Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6
[linux-2.6.git] / fs / pipe.c
1 /*
2  *  linux/fs/pipe.c
3  *
4  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
5  */
6
7 #include <linux/mm.h>
8 #include <linux/file.h>
9 #include <linux/poll.h>
10 #include <linux/slab.h>
11 #include <linux/module.h>
12 #include <linux/init.h>
13 #include <linux/fs.h>
14 #include <linux/mount.h>
15 #include <linux/pipe_fs_i.h>
16 #include <linux/uio.h>
17 #include <linux/highmem.h>
18 #include <linux/pagemap.h>
19 #include <linux/audit.h>
20 #include <linux/syscalls.h>
21
22 #include <asm/uaccess.h>
23 #include <asm/ioctls.h>
24
25 /*
26  * We use a start+len construction, which provides full use of the 
27  * allocated memory.
28  * -- Florian Coosmann (FGC)
29  * 
30  * Reads with count = 0 should always return 0.
31  * -- Julian Bradfield 1999-06-07.
32  *
33  * FIFOs and Pipes now generate SIGIO for both readers and writers.
34  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
35  *
36  * pipe_read & write cleanup
37  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
38  */
39
40 /* Drop the inode semaphore and wait for a pipe event, atomically */
41 void pipe_wait(struct pipe_inode_info *pipe)
42 {
43         DEFINE_WAIT(wait);
44
45         /*
46          * Pipes are system-local resources, so sleeping on them
47          * is considered a noninteractive wait:
48          */
49         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
50         if (pipe->inode)
51                 mutex_unlock(&pipe->inode->i_mutex);
52         schedule();
53         finish_wait(&pipe->wait, &wait);
54         if (pipe->inode)
55                 mutex_lock(&pipe->inode->i_mutex);
56 }
57
58 static int
59 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
60                         int atomic)
61 {
62         unsigned long copy;
63
64         while (len > 0) {
65                 while (!iov->iov_len)
66                         iov++;
67                 copy = min_t(unsigned long, len, iov->iov_len);
68
69                 if (atomic) {
70                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
71                                 return -EFAULT;
72                 } else {
73                         if (copy_from_user(to, iov->iov_base, copy))
74                                 return -EFAULT;
75                 }
76                 to += copy;
77                 len -= copy;
78                 iov->iov_base += copy;
79                 iov->iov_len -= copy;
80         }
81         return 0;
82 }
83
84 static int
85 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
86                       int atomic)
87 {
88         unsigned long copy;
89
90         while (len > 0) {
91                 while (!iov->iov_len)
92                         iov++;
93                 copy = min_t(unsigned long, len, iov->iov_len);
94
95                 if (atomic) {
96                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
97                                 return -EFAULT;
98                 } else {
99                         if (copy_to_user(iov->iov_base, from, copy))
100                                 return -EFAULT;
101                 }
102                 from += copy;
103                 len -= copy;
104                 iov->iov_base += copy;
105                 iov->iov_len -= copy;
106         }
107         return 0;
108 }
109
110 /*
111  * Attempt to pre-fault in the user memory, so we can use atomic copies.
112  * Returns the number of bytes not faulted in.
113  */
114 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
115 {
116         while (!iov->iov_len)
117                 iov++;
118
119         while (len > 0) {
120                 unsigned long this_len;
121
122                 this_len = min_t(unsigned long, len, iov->iov_len);
123                 if (fault_in_pages_writeable(iov->iov_base, this_len))
124                         break;
125
126                 len -= this_len;
127                 iov++;
128         }
129
130         return len;
131 }
132
133 /*
134  * Pre-fault in the user memory, so we can use atomic copies.
135  */
136 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
137 {
138         while (!iov->iov_len)
139                 iov++;
140
141         while (len > 0) {
142                 unsigned long this_len;
143
144                 this_len = min_t(unsigned long, len, iov->iov_len);
145                 fault_in_pages_readable(iov->iov_base, this_len);
146                 len -= this_len;
147                 iov++;
148         }
149 }
150
151 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
152                                   struct pipe_buffer *buf)
153 {
154         struct page *page = buf->page;
155
156         /*
157          * If nobody else uses this page, and we don't already have a
158          * temporary page, let's keep track of it as a one-deep
159          * allocation cache. (Otherwise just release our reference to it)
160          */
161         if (page_count(page) == 1 && !pipe->tmp_page)
162                 pipe->tmp_page = page;
163         else
164                 page_cache_release(page);
165 }
166
167 /**
168  * generic_pipe_buf_map - virtually map a pipe buffer
169  * @pipe:       the pipe that the buffer belongs to
170  * @buf:        the buffer that should be mapped
171  * @atomic:     whether to use an atomic map
172  *
173  * Description:
174  *      This function returns a kernel virtual address mapping for the
175  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
176  *      and the caller has to be careful not to fault before calling
177  *      the unmap function.
178  *
179  *      Note that this function occupies KM_USER0 if @atomic != 0.
180  */
181 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
182                            struct pipe_buffer *buf, int atomic)
183 {
184         if (atomic) {
185                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
186                 return kmap_atomic(buf->page, KM_USER0);
187         }
188
189         return kmap(buf->page);
190 }
191
192 /**
193  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
194  * @pipe:       the pipe that the buffer belongs to
195  * @buf:        the buffer that should be unmapped
196  * @map_data:   the data that the mapping function returned
197  *
198  * Description:
199  *      This function undoes the mapping that ->map() provided.
200  */
201 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
202                             struct pipe_buffer *buf, void *map_data)
203 {
204         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
205                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
206                 kunmap_atomic(map_data, KM_USER0);
207         } else
208                 kunmap(buf->page);
209 }
210
211 /**
212  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
213  * @pipe:       the pipe that the buffer belongs to
214  * @buf:        the buffer to attempt to steal
215  *
216  * Description:
217  *      This function attempts to steal the &struct page attached to
218  *      @buf. If successful, this function returns 0 and returns with
219  *      the page locked. The caller may then reuse the page for whatever
220  *      he wishes; the typical use is insertion into a different file
221  *      page cache.
222  */
223 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
224                            struct pipe_buffer *buf)
225 {
226         struct page *page = buf->page;
227
228         /*
229          * A reference of one is golden, that means that the owner of this
230          * page is the only one holding a reference to it. lock the page
231          * and return OK.
232          */
233         if (page_count(page) == 1) {
234                 lock_page(page);
235                 return 0;
236         }
237
238         return 1;
239 }
240
241 /**
242  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
243  * @pipe:       the pipe that the buffer belongs to
244  * @buf:        the buffer to get a reference to
245  *
246  * Description:
247  *      This function grabs an extra reference to @buf. It's used in
248  *      in the tee() system call, when we duplicate the buffers in one
249  *      pipe into another.
250  */
251 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
252 {
253         page_cache_get(buf->page);
254 }
255
256 /**
257  * generic_pipe_buf_confirm - verify contents of the pipe buffer
258  * @info:       the pipe that the buffer belongs to
259  * @buf:        the buffer to confirm
260  *
261  * Description:
262  *      This function does nothing, because the generic pipe code uses
263  *      pages that are always good when inserted into the pipe.
264  */
265 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
266                              struct pipe_buffer *buf)
267 {
268         return 0;
269 }
270
271 static const struct pipe_buf_operations anon_pipe_buf_ops = {
272         .can_merge = 1,
273         .map = generic_pipe_buf_map,
274         .unmap = generic_pipe_buf_unmap,
275         .confirm = generic_pipe_buf_confirm,
276         .release = anon_pipe_buf_release,
277         .steal = generic_pipe_buf_steal,
278         .get = generic_pipe_buf_get,
279 };
280
281 static ssize_t
282 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
283            unsigned long nr_segs, loff_t pos)
284 {
285         struct file *filp = iocb->ki_filp;
286         struct inode *inode = filp->f_path.dentry->d_inode;
287         struct pipe_inode_info *pipe;
288         int do_wakeup;
289         ssize_t ret;
290         struct iovec *iov = (struct iovec *)_iov;
291         size_t total_len;
292
293         total_len = iov_length(iov, nr_segs);
294         /* Null read succeeds. */
295         if (unlikely(total_len == 0))
296                 return 0;
297
298         do_wakeup = 0;
299         ret = 0;
300         mutex_lock(&inode->i_mutex);
301         pipe = inode->i_pipe;
302         for (;;) {
303                 int bufs = pipe->nrbufs;
304                 if (bufs) {
305                         int curbuf = pipe->curbuf;
306                         struct pipe_buffer *buf = pipe->bufs + curbuf;
307                         const struct pipe_buf_operations *ops = buf->ops;
308                         void *addr;
309                         size_t chars = buf->len;
310                         int error, atomic;
311
312                         if (chars > total_len)
313                                 chars = total_len;
314
315                         error = ops->confirm(pipe, buf);
316                         if (error) {
317                                 if (!ret)
318                                         error = ret;
319                                 break;
320                         }
321
322                         atomic = !iov_fault_in_pages_write(iov, chars);
323 redo:
324                         addr = ops->map(pipe, buf, atomic);
325                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
326                         ops->unmap(pipe, buf, addr);
327                         if (unlikely(error)) {
328                                 /*
329                                  * Just retry with the slow path if we failed.
330                                  */
331                                 if (atomic) {
332                                         atomic = 0;
333                                         goto redo;
334                                 }
335                                 if (!ret)
336                                         ret = error;
337                                 break;
338                         }
339                         ret += chars;
340                         buf->offset += chars;
341                         buf->len -= chars;
342                         if (!buf->len) {
343                                 buf->ops = NULL;
344                                 ops->release(pipe, buf);
345                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
346                                 pipe->curbuf = curbuf;
347                                 pipe->nrbufs = --bufs;
348                                 do_wakeup = 1;
349                         }
350                         total_len -= chars;
351                         if (!total_len)
352                                 break;  /* common path: read succeeded */
353                 }
354                 if (bufs)       /* More to do? */
355                         continue;
356                 if (!pipe->writers)
357                         break;
358                 if (!pipe->waiting_writers) {
359                         /* syscall merging: Usually we must not sleep
360                          * if O_NONBLOCK is set, or if we got some data.
361                          * But if a writer sleeps in kernel space, then
362                          * we can wait for that data without violating POSIX.
363                          */
364                         if (ret)
365                                 break;
366                         if (filp->f_flags & O_NONBLOCK) {
367                                 ret = -EAGAIN;
368                                 break;
369                         }
370                 }
371                 if (signal_pending(current)) {
372                         if (!ret)
373                                 ret = -ERESTARTSYS;
374                         break;
375                 }
376                 if (do_wakeup) {
377                         wake_up_interruptible_sync(&pipe->wait);
378                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
379                 }
380                 pipe_wait(pipe);
381         }
382         mutex_unlock(&inode->i_mutex);
383
384         /* Signal writers asynchronously that there is more room. */
385         if (do_wakeup) {
386                 wake_up_interruptible_sync(&pipe->wait);
387                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
388         }
389         if (ret > 0)
390                 file_accessed(filp);
391         return ret;
392 }
393
394 static ssize_t
395 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
396             unsigned long nr_segs, loff_t ppos)
397 {
398         struct file *filp = iocb->ki_filp;
399         struct inode *inode = filp->f_path.dentry->d_inode;
400         struct pipe_inode_info *pipe;
401         ssize_t ret;
402         int do_wakeup;
403         struct iovec *iov = (struct iovec *)_iov;
404         size_t total_len;
405         ssize_t chars;
406
407         total_len = iov_length(iov, nr_segs);
408         /* Null write succeeds. */
409         if (unlikely(total_len == 0))
410                 return 0;
411
412         do_wakeup = 0;
413         ret = 0;
414         mutex_lock(&inode->i_mutex);
415         pipe = inode->i_pipe;
416
417         if (!pipe->readers) {
418                 send_sig(SIGPIPE, current, 0);
419                 ret = -EPIPE;
420                 goto out;
421         }
422
423         /* We try to merge small writes */
424         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
425         if (pipe->nrbufs && chars != 0) {
426                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
427                                                         (PIPE_BUFFERS-1);
428                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
429                 const struct pipe_buf_operations *ops = buf->ops;
430                 int offset = buf->offset + buf->len;
431
432                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
433                         int error, atomic = 1;
434                         void *addr;
435
436                         error = ops->confirm(pipe, buf);
437                         if (error)
438                                 goto out;
439
440                         iov_fault_in_pages_read(iov, chars);
441 redo1:
442                         addr = ops->map(pipe, buf, atomic);
443                         error = pipe_iov_copy_from_user(offset + addr, iov,
444                                                         chars, atomic);
445                         ops->unmap(pipe, buf, addr);
446                         ret = error;
447                         do_wakeup = 1;
448                         if (error) {
449                                 if (atomic) {
450                                         atomic = 0;
451                                         goto redo1;
452                                 }
453                                 goto out;
454                         }
455                         buf->len += chars;
456                         total_len -= chars;
457                         ret = chars;
458                         if (!total_len)
459                                 goto out;
460                 }
461         }
462
463         for (;;) {
464                 int bufs;
465
466                 if (!pipe->readers) {
467                         send_sig(SIGPIPE, current, 0);
468                         if (!ret)
469                                 ret = -EPIPE;
470                         break;
471                 }
472                 bufs = pipe->nrbufs;
473                 if (bufs < PIPE_BUFFERS) {
474                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS-1);
475                         struct pipe_buffer *buf = pipe->bufs + newbuf;
476                         struct page *page = pipe->tmp_page;
477                         char *src;
478                         int error, atomic = 1;
479
480                         if (!page) {
481                                 page = alloc_page(GFP_HIGHUSER);
482                                 if (unlikely(!page)) {
483                                         ret = ret ? : -ENOMEM;
484                                         break;
485                                 }
486                                 pipe->tmp_page = page;
487                         }
488                         /* Always wake up, even if the copy fails. Otherwise
489                          * we lock up (O_NONBLOCK-)readers that sleep due to
490                          * syscall merging.
491                          * FIXME! Is this really true?
492                          */
493                         do_wakeup = 1;
494                         chars = PAGE_SIZE;
495                         if (chars > total_len)
496                                 chars = total_len;
497
498                         iov_fault_in_pages_read(iov, chars);
499 redo2:
500                         if (atomic)
501                                 src = kmap_atomic(page, KM_USER0);
502                         else
503                                 src = kmap(page);
504
505                         error = pipe_iov_copy_from_user(src, iov, chars,
506                                                         atomic);
507                         if (atomic)
508                                 kunmap_atomic(src, KM_USER0);
509                         else
510                                 kunmap(page);
511
512                         if (unlikely(error)) {
513                                 if (atomic) {
514                                         atomic = 0;
515                                         goto redo2;
516                                 }
517                                 if (!ret)
518                                         ret = error;
519                                 break;
520                         }
521                         ret += chars;
522
523                         /* Insert it into the buffer array */
524                         buf->page = page;
525                         buf->ops = &anon_pipe_buf_ops;
526                         buf->offset = 0;
527                         buf->len = chars;
528                         pipe->nrbufs = ++bufs;
529                         pipe->tmp_page = NULL;
530
531                         total_len -= chars;
532                         if (!total_len)
533                                 break;
534                 }
535                 if (bufs < PIPE_BUFFERS)
536                         continue;
537                 if (filp->f_flags & O_NONBLOCK) {
538                         if (!ret)
539                                 ret = -EAGAIN;
540                         break;
541                 }
542                 if (signal_pending(current)) {
543                         if (!ret)
544                                 ret = -ERESTARTSYS;
545                         break;
546                 }
547                 if (do_wakeup) {
548                         wake_up_interruptible_sync(&pipe->wait);
549                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
550                         do_wakeup = 0;
551                 }
552                 pipe->waiting_writers++;
553                 pipe_wait(pipe);
554                 pipe->waiting_writers--;
555         }
556 out:
557         mutex_unlock(&inode->i_mutex);
558         if (do_wakeup) {
559                 wake_up_interruptible_sync(&pipe->wait);
560                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
561         }
562         if (ret > 0)
563                 file_update_time(filp);
564         return ret;
565 }
566
567 static ssize_t
568 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
569 {
570         return -EBADF;
571 }
572
573 static ssize_t
574 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
575            loff_t *ppos)
576 {
577         return -EBADF;
578 }
579
580 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
581 {
582         struct inode *inode = filp->f_path.dentry->d_inode;
583         struct pipe_inode_info *pipe;
584         int count, buf, nrbufs;
585
586         switch (cmd) {
587                 case FIONREAD:
588                         mutex_lock(&inode->i_mutex);
589                         pipe = inode->i_pipe;
590                         count = 0;
591                         buf = pipe->curbuf;
592                         nrbufs = pipe->nrbufs;
593                         while (--nrbufs >= 0) {
594                                 count += pipe->bufs[buf].len;
595                                 buf = (buf+1) & (PIPE_BUFFERS-1);
596                         }
597                         mutex_unlock(&inode->i_mutex);
598
599                         return put_user(count, (int __user *)arg);
600                 default:
601                         return -EINVAL;
602         }
603 }
604
605 /* No kernel lock held - fine */
606 static unsigned int
607 pipe_poll(struct file *filp, poll_table *wait)
608 {
609         unsigned int mask;
610         struct inode *inode = filp->f_path.dentry->d_inode;
611         struct pipe_inode_info *pipe = inode->i_pipe;
612         int nrbufs;
613
614         poll_wait(filp, &pipe->wait, wait);
615
616         /* Reading only -- no need for acquiring the semaphore.  */
617         nrbufs = pipe->nrbufs;
618         mask = 0;
619         if (filp->f_mode & FMODE_READ) {
620                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
621                 if (!pipe->writers && filp->f_version != pipe->w_counter)
622                         mask |= POLLHUP;
623         }
624
625         if (filp->f_mode & FMODE_WRITE) {
626                 mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
627                 /*
628                  * Most Unices do not set POLLERR for FIFOs but on Linux they
629                  * behave exactly like pipes for poll().
630                  */
631                 if (!pipe->readers)
632                         mask |= POLLERR;
633         }
634
635         return mask;
636 }
637
638 static int
639 pipe_release(struct inode *inode, int decr, int decw)
640 {
641         struct pipe_inode_info *pipe;
642
643         mutex_lock(&inode->i_mutex);
644         pipe = inode->i_pipe;
645         pipe->readers -= decr;
646         pipe->writers -= decw;
647
648         if (!pipe->readers && !pipe->writers) {
649                 free_pipe_info(inode);
650         } else {
651                 wake_up_interruptible_sync(&pipe->wait);
652                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
653                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
654         }
655         mutex_unlock(&inode->i_mutex);
656
657         return 0;
658 }
659
660 static int
661 pipe_read_fasync(int fd, struct file *filp, int on)
662 {
663         struct inode *inode = filp->f_path.dentry->d_inode;
664         int retval;
665
666         mutex_lock(&inode->i_mutex);
667         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
668         mutex_unlock(&inode->i_mutex);
669
670         return retval;
671 }
672
673
674 static int
675 pipe_write_fasync(int fd, struct file *filp, int on)
676 {
677         struct inode *inode = filp->f_path.dentry->d_inode;
678         int retval;
679
680         mutex_lock(&inode->i_mutex);
681         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
682         mutex_unlock(&inode->i_mutex);
683
684         return retval;
685 }
686
687
688 static int
689 pipe_rdwr_fasync(int fd, struct file *filp, int on)
690 {
691         struct inode *inode = filp->f_path.dentry->d_inode;
692         struct pipe_inode_info *pipe = inode->i_pipe;
693         int retval;
694
695         mutex_lock(&inode->i_mutex);
696         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
697         if (retval >= 0) {
698                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
699                 if (retval < 0) /* this can happen only if on == T */
700                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
701         }
702         mutex_unlock(&inode->i_mutex);
703         return retval;
704 }
705
706
707 static int
708 pipe_read_release(struct inode *inode, struct file *filp)
709 {
710         return pipe_release(inode, 1, 0);
711 }
712
713 static int
714 pipe_write_release(struct inode *inode, struct file *filp)
715 {
716         return pipe_release(inode, 0, 1);
717 }
718
719 static int
720 pipe_rdwr_release(struct inode *inode, struct file *filp)
721 {
722         int decr, decw;
723
724         decr = (filp->f_mode & FMODE_READ) != 0;
725         decw = (filp->f_mode & FMODE_WRITE) != 0;
726         return pipe_release(inode, decr, decw);
727 }
728
729 static int
730 pipe_read_open(struct inode *inode, struct file *filp)
731 {
732         /* We could have perhaps used atomic_t, but this and friends
733            below are the only places.  So it doesn't seem worthwhile.  */
734         mutex_lock(&inode->i_mutex);
735         inode->i_pipe->readers++;
736         mutex_unlock(&inode->i_mutex);
737
738         return 0;
739 }
740
741 static int
742 pipe_write_open(struct inode *inode, struct file *filp)
743 {
744         mutex_lock(&inode->i_mutex);
745         inode->i_pipe->writers++;
746         mutex_unlock(&inode->i_mutex);
747
748         return 0;
749 }
750
751 static int
752 pipe_rdwr_open(struct inode *inode, struct file *filp)
753 {
754         mutex_lock(&inode->i_mutex);
755         if (filp->f_mode & FMODE_READ)
756                 inode->i_pipe->readers++;
757         if (filp->f_mode & FMODE_WRITE)
758                 inode->i_pipe->writers++;
759         mutex_unlock(&inode->i_mutex);
760
761         return 0;
762 }
763
764 /*
765  * The file_operations structs are not static because they
766  * are also used in linux/fs/fifo.c to do operations on FIFOs.
767  *
768  * Pipes reuse fifos' file_operations structs.
769  */
770 const struct file_operations read_pipefifo_fops = {
771         .llseek         = no_llseek,
772         .read           = do_sync_read,
773         .aio_read       = pipe_read,
774         .write          = bad_pipe_w,
775         .poll           = pipe_poll,
776         .unlocked_ioctl = pipe_ioctl,
777         .open           = pipe_read_open,
778         .release        = pipe_read_release,
779         .fasync         = pipe_read_fasync,
780 };
781
782 const struct file_operations write_pipefifo_fops = {
783         .llseek         = no_llseek,
784         .read           = bad_pipe_r,
785         .write          = do_sync_write,
786         .aio_write      = pipe_write,
787         .poll           = pipe_poll,
788         .unlocked_ioctl = pipe_ioctl,
789         .open           = pipe_write_open,
790         .release        = pipe_write_release,
791         .fasync         = pipe_write_fasync,
792 };
793
794 const struct file_operations rdwr_pipefifo_fops = {
795         .llseek         = no_llseek,
796         .read           = do_sync_read,
797         .aio_read       = pipe_read,
798         .write          = do_sync_write,
799         .aio_write      = pipe_write,
800         .poll           = pipe_poll,
801         .unlocked_ioctl = pipe_ioctl,
802         .open           = pipe_rdwr_open,
803         .release        = pipe_rdwr_release,
804         .fasync         = pipe_rdwr_fasync,
805 };
806
807 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
808 {
809         struct pipe_inode_info *pipe;
810
811         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
812         if (pipe) {
813                 init_waitqueue_head(&pipe->wait);
814                 pipe->r_counter = pipe->w_counter = 1;
815                 pipe->inode = inode;
816         }
817
818         return pipe;
819 }
820
821 void __free_pipe_info(struct pipe_inode_info *pipe)
822 {
823         int i;
824
825         for (i = 0; i < PIPE_BUFFERS; i++) {
826                 struct pipe_buffer *buf = pipe->bufs + i;
827                 if (buf->ops)
828                         buf->ops->release(pipe, buf);
829         }
830         if (pipe->tmp_page)
831                 __free_page(pipe->tmp_page);
832         kfree(pipe);
833 }
834
835 void free_pipe_info(struct inode *inode)
836 {
837         __free_pipe_info(inode->i_pipe);
838         inode->i_pipe = NULL;
839 }
840
841 static struct vfsmount *pipe_mnt __read_mostly;
842 static int pipefs_delete_dentry(struct dentry *dentry)
843 {
844         /*
845          * At creation time, we pretended this dentry was hashed
846          * (by clearing DCACHE_UNHASHED bit in d_flags)
847          * At delete time, we restore the truth : not hashed.
848          * (so that dput() can proceed correctly)
849          */
850         dentry->d_flags |= DCACHE_UNHASHED;
851         return 0;
852 }
853
854 /*
855  * pipefs_dname() is called from d_path().
856  */
857 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
858 {
859         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
860                                 dentry->d_inode->i_ino);
861 }
862
863 static const struct dentry_operations pipefs_dentry_operations = {
864         .d_delete       = pipefs_delete_dentry,
865         .d_dname        = pipefs_dname,
866 };
867
868 static struct inode * get_pipe_inode(void)
869 {
870         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
871         struct pipe_inode_info *pipe;
872
873         if (!inode)
874                 goto fail_inode;
875
876         pipe = alloc_pipe_info(inode);
877         if (!pipe)
878                 goto fail_iput;
879         inode->i_pipe = pipe;
880
881         pipe->readers = pipe->writers = 1;
882         inode->i_fop = &rdwr_pipefifo_fops;
883
884         /*
885          * Mark the inode dirty from the very beginning,
886          * that way it will never be moved to the dirty
887          * list because "mark_inode_dirty()" will think
888          * that it already _is_ on the dirty list.
889          */
890         inode->i_state = I_DIRTY;
891         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
892         inode->i_uid = current_fsuid();
893         inode->i_gid = current_fsgid();
894         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
895
896         return inode;
897
898 fail_iput:
899         iput(inode);
900
901 fail_inode:
902         return NULL;
903 }
904
905 struct file *create_write_pipe(int flags)
906 {
907         int err;
908         struct inode *inode;
909         struct file *f;
910         struct dentry *dentry;
911         struct qstr name = { .name = "" };
912
913         err = -ENFILE;
914         inode = get_pipe_inode();
915         if (!inode)
916                 goto err;
917
918         err = -ENOMEM;
919         dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
920         if (!dentry)
921                 goto err_inode;
922
923         dentry->d_op = &pipefs_dentry_operations;
924         /*
925          * We dont want to publish this dentry into global dentry hash table.
926          * We pretend dentry is already hashed, by unsetting DCACHE_UNHASHED
927          * This permits a working /proc/$pid/fd/XXX on pipes
928          */
929         dentry->d_flags &= ~DCACHE_UNHASHED;
930         d_instantiate(dentry, inode);
931
932         err = -ENFILE;
933         f = alloc_file(pipe_mnt, dentry, FMODE_WRITE, &write_pipefifo_fops);
934         if (!f)
935                 goto err_dentry;
936         f->f_mapping = inode->i_mapping;
937
938         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
939         f->f_version = 0;
940
941         return f;
942
943  err_dentry:
944         free_pipe_info(inode);
945         dput(dentry);
946         return ERR_PTR(err);
947
948  err_inode:
949         free_pipe_info(inode);
950         iput(inode);
951  err:
952         return ERR_PTR(err);
953 }
954
955 void free_write_pipe(struct file *f)
956 {
957         free_pipe_info(f->f_dentry->d_inode);
958         path_put(&f->f_path);
959         put_filp(f);
960 }
961
962 struct file *create_read_pipe(struct file *wrf, int flags)
963 {
964         struct file *f = get_empty_filp();
965         if (!f)
966                 return ERR_PTR(-ENFILE);
967
968         /* Grab pipe from the writer */
969         f->f_path = wrf->f_path;
970         path_get(&wrf->f_path);
971         f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
972
973         f->f_pos = 0;
974         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
975         f->f_op = &read_pipefifo_fops;
976         f->f_mode = FMODE_READ;
977         f->f_version = 0;
978
979         return f;
980 }
981
982 int do_pipe_flags(int *fd, int flags)
983 {
984         struct file *fw, *fr;
985         int error;
986         int fdw, fdr;
987
988         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
989                 return -EINVAL;
990
991         fw = create_write_pipe(flags);
992         if (IS_ERR(fw))
993                 return PTR_ERR(fw);
994         fr = create_read_pipe(fw, flags);
995         error = PTR_ERR(fr);
996         if (IS_ERR(fr))
997                 goto err_write_pipe;
998
999         error = get_unused_fd_flags(flags);
1000         if (error < 0)
1001                 goto err_read_pipe;
1002         fdr = error;
1003
1004         error = get_unused_fd_flags(flags);
1005         if (error < 0)
1006                 goto err_fdr;
1007         fdw = error;
1008
1009         audit_fd_pair(fdr, fdw);
1010         fd_install(fdr, fr);
1011         fd_install(fdw, fw);
1012         fd[0] = fdr;
1013         fd[1] = fdw;
1014
1015         return 0;
1016
1017  err_fdr:
1018         put_unused_fd(fdr);
1019  err_read_pipe:
1020         path_put(&fr->f_path);
1021         put_filp(fr);
1022  err_write_pipe:
1023         free_write_pipe(fw);
1024         return error;
1025 }
1026
1027 /*
1028  * sys_pipe() is the normal C calling standard for creating
1029  * a pipe. It's not the way Unix traditionally does this, though.
1030  */
1031 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1032 {
1033         int fd[2];
1034         int error;
1035
1036         error = do_pipe_flags(fd, flags);
1037         if (!error) {
1038                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1039                         sys_close(fd[0]);
1040                         sys_close(fd[1]);
1041                         error = -EFAULT;
1042                 }
1043         }
1044         return error;
1045 }
1046
1047 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1048 {
1049         return sys_pipe2(fildes, 0);
1050 }
1051
1052 /*
1053  * pipefs should _never_ be mounted by userland - too much of security hassle,
1054  * no real gain from having the whole whorehouse mounted. So we don't need
1055  * any operations on the root directory. However, we need a non-trivial
1056  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1057  */
1058 static int pipefs_get_sb(struct file_system_type *fs_type,
1059                          int flags, const char *dev_name, void *data,
1060                          struct vfsmount *mnt)
1061 {
1062         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1063 }
1064
1065 static struct file_system_type pipe_fs_type = {
1066         .name           = "pipefs",
1067         .get_sb         = pipefs_get_sb,
1068         .kill_sb        = kill_anon_super,
1069 };
1070
1071 static int __init init_pipe_fs(void)
1072 {
1073         int err = register_filesystem(&pipe_fs_type);
1074
1075         if (!err) {
1076                 pipe_mnt = kern_mount(&pipe_fs_type);
1077                 if (IS_ERR(pipe_mnt)) {
1078                         err = PTR_ERR(pipe_mnt);
1079                         unregister_filesystem(&pipe_fs_type);
1080                 }
1081         }
1082         return err;
1083 }
1084
1085 static void __exit exit_pipe_fs(void)
1086 {
1087         unregister_filesystem(&pipe_fs_type);
1088         mntput(pipe_mnt);
1089 }
1090
1091 fs_initcall(init_pipe_fs);
1092 module_exit(exit_pipe_fs);