sched: Reset sched stats on fork()
[linux-2.6] / 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 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
41 {
42         if (pipe->inode)
43                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
44 }
45
46 void pipe_lock(struct pipe_inode_info *pipe)
47 {
48         /*
49          * pipe_lock() nests non-pipe inode locks (for writing to a file)
50          */
51         pipe_lock_nested(pipe, I_MUTEX_PARENT);
52 }
53 EXPORT_SYMBOL(pipe_lock);
54
55 void pipe_unlock(struct pipe_inode_info *pipe)
56 {
57         if (pipe->inode)
58                 mutex_unlock(&pipe->inode->i_mutex);
59 }
60 EXPORT_SYMBOL(pipe_unlock);
61
62 void pipe_double_lock(struct pipe_inode_info *pipe1,
63                       struct pipe_inode_info *pipe2)
64 {
65         BUG_ON(pipe1 == pipe2);
66
67         if (pipe1 < pipe2) {
68                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
69                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
70         } else {
71                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
72                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
73         }
74 }
75
76 /* Drop the inode semaphore and wait for a pipe event, atomically */
77 void pipe_wait(struct pipe_inode_info *pipe)
78 {
79         DEFINE_WAIT(wait);
80
81         /*
82          * Pipes are system-local resources, so sleeping on them
83          * is considered a noninteractive wait:
84          */
85         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
86         pipe_unlock(pipe);
87         schedule();
88         finish_wait(&pipe->wait, &wait);
89         pipe_lock(pipe);
90 }
91
92 static int
93 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
94                         int atomic)
95 {
96         unsigned long copy;
97
98         while (len > 0) {
99                 while (!iov->iov_len)
100                         iov++;
101                 copy = min_t(unsigned long, len, iov->iov_len);
102
103                 if (atomic) {
104                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
105                                 return -EFAULT;
106                 } else {
107                         if (copy_from_user(to, iov->iov_base, copy))
108                                 return -EFAULT;
109                 }
110                 to += copy;
111                 len -= copy;
112                 iov->iov_base += copy;
113                 iov->iov_len -= copy;
114         }
115         return 0;
116 }
117
118 static int
119 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
120                       int atomic)
121 {
122         unsigned long copy;
123
124         while (len > 0) {
125                 while (!iov->iov_len)
126                         iov++;
127                 copy = min_t(unsigned long, len, iov->iov_len);
128
129                 if (atomic) {
130                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
131                                 return -EFAULT;
132                 } else {
133                         if (copy_to_user(iov->iov_base, from, copy))
134                                 return -EFAULT;
135                 }
136                 from += copy;
137                 len -= copy;
138                 iov->iov_base += copy;
139                 iov->iov_len -= copy;
140         }
141         return 0;
142 }
143
144 /*
145  * Attempt to pre-fault in the user memory, so we can use atomic copies.
146  * Returns the number of bytes not faulted in.
147  */
148 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
149 {
150         while (!iov->iov_len)
151                 iov++;
152
153         while (len > 0) {
154                 unsigned long this_len;
155
156                 this_len = min_t(unsigned long, len, iov->iov_len);
157                 if (fault_in_pages_writeable(iov->iov_base, this_len))
158                         break;
159
160                 len -= this_len;
161                 iov++;
162         }
163
164         return len;
165 }
166
167 /*
168  * Pre-fault in the user memory, so we can use atomic copies.
169  */
170 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
171 {
172         while (!iov->iov_len)
173                 iov++;
174
175         while (len > 0) {
176                 unsigned long this_len;
177
178                 this_len = min_t(unsigned long, len, iov->iov_len);
179                 fault_in_pages_readable(iov->iov_base, this_len);
180                 len -= this_len;
181                 iov++;
182         }
183 }
184
185 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
186                                   struct pipe_buffer *buf)
187 {
188         struct page *page = buf->page;
189
190         /*
191          * If nobody else uses this page, and we don't already have a
192          * temporary page, let's keep track of it as a one-deep
193          * allocation cache. (Otherwise just release our reference to it)
194          */
195         if (page_count(page) == 1 && !pipe->tmp_page)
196                 pipe->tmp_page = page;
197         else
198                 page_cache_release(page);
199 }
200
201 /**
202  * generic_pipe_buf_map - virtually map a pipe buffer
203  * @pipe:       the pipe that the buffer belongs to
204  * @buf:        the buffer that should be mapped
205  * @atomic:     whether to use an atomic map
206  *
207  * Description:
208  *      This function returns a kernel virtual address mapping for the
209  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
210  *      and the caller has to be careful not to fault before calling
211  *      the unmap function.
212  *
213  *      Note that this function occupies KM_USER0 if @atomic != 0.
214  */
215 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
216                            struct pipe_buffer *buf, int atomic)
217 {
218         if (atomic) {
219                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
220                 return kmap_atomic(buf->page, KM_USER0);
221         }
222
223         return kmap(buf->page);
224 }
225
226 /**
227  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
228  * @pipe:       the pipe that the buffer belongs to
229  * @buf:        the buffer that should be unmapped
230  * @map_data:   the data that the mapping function returned
231  *
232  * Description:
233  *      This function undoes the mapping that ->map() provided.
234  */
235 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
236                             struct pipe_buffer *buf, void *map_data)
237 {
238         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
239                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
240                 kunmap_atomic(map_data, KM_USER0);
241         } else
242                 kunmap(buf->page);
243 }
244
245 /**
246  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
247  * @pipe:       the pipe that the buffer belongs to
248  * @buf:        the buffer to attempt to steal
249  *
250  * Description:
251  *      This function attempts to steal the &struct page attached to
252  *      @buf. If successful, this function returns 0 and returns with
253  *      the page locked. The caller may then reuse the page for whatever
254  *      he wishes; the typical use is insertion into a different file
255  *      page cache.
256  */
257 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
258                            struct pipe_buffer *buf)
259 {
260         struct page *page = buf->page;
261
262         /*
263          * A reference of one is golden, that means that the owner of this
264          * page is the only one holding a reference to it. lock the page
265          * and return OK.
266          */
267         if (page_count(page) == 1) {
268                 lock_page(page);
269                 return 0;
270         }
271
272         return 1;
273 }
274
275 /**
276  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
277  * @pipe:       the pipe that the buffer belongs to
278  * @buf:        the buffer to get a reference to
279  *
280  * Description:
281  *      This function grabs an extra reference to @buf. It's used in
282  *      in the tee() system call, when we duplicate the buffers in one
283  *      pipe into another.
284  */
285 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
286 {
287         page_cache_get(buf->page);
288 }
289
290 /**
291  * generic_pipe_buf_confirm - verify contents of the pipe buffer
292  * @info:       the pipe that the buffer belongs to
293  * @buf:        the buffer to confirm
294  *
295  * Description:
296  *      This function does nothing, because the generic pipe code uses
297  *      pages that are always good when inserted into the pipe.
298  */
299 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
300                              struct pipe_buffer *buf)
301 {
302         return 0;
303 }
304
305 /**
306  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
307  * @pipe:       the pipe that the buffer belongs to
308  * @buf:        the buffer to put a reference to
309  *
310  * Description:
311  *      This function releases a reference to @buf.
312  */
313 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
314                               struct pipe_buffer *buf)
315 {
316         page_cache_release(buf->page);
317 }
318
319 static const struct pipe_buf_operations anon_pipe_buf_ops = {
320         .can_merge = 1,
321         .map = generic_pipe_buf_map,
322         .unmap = generic_pipe_buf_unmap,
323         .confirm = generic_pipe_buf_confirm,
324         .release = anon_pipe_buf_release,
325         .steal = generic_pipe_buf_steal,
326         .get = generic_pipe_buf_get,
327 };
328
329 static ssize_t
330 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
331            unsigned long nr_segs, loff_t pos)
332 {
333         struct file *filp = iocb->ki_filp;
334         struct inode *inode = filp->f_path.dentry->d_inode;
335         struct pipe_inode_info *pipe;
336         int do_wakeup;
337         ssize_t ret;
338         struct iovec *iov = (struct iovec *)_iov;
339         size_t total_len;
340
341         total_len = iov_length(iov, nr_segs);
342         /* Null read succeeds. */
343         if (unlikely(total_len == 0))
344                 return 0;
345
346         do_wakeup = 0;
347         ret = 0;
348         mutex_lock(&inode->i_mutex);
349         pipe = inode->i_pipe;
350         for (;;) {
351                 int bufs = pipe->nrbufs;
352                 if (bufs) {
353                         int curbuf = pipe->curbuf;
354                         struct pipe_buffer *buf = pipe->bufs + curbuf;
355                         const struct pipe_buf_operations *ops = buf->ops;
356                         void *addr;
357                         size_t chars = buf->len;
358                         int error, atomic;
359
360                         if (chars > total_len)
361                                 chars = total_len;
362
363                         error = ops->confirm(pipe, buf);
364                         if (error) {
365                                 if (!ret)
366                                         error = ret;
367                                 break;
368                         }
369
370                         atomic = !iov_fault_in_pages_write(iov, chars);
371 redo:
372                         addr = ops->map(pipe, buf, atomic);
373                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
374                         ops->unmap(pipe, buf, addr);
375                         if (unlikely(error)) {
376                                 /*
377                                  * Just retry with the slow path if we failed.
378                                  */
379                                 if (atomic) {
380                                         atomic = 0;
381                                         goto redo;
382                                 }
383                                 if (!ret)
384                                         ret = error;
385                                 break;
386                         }
387                         ret += chars;
388                         buf->offset += chars;
389                         buf->len -= chars;
390                         if (!buf->len) {
391                                 buf->ops = NULL;
392                                 ops->release(pipe, buf);
393                                 curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
394                                 pipe->curbuf = curbuf;
395                                 pipe->nrbufs = --bufs;
396                                 do_wakeup = 1;
397                         }
398                         total_len -= chars;
399                         if (!total_len)
400                                 break;  /* common path: read succeeded */
401                 }
402                 if (bufs)       /* More to do? */
403                         continue;
404                 if (!pipe->writers)
405                         break;
406                 if (!pipe->waiting_writers) {
407                         /* syscall merging: Usually we must not sleep
408                          * if O_NONBLOCK is set, or if we got some data.
409                          * But if a writer sleeps in kernel space, then
410                          * we can wait for that data without violating POSIX.
411                          */
412                         if (ret)
413                                 break;
414                         if (filp->f_flags & O_NONBLOCK) {
415                                 ret = -EAGAIN;
416                                 break;
417                         }
418                 }
419                 if (signal_pending(current)) {
420                         if (!ret)
421                                 ret = -ERESTARTSYS;
422                         break;
423                 }
424                 if (do_wakeup) {
425                         wake_up_interruptible_sync(&pipe->wait);
426                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
427                 }
428                 pipe_wait(pipe);
429         }
430         mutex_unlock(&inode->i_mutex);
431
432         /* Signal writers asynchronously that there is more room. */
433         if (do_wakeup) {
434                 wake_up_interruptible_sync(&pipe->wait);
435                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
436         }
437         if (ret > 0)
438                 file_accessed(filp);
439         return ret;
440 }
441
442 static ssize_t
443 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
444             unsigned long nr_segs, loff_t ppos)
445 {
446         struct file *filp = iocb->ki_filp;
447         struct inode *inode = filp->f_path.dentry->d_inode;
448         struct pipe_inode_info *pipe;
449         ssize_t ret;
450         int do_wakeup;
451         struct iovec *iov = (struct iovec *)_iov;
452         size_t total_len;
453         ssize_t chars;
454
455         total_len = iov_length(iov, nr_segs);
456         /* Null write succeeds. */
457         if (unlikely(total_len == 0))
458                 return 0;
459
460         do_wakeup = 0;
461         ret = 0;
462         mutex_lock(&inode->i_mutex);
463         pipe = inode->i_pipe;
464
465         if (!pipe->readers) {
466                 send_sig(SIGPIPE, current, 0);
467                 ret = -EPIPE;
468                 goto out;
469         }
470
471         /* We try to merge small writes */
472         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
473         if (pipe->nrbufs && chars != 0) {
474                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
475                                                         (PIPE_BUFFERS-1);
476                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
477                 const struct pipe_buf_operations *ops = buf->ops;
478                 int offset = buf->offset + buf->len;
479
480                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
481                         int error, atomic = 1;
482                         void *addr;
483
484                         error = ops->confirm(pipe, buf);
485                         if (error)
486                                 goto out;
487
488                         iov_fault_in_pages_read(iov, chars);
489 redo1:
490                         addr = ops->map(pipe, buf, atomic);
491                         error = pipe_iov_copy_from_user(offset + addr, iov,
492                                                         chars, atomic);
493                         ops->unmap(pipe, buf, addr);
494                         ret = error;
495                         do_wakeup = 1;
496                         if (error) {
497                                 if (atomic) {
498                                         atomic = 0;
499                                         goto redo1;
500                                 }
501                                 goto out;
502                         }
503                         buf->len += chars;
504                         total_len -= chars;
505                         ret = chars;
506                         if (!total_len)
507                                 goto out;
508                 }
509         }
510
511         for (;;) {
512                 int bufs;
513
514                 if (!pipe->readers) {
515                         send_sig(SIGPIPE, current, 0);
516                         if (!ret)
517                                 ret = -EPIPE;
518                         break;
519                 }
520                 bufs = pipe->nrbufs;
521                 if (bufs < PIPE_BUFFERS) {
522                         int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS-1);
523                         struct pipe_buffer *buf = pipe->bufs + newbuf;
524                         struct page *page = pipe->tmp_page;
525                         char *src;
526                         int error, atomic = 1;
527
528                         if (!page) {
529                                 page = alloc_page(GFP_HIGHUSER);
530                                 if (unlikely(!page)) {
531                                         ret = ret ? : -ENOMEM;
532                                         break;
533                                 }
534                                 pipe->tmp_page = page;
535                         }
536                         /* Always wake up, even if the copy fails. Otherwise
537                          * we lock up (O_NONBLOCK-)readers that sleep due to
538                          * syscall merging.
539                          * FIXME! Is this really true?
540                          */
541                         do_wakeup = 1;
542                         chars = PAGE_SIZE;
543                         if (chars > total_len)
544                                 chars = total_len;
545
546                         iov_fault_in_pages_read(iov, chars);
547 redo2:
548                         if (atomic)
549                                 src = kmap_atomic(page, KM_USER0);
550                         else
551                                 src = kmap(page);
552
553                         error = pipe_iov_copy_from_user(src, iov, chars,
554                                                         atomic);
555                         if (atomic)
556                                 kunmap_atomic(src, KM_USER0);
557                         else
558                                 kunmap(page);
559
560                         if (unlikely(error)) {
561                                 if (atomic) {
562                                         atomic = 0;
563                                         goto redo2;
564                                 }
565                                 if (!ret)
566                                         ret = error;
567                                 break;
568                         }
569                         ret += chars;
570
571                         /* Insert it into the buffer array */
572                         buf->page = page;
573                         buf->ops = &anon_pipe_buf_ops;
574                         buf->offset = 0;
575                         buf->len = chars;
576                         pipe->nrbufs = ++bufs;
577                         pipe->tmp_page = NULL;
578
579                         total_len -= chars;
580                         if (!total_len)
581                                 break;
582                 }
583                 if (bufs < PIPE_BUFFERS)
584                         continue;
585                 if (filp->f_flags & O_NONBLOCK) {
586                         if (!ret)
587                                 ret = -EAGAIN;
588                         break;
589                 }
590                 if (signal_pending(current)) {
591                         if (!ret)
592                                 ret = -ERESTARTSYS;
593                         break;
594                 }
595                 if (do_wakeup) {
596                         wake_up_interruptible_sync(&pipe->wait);
597                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
598                         do_wakeup = 0;
599                 }
600                 pipe->waiting_writers++;
601                 pipe_wait(pipe);
602                 pipe->waiting_writers--;
603         }
604 out:
605         mutex_unlock(&inode->i_mutex);
606         if (do_wakeup) {
607                 wake_up_interruptible_sync(&pipe->wait);
608                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
609         }
610         if (ret > 0)
611                 file_update_time(filp);
612         return ret;
613 }
614
615 static ssize_t
616 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
617 {
618         return -EBADF;
619 }
620
621 static ssize_t
622 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
623            loff_t *ppos)
624 {
625         return -EBADF;
626 }
627
628 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
629 {
630         struct inode *inode = filp->f_path.dentry->d_inode;
631         struct pipe_inode_info *pipe;
632         int count, buf, nrbufs;
633
634         switch (cmd) {
635                 case FIONREAD:
636                         mutex_lock(&inode->i_mutex);
637                         pipe = inode->i_pipe;
638                         count = 0;
639                         buf = pipe->curbuf;
640                         nrbufs = pipe->nrbufs;
641                         while (--nrbufs >= 0) {
642                                 count += pipe->bufs[buf].len;
643                                 buf = (buf+1) & (PIPE_BUFFERS-1);
644                         }
645                         mutex_unlock(&inode->i_mutex);
646
647                         return put_user(count, (int __user *)arg);
648                 default:
649                         return -EINVAL;
650         }
651 }
652
653 /* No kernel lock held - fine */
654 static unsigned int
655 pipe_poll(struct file *filp, poll_table *wait)
656 {
657         unsigned int mask;
658         struct inode *inode = filp->f_path.dentry->d_inode;
659         struct pipe_inode_info *pipe = inode->i_pipe;
660         int nrbufs;
661
662         poll_wait(filp, &pipe->wait, wait);
663
664         /* Reading only -- no need for acquiring the semaphore.  */
665         nrbufs = pipe->nrbufs;
666         mask = 0;
667         if (filp->f_mode & FMODE_READ) {
668                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
669                 if (!pipe->writers && filp->f_version != pipe->w_counter)
670                         mask |= POLLHUP;
671         }
672
673         if (filp->f_mode & FMODE_WRITE) {
674                 mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
675                 /*
676                  * Most Unices do not set POLLERR for FIFOs but on Linux they
677                  * behave exactly like pipes for poll().
678                  */
679                 if (!pipe->readers)
680                         mask |= POLLERR;
681         }
682
683         return mask;
684 }
685
686 static int
687 pipe_release(struct inode *inode, int decr, int decw)
688 {
689         struct pipe_inode_info *pipe;
690
691         mutex_lock(&inode->i_mutex);
692         pipe = inode->i_pipe;
693         pipe->readers -= decr;
694         pipe->writers -= decw;
695
696         if (!pipe->readers && !pipe->writers) {
697                 free_pipe_info(inode);
698         } else {
699                 wake_up_interruptible_sync(&pipe->wait);
700                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
701                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
702         }
703         mutex_unlock(&inode->i_mutex);
704
705         return 0;
706 }
707
708 static int
709 pipe_read_fasync(int fd, struct file *filp, int on)
710 {
711         struct inode *inode = filp->f_path.dentry->d_inode;
712         int retval;
713
714         mutex_lock(&inode->i_mutex);
715         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
716         mutex_unlock(&inode->i_mutex);
717
718         return retval;
719 }
720
721
722 static int
723 pipe_write_fasync(int fd, struct file *filp, int on)
724 {
725         struct inode *inode = filp->f_path.dentry->d_inode;
726         int retval;
727
728         mutex_lock(&inode->i_mutex);
729         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
730         mutex_unlock(&inode->i_mutex);
731
732         return retval;
733 }
734
735
736 static int
737 pipe_rdwr_fasync(int fd, struct file *filp, int on)
738 {
739         struct inode *inode = filp->f_path.dentry->d_inode;
740         struct pipe_inode_info *pipe = inode->i_pipe;
741         int retval;
742
743         mutex_lock(&inode->i_mutex);
744         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
745         if (retval >= 0) {
746                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
747                 if (retval < 0) /* this can happen only if on == T */
748                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
749         }
750         mutex_unlock(&inode->i_mutex);
751         return retval;
752 }
753
754
755 static int
756 pipe_read_release(struct inode *inode, struct file *filp)
757 {
758         return pipe_release(inode, 1, 0);
759 }
760
761 static int
762 pipe_write_release(struct inode *inode, struct file *filp)
763 {
764         return pipe_release(inode, 0, 1);
765 }
766
767 static int
768 pipe_rdwr_release(struct inode *inode, struct file *filp)
769 {
770         int decr, decw;
771
772         decr = (filp->f_mode & FMODE_READ) != 0;
773         decw = (filp->f_mode & FMODE_WRITE) != 0;
774         return pipe_release(inode, decr, decw);
775 }
776
777 static int
778 pipe_read_open(struct inode *inode, struct file *filp)
779 {
780         /* We could have perhaps used atomic_t, but this and friends
781            below are the only places.  So it doesn't seem worthwhile.  */
782         mutex_lock(&inode->i_mutex);
783         inode->i_pipe->readers++;
784         mutex_unlock(&inode->i_mutex);
785
786         return 0;
787 }
788
789 static int
790 pipe_write_open(struct inode *inode, struct file *filp)
791 {
792         mutex_lock(&inode->i_mutex);
793         inode->i_pipe->writers++;
794         mutex_unlock(&inode->i_mutex);
795
796         return 0;
797 }
798
799 static int
800 pipe_rdwr_open(struct inode *inode, struct file *filp)
801 {
802         mutex_lock(&inode->i_mutex);
803         if (filp->f_mode & FMODE_READ)
804                 inode->i_pipe->readers++;
805         if (filp->f_mode & FMODE_WRITE)
806                 inode->i_pipe->writers++;
807         mutex_unlock(&inode->i_mutex);
808
809         return 0;
810 }
811
812 /*
813  * The file_operations structs are not static because they
814  * are also used in linux/fs/fifo.c to do operations on FIFOs.
815  *
816  * Pipes reuse fifos' file_operations structs.
817  */
818 const struct file_operations read_pipefifo_fops = {
819         .llseek         = no_llseek,
820         .read           = do_sync_read,
821         .aio_read       = pipe_read,
822         .write          = bad_pipe_w,
823         .poll           = pipe_poll,
824         .unlocked_ioctl = pipe_ioctl,
825         .open           = pipe_read_open,
826         .release        = pipe_read_release,
827         .fasync         = pipe_read_fasync,
828 };
829
830 const struct file_operations write_pipefifo_fops = {
831         .llseek         = no_llseek,
832         .read           = bad_pipe_r,
833         .write          = do_sync_write,
834         .aio_write      = pipe_write,
835         .poll           = pipe_poll,
836         .unlocked_ioctl = pipe_ioctl,
837         .open           = pipe_write_open,
838         .release        = pipe_write_release,
839         .fasync         = pipe_write_fasync,
840 };
841
842 const struct file_operations rdwr_pipefifo_fops = {
843         .llseek         = no_llseek,
844         .read           = do_sync_read,
845         .aio_read       = pipe_read,
846         .write          = do_sync_write,
847         .aio_write      = pipe_write,
848         .poll           = pipe_poll,
849         .unlocked_ioctl = pipe_ioctl,
850         .open           = pipe_rdwr_open,
851         .release        = pipe_rdwr_release,
852         .fasync         = pipe_rdwr_fasync,
853 };
854
855 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
856 {
857         struct pipe_inode_info *pipe;
858
859         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
860         if (pipe) {
861                 init_waitqueue_head(&pipe->wait);
862                 pipe->r_counter = pipe->w_counter = 1;
863                 pipe->inode = inode;
864         }
865
866         return pipe;
867 }
868
869 void __free_pipe_info(struct pipe_inode_info *pipe)
870 {
871         int i;
872
873         for (i = 0; i < PIPE_BUFFERS; i++) {
874                 struct pipe_buffer *buf = pipe->bufs + i;
875                 if (buf->ops)
876                         buf->ops->release(pipe, buf);
877         }
878         if (pipe->tmp_page)
879                 __free_page(pipe->tmp_page);
880         kfree(pipe);
881 }
882
883 void free_pipe_info(struct inode *inode)
884 {
885         __free_pipe_info(inode->i_pipe);
886         inode->i_pipe = NULL;
887 }
888
889 static struct vfsmount *pipe_mnt __read_mostly;
890 static int pipefs_delete_dentry(struct dentry *dentry)
891 {
892         /*
893          * At creation time, we pretended this dentry was hashed
894          * (by clearing DCACHE_UNHASHED bit in d_flags)
895          * At delete time, we restore the truth : not hashed.
896          * (so that dput() can proceed correctly)
897          */
898         dentry->d_flags |= DCACHE_UNHASHED;
899         return 0;
900 }
901
902 /*
903  * pipefs_dname() is called from d_path().
904  */
905 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
906 {
907         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
908                                 dentry->d_inode->i_ino);
909 }
910
911 static const struct dentry_operations pipefs_dentry_operations = {
912         .d_delete       = pipefs_delete_dentry,
913         .d_dname        = pipefs_dname,
914 };
915
916 static struct inode * get_pipe_inode(void)
917 {
918         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
919         struct pipe_inode_info *pipe;
920
921         if (!inode)
922                 goto fail_inode;
923
924         pipe = alloc_pipe_info(inode);
925         if (!pipe)
926                 goto fail_iput;
927         inode->i_pipe = pipe;
928
929         pipe->readers = pipe->writers = 1;
930         inode->i_fop = &rdwr_pipefifo_fops;
931
932         /*
933          * Mark the inode dirty from the very beginning,
934          * that way it will never be moved to the dirty
935          * list because "mark_inode_dirty()" will think
936          * that it already _is_ on the dirty list.
937          */
938         inode->i_state = I_DIRTY;
939         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
940         inode->i_uid = current_fsuid();
941         inode->i_gid = current_fsgid();
942         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
943
944         return inode;
945
946 fail_iput:
947         iput(inode);
948
949 fail_inode:
950         return NULL;
951 }
952
953 struct file *create_write_pipe(int flags)
954 {
955         int err;
956         struct inode *inode;
957         struct file *f;
958         struct dentry *dentry;
959         struct qstr name = { .name = "" };
960
961         err = -ENFILE;
962         inode = get_pipe_inode();
963         if (!inode)
964                 goto err;
965
966         err = -ENOMEM;
967         dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
968         if (!dentry)
969                 goto err_inode;
970
971         dentry->d_op = &pipefs_dentry_operations;
972         /*
973          * We dont want to publish this dentry into global dentry hash table.
974          * We pretend dentry is already hashed, by unsetting DCACHE_UNHASHED
975          * This permits a working /proc/$pid/fd/XXX on pipes
976          */
977         dentry->d_flags &= ~DCACHE_UNHASHED;
978         d_instantiate(dentry, inode);
979
980         err = -ENFILE;
981         f = alloc_file(pipe_mnt, dentry, FMODE_WRITE, &write_pipefifo_fops);
982         if (!f)
983                 goto err_dentry;
984         f->f_mapping = inode->i_mapping;
985
986         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
987         f->f_version = 0;
988
989         return f;
990
991  err_dentry:
992         free_pipe_info(inode);
993         dput(dentry);
994         return ERR_PTR(err);
995
996  err_inode:
997         free_pipe_info(inode);
998         iput(inode);
999  err:
1000         return ERR_PTR(err);
1001 }
1002
1003 void free_write_pipe(struct file *f)
1004 {
1005         free_pipe_info(f->f_dentry->d_inode);
1006         path_put(&f->f_path);
1007         put_filp(f);
1008 }
1009
1010 struct file *create_read_pipe(struct file *wrf, int flags)
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 = wrf->f_path;
1018         path_get(&wrf->f_path);
1019         f->f_mapping = wrf->f_path.dentry->d_inode->i_mapping;
1020
1021         f->f_pos = 0;
1022         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1023         f->f_op = &read_pipefifo_fops;
1024         f->f_mode = FMODE_READ;
1025         f->f_version = 0;
1026
1027         return f;
1028 }
1029
1030 int do_pipe_flags(int *fd, int flags)
1031 {
1032         struct file *fw, *fr;
1033         int error;
1034         int fdw, fdr;
1035
1036         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1037                 return -EINVAL;
1038
1039         fw = create_write_pipe(flags);
1040         if (IS_ERR(fw))
1041                 return PTR_ERR(fw);
1042         fr = create_read_pipe(fw, flags);
1043         error = PTR_ERR(fr);
1044         if (IS_ERR(fr))
1045                 goto err_write_pipe;
1046
1047         error = get_unused_fd_flags(flags);
1048         if (error < 0)
1049                 goto err_read_pipe;
1050         fdr = error;
1051
1052         error = get_unused_fd_flags(flags);
1053         if (error < 0)
1054                 goto err_fdr;
1055         fdw = error;
1056
1057         audit_fd_pair(fdr, fdw);
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_fdr:
1066         put_unused_fd(fdr);
1067  err_read_pipe:
1068         path_put(&fr->f_path);
1069         put_filp(fr);
1070  err_write_pipe:
1071         free_write_pipe(fw);
1072         return error;
1073 }
1074
1075 /*
1076  * sys_pipe() is the normal C calling standard for creating
1077  * a pipe. It's not the way Unix traditionally does this, though.
1078  */
1079 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1080 {
1081         int fd[2];
1082         int error;
1083
1084         error = do_pipe_flags(fd, flags);
1085         if (!error) {
1086                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1087                         sys_close(fd[0]);
1088                         sys_close(fd[1]);
1089                         error = -EFAULT;
1090                 }
1091         }
1092         return error;
1093 }
1094
1095 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1096 {
1097         return sys_pipe2(fildes, 0);
1098 }
1099
1100 /*
1101  * pipefs should _never_ be mounted by userland - too much of security hassle,
1102  * no real gain from having the whole whorehouse mounted. So we don't need
1103  * any operations on the root directory. However, we need a non-trivial
1104  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1105  */
1106 static int pipefs_get_sb(struct file_system_type *fs_type,
1107                          int flags, const char *dev_name, void *data,
1108                          struct vfsmount *mnt)
1109 {
1110         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1111 }
1112
1113 static struct file_system_type pipe_fs_type = {
1114         .name           = "pipefs",
1115         .get_sb         = pipefs_get_sb,
1116         .kill_sb        = kill_anon_super,
1117 };
1118
1119 static int __init init_pipe_fs(void)
1120 {
1121         int err = register_filesystem(&pipe_fs_type);
1122
1123         if (!err) {
1124                 pipe_mnt = kern_mount(&pipe_fs_type);
1125                 if (IS_ERR(pipe_mnt)) {
1126                         err = PTR_ERR(pipe_mnt);
1127                         unregister_filesystem(&pipe_fs_type);
1128                 }
1129         }
1130         return err;
1131 }
1132
1133 static void __exit exit_pipe_fs(void)
1134 {
1135         unregister_filesystem(&pipe_fs_type);
1136         mntput(pipe_mnt);
1137 }
1138
1139 fs_initcall(init_pipe_fs);
1140 module_exit(exit_pipe_fs);