Merge master.kernel.org:/pub/scm/linux/kernel/git/jejb/scsi-misc-2.6
[sfrench/cifs-2.6.git] / fs / fuse / dev.c
1 /*
2   FUSE: Filesystem in Userspace
3   Copyright (C) 2001-2006  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 void fuse_request_free(struct fuse_req *req)
51 {
52         kmem_cache_free(fuse_req_cachep, req);
53 }
54
55 static void block_sigs(sigset_t *oldset)
56 {
57         sigset_t mask;
58
59         siginitsetinv(&mask, sigmask(SIGKILL));
60         sigprocmask(SIG_BLOCK, &mask, oldset);
61 }
62
63 static void restore_sigs(sigset_t *oldset)
64 {
65         sigprocmask(SIG_SETMASK, oldset, NULL);
66 }
67
68 static void __fuse_get_request(struct fuse_req *req)
69 {
70         atomic_inc(&req->count);
71 }
72
73 /* Must be called with > 1 refcount */
74 static void __fuse_put_request(struct fuse_req *req)
75 {
76         BUG_ON(atomic_read(&req->count) < 2);
77         atomic_dec(&req->count);
78 }
79
80 static void fuse_req_init_context(struct fuse_req *req)
81 {
82         req->in.h.uid = current->fsuid;
83         req->in.h.gid = current->fsgid;
84         req->in.h.pid = current->pid;
85 }
86
87 struct fuse_req *fuse_get_req(struct fuse_conn *fc)
88 {
89         struct fuse_req *req;
90         sigset_t oldset;
91         int intr;
92         int err;
93
94         atomic_inc(&fc->num_waiting);
95         block_sigs(&oldset);
96         intr = wait_event_interruptible(fc->blocked_waitq, !fc->blocked);
97         restore_sigs(&oldset);
98         err = -EINTR;
99         if (intr)
100                 goto out;
101
102         err = -ENOTCONN;
103         if (!fc->connected)
104                 goto out;
105
106         req = fuse_request_alloc();
107         err = -ENOMEM;
108         if (!req)
109                 goto out;
110
111         fuse_req_init_context(req);
112         req->waiting = 1;
113         return req;
114
115  out:
116         atomic_dec(&fc->num_waiting);
117         return ERR_PTR(err);
118 }
119
120 /*
121  * Return request in fuse_file->reserved_req.  However that may
122  * currently be in use.  If that is the case, wait for it to become
123  * available.
124  */
125 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
126                                          struct file *file)
127 {
128         struct fuse_req *req = NULL;
129         struct fuse_file *ff = file->private_data;
130
131         do {
132                 wait_event(fc->blocked_waitq, ff->reserved_req);
133                 spin_lock(&fc->lock);
134                 if (ff->reserved_req) {
135                         req = ff->reserved_req;
136                         ff->reserved_req = NULL;
137                         get_file(file);
138                         req->stolen_file = file;
139                 }
140                 spin_unlock(&fc->lock);
141         } while (!req);
142
143         return req;
144 }
145
146 /*
147  * Put stolen request back into fuse_file->reserved_req
148  */
149 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
150 {
151         struct file *file = req->stolen_file;
152         struct fuse_file *ff = file->private_data;
153
154         spin_lock(&fc->lock);
155         fuse_request_init(req);
156         BUG_ON(ff->reserved_req);
157         ff->reserved_req = req;
158         wake_up(&fc->blocked_waitq);
159         spin_unlock(&fc->lock);
160         fput(file);
161 }
162
163 /*
164  * Gets a requests for a file operation, always succeeds
165  *
166  * This is used for sending the FLUSH request, which must get to
167  * userspace, due to POSIX locks which may need to be unlocked.
168  *
169  * If allocation fails due to OOM, use the reserved request in
170  * fuse_file.
171  *
172  * This is very unlikely to deadlock accidentally, since the
173  * filesystem should not have it's own file open.  If deadlock is
174  * intentional, it can still be broken by "aborting" the filesystem.
175  */
176 struct fuse_req *fuse_get_req_nofail(struct fuse_conn *fc, struct file *file)
177 {
178         struct fuse_req *req;
179
180         atomic_inc(&fc->num_waiting);
181         wait_event(fc->blocked_waitq, !fc->blocked);
182         req = fuse_request_alloc();
183         if (!req)
184                 req = get_reserved_req(fc, file);
185
186         fuse_req_init_context(req);
187         req->waiting = 1;
188         return req;
189 }
190
191 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
192 {
193         if (atomic_dec_and_test(&req->count)) {
194                 if (req->waiting)
195                         atomic_dec(&fc->num_waiting);
196
197                 if (req->stolen_file)
198                         put_reserved_req(fc, req);
199                 else
200                         fuse_request_free(req);
201         }
202 }
203
204 /*
205  * This function is called when a request is finished.  Either a reply
206  * has arrived or it was aborted (and not yet sent) or some error
207  * occurred during communication with userspace, or the device file
208  * was closed.  The requester thread is woken up (if still waiting),
209  * the 'end' callback is called if given, else the reference to the
210  * request is released
211  *
212  * Called with fc->lock, unlocks it
213  */
214 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
215         __releases(fc->lock)
216 {
217         void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
218         req->end = NULL;
219         list_del(&req->list);
220         list_del(&req->intr_entry);
221         req->state = FUSE_REQ_FINISHED;
222         if (req->background) {
223                 if (fc->num_background == FUSE_MAX_BACKGROUND) {
224                         fc->blocked = 0;
225                         wake_up_all(&fc->blocked_waitq);
226                 }
227                 fc->num_background--;
228         }
229         spin_unlock(&fc->lock);
230         dput(req->dentry);
231         mntput(req->vfsmount);
232         if (req->file)
233                 fput(req->file);
234         wake_up(&req->waitq);
235         if (end)
236                 end(fc, req);
237         else
238                 fuse_put_request(fc, req);
239 }
240
241 static void wait_answer_interruptible(struct fuse_conn *fc,
242                                       struct fuse_req *req)
243 {
244         if (signal_pending(current))
245                 return;
246
247         spin_unlock(&fc->lock);
248         wait_event_interruptible(req->waitq, req->state == FUSE_REQ_FINISHED);
249         spin_lock(&fc->lock);
250 }
251
252 static void queue_interrupt(struct fuse_conn *fc, struct fuse_req *req)
253 {
254         list_add_tail(&req->intr_entry, &fc->interrupts);
255         wake_up(&fc->waitq);
256         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
257 }
258
259 /* Called with fc->lock held.  Releases, and then reacquires it. */
260 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
261 {
262         if (!fc->no_interrupt) {
263                 /* Any signal may interrupt this */
264                 wait_answer_interruptible(fc, req);
265
266                 if (req->aborted)
267                         goto aborted;
268                 if (req->state == FUSE_REQ_FINISHED)
269                         return;
270
271                 req->interrupted = 1;
272                 if (req->state == FUSE_REQ_SENT)
273                         queue_interrupt(fc, req);
274         }
275
276         if (req->force) {
277                 spin_unlock(&fc->lock);
278                 wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
279                 spin_lock(&fc->lock);
280         } else {
281                 sigset_t oldset;
282
283                 /* Only fatal signals may interrupt this */
284                 block_sigs(&oldset);
285                 wait_answer_interruptible(fc, req);
286                 restore_sigs(&oldset);
287         }
288
289         if (req->aborted)
290                 goto aborted;
291         if (req->state == FUSE_REQ_FINISHED)
292                 return;
293
294         req->out.h.error = -EINTR;
295         req->aborted = 1;
296
297  aborted:
298         if (req->locked) {
299                 /* This is uninterruptible sleep, because data is
300                    being copied to/from the buffers of req.  During
301                    locked state, there mustn't be any filesystem
302                    operation (e.g. page fault), since that could lead
303                    to deadlock */
304                 spin_unlock(&fc->lock);
305                 wait_event(req->waitq, !req->locked);
306                 spin_lock(&fc->lock);
307         }
308         if (req->state == FUSE_REQ_PENDING) {
309                 list_del(&req->list);
310                 __fuse_put_request(req);
311         } else if (req->state == FUSE_REQ_SENT) {
312                 spin_unlock(&fc->lock);
313                 wait_event(req->waitq, req->state == FUSE_REQ_FINISHED);
314                 spin_lock(&fc->lock);
315         }
316 }
317
318 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
319 {
320         unsigned nbytes = 0;
321         unsigned i;
322
323         for (i = 0; i < numargs; i++)
324                 nbytes += args[i].size;
325
326         return nbytes;
327 }
328
329 static u64 fuse_get_unique(struct fuse_conn *fc)
330  {
331         fc->reqctr++;
332         /* zero is special */
333         if (fc->reqctr == 0)
334                 fc->reqctr = 1;
335
336         return fc->reqctr;
337 }
338
339 static void queue_request(struct fuse_conn *fc, struct fuse_req *req)
340 {
341         req->in.h.unique = fuse_get_unique(fc);
342         req->in.h.len = sizeof(struct fuse_in_header) +
343                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
344         list_add_tail(&req->list, &fc->pending);
345         req->state = FUSE_REQ_PENDING;
346         if (!req->waiting) {
347                 req->waiting = 1;
348                 atomic_inc(&fc->num_waiting);
349         }
350         wake_up(&fc->waitq);
351         kill_fasync(&fc->fasync, SIGIO, POLL_IN);
352 }
353
354 void request_send(struct fuse_conn *fc, struct fuse_req *req)
355 {
356         req->isreply = 1;
357         spin_lock(&fc->lock);
358         if (!fc->connected)
359                 req->out.h.error = -ENOTCONN;
360         else if (fc->conn_error)
361                 req->out.h.error = -ECONNREFUSED;
362         else {
363                 queue_request(fc, req);
364                 /* acquire extra reference, since request is still needed
365                    after request_end() */
366                 __fuse_get_request(req);
367
368                 request_wait_answer(fc, req);
369         }
370         spin_unlock(&fc->lock);
371 }
372
373 static void request_send_nowait(struct fuse_conn *fc, struct fuse_req *req)
374 {
375         spin_lock(&fc->lock);
376         if (fc->connected) {
377                 req->background = 1;
378                 fc->num_background++;
379                 if (fc->num_background == FUSE_MAX_BACKGROUND)
380                         fc->blocked = 1;
381
382                 queue_request(fc, req);
383                 spin_unlock(&fc->lock);
384         } else {
385                 req->out.h.error = -ENOTCONN;
386                 request_end(fc, req);
387         }
388 }
389
390 void request_send_noreply(struct fuse_conn *fc, struct fuse_req *req)
391 {
392         req->isreply = 0;
393         request_send_nowait(fc, req);
394 }
395
396 void request_send_background(struct fuse_conn *fc, struct fuse_req *req)
397 {
398         req->isreply = 1;
399         request_send_nowait(fc, req);
400 }
401
402 /*
403  * Lock the request.  Up to the next unlock_request() there mustn't be
404  * anything that could cause a page-fault.  If the request was already
405  * aborted bail out.
406  */
407 static int lock_request(struct fuse_conn *fc, struct fuse_req *req)
408 {
409         int err = 0;
410         if (req) {
411                 spin_lock(&fc->lock);
412                 if (req->aborted)
413                         err = -ENOENT;
414                 else
415                         req->locked = 1;
416                 spin_unlock(&fc->lock);
417         }
418         return err;
419 }
420
421 /*
422  * Unlock request.  If it was aborted during being locked, the
423  * requester thread is currently waiting for it to be unlocked, so
424  * wake it up.
425  */
426 static void unlock_request(struct fuse_conn *fc, struct fuse_req *req)
427 {
428         if (req) {
429                 spin_lock(&fc->lock);
430                 req->locked = 0;
431                 if (req->aborted)
432                         wake_up(&req->waitq);
433                 spin_unlock(&fc->lock);
434         }
435 }
436
437 struct fuse_copy_state {
438         struct fuse_conn *fc;
439         int write;
440         struct fuse_req *req;
441         const struct iovec *iov;
442         unsigned long nr_segs;
443         unsigned long seglen;
444         unsigned long addr;
445         struct page *pg;
446         void *mapaddr;
447         void *buf;
448         unsigned len;
449 };
450
451 static void fuse_copy_init(struct fuse_copy_state *cs, struct fuse_conn *fc,
452                            int write, struct fuse_req *req,
453                            const struct iovec *iov, unsigned long nr_segs)
454 {
455         memset(cs, 0, sizeof(*cs));
456         cs->fc = fc;
457         cs->write = write;
458         cs->req = req;
459         cs->iov = iov;
460         cs->nr_segs = nr_segs;
461 }
462
463 /* Unmap and put previous page of userspace buffer */
464 static void fuse_copy_finish(struct fuse_copy_state *cs)
465 {
466         if (cs->mapaddr) {
467                 kunmap_atomic(cs->mapaddr, KM_USER0);
468                 if (cs->write) {
469                         flush_dcache_page(cs->pg);
470                         set_page_dirty_lock(cs->pg);
471                 }
472                 put_page(cs->pg);
473                 cs->mapaddr = NULL;
474         }
475 }
476
477 /*
478  * Get another pagefull of userspace buffer, and map it to kernel
479  * address space, and lock request
480  */
481 static int fuse_copy_fill(struct fuse_copy_state *cs)
482 {
483         unsigned long offset;
484         int err;
485
486         unlock_request(cs->fc, cs->req);
487         fuse_copy_finish(cs);
488         if (!cs->seglen) {
489                 BUG_ON(!cs->nr_segs);
490                 cs->seglen = cs->iov[0].iov_len;
491                 cs->addr = (unsigned long) cs->iov[0].iov_base;
492                 cs->iov ++;
493                 cs->nr_segs --;
494         }
495         down_read(&current->mm->mmap_sem);
496         err = get_user_pages(current, current->mm, cs->addr, 1, cs->write, 0,
497                              &cs->pg, NULL);
498         up_read(&current->mm->mmap_sem);
499         if (err < 0)
500                 return err;
501         BUG_ON(err != 1);
502         offset = cs->addr % PAGE_SIZE;
503         cs->mapaddr = kmap_atomic(cs->pg, KM_USER0);
504         cs->buf = cs->mapaddr + offset;
505         cs->len = min(PAGE_SIZE - offset, cs->seglen);
506         cs->seglen -= cs->len;
507         cs->addr += cs->len;
508
509         return lock_request(cs->fc, cs->req);
510 }
511
512 /* Do as much copy to/from userspace buffer as we can */
513 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
514 {
515         unsigned ncpy = min(*size, cs->len);
516         if (val) {
517                 if (cs->write)
518                         memcpy(cs->buf, *val, ncpy);
519                 else
520                         memcpy(*val, cs->buf, ncpy);
521                 *val += ncpy;
522         }
523         *size -= ncpy;
524         cs->len -= ncpy;
525         cs->buf += ncpy;
526         return ncpy;
527 }
528
529 /*
530  * Copy a page in the request to/from the userspace buffer.  Must be
531  * done atomically
532  */
533 static int fuse_copy_page(struct fuse_copy_state *cs, struct page *page,
534                           unsigned offset, unsigned count, int zeroing)
535 {
536         if (page && zeroing && count < PAGE_SIZE) {
537                 void *mapaddr = kmap_atomic(page, KM_USER1);
538                 memset(mapaddr, 0, PAGE_SIZE);
539                 kunmap_atomic(mapaddr, KM_USER1);
540         }
541         while (count) {
542                 int err;
543                 if (!cs->len && (err = fuse_copy_fill(cs)))
544                         return err;
545                 if (page) {
546                         void *mapaddr = kmap_atomic(page, KM_USER1);
547                         void *buf = mapaddr + offset;
548                         offset += fuse_copy_do(cs, &buf, &count);
549                         kunmap_atomic(mapaddr, KM_USER1);
550                 } else
551                         offset += fuse_copy_do(cs, NULL, &count);
552         }
553         if (page && !cs->write)
554                 flush_dcache_page(page);
555         return 0;
556 }
557
558 /* Copy pages in the request to/from userspace buffer */
559 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
560                            int zeroing)
561 {
562         unsigned i;
563         struct fuse_req *req = cs->req;
564         unsigned offset = req->page_offset;
565         unsigned count = min(nbytes, (unsigned) PAGE_SIZE - offset);
566
567         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
568                 struct page *page = req->pages[i];
569                 int err = fuse_copy_page(cs, page, offset, count, zeroing);
570                 if (err)
571                         return err;
572
573                 nbytes -= count;
574                 count = min(nbytes, (unsigned) PAGE_SIZE);
575                 offset = 0;
576         }
577         return 0;
578 }
579
580 /* Copy a single argument in the request to/from userspace buffer */
581 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
582 {
583         while (size) {
584                 int err;
585                 if (!cs->len && (err = fuse_copy_fill(cs)))
586                         return err;
587                 fuse_copy_do(cs, &val, &size);
588         }
589         return 0;
590 }
591
592 /* Copy request arguments to/from userspace buffer */
593 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
594                           unsigned argpages, struct fuse_arg *args,
595                           int zeroing)
596 {
597         int err = 0;
598         unsigned i;
599
600         for (i = 0; !err && i < numargs; i++)  {
601                 struct fuse_arg *arg = &args[i];
602                 if (i == numargs - 1 && argpages)
603                         err = fuse_copy_pages(cs, arg->size, zeroing);
604                 else
605                         err = fuse_copy_one(cs, arg->value, arg->size);
606         }
607         return err;
608 }
609
610 static int request_pending(struct fuse_conn *fc)
611 {
612         return !list_empty(&fc->pending) || !list_empty(&fc->interrupts);
613 }
614
615 /* Wait until a request is available on the pending list */
616 static void request_wait(struct fuse_conn *fc)
617 {
618         DECLARE_WAITQUEUE(wait, current);
619
620         add_wait_queue_exclusive(&fc->waitq, &wait);
621         while (fc->connected && !request_pending(fc)) {
622                 set_current_state(TASK_INTERRUPTIBLE);
623                 if (signal_pending(current))
624                         break;
625
626                 spin_unlock(&fc->lock);
627                 schedule();
628                 spin_lock(&fc->lock);
629         }
630         set_current_state(TASK_RUNNING);
631         remove_wait_queue(&fc->waitq, &wait);
632 }
633
634 /*
635  * Transfer an interrupt request to userspace
636  *
637  * Unlike other requests this is assembled on demand, without a need
638  * to allocate a separate fuse_req structure.
639  *
640  * Called with fc->lock held, releases it
641  */
642 static int fuse_read_interrupt(struct fuse_conn *fc, struct fuse_req *req,
643                                const struct iovec *iov, unsigned long nr_segs)
644         __releases(fc->lock)
645 {
646         struct fuse_copy_state cs;
647         struct fuse_in_header ih;
648         struct fuse_interrupt_in arg;
649         unsigned reqsize = sizeof(ih) + sizeof(arg);
650         int err;
651
652         list_del_init(&req->intr_entry);
653         req->intr_unique = fuse_get_unique(fc);
654         memset(&ih, 0, sizeof(ih));
655         memset(&arg, 0, sizeof(arg));
656         ih.len = reqsize;
657         ih.opcode = FUSE_INTERRUPT;
658         ih.unique = req->intr_unique;
659         arg.unique = req->in.h.unique;
660
661         spin_unlock(&fc->lock);
662         if (iov_length(iov, nr_segs) < reqsize)
663                 return -EINVAL;
664
665         fuse_copy_init(&cs, fc, 1, NULL, iov, nr_segs);
666         err = fuse_copy_one(&cs, &ih, sizeof(ih));
667         if (!err)
668                 err = fuse_copy_one(&cs, &arg, sizeof(arg));
669         fuse_copy_finish(&cs);
670
671         return err ? err : reqsize;
672 }
673
674 /*
675  * Read a single request into the userspace filesystem's buffer.  This
676  * function waits until a request is available, then removes it from
677  * the pending list and copies request data to userspace buffer.  If
678  * no reply is needed (FORGET) or request has been aborted or there
679  * was an error during the copying then it's finished by calling
680  * request_end().  Otherwise add it to the processing list, and set
681  * the 'sent' flag.
682  */
683 static ssize_t fuse_dev_read(struct kiocb *iocb, const struct iovec *iov,
684                               unsigned long nr_segs, loff_t pos)
685 {
686         int err;
687         struct fuse_req *req;
688         struct fuse_in *in;
689         struct fuse_copy_state cs;
690         unsigned reqsize;
691         struct file *file = iocb->ki_filp;
692         struct fuse_conn *fc = fuse_get_conn(file);
693         if (!fc)
694                 return -EPERM;
695
696  restart:
697         spin_lock(&fc->lock);
698         err = -EAGAIN;
699         if ((file->f_flags & O_NONBLOCK) && fc->connected &&
700             !request_pending(fc))
701                 goto err_unlock;
702
703         request_wait(fc);
704         err = -ENODEV;
705         if (!fc->connected)
706                 goto err_unlock;
707         err = -ERESTARTSYS;
708         if (!request_pending(fc))
709                 goto err_unlock;
710
711         if (!list_empty(&fc->interrupts)) {
712                 req = list_entry(fc->interrupts.next, struct fuse_req,
713                                  intr_entry);
714                 return fuse_read_interrupt(fc, req, iov, nr_segs);
715         }
716
717         req = list_entry(fc->pending.next, struct fuse_req, list);
718         req->state = FUSE_REQ_READING;
719         list_move(&req->list, &fc->io);
720
721         in = &req->in;
722         reqsize = in->h.len;
723         /* If request is too large, reply with an error and restart the read */
724         if (iov_length(iov, nr_segs) < reqsize) {
725                 req->out.h.error = -EIO;
726                 /* SETXATTR is special, since it may contain too large data */
727                 if (in->h.opcode == FUSE_SETXATTR)
728                         req->out.h.error = -E2BIG;
729                 request_end(fc, req);
730                 goto restart;
731         }
732         spin_unlock(&fc->lock);
733         fuse_copy_init(&cs, fc, 1, req, iov, nr_segs);
734         err = fuse_copy_one(&cs, &in->h, sizeof(in->h));
735         if (!err)
736                 err = fuse_copy_args(&cs, in->numargs, in->argpages,
737                                      (struct fuse_arg *) in->args, 0);
738         fuse_copy_finish(&cs);
739         spin_lock(&fc->lock);
740         req->locked = 0;
741         if (!err && req->aborted)
742                 err = -ENOENT;
743         if (err) {
744                 if (!req->aborted)
745                         req->out.h.error = -EIO;
746                 request_end(fc, req);
747                 return err;
748         }
749         if (!req->isreply)
750                 request_end(fc, req);
751         else {
752                 req->state = FUSE_REQ_SENT;
753                 list_move_tail(&req->list, &fc->processing);
754                 if (req->interrupted)
755                         queue_interrupt(fc, req);
756                 spin_unlock(&fc->lock);
757         }
758         return reqsize;
759
760  err_unlock:
761         spin_unlock(&fc->lock);
762         return err;
763 }
764
765 /* Look up request on processing list by unique ID */
766 static struct fuse_req *request_find(struct fuse_conn *fc, u64 unique)
767 {
768         struct list_head *entry;
769
770         list_for_each(entry, &fc->processing) {
771                 struct fuse_req *req;
772                 req = list_entry(entry, struct fuse_req, list);
773                 if (req->in.h.unique == unique || req->intr_unique == unique)
774                         return req;
775         }
776         return NULL;
777 }
778
779 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
780                          unsigned nbytes)
781 {
782         unsigned reqsize = sizeof(struct fuse_out_header);
783
784         if (out->h.error)
785                 return nbytes != reqsize ? -EINVAL : 0;
786
787         reqsize += len_args(out->numargs, out->args);
788
789         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
790                 return -EINVAL;
791         else if (reqsize > nbytes) {
792                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
793                 unsigned diffsize = reqsize - nbytes;
794                 if (diffsize > lastarg->size)
795                         return -EINVAL;
796                 lastarg->size -= diffsize;
797         }
798         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
799                               out->page_zeroing);
800 }
801
802 /*
803  * Write a single reply to a request.  First the header is copied from
804  * the write buffer.  The request is then searched on the processing
805  * list by the unique ID found in the header.  If found, then remove
806  * it from the list and copy the rest of the buffer to the request.
807  * The request is finished by calling request_end()
808  */
809 static ssize_t fuse_dev_write(struct kiocb *iocb, const struct iovec *iov,
810                                unsigned long nr_segs, loff_t pos)
811 {
812         int err;
813         unsigned nbytes = iov_length(iov, nr_segs);
814         struct fuse_req *req;
815         struct fuse_out_header oh;
816         struct fuse_copy_state cs;
817         struct fuse_conn *fc = fuse_get_conn(iocb->ki_filp);
818         if (!fc)
819                 return -EPERM;
820
821         fuse_copy_init(&cs, fc, 0, NULL, iov, nr_segs);
822         if (nbytes < sizeof(struct fuse_out_header))
823                 return -EINVAL;
824
825         err = fuse_copy_one(&cs, &oh, sizeof(oh));
826         if (err)
827                 goto err_finish;
828         err = -EINVAL;
829         if (!oh.unique || oh.error <= -1000 || oh.error > 0 ||
830             oh.len != nbytes)
831                 goto err_finish;
832
833         spin_lock(&fc->lock);
834         err = -ENOENT;
835         if (!fc->connected)
836                 goto err_unlock;
837
838         req = request_find(fc, oh.unique);
839         if (!req)
840                 goto err_unlock;
841
842         if (req->aborted) {
843                 spin_unlock(&fc->lock);
844                 fuse_copy_finish(&cs);
845                 spin_lock(&fc->lock);
846                 request_end(fc, req);
847                 return -ENOENT;
848         }
849         /* Is it an interrupt reply? */
850         if (req->intr_unique == oh.unique) {
851                 err = -EINVAL;
852                 if (nbytes != sizeof(struct fuse_out_header))
853                         goto err_unlock;
854
855                 if (oh.error == -ENOSYS)
856                         fc->no_interrupt = 1;
857                 else if (oh.error == -EAGAIN)
858                         queue_interrupt(fc, req);
859
860                 spin_unlock(&fc->lock);
861                 fuse_copy_finish(&cs);
862                 return nbytes;
863         }
864
865         req->state = FUSE_REQ_WRITING;
866         list_move(&req->list, &fc->io);
867         req->out.h = oh;
868         req->locked = 1;
869         cs.req = req;
870         spin_unlock(&fc->lock);
871
872         err = copy_out_args(&cs, &req->out, nbytes);
873         fuse_copy_finish(&cs);
874
875         spin_lock(&fc->lock);
876         req->locked = 0;
877         if (!err) {
878                 if (req->aborted)
879                         err = -ENOENT;
880         } else if (!req->aborted)
881                 req->out.h.error = -EIO;
882         request_end(fc, req);
883
884         return err ? err : nbytes;
885
886  err_unlock:
887         spin_unlock(&fc->lock);
888  err_finish:
889         fuse_copy_finish(&cs);
890         return err;
891 }
892
893 static unsigned fuse_dev_poll(struct file *file, poll_table *wait)
894 {
895         unsigned mask = POLLOUT | POLLWRNORM;
896         struct fuse_conn *fc = fuse_get_conn(file);
897         if (!fc)
898                 return POLLERR;
899
900         poll_wait(file, &fc->waitq, wait);
901
902         spin_lock(&fc->lock);
903         if (!fc->connected)
904                 mask = POLLERR;
905         else if (request_pending(fc))
906                 mask |= POLLIN | POLLRDNORM;
907         spin_unlock(&fc->lock);
908
909         return mask;
910 }
911
912 /*
913  * Abort all requests on the given list (pending or processing)
914  *
915  * This function releases and reacquires fc->lock
916  */
917 static void end_requests(struct fuse_conn *fc, struct list_head *head)
918 {
919         while (!list_empty(head)) {
920                 struct fuse_req *req;
921                 req = list_entry(head->next, struct fuse_req, list);
922                 req->out.h.error = -ECONNABORTED;
923                 request_end(fc, req);
924                 spin_lock(&fc->lock);
925         }
926 }
927
928 /*
929  * Abort requests under I/O
930  *
931  * The requests are set to aborted and finished, and the request
932  * waiter is woken up.  This will make request_wait_answer() wait
933  * until the request is unlocked and then return.
934  *
935  * If the request is asynchronous, then the end function needs to be
936  * called after waiting for the request to be unlocked (if it was
937  * locked).
938  */
939 static void end_io_requests(struct fuse_conn *fc)
940 {
941         while (!list_empty(&fc->io)) {
942                 struct fuse_req *req =
943                         list_entry(fc->io.next, struct fuse_req, list);
944                 void (*end) (struct fuse_conn *, struct fuse_req *) = req->end;
945
946                 req->aborted = 1;
947                 req->out.h.error = -ECONNABORTED;
948                 req->state = FUSE_REQ_FINISHED;
949                 list_del_init(&req->list);
950                 wake_up(&req->waitq);
951                 if (end) {
952                         req->end = NULL;
953                         /* The end function will consume this reference */
954                         __fuse_get_request(req);
955                         spin_unlock(&fc->lock);
956                         wait_event(req->waitq, !req->locked);
957                         end(fc, req);
958                         spin_lock(&fc->lock);
959                 }
960         }
961 }
962
963 /*
964  * Abort all requests.
965  *
966  * Emergency exit in case of a malicious or accidental deadlock, or
967  * just a hung filesystem.
968  *
969  * The same effect is usually achievable through killing the
970  * filesystem daemon and all users of the filesystem.  The exception
971  * is the combination of an asynchronous request and the tricky
972  * deadlock (see Documentation/filesystems/fuse.txt).
973  *
974  * During the aborting, progression of requests from the pending and
975  * processing lists onto the io list, and progression of new requests
976  * onto the pending list is prevented by req->connected being false.
977  *
978  * Progression of requests under I/O to the processing list is
979  * prevented by the req->aborted flag being true for these requests.
980  * For this reason requests on the io list must be aborted first.
981  */
982 void fuse_abort_conn(struct fuse_conn *fc)
983 {
984         spin_lock(&fc->lock);
985         if (fc->connected) {
986                 fc->connected = 0;
987                 fc->blocked = 0;
988                 end_io_requests(fc);
989                 end_requests(fc, &fc->pending);
990                 end_requests(fc, &fc->processing);
991                 wake_up_all(&fc->waitq);
992                 wake_up_all(&fc->blocked_waitq);
993                 kill_fasync(&fc->fasync, SIGIO, POLL_IN);
994         }
995         spin_unlock(&fc->lock);
996 }
997
998 static int fuse_dev_release(struct inode *inode, struct file *file)
999 {
1000         struct fuse_conn *fc = fuse_get_conn(file);
1001         if (fc) {
1002                 spin_lock(&fc->lock);
1003                 fc->connected = 0;
1004                 end_requests(fc, &fc->pending);
1005                 end_requests(fc, &fc->processing);
1006                 spin_unlock(&fc->lock);
1007                 fasync_helper(-1, file, 0, &fc->fasync);
1008                 fuse_conn_put(fc);
1009         }
1010
1011         return 0;
1012 }
1013
1014 static int fuse_dev_fasync(int fd, struct file *file, int on)
1015 {
1016         struct fuse_conn *fc = fuse_get_conn(file);
1017         if (!fc)
1018                 return -EPERM;
1019
1020         /* No locking - fasync_helper does its own locking */
1021         return fasync_helper(fd, file, on, &fc->fasync);
1022 }
1023
1024 const struct file_operations fuse_dev_operations = {
1025         .owner          = THIS_MODULE,
1026         .llseek         = no_llseek,
1027         .read           = do_sync_read,
1028         .aio_read       = fuse_dev_read,
1029         .write          = do_sync_write,
1030         .aio_write      = fuse_dev_write,
1031         .poll           = fuse_dev_poll,
1032         .release        = fuse_dev_release,
1033         .fasync         = fuse_dev_fasync,
1034 };
1035
1036 static struct miscdevice fuse_miscdevice = {
1037         .minor = FUSE_MINOR,
1038         .name  = "fuse",
1039         .fops = &fuse_dev_operations,
1040 };
1041
1042 int __init fuse_dev_init(void)
1043 {
1044         int err = -ENOMEM;
1045         fuse_req_cachep = kmem_cache_create("fuse_request",
1046                                             sizeof(struct fuse_req),
1047                                             0, 0, NULL);
1048         if (!fuse_req_cachep)
1049                 goto out;
1050
1051         err = misc_register(&fuse_miscdevice);
1052         if (err)
1053                 goto out_cache_clean;
1054
1055         return 0;
1056
1057  out_cache_clean:
1058         kmem_cache_destroy(fuse_req_cachep);
1059  out:
1060         return err;
1061 }
1062
1063 void fuse_dev_cleanup(void)
1064 {
1065         misc_deregister(&fuse_miscdevice);
1066         kmem_cache_destroy(fuse_req_cachep);
1067 }