sched: cleanup, remove the TASK_NONINTERACTIVE flag
[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
21 #include <asm/uaccess.h>
22 #include <asm/ioctls.h>
23
24 /*
25  * We use a start+len construction, which provides full use of the 
26  * allocated memory.
27  * -- Florian Coosmann (FGC)
28  * 
29  * Reads with count = 0 should always return 0.
30  * -- Julian Bradfield 1999-06-07.
31  *
32  * FIFOs and Pipes now generate SIGIO for both readers and writers.
33  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
34  *
35  * pipe_read & write cleanup
36  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
37  */
38
39 /* Drop the inode semaphore and wait for a pipe event, atomically */
40 void pipe_wait(struct pipe_inode_info *pipe)
41 {
42         DEFINE_WAIT(wait);
43
44         /*
45          * Pipes are system-local resources, so sleeping on them
46          * is considered a noninteractive wait:
47          */
48         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
49         if (pipe->inode)
50                 mutex_unlock(&pipe->inode->i_mutex);
51         schedule();
52         finish_wait(&pipe->wait, &wait);
53         if (pipe->inode)
54                 mutex_lock(&pipe->inode->i_mutex);
55 }
56
57 static int
58 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
59                         int atomic)
60 {
61         unsigned long copy;
62
63         while (len > 0) {
64                 while (!iov->iov_len)
65                         iov++;
66                 copy = min_t(unsigned long, len, iov->iov_len);
67
68                 if (atomic) {
69                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
70                                 return -EFAULT;
71                 } else {
72                         if (copy_from_user(to, iov->iov_base, copy))
73                                 return -EFAULT;
74                 }
75                 to += copy;
76                 len -= copy;
77                 iov->iov_base += copy;
78                 iov->iov_len -= copy;
79         }
80         return 0;
81 }
82
83 static int
84 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
85                       int atomic)
86 {
87         unsigned long copy;
88
89         while (len > 0) {
90                 while (!iov->iov_len)
91                         iov++;
92                 copy = min_t(unsigned long, len, iov->iov_len);
93
94                 if (atomic) {
95                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
96                                 return -EFAULT;
97                 } else {
98                         if (copy_to_user(iov->iov_base, from, copy))
99                                 return -EFAULT;
100                 }
101                 from += copy;
102                 len -= copy;
103                 iov->iov_base += copy;
104                 iov->iov_len -= copy;
105         }
106         return 0;
107 }
108
109 /*
110  * Attempt to pre-fault in the user memory, so we can use atomic copies.
111  * Returns the number of bytes not faulted in.
112  */
113 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
114 {
115         while (!iov->iov_len)
116                 iov++;
117
118         while (len > 0) {
119                 unsigned long this_len;
120
121                 this_len = min_t(unsigned long, len, iov->iov_len);
122                 if (fault_in_pages_writeable(iov->iov_base, this_len))
123                         break;
124
125                 len -= this_len;
126                 iov++;
127         }
128
129         return len;
130 }
131
132 /*
133  * Pre-fault in the user memory, so we can use atomic copies.
134  */
135 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
136 {
137         while (!iov->iov_len)
138                 iov++;
139
140         while (len > 0) {
141                 unsigned long this_len;
142
143                 this_len = min_t(unsigned long, len, iov->iov_len);
144                 fault_in_pages_readable(iov->iov_base, this_len);
145                 len -= this_len;
146                 iov++;
147         }
148 }
149
150 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
151                                   struct pipe_buffer *buf)
152 {
153         struct page *page = buf->page;
154
155         /*
156          * If nobody else uses this page, and we don't already have a
157          * temporary page, let's keep track of it as a one-deep
158          * allocation cache. (Otherwise just release our reference to it)
159          */
160         if (page_count(page) == 1 && !pipe->tmp_page)
161                 pipe->tmp_page = page;
162         else
163                 page_cache_release(page);
164 }
165
166 /**
167  * generic_pipe_buf_map - virtually map a pipe buffer
168  * @pipe:       the pipe that the buffer belongs to
169  * @buf:        the buffer that should be mapped
170  * @atomic:     whether to use an atomic map
171  *
172  * Description:
173  *      This function returns a kernel virtual address mapping for the
174  *      passed in @pipe_buffer. If @atomic is set, an atomic map is provided
175  *      and the caller has to be careful not to fault before calling
176  *      the unmap function.
177  *
178  *      Note that this function occupies KM_USER0 if @atomic != 0.
179  */
180 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
181                            struct pipe_buffer *buf, int atomic)
182 {
183         if (atomic) {
184                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
185                 return kmap_atomic(buf->page, KM_USER0);
186         }
187
188         return kmap(buf->page);
189 }
190
191 /**
192  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
193  * @pipe:       the pipe that the buffer belongs to
194  * @buf:        the buffer that should be unmapped
195  * @map_data:   the data that the mapping function returned
196  *
197  * Description:
198  *      This function undoes the mapping that ->map() provided.
199  */
200 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
201                             struct pipe_buffer *buf, void *map_data)
202 {
203         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
204                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
205                 kunmap_atomic(map_data, KM_USER0);
206         } else
207                 kunmap(buf->page);
208 }
209
210 /**
211  * generic_pipe_buf_steal - attempt to take ownership of a @pipe_buffer
212  * @pipe:       the pipe that the buffer belongs to
213  * @buf:        the buffer to attempt to steal
214  *
215  * Description:
216  *      This function attempts to steal the @struct page attached to
217  *      @buf. If successful, this function returns 0 and returns with
218  *      the page locked. The caller may then reuse the page for whatever
219  *      he wishes, the typical use is insertion into a different file
220  *      page cache.
221  */
222 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
223                            struct pipe_buffer *buf)
224 {
225         struct page *page = buf->page;
226
227         /*
228          * A reference of one is golden, that means that the owner of this
229          * page is the only one holding a reference to it. lock the page
230          * and return OK.
231          */
232         if (page_count(page) == 1) {
233                 lock_page(page);
234                 return 0;
235         }
236
237         return 1;
238 }
239
240 /**
241  * generic_pipe_buf_get - get a reference to a @struct pipe_buffer
242  * @pipe:       the pipe that the buffer belongs to
243  * @buf:        the buffer to get a reference to
244  *
245  * Description:
246  *      This function grabs an extra reference to @buf. It's used in
247  *      in the tee() system call, when we duplicate the buffers in one
248  *      pipe into another.
249  */
250 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
251 {
252         page_cache_get(buf->page);
253 }
254
255 /**
256  * generic_pipe_buf_confirm - verify contents of the pipe buffer
257  * @info:       the pipe that the buffer belongs to
258  * @buf:        the buffer to confirm
259  *
260  * Description:
261  *      This function does nothing, because the generic pipe code uses
262  *      pages that are always good when inserted into the pipe.
263  */
264 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
265                              struct pipe_buffer *buf)
266 {
267         return 0;
268 }
269
270 static const struct pipe_buf_operations anon_pipe_buf_ops = {
271         .can_merge = 1,
272         .map = generic_pipe_buf_map,
273         .unmap = generic_pipe_buf_unmap,
274         .confirm = generic_pipe_buf_confirm,
275         .release = anon_pipe_buf_release,
276         .steal = generic_pipe_buf_steal,
277         .get = generic_pipe_buf_get,
278 };
279
280 static ssize_t
281 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
282            unsigned long nr_segs, loff_t pos)
283 {
284         struct file *filp = iocb->ki_filp;
285         struct inode *inode = filp->f_path.dentry->d_inode;
286         struct pipe_inode_info *pipe;
287         int do_wakeup;
288         ssize_t ret;
289         struct iovec *iov = (struct iovec *)_iov;
290         size_t total_len;
291
292         total_len = iov_length(iov, nr_segs);
293         /* Null read succeeds. */
294         if (unlikely(total_len == 0))
295                 return 0;
296
297         do_wakeup = 0;
298         ret = 0;
299         mutex_lock(&inode->i_mutex);
300         pipe = inode->i_pipe;
301         for (;;) {
302                 int bufs = pipe->nrbufs;
303                 if (bufs) {
304                         int curbuf = pipe->curbuf;
305                         struct pipe_buffer *buf = pipe->bufs + curbuf;
306                         const struct pipe_buf_operations *ops = buf->ops;
307                         void *addr;
308                         size_t chars = buf->len;
309                         int error, atomic;
310
311                         if (chars > total_len)
312                                 chars = total_len;
313
314                         error = ops->confirm(pipe, buf);
315                         if (error) {
316                                 if (!ret)
317                                         error = ret;
318                                 break;
319                         }
320
321                         atomic = !iov_fault_in_pages_write(iov, chars);
322 redo:
323                         addr = ops->map(pipe, buf, atomic);
324                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
325                         ops->unmap(pipe, buf, addr);
326                         if (unlikely(error)) {
327                                 /*
328                                  * Just retry with the slow path if we failed.
329                                  */
330                                 if (atomic) {
331                                         atomic = 0;
332                                         goto redo;
333                                 }
334                                 if (!ret)
335                                         ret = error;
336                                 break;
337                         }
338                         ret += chars;
339                         buf->offset += chars;
340                         buf->len -= chars;
341                         if (!buf->len) {
342                                 buf->ops = NULL;
343                                 ops->release(pipe, buf);
344                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
345                                 pipe->curbuf = curbuf;
346                                 pipe->nrbufs = --bufs;
347                                 do_wakeup = 1;
348                         }
349                         total_len -= chars;
350                         if (!total_len)
351                                 break;  /* common path: read succeeded */
352                 }
353                 if (bufs)       /* More to do? */
354                         continue;
355                 if (!pipe->writers)
356                         break;
357                 if (!pipe->waiting_writers) {
358                         /* syscall merging: Usually we must not sleep
359                          * if O_NONBLOCK is set, or if we got some data.
360                          * But if a writer sleeps in kernel space, then
361                          * we can wait for that data without violating POSIX.
362                          */
363                         if (ret)
364                                 break;
365                         if (filp->f_flags & O_NONBLOCK) {
366                                 ret = -EAGAIN;
367                                 break;
368                         }
369                 }
370                 if (signal_pending(current)) {
371                         if (!ret)
372                                 ret = -ERESTARTSYS;
373                         break;
374                 }
375                 if (do_wakeup) {
376                         wake_up_interruptible_sync(&pipe->wait);
377                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
378                 }
379                 pipe_wait(pipe);
380         }
381         mutex_unlock(&inode->i_mutex);
382
383         /* Signal writers asynchronously that there is more room. */
384         if (do_wakeup) {
385                 wake_up_interruptible(&pipe->wait);
386                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
387         }
388         if (ret > 0)
389                 file_accessed(filp);
390         return ret;
391 }
392
393 static ssize_t
394 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
395             unsigned long nr_segs, loff_t ppos)
396 {
397         struct file *filp = iocb->ki_filp;
398         struct inode *inode = filp->f_path.dentry->d_inode;
399         struct pipe_inode_info *pipe;
400         ssize_t ret;
401         int do_wakeup;
402         struct iovec *iov = (struct iovec *)_iov;
403         size_t total_len;
404         ssize_t chars;
405
406         total_len = iov_length(iov, nr_segs);
407         /* Null write succeeds. */
408         if (unlikely(total_len == 0))
409                 return 0;
410
411         do_wakeup = 0;
412         ret = 0;
413         mutex_lock(&inode->i_mutex);
414         pipe = inode->i_pipe;
415
416         if (!pipe->readers) {
417                 send_sig(SIGPIPE, current, 0);
418                 ret = -EPIPE;
419                 goto out;
420         }
421
422         /* We try to merge small writes */
423         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
424         if (pipe->nrbufs && chars != 0) {
425                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
426                                                         (PIPE_BUFFERS-1);
427                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
428                 const struct pipe_buf_operations *ops = buf->ops;
429                 int offset = buf->offset + buf->len;
430
431                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
432                         int error, atomic = 1;
433                         void *addr;
434
435                         error = ops->confirm(pipe, buf);
436                         if (error)
437                                 goto out;
438
439                         iov_fault_in_pages_read(iov, chars);
440 redo1:
441                         addr = ops->map(pipe, buf, atomic);
442                         error = pipe_iov_copy_from_user(offset + addr, iov,
443                                                         chars, atomic);
444                         ops->unmap(pipe, buf, addr);
445                         ret = error;
446                         do_wakeup = 1;
447                         if (error) {
448                                 if (atomic) {
449                                         atomic = 0;
450                                         goto redo1;
451                                 }
452                                 goto out;
453                         }
454                         buf->len += chars;
455                         total_len -= chars;
456                         ret = chars;
457                         if (!total_len)
458                                 goto out;
459                 }
460         }
461
462         for (;;) {
463                 int bufs;
464
465                 if (!pipe->readers) {
466                         send_sig(SIGPIPE, current, 0);
467                         if (!ret)
468                                 ret = -EPIPE;
469                         break;
470                 }
471                 bufs = pipe->nrbufs;
472                 if (bufs < PIPE_BUFFERS) {
473                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS-1);
474                         struct pipe_buffer *buf = pipe->bufs + newbuf;
475                         struct page *page = pipe->tmp_page;
476                         char *src;
477                         int error, atomic = 1;
478
479                         if (!page) {
480                                 page = alloc_page(GFP_HIGHUSER);
481                                 if (unlikely(!page)) {
482                                         ret = ret ? : -ENOMEM;
483                                         break;
484                                 }
485                                 pipe->tmp_page = page;
486                         }
487                         /* Always wake up, even if the copy fails. Otherwise
488                          * we lock up (O_NONBLOCK-)readers that sleep due to
489                          * syscall merging.
490                          * FIXME! Is this really true?
491                          */
492                         do_wakeup = 1;
493                         chars = PAGE_SIZE;
494                         if (chars > total_len)
495                                 chars = total_len;
496
497                         iov_fault_in_pages_read(iov, chars);
498 redo2:
499                         if (atomic)
500                                 src = kmap_atomic(page, KM_USER0);
501                         else
502                                 src = kmap(page);
503
504                         error = pipe_iov_copy_from_user(src, iov, chars,
505                                                         atomic);
506                         if (atomic)
507                                 kunmap_atomic(src, KM_USER0);
508                         else
509                                 kunmap(page);
510
511                         if (unlikely(error)) {
512                                 if (atomic) {
513                                         atomic = 0;
514                                         goto redo2;
515                                 }
516                                 if (!ret)
517                                         ret = error;
518                                 break;
519                         }
520                         ret += chars;
521
522                         /* Insert it into the buffer array */
523                         buf->page = page;
524                         buf->ops = &anon_pipe_buf_ops;
525                         buf->offset = 0;
526                         buf->len = chars;
527                         pipe->nrbufs = ++bufs;
528                         pipe->tmp_page = NULL;
529
530                         total_len -= chars;
531                         if (!total_len)
532                                 break;
533                 }
534                 if (bufs < PIPE_BUFFERS)
535                         continue;
536                 if (filp->f_flags & O_NONBLOCK) {
537                         if (!ret)
538                                 ret = -EAGAIN;
539                         break;
540                 }
541                 if (signal_pending(current)) {
542                         if (!ret)
543                                 ret = -ERESTARTSYS;
544                         break;
545                 }
546                 if (do_wakeup) {
547                         wake_up_interruptible_sync(&pipe->wait);
548                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
549                         do_wakeup = 0;
550                 }
551                 pipe->waiting_writers++;
552                 pipe_wait(pipe);
553                 pipe->waiting_writers--;
554         }
555 out:
556         mutex_unlock(&inode->i_mutex);
557         if (do_wakeup) {
558                 wake_up_interruptible(&pipe->wait);
559                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
560         }
561         if (ret > 0)
562                 file_update_time(filp);
563         return ret;
564 }
565
566 static ssize_t
567 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
568 {
569         return -EBADF;
570 }
571
572 static ssize_t
573 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
574            loff_t *ppos)
575 {
576         return -EBADF;
577 }
578
579 static int
580 pipe_ioctl(struct inode *pino, struct file *filp,
581            unsigned int cmd, unsigned long arg)
582 {
583         struct inode *inode = filp->f_path.dentry->d_inode;
584         struct pipe_inode_info *pipe;
585         int count, buf, nrbufs;
586
587         switch (cmd) {
588                 case FIONREAD:
589                         mutex_lock(&inode->i_mutex);
590                         pipe = inode->i_pipe;
591                         count = 0;
592                         buf = pipe->curbuf;
593                         nrbufs = pipe->nrbufs;
594                         while (--nrbufs >= 0) {
595                                 count += pipe->bufs[buf].len;
596                                 buf = (buf+1) & (PIPE_BUFFERS-1);
597                         }
598                         mutex_unlock(&inode->i_mutex);
599
600                         return put_user(count, (int __user *)arg);
601                 default:
602                         return -EINVAL;
603         }
604 }
605
606 /* No kernel lock held - fine */
607 static unsigned int
608 pipe_poll(struct file *filp, poll_table *wait)
609 {
610         unsigned int mask;
611         struct inode *inode = filp->f_path.dentry->d_inode;
612         struct pipe_inode_info *pipe = inode->i_pipe;
613         int nrbufs;
614
615         poll_wait(filp, &pipe->wait, wait);
616
617         /* Reading only -- no need for acquiring the semaphore.  */
618         nrbufs = pipe->nrbufs;
619         mask = 0;
620         if (filp->f_mode & FMODE_READ) {
621                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
622                 if (!pipe->writers && filp->f_version != pipe->w_counter)
623                         mask |= POLLHUP;
624         }
625
626         if (filp->f_mode & FMODE_WRITE) {
627                 mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
628                 /*
629                  * Most Unices do not set POLLERR for FIFOs but on Linux they
630                  * behave exactly like pipes for poll().
631                  */
632                 if (!pipe->readers)
633                         mask |= POLLERR;
634         }
635
636         return mask;
637 }
638
639 static int
640 pipe_release(struct inode *inode, int decr, int decw)
641 {
642         struct pipe_inode_info *pipe;
643
644         mutex_lock(&inode->i_mutex);
645         pipe = inode->i_pipe;
646         pipe->readers -= decr;
647         pipe->writers -= decw;
648
649         if (!pipe->readers && !pipe->writers) {
650                 free_pipe_info(inode);
651         } else {
652                 wake_up_interruptible(&pipe->wait);
653                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
654                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
655         }
656         mutex_unlock(&inode->i_mutex);
657
658         return 0;
659 }
660
661 static int
662 pipe_read_fasync(int fd, struct file *filp, int on)
663 {
664         struct inode *inode = filp->f_path.dentry->d_inode;
665         int retval;
666
667         mutex_lock(&inode->i_mutex);
668         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
669         mutex_unlock(&inode->i_mutex);
670
671         if (retval < 0)
672                 return retval;
673
674         return 0;
675 }
676
677
678 static int
679 pipe_write_fasync(int fd, struct file *filp, int on)
680 {
681         struct inode *inode = filp->f_path.dentry->d_inode;
682         int retval;
683
684         mutex_lock(&inode->i_mutex);
685         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
686         mutex_unlock(&inode->i_mutex);
687
688         if (retval < 0)
689                 return retval;
690
691         return 0;
692 }
693
694
695 static int
696 pipe_rdwr_fasync(int fd, struct file *filp, int on)
697 {
698         struct inode *inode = filp->f_path.dentry->d_inode;
699         struct pipe_inode_info *pipe = inode->i_pipe;
700         int retval;
701
702         mutex_lock(&inode->i_mutex);
703
704         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
705
706         if (retval >= 0)
707                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
708
709         mutex_unlock(&inode->i_mutex);
710
711         if (retval < 0)
712                 return retval;
713
714         return 0;
715 }
716
717
718 static int
719 pipe_read_release(struct inode *inode, struct file *filp)
720 {
721         pipe_read_fasync(-1, filp, 0);
722         return pipe_release(inode, 1, 0);
723 }
724
725 static int
726 pipe_write_release(struct inode *inode, struct file *filp)
727 {
728         pipe_write_fasync(-1, filp, 0);
729         return pipe_release(inode, 0, 1);
730 }
731
732 static int
733 pipe_rdwr_release(struct inode *inode, struct file *filp)
734 {
735         int decr, decw;
736
737         pipe_rdwr_fasync(-1, filp, 0);
738         decr = (filp->f_mode & FMODE_READ) != 0;
739         decw = (filp->f_mode & FMODE_WRITE) != 0;
740         return pipe_release(inode, decr, decw);
741 }
742
743 static int
744 pipe_read_open(struct inode *inode, struct file *filp)
745 {
746         /* We could have perhaps used atomic_t, but this and friends
747            below are the only places.  So it doesn't seem worthwhile.  */
748         mutex_lock(&inode->i_mutex);
749         inode->i_pipe->readers++;
750         mutex_unlock(&inode->i_mutex);
751
752         return 0;
753 }
754
755 static int
756 pipe_write_open(struct inode *inode, struct file *filp)
757 {
758         mutex_lock(&inode->i_mutex);
759         inode->i_pipe->writers++;
760         mutex_unlock(&inode->i_mutex);
761
762         return 0;
763 }
764
765 static int
766 pipe_rdwr_open(struct inode *inode, struct file *filp)
767 {
768         mutex_lock(&inode->i_mutex);
769         if (filp->f_mode & FMODE_READ)
770                 inode->i_pipe->readers++;
771         if (filp->f_mode & FMODE_WRITE)
772                 inode->i_pipe->writers++;
773         mutex_unlock(&inode->i_mutex);
774
775         return 0;
776 }
777
778 /*
779  * The file_operations structs are not static because they
780  * are also used in linux/fs/fifo.c to do operations on FIFOs.
781  */
782 const struct file_operations read_fifo_fops = {
783         .llseek         = no_llseek,
784         .read           = do_sync_read,
785         .aio_read       = pipe_read,
786         .write          = bad_pipe_w,
787         .poll           = pipe_poll,
788         .ioctl          = pipe_ioctl,
789         .open           = pipe_read_open,
790         .release        = pipe_read_release,
791         .fasync         = pipe_read_fasync,
792 };
793
794 const struct file_operations write_fifo_fops = {
795         .llseek         = no_llseek,
796         .read           = bad_pipe_r,
797         .write          = do_sync_write,
798         .aio_write      = pipe_write,
799         .poll           = pipe_poll,
800         .ioctl          = pipe_ioctl,
801         .open           = pipe_write_open,
802         .release        = pipe_write_release,
803         .fasync         = pipe_write_fasync,
804 };
805
806 const struct file_operations rdwr_fifo_fops = {
807         .llseek         = no_llseek,
808         .read           = do_sync_read,
809         .aio_read       = pipe_read,
810         .write          = do_sync_write,
811         .aio_write      = pipe_write,
812         .poll           = pipe_poll,
813         .ioctl          = pipe_ioctl,
814         .open           = pipe_rdwr_open,
815         .release        = pipe_rdwr_release,
816         .fasync         = pipe_rdwr_fasync,
817 };
818
819 static const struct file_operations read_pipe_fops = {
820         .llseek         = no_llseek,
821         .read           = do_sync_read,
822         .aio_read       = pipe_read,
823         .write          = bad_pipe_w,
824         .poll           = pipe_poll,
825         .ioctl          = pipe_ioctl,
826         .open           = pipe_read_open,
827         .release        = pipe_read_release,
828         .fasync         = pipe_read_fasync,
829 };
830
831 static const struct file_operations write_pipe_fops = {
832         .llseek         = no_llseek,
833         .read           = bad_pipe_r,
834         .write          = do_sync_write,
835         .aio_write      = pipe_write,
836         .poll           = pipe_poll,
837         .ioctl          = pipe_ioctl,
838         .open           = pipe_write_open,
839         .release        = pipe_write_release,
840         .fasync         = pipe_write_fasync,
841 };
842
843 static const struct file_operations rdwr_pipe_fops = {
844         .llseek         = no_llseek,
845         .read           = do_sync_read,
846         .aio_read       = pipe_read,
847         .write          = do_sync_write,
848         .aio_write      = pipe_write,
849         .poll           = pipe_poll,
850         .ioctl          = pipe_ioctl,
851         .open           = pipe_rdwr_open,
852         .release        = pipe_rdwr_release,
853         .fasync         = pipe_rdwr_fasync,
854 };
855
856 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
857 {
858         struct pipe_inode_info *pipe;
859
860         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
861         if (pipe) {
862                 init_waitqueue_head(&pipe->wait);
863                 pipe->r_counter = pipe->w_counter = 1;
864                 pipe->inode = inode;
865         }
866
867         return pipe;
868 }
869
870 void __free_pipe_info(struct pipe_inode_info *pipe)
871 {
872         int i;
873
874         for (i = 0; i < PIPE_BUFFERS; i++) {
875                 struct pipe_buffer *buf = pipe->bufs + i;
876                 if (buf->ops)
877                         buf->ops->release(pipe, buf);
878         }
879         if (pipe->tmp_page)
880                 __free_page(pipe->tmp_page);
881         kfree(pipe);
882 }
883
884 void free_pipe_info(struct inode *inode)
885 {
886         __free_pipe_info(inode->i_pipe);
887         inode->i_pipe = NULL;
888 }
889
890 static struct vfsmount *pipe_mnt __read_mostly;
891 static int pipefs_delete_dentry(struct dentry *dentry)
892 {
893         /*
894          * At creation time, we pretended this dentry was hashed
895          * (by clearing DCACHE_UNHASHED bit in d_flags)
896          * At delete time, we restore the truth : not hashed.
897          * (so that dput() can proceed correctly)
898          */
899         dentry->d_flags |= DCACHE_UNHASHED;
900         return 0;
901 }
902
903 /*
904  * pipefs_dname() is called from d_path().
905  */
906 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
907 {
908         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
909                                 dentry->d_inode->i_ino);
910 }
911
912 static struct dentry_operations pipefs_dentry_operations = {
913         .d_delete       = pipefs_delete_dentry,
914         .d_dname        = pipefs_dname,
915 };
916
917 static struct inode * get_pipe_inode(void)
918 {
919         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
920         struct pipe_inode_info *pipe;
921
922         if (!inode)
923                 goto fail_inode;
924
925         pipe = alloc_pipe_info(inode);
926         if (!pipe)
927                 goto fail_iput;
928         inode->i_pipe = pipe;
929
930         pipe->readers = pipe->writers = 1;
931         inode->i_fop = &rdwr_pipe_fops;
932
933         /*
934          * Mark the inode dirty from the very beginning,
935          * that way it will never be moved to the dirty
936          * list because "mark_inode_dirty()" will think
937          * that it already _is_ on the dirty list.
938          */
939         inode->i_state = I_DIRTY;
940         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
941         inode->i_uid = current->fsuid;
942         inode->i_gid = current->fsgid;
943         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
944
945         return inode;
946
947 fail_iput:
948         iput(inode);
949
950 fail_inode:
951         return NULL;
952 }
953
954 struct file *create_write_pipe(void)
955 {
956         int err;
957         struct inode *inode;
958         struct file *f;
959         struct dentry *dentry;
960         struct qstr name = { .name = "" };
961
962         f = get_empty_filp();
963         if (!f)
964                 return ERR_PTR(-ENFILE);
965         err = -ENFILE;
966         inode = get_pipe_inode();
967         if (!inode)
968                 goto err_file;
969
970         err = -ENOMEM;
971         dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
972         if (!dentry)
973                 goto err_inode;
974
975         dentry->d_op = &pipefs_dentry_operations;
976         /*
977          * We dont want to publish this dentry into global dentry hash table.
978          * We pretend dentry is already hashed, by unsetting DCACHE_UNHASHED
979          * This permits a working /proc/$pid/fd/XXX on pipes
980          */
981         dentry->d_flags &= ~DCACHE_UNHASHED;
982         d_instantiate(dentry, inode);
983         f->f_path.mnt = mntget(pipe_mnt);
984         f->f_path.dentry = dentry;
985         f->f_mapping = inode->i_mapping;
986
987         f->f_flags = O_WRONLY;
988         f->f_op = &write_pipe_fops;
989         f->f_mode = FMODE_WRITE;
990         f->f_version = 0;
991
992         return f;
993
994  err_inode:
995         free_pipe_info(inode);
996         iput(inode);
997  err_file:
998         put_filp(f);
999         return ERR_PTR(err);
1000 }
1001
1002 void free_write_pipe(struct file *f)
1003 {
1004         free_pipe_info(f->f_dentry->d_inode);
1005         dput(f->f_path.dentry);
1006         mntput(f->f_path.mnt);
1007         put_filp(f);
1008 }
1009
1010 struct file *create_read_pipe(struct file *wrf)
1011 {
1012         struct file *f = get_empty_filp();
1013         if (!f)
1014                 return ERR_PTR(-ENFILE);
1015
1016         /* Grab pipe from the writer */
1017         f->f_path.mnt = mntget(wrf->f_path.mnt);
1018         f->f_path.dentry = dget(wrf->f_path.dentry);
1019         f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
1020
1021         f->f_pos = 0;
1022         f->f_flags = O_RDONLY;
1023         f->f_op = &read_pipe_fops;
1024         f->f_mode = FMODE_READ;
1025         f->f_version = 0;
1026
1027         return f;
1028 }
1029
1030 int do_pipe(int *fd)
1031 {
1032         struct file *fw, *fr;
1033         int error;
1034         int fdw, fdr;
1035
1036         fw = create_write_pipe();
1037         if (IS_ERR(fw))
1038                 return PTR_ERR(fw);
1039         fr = create_read_pipe(fw);
1040         error = PTR_ERR(fr);
1041         if (IS_ERR(fr))
1042                 goto err_write_pipe;
1043
1044         error = get_unused_fd();
1045         if (error < 0)
1046                 goto err_read_pipe;
1047         fdr = error;
1048
1049         error = get_unused_fd();
1050         if (error < 0)
1051                 goto err_fdr;
1052         fdw = error;
1053
1054         error = audit_fd_pair(fdr, fdw);
1055         if (error < 0)
1056                 goto err_fdw;
1057
1058         fd_install(fdr, fr);
1059         fd_install(fdw, fw);
1060         fd[0] = fdr;
1061         fd[1] = fdw;
1062
1063         return 0;
1064
1065  err_fdw:
1066         put_unused_fd(fdw);
1067  err_fdr:
1068         put_unused_fd(fdr);
1069  err_read_pipe:
1070         dput(fr->f_dentry);
1071         mntput(fr->f_vfsmnt);
1072         put_filp(fr);
1073  err_write_pipe:
1074         free_write_pipe(fw);
1075         return error;
1076 }
1077
1078 /*
1079  * pipefs should _never_ be mounted by userland - too much of security hassle,
1080  * no real gain from having the whole whorehouse mounted. So we don't need
1081  * any operations on the root directory. However, we need a non-trivial
1082  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1083  */
1084 static int pipefs_get_sb(struct file_system_type *fs_type,
1085                          int flags, const char *dev_name, void *data,
1086                          struct vfsmount *mnt)
1087 {
1088         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1089 }
1090
1091 static struct file_system_type pipe_fs_type = {
1092         .name           = "pipefs",
1093         .get_sb         = pipefs_get_sb,
1094         .kill_sb        = kill_anon_super,
1095 };
1096
1097 static int __init init_pipe_fs(void)
1098 {
1099         int err = register_filesystem(&pipe_fs_type);
1100
1101         if (!err) {
1102                 pipe_mnt = kern_mount(&pipe_fs_type);
1103                 if (IS_ERR(pipe_mnt)) {
1104                         err = PTR_ERR(pipe_mnt);
1105                         unregister_filesystem(&pipe_fs_type);
1106                 }
1107         }
1108         return err;
1109 }
1110
1111 static void __exit exit_pipe_fs(void)
1112 {
1113         unregister_filesystem(&pipe_fs_type);
1114         mntput(pipe_mnt);
1115 }
1116
1117 fs_initcall(init_pipe_fs);
1118 module_exit(exit_pipe_fs);