188         atomic_inc(&fc->num_waiting);
189         wait_event(fc->blocked_waitq, !fc->blocked);
190         req = fuse_request_alloc();
191         if (!req)
192                 req = get_reserved_req(fc, file);
193
194         fuse_req_init_context(req);
195         req->waiting = 1;
196         return req;
197 }
198
199 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
200 {
201         if (atomic_dec_and_test(&req->count)) {
202                 if (req->waiting)
203                         atomic_dec(&fc->num_waiting);
204
205                 if (req->stolen_file)
206                         put_reserved_req(fc, req);
207                 else
208                         fuse_request_free(req);
209         }
210 }
211
212 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
213 {
214         unsigned nbytes = 0;
215         unsigned i;
216
217         for (i = 0; i < numargs; i++)
218                 nbytes += args[i].size;
219
220         return nbytes;
221 }
222
223 static u64 fuse_get_unique(struct fuse_conn *fc)
224 {
225         fc->reqctr++;
226         /* zero is special */
227         if (fc->reqctr == 0)
228                 fc->reqctr = 1;
229
230         return fc->reqctr;
231 }
232
233 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
234 {
235         req->in.h.unique = fuse_get_unique(fc);
236         req->in.h.len = sizeof(struct fuse_in_header) +
237                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
238         list_add_tail(&req->list, &fc->pending);
239         req->state = FUSE_REQ_PENDING;
240         if (!req->waiting) {
241                 req->waiting = 1;
242                 atomic_inc(&fc->num_waiting);
243         }
244         wake_up(&fc->waitq);
245         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
246 }
247
248 static void flush_bg_queue(struct fuse_conn *fc)
249 {
250         while (fc->active_background < FUSE_MAX_BACKGROUND &&
251                !list_empty(&fc->bg_queue)) {
252                 struct fuse_req *req;
253
254                 req = list_entry(fc->bg_queue.next, struct fuse_req, list);
255                 list_del(&req->list);
256                 fc->active_background++;
257                 queue_request(fc, req);
258         }
259 }
260
261 /*
262  * This function is called when a request is finished.  Either a reply
263  * has arrived or it was aborted (and not yet sent) or some error
264  * occurred during communication with userspace, or the device file
265  * was closed.  The requester thread is woken up (if still waiting),
266  * the 'end' callback is called if given, else the reference to the
267  * request is released
268  *
269  * Called with fc->lock, unlocks it
270  */
271 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
272 __releases(&fc->lock)
273 {
274         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
275         req->end = NULL;
276         list_del(&req->list);
277         list_del(&req->intr_entry);
278         req->state = FUSE_REQ_FINISHED;
279         if (req->background) {
280                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
281                         fc->blocked = 0;
282                         wake_up_all(&fc->blocked_waitq);
283                 }
284                 if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
285                         clear_bdi_congested(&fc->bdi, READ);
286                         clear_bdi_congested(&fc->bdi, WRITE);
287                 }
288                 fc->num_background--;
289                 fc->active_background--;
290                 flush_bg_queue(fc);
291         }
292         spin_unlock(&fc->lock);
293         wake_up(&req->waitq);
294         if (end)
295                 end(fc, req);
296         fuse_put_request(fc, req);
297 }
298
299 static void wait_answer_interruptible(struct fuse_conn *fc,
300                                       struct fuse_req *req)
301 __releases(&fc->lock)
302 __acquires(&fc->lock)
303 {
304         if (signal_pending(current))
305                 return;
306
307         spin_unlock(&fc->lock);
308         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
309         spin_lock(&fc->lock);
310 }
311
312 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
313 {
314         list_add_tail(&req->intr_entry, &fc->interrupts);
315         wake_up(&fc->waitq);
316         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
317 }
318
319 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
320 __releases(&fc->lock)
321 __acquires(&fc->lock)
322 {
323         if (!fc->no_interrupt) {
324                 /* Any signal may interrupt this */
325                 wait_answer_interruptible(fc, req);
326
327                 if (req->aborted)
328                         goto aborted;
329                 if (req->state == FUSE_REQ_FINISHED)
330                         return;
331
332                 req->interrupted = 1;
333                 if (req->state == FUSE_REQ_SENT)
334                         queue_interrupt(fc, req);
335         }
336
337         if (!req->force) {
338                 sigset_t oldset;
339
340                 /* Only fatal signals may interrupt this */
341                 block_sigs(&oldset);
342                 wait_answer_interruptible(fc, req);
343                 restore_sigs(&oldset);
344
345                 if (req->aborted)
346                         goto aborted;
347                 if (req->state == FUSE_REQ_FINISHED)
348                         return;
349
350                 /* Request is not yet in userspace, bail out */
351                 if (req->state == FUSE_REQ_PENDING) {
352                         list_del(&req->list);
353                         __fuse_put_request(req);
354                         req->out.h.error = -EINTR;
355                         return;
356                 }
357         }
358
359         /*
360          * Either request is already in userspace, or it was forced.
361          * Wait it out.
362          */
363         spin_unlock(&fc->lock);
364         wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
365         spin_lock(&fc->lock);
366
367         if (!req->aborted)
368                 return;
369
370  aborted:
371         BUG_ON(req->state != FUSE_REQ_FINISHED);
372         if (req->locked) {
373                 /* This is uninterruptible sleep, because data is
374                    being copied to/from the buffers of req.  During
375                    locked state, there mustn't be any filesystem
376                    operation (e.g. page fault), since that could lead
377                    to deadlock */
378                 spin_unlock(&fc->lock);
379                 wait_event(req->waitq, !req->locked);
380                 spin_lock(&fc->lock);
381         }
382 }
383
384 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
385 {
386         req->isreply = 1;
387         spin_lock(&fc->lock);
388         if (!fc->connected)
389                 req->out.h.error = -ENOTCONN;
390         else if (fc->conn_error)
391                 req->out.h.error = -ECONNREFUSED;
392         else {
393                 queue_request(fc, req);
394                 /* acquire extra reference, since request is still needed
395                    after request_end() */
396                 __fuse_get_request(req);
397
398                 request_wait_answer(fc, req);
399         }
400         spin_unlock(&fc->lock);
401 }
402
403 static void fuse_request_send_nowait_locked(struct fuse_conn *fc,
404                                             struct fuse_req *req)
405 {
406         req->background = 1;
407         fc->num_background++;
408         if (fc->num_background == FUSE_MAX_BACKGROUND)
409                 fc->blocked = 1;
410         if (fc->num_background == FUSE_CONGESTION_THRESHOLD) {
411                 set_bdi_congested(&fc->bdi, READ);
412                 set_bdi_congested(&fc->bdi, WRITE);
413         }
414         list_add_tail(&req->list, &fc->bg_queue);
415         flush_bg_queue(fc);
416 }
417
418 static void fuse_request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
419 {
420         spin_lock(&fc->lock);
421         if (fc->connected) {
422                 fuse_request_send_nowait_locked(fc, req);
423                 spin_unlock(&fc->lock);
424         } else {
425                 req->out.h.error = -ENOTCONN;
426                 request_end(fc, req);
427         }
428 }
429
430 void fuse_request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
431 {
432         req->isreply = 0;
433         fuse_request_send_nowait(fc, req);
434 }
435
436 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
437 {
438         req->isreply = 1;
439         fuse_request_send_nowait(fc, req);
440 }
441
442 /*
443  * Called under fc->lock
444  *
445  * fc->connected must have been checked previously
446  */
447 void fuse_request_send_background_locked(struct fuse_conn *fc,
448                                          struct fuse_req *req)
449 {
450         req->isreply = 1;
451         fuse_request_send_nowait_locked(fc, req);
452 }
453
454 /*
455  * Lock the request.  Up to the next unlock_request() there mustn't be
456  * anything that could cause a page-fault.  If the request was already
457  * aborted bail out.
458  */
459 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
460 {
461         int err = 0;
462         if (req) {
463                 spin_lock(&fc->lock);
464                 if (req->aborted)
465                         err = -ENOENT;
466                 else
467                         req->locked = 1;
468                 spin_unlock(&fc->lock);
469         }
470         return err;
471 }
472
473 /*
474  * Unlock request.  If it was aborted during being locked, the
475  * requester thread is currently waiting for it to be unlocked, so
476  * wake it up.
477  */
478 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
479 {
480         if (req) {
481                 spin_lock(&fc->lock);
482                 req->locked = 0;
483                 if (req->aborted)
484                         wake_up(&req->waitq);
485                 spin_unlock(&fc->lock);
486         }
487 }
488
489 struct fuse_copy_state {
490         struct fuse_conn *fc;
491         int write;
492         struct fuse_req *req;
493         const struct iovec *iov;
494         unsigned long nr_segs;
495         unsigned long seglen;
496         unsigned long addr;
497         struct page *pg;
498         void *mapaddr;
499         void *buf;
500         unsigned len;
501 };
502
503 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
504                            int write, struct fuse_req *req,
505                            const struct iovec *iov, unsigned long nr_segs)
506 {
507         memset(cs, 0, sizeof(*cs));
508         cs->fc = fc;
509         cs->write = write;
510         cs->req = req;
511         cs->iov = iov;
512         cs->nr_segs = nr_segs;
513 }
514
515 /* Unmap and put previous page of userspace buffer */
516 static void fuse_copy_finish(struct fuse_copy_state *cs)
517 {
518         if (cs->mapaddr) {
519                 kunmap_atomic(cs->mapaddr, KM_USER0);
520                 if (cs->write) {
521                         flush_dcache_page(cs->pg);
522                         set_page_dirty_lock(cs->pg);
523                 }
524                 put_page(cs->pg);
525                 cs->mapaddr = NULL;
526         }
527 }
528
529 /*
530  * Get another pagefull of userspace buffer, and map it to kernel
531  * address space, and lock request
532  */
533 static int fuse_copy_fill(struct fuse_copy_state *cs)
534 {
535         unsigned long offset;
536         int err;
537
538         unlock_request(cs->fc, cs->req);
539         fuse_copy_finish(cs);
540         if (!cs->seglen) {
541                 BUG_ON(!cs->nr_segs);
542                 cs->seglen = cs->iov[0].iov_len;
543                 cs->addr = (unsigned long) cs->iov[0].iov_base;
544                 cs->iov++;
545                 cs->nr_segs--;
546         }
547         down_read(&current->mm->mmap_sem);
548         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
549                              &cs->pg, NULL);
550         up_read(&current->mm->mmap_sem);
551         if (err < 0)
552                 return err;
553         BUG_ON(err != 1);
554         offset = cs->addr % PAGE_SIZE;
555         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
556         cs->buf = cs->mapaddr + offset;
557         cs->len = min(PAGE_SIZE - offset, cs->seglen);
558         cs->seglen -= cs->len;
559         cs->addr += cs->len;
560
561         return lock_request(cs->fc, cs->req);
562 }
563
564 /* Do as much copy to/from userspace buffer as we can */
565 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
566 {
567         unsigned ncpy = min(*size, cs->len);
568         if (val) {
569                 if (cs->write)
570                         memcpy(cs->buf, *val, ncpy);
571                 else
572                         memcpy(*val, cs->buf, ncpy);
573                 *val += ncpy;
574         }
575         *size -= ncpy;
576         cs->len -= ncpy;
577         cs->buf += ncpy;
578         return ncpy;
579 }
580
581 /*
582  * Copy a page in the request to/from the userspace buffer.  Must be
583  * done atomically
584  */
585 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
586                           unsigned offset, unsigned count, int zeroing)
587 {
588         if (page && zeroing && count < PAGE_SIZE) {
589                 void *mapaddr = kmap_atomic(page, KM_USER1);
590                 memset(mapaddr, 0, PAGE_SIZE);
591                 kunmap_atomic(mapaddr, KM_USER1);
592         }
593         while (count) {
594                 if (!cs->len) {
595                         int err = fuse_copy_fill(cs);
596                         if (err)
597                                 return err;
598                 }
599                 if (page) {
600                         void *mapaddr = kmap_atomic(page, KM_USER1);
601                         void *buf = mapaddr + offset;
602                         offset += fuse_copy_do(cs, &buf, &count);
603                         kunmap_atomic(mapaddr, KM_USER1);
604                 } else
605                         offset += fuse_copy_do(cs, NULL, &count);
606         }
607         if (page && !cs->write)
608                 flush_dcache_page(page);
609         return 0;
610 }
611
612 /* Copy pages in the request to/from userspace buffer */
613 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
614                            int zeroing)
615 {
616         unsigned i;
617         struct fuse_req *req = cs->req;
618         unsigned offset = req->page_offset;
619         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
620
621         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
622                 struct page *page = req->pages[i];
623                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
624                 if (err)
625                         return err;
626
627                 nbytes -= count;
628                 count = min(nbytes, (unsigned) PAGE_SIZE);
629                 offset = 0;
630         }
631         return 0;
632 }
633
634 /* Copy a single argument in the request to/from userspace buffer */
635 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
636 {
637         while (size) {
638                 if (!cs->len) {
639                         int err = fuse_copy_fill(cs);
640                         if (err)
641                                 return err;
642                 }
643                 fuse_copy_do(cs, &val, &size);
644         }
645         return 0;
646 }
647
648 /* Copy request arguments to/from userspace buffer */
649 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
650                           unsigned argpages, struct fuse_arg *args,
651                           int zeroing)
652 {
653         int err = 0;
654         unsigned i;
655
656         for (i = 0; !err && i < numargs; i++)  {
657                 struct fuse_arg *arg = &args[i];
658                 if (i == numargs - 1 && argpages)
659                         err = fuse_copy_pages(cs, arg->size, zeroing);
660                 else
661                         err = fuse_copy_one(cs, arg->value, arg->size);
662         }
663         return err;
664 }
665
666 static int request_pending(struct fuse_conn *fc)
667 {
668         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
669 }
670
671 /* Wait until a request is available on the pending list */
672 static void request_wait(struct fuse_conn *fc)
673 __releases(&fc->lock)
674 __acquires(&fc->lock)
675 {
676         DECLARE_WAITQUEUE(wait, current);
677
678         add_wait_queue_exclusive(&fc->waitq, &wait);
679         while (fc->connected && !request_pending(fc)) {
680                 set_current_state(TASK_INTERRUPTIBLE);
681                 if (signal_pending(current))
682                         break;
683
684                 spin_unlock(&fc->lock);
685                 schedule();
686                 spin_lock(&fc->lock);
687         }
688         set_current_state(TASK_RUNNING);
689         remove_wait_queue(&fc->waitq, &wait);
690 }
691
692 /*
693  * Transfer an interrupt request to userspace
694  *
695  * Unlike other requests this is assembled on demand, without a need
696  * to allocate a separate fuse_req structure.
697  *
698  * Called with fc->lock held, releases it
699  */
700 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
701                                const struct iovec *iov, unsigned long nr_segs)
702 __releases(&fc->lock)
703 {
704         struct fuse_copy_state cs;
705         struct fuse_in_header ih;
706         struct fuse_interrupt_in arg;
707         unsigned reqsize = sizeof(ih) + sizeof(arg);
708         int err;
709
710         list_del_init(&req->intr_entry);
711         req->intr_unique = fuse_get_unique(fc);
712         memset(&ih, 0, sizeof(ih));
713         memset(&arg, 0, sizeof(arg));
714         ih.len = reqsize;
715         ih.opcode = FUSE_INTERRUPT;
716         ih.unique = req->intr_unique;
717         arg.unique = req->in.h.unique;
718
719         spin_unlock(&fc->lock);
720         if (iov_length(iov, nr_segs) < reqsize)
721                 return -EINVAL;
722
723         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
724         err = fuse_copy_one(&cs, &ih, sizeof(ih));
725         if (!err)
726                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
727         fuse_copy_finish(&cs);
728
729         return err ? err : reqsize;
730 }
731
732 /*
733  * Read a single request into the userspace filesystem's buffer.  This
734  * function waits until a request is available, then removes it from
735  * the pending list and copies request data to userspace buffer.  If
736  * no reply is needed (FORGET) or request has been aborted or there
737  * was an error during the copying then it's finished by calling
738  * request_end().  Otherwise add it to the processing list, and set
739  * the 'sent' flag.
740  */
741 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
742                               unsigned long nr_segs, loff_t pos)
743 {
744         int err;
745         struct fuse_req *req;
746         struct fuse_in *in;
747         struct fuse_copy_state cs;
748         unsigned reqsize;
749         struct file *file = iocb->ki_filp;
750         struct fuse_conn *fc = fuse_get_conn(file);
751         if (!fc)
752                 return -EPERM;
753
754  restart:
755         spin_lock(&fc->lock);
756         err = -EAGAIN;
757         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
758             !request_pending(fc))
759                 goto err_unlock;
760
761         request_wait(fc);
762         err = -ENODEV;
763         if (!fc->connected)
764                 goto err_unlock;
765         err = -ERESTARTSYS;
766         if (!request_pending(fc))
767                 goto err_unlock;
768
769         if (!list_empty(&fc->interrupts)) {
770                 req = list_entry(fc->interrupts.next, struct fuse_req,
771                                  intr_entry);
772                 return fuse_read_interrupt(fc, req, iov, nr_segs);
773         }
774
775         req = list_entry(fc->pending.next, struct fuse_req, list);
776         req->state = FUSE_REQ_READING;
777         list_move(&req->list, &fc->io);
778
779         in = &req->in;
780         reqsize = in->h.len;
781         /* If request is too large, reply with an error and restart the read */
782         if (iov_length(iov, nr_segs) < reqsize) {
783                 req->out.h.error = -EIO;
784                 /* SETXATTR is special, since it may contain too large data */
785                 if (in->h.opcode == FUSE_SETXATTR)
786                         req->out.h.error = -E2BIG;
787                 request_end(fc, req);
788                 goto restart;
789         }
790         spin_unlock(&fc->lock);
791         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
792         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
793         if (!err)
794                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
795                                      (struct fuse_arg *) in->args, 0);
796         fuse_copy_finish(&cs);
797         spin_lock(&fc->lock);
798         req->locked = 0;
799         if (req->aborted) {
800                 request_end(fc, req);
801                 return -ENODEV;
802         }
803         if (err) {
804                 req->out.h.error = -EIO;
805                 request_end(fc, req);
806                 return err;
807         }
808         if (!req->isreply)
809                 request_end(fc, req);
810         else {
811                 req->state = FUSE_REQ_SENT;
812                 list_move_tail(&req->list, &fc->processing);
813                 if (req->interrupted)
814                         queue_interrupt(fc, req);
815                 spin_unlock(&fc->lock);
816         }
817         return reqsize;
818
819  err_unlock:
820         spin_unlock(&fc->lock);
821         return err;
822 }
823
824 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
825                             struct fuse_copy_state *cs)
826 {
827         struct fuse_notify_poll_wakeup_out outarg;
828         int err;
829
830         if (size != sizeof(outarg))
831                 return -EINVAL;
832
833         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
834         if (err)
835                 return err;
836
837         return fuse_notify_poll_wakeup(fc, &outarg);
838 }
839
840 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
841                        unsigned int size, struct fuse_copy_state *cs)
842 {
843         switch (code) {
844         case FUSE_NOTIFY_POLL:
845                 return fuse_notify_poll(fc, size, cs);
846
847         default:
848                 return -EINVAL;
849         }
850 }
851
852 /* Look up request on processing list by unique ID */
853 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
854 {
855         struct list_head *entry;
856
857         list_for_each(entry, &fc->processing) {
858                 struct fuse_req *req;
859                 req = list_entry(entry, struct fuse_req, list);
860                 if (req->in.h.unique == unique || req->intr_unique == unique)
861                         return req;
862         }
863         return NULL;
864 }
865
866 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
867                          unsigned nbytes)
868 {
869         unsigned reqsize = sizeof(struct fuse_out_header);
870
871         if (out->h.error)
872                 return nbytes != reqsize ? -EINVAL : 0;
873
874         reqsize += len_args(out->numargs, out->args);
875
876         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
877                 return -EINVAL;
878         else if (reqsize > nbytes) {
879                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
880                 unsigned diffsize = reqsize - nbytes;
881                 if (diffsize > lastarg->size)
882                         return -EINVAL;
883                 lastarg->size -= diffsize;
884         }
885         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
886                               out->page_zeroing);
887 }
888
889 /*
890  * Write a single reply to a request.  First the header is copied from
891  * the write buffer.  The request is then searched on the processing
892  * list by the unique ID found in the header.  If found, then remove
893  * it from the list and copy the rest of the buffer to the request.
894  * The request is finished by calling request_end()
895  */
896 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
897                                unsigned long nr_segs, loff_t pos)
898 {
899         int err;
900         unsigned nbytes = iov_length(iov, nr_segs);
901         struct fuse_req *req;
902         struct fuse_out_header oh;
903         struct fuse_copy_state cs;
904         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
905         if (!fc)
906                 return -EPERM;
907
908         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
909         if (nbytes < sizeof(struct fuse_out_header))
910                 return -EINVAL;
911
912         err = fuse_copy_one(&cs, &oh, sizeof(oh));
913         if (err)
914                 goto err_finish;
915
916         err = -EINVAL;
917         if (oh.len != nbytes)
918                 goto err_finish;
919
920         /*
921          * Zero oh.unique indicates unsolicited notification message
922          * and error contains notification code.
923          */
924         if (!oh.unique) {
925                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), &cs);
926                 fuse_copy_finish(&cs);
927                 return err ? err : nbytes;
928         }
929
930         err = -EINVAL;
931         if (oh.error <= -1000 || oh.error > 0)
932                 goto err_finish;
933
934         spin_lock(&fc->lock);
935         err = -ENOENT;
936         if (!fc->connected)
937                 goto err_unlock;
938
939         req = request_find(fc, oh.unique);
940         if (!req)
941                 goto err_unlock;
942
943         if (req->aborted) {
944                 spin_unlock(&fc->lock);
945                 fuse_copy_finish(&cs);
946                 spin_lock(&fc->lock);
947                 request_end(fc, req);
948                 return -ENOENT;
949         }
950         /* Is it an interrupt reply? */
951         if (req->intr_unique == oh.unique) {
952                 err = -EINVAL;
953                 if (nbytes != sizeof(struct fuse_out_header))
954                         goto err_unlock;
955
956                 if (oh.error == -ENOSYS)
957                         fc->no_interrupt = 1;
958                 else if (oh.error == -EAGAIN)
959                         queue_interrupt(fc, req);
960
961                 spin_unlock(&fc->lock);
962                 fuse_copy_finish(&cs);
963                 return nbytes;
964         }
965
966         req->state = FUSE_REQ_WRITING;
967         list_move(&req->list, &fc->io);
968         req->out.h = oh;
969         req->locked = 1;
970         cs.req = req;
971         spin_unlock(&fc->lock);
972
973         err = copy_out_args(&cs, &req->out, nbytes);
974         fuse_copy_finish(&cs);
975
976         spin_lock(&fc->lock);
977         req->locked = 0;
978         if (!err) {
979                 if (req->aborted)
980                         err = -ENOENT;
981         } else if (!req->aborted)
982                 req->out.h.error = -EIO;
983         request_end(fc, req);
984
985         return err ? err : nbytes;
986
987  err_unlock:
988         spin_unlock(&fc->lock);
989  err_finish:
990         fuse_copy_finish(&cs);
991         return err;
992 }
993
994 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
995 {
996         unsigned mask = POLLOUT | POLLWRNORM;
997         struct fuse_conn *fc = fuse_get_conn(file);
998         if (!fc)
999                 return POLLERR;
1000
1001         poll_wait(file, &fc->waitq, wait);
1002
1003         spin_lock(&fc->lock);
1004         if (!fc->connected)
1005                 mask = POLLERR;
1006         else if (request_pending(fc))
1007                 mask |= POLLIN | POLLRDNORM;
1008         spin_unlock(&fc->lock);
1009
1010         return mask;
1011 }
1012
1013 /*
1014  * Abort all requests on the given list (pending or processing)
1015  *
1016  * This function releases and reacquires fc->lock
1017  */
1018 static void end_requests(struct fuse_conn *fc, struct list_head *head)
1019 __releases(&fc->lock)
1020 __acquires(&fc->lock)
1021 {
1022         while (!list_empty(head)) {
1023                 struct fuse_req *req;
1024                 req = list_entry(head->next, struct fuse_req, list);
1025                 req->out.h.error = -ECONNABORTED;
1026                 request_end(fc, req);
1027                 spin_lock(&fc->lock);
1028         }
1029 }
1030
1031 /*
1032  * Abort requests under I/O
1033  *
1034  * The requests are set to aborted and finished, and the request
1035  * waiter is woken up.  This will make request_wait_answer() wait
1036  * until the request is unlocked and then return.
1037  *
1038  * If the request is asynchronous, then the end function needs to be
1039  * called after waiting for the request to be unlocked (if it was
1040  * locked).
1041  */
1042 static void end_io_requests(struct fuse_conn *fc)
1043 __releases(&fc->lock)
1044 __acquires(&fc->lock)
1045 {
1046         while (!list_empty(&fc->io)) {
1047                 struct fuse_req *req =
1048                         list_entry(fc->io.next, struct fuse_req, list);
1049                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
1050
1051                 req->aborted = 1;
1052                 req->out.h.error = -ECONNABORTED;
1053                 req->state = FUSE_REQ_FINISHED;
1054                 list_del_init(&req->list);
1055                 wake_up(&req->waitq);
1056                 if (end) {
1057                         req->end = NULL;
1058                         __fuse_get_request(req);
1059                         spin_unlock(&fc->lock);
1060                         wait_event(req->waitq, !req->locked);
1061                         end(fc, req);
1062                         fuse_put_request(fc, req);
1063                         spin_lock(&fc->lock);
1064                 }
1065         }
1066 }
1067
1068 /*
1069  * Abort all requests.
1070  *
1071  * Emergency exit in case of a malicious or accidental deadlock, or
1072  * just a hung filesystem.
1073  *
1074  * The same effect is usually achievable through killing the
1075  * filesystem daemon and all users of the filesystem.  The exception
1076  * is the combination of an asynchronous request and the tricky
1077  * deadlock (see Documentation/filesystems/fuse.txt).
1078  *
1079  * During the aborting, progression of requests from the pending and
1080  * processing lists onto the io list, and progression of new requests
1081  * onto the pending list is prevented by req->connected being false.
1082  *
1083  * Progression of requests under I/O to the processing list is
1084  * prevented by the req->aborted flag being true for these requests.
1085  * For this reason requests on the io list must be aborted first.
1086  */
1087 void fuse_abort_conn(struct fuse_conn *fc)
1088 {
1089         spin_lock(&fc->lock);
1090         if (fc->connected) {
1091                 fc->connected = 0;
1092                 fc->blocked = 0;
1093                 end_io_requests(fc);
1094                 end_requests(fc, &fc->pending);
1095                 end_requests(fc, &fc->processing);
1096                 wake_up_all(&fc->waitq);
1097                 wake_up_all(&fc->blocked_waitq);
1098                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
1099         }
1100         spin_unlock(&fc->lock);
1101 }
1102
1103 static int fuse_dev_release(struct inode *inode, struct file *file)
1104 {
1105         struct fuse_conn *fc = fuse_get_conn(file);
1106         if (fc) {
1107                 spin_lock(&fc->lock);
1108                 fc->connected = 0;
1109                 end_requests(fc, &fc->pending);
1110                 end_requests(fc, &fc->processing);
1111                 spin_unlock(&fc->lock);
1112                 fuse_conn_put(fc);
1113         }
1114
1115         return 0;
1116 }
1117
1118 static int fuse_dev_fasync(int fd, struct file *file, int on)
1119 {
1120         struct fuse_conn *fc = fuse_get_conn(file);
1121         if (!fc)
1122                 return -EPERM;
1123
1124         /* No locking - fasync_helper does its own locking */
1125         return fasync_helper(fd, file, on, &fc->fasync);
1126 }
1127
1128 const struct file_operations fuse_dev_operations = {
1129         .owner          = THIS_MODULE,
1130         .llseek         = no_llseek,
1131         .read           = do_sync_read,
1132         .aio_read       = fuse_dev_read,
1133         .write          = do_sync_write,
1134         .aio_write      = fuse_dev_write,
1135         .poll           = fuse_dev_poll,
1136         .release        = fuse_dev_release,
1137         .fasync         = fuse_dev_fasync,
1138 };
1139
1140 static struct miscdevice fuse_miscdevice = {
1141         .minor = FUSE_MINOR,
1142         .name  = "fuse",
1143         .fops = &fuse_dev_operations,
1144 };
1145
1146 int __init fuse_dev_init(void)
1147 {
1148         int err = -ENOMEM;
1149         fuse_req_cachep = kmem_cache_create("fuse_request",
1150                                             sizeof(struct fuse_req),
1151                                             0, 0, NULL);
1152         if (!fuse_req_cachep)
1153                 goto out;
1154
1155         err = misc_register(&fuse_miscdevice);
1156         if (err)
1157                 goto out_cache_clean;
1158
1159         return 0;
1160
1161  out_cache_clean:
1162         kmem_cache_destroy(fuse_req_cachep);
1163  out:
1164         return err;
1165 }
1166
1167 void fuse_dev_cleanup(void)
1168 {
1169         misc_deregister(&fuse_miscdevice);
1170         kmem_cache_destroy(fuse_req_cachep);
1171 }