60da077400f1ddcc7c97abf3bc524912cc22c579
[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/log2.h>
15 #include <linux/mount.h>
16 #include <linux/pipe_fs_i.h>
17 #include <linux/uio.h>
18 #include <linux/highmem.h>
19 #include <linux/pagemap.h>
20 #include <linux/audit.h>
21 #include <linux/syscalls.h>
22 #include <linux/fcntl.h>
23
24 #include <asm/uaccess.h>
25 #include <asm/ioctls.h>
26
27 /*
28  * The max size that a non-root user is allowed to grow the pipe. Can
29  * be set by root in /proc/sys/fs/pipe-max-pages
30  */
31 unsigned int pipe_max_pages = PIPE_DEF_BUFFERS * 16;
32
33 /*
34  * We use a start+len construction, which provides full use of the 
35  * allocated memory.
36  * -- Florian Coosmann (FGC)
37  * 
38  * Reads with count = 0 should always return 0.
39  * -- Julian Bradfield 1999-06-07.
40  *
41  * FIFOs and Pipes now generate SIGIO for both readers and writers.
42  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
43  *
44  * pipe_read & write cleanup
45  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
46  */
47
48 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
49 {
50         if (pipe->inode)
51                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
52 }
53
54 void pipe_lock(struct pipe_inode_info *pipe)
55 {
56         /*
57          * pipe_lock() nests non-pipe inode locks (for writing to a file)
58          */
59         pipe_lock_nested(pipe, I_MUTEX_PARENT);
60 }
61 EXPORT_SYMBOL(pipe_lock);
62
63 void pipe_unlock(struct pipe_inode_info *pipe)
64 {
65         if (pipe->inode)
66                 mutex_unlock(&pipe->inode->i_mutex);
67 }
68 EXPORT_SYMBOL(pipe_unlock);
69
70 void pipe_double_lock(struct pipe_inode_info *pipe1,
71                       struct pipe_inode_info *pipe2)
72 {
73         BUG_ON(pipe1 == pipe2);
74
75         if (pipe1 < pipe2) {
76                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
77                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
78         } else {
79                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
80                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
81         }
82 }
83
84 /* Drop the inode semaphore and wait for a pipe event, atomically */
85 void pipe_wait(struct pipe_inode_info *pipe)
86 {
87         DEFINE_WAIT(wait);
88
89         /*
90          * Pipes are system-local resources, so sleeping on them
91          * is considered a noninteractive wait:
92          */
93         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
94         pipe_unlock(pipe);
95         schedule();
96         finish_wait(&pipe->wait, &wait);
97         pipe_lock(pipe);
98 }
99
100 static int
101 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
102                         int atomic)
103 {
104         unsigned long copy;
105
106         while (len > 0) {
107                 while (!iov->iov_len)
108                         iov++;
109                 copy = min_t(unsigned long, len, iov->iov_len);
110
111                 if (atomic) {
112                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
113                                 return -EFAULT;
114                 } else {
115                         if (copy_from_user(to, iov->iov_base, copy))
116                                 return -EFAULT;
117                 }
118                 to += copy;
119                 len -= copy;
120                 iov->iov_base += copy;
121                 iov->iov_len -= copy;
122         }
123         return 0;
124 }
125
126 static int
127 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
128                       int atomic)
129 {
130         unsigned long copy;
131
132         while (len > 0) {
133                 while (!iov->iov_len)
134                         iov++;
135                 copy = min_t(unsigned long, len, iov->iov_len);
136
137                 if (atomic) {
138                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
139                                 return -EFAULT;
140                 } else {
141                         if (copy_to_user(iov->iov_base, from, copy))
142                                 return -EFAULT;
143                 }
144                 from += copy;
145                 len -= copy;
146                 iov->iov_base += copy;
147                 iov->iov_len -= copy;
148         }
149         return 0;
150 }
151
152 /*
153  * Attempt to pre-fault in the user memory, so we can use atomic copies.
154  * Returns the number of bytes not faulted in.
155  */
156 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
157 {
158         while (!iov->iov_len)
159                 iov++;
160
161         while (len > 0) {
162                 unsigned long this_len;
163
164                 this_len = min_t(unsigned long, len, iov->iov_len);
165                 if (fault_in_pages_writeable(iov->iov_base, this_len))
166                         break;
167
168                 len -= this_len;
169                 iov++;
170         }
171
172         return len;
173 }
174
175 /*
176  * Pre-fault in the user memory, so we can use atomic copies.
177  */
178 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
179 {
180         while (!iov->iov_len)
181                 iov++;
182
183         while (len > 0) {
184                 unsigned long this_len;
185
186                 this_len = min_t(unsigned long, len, iov->iov_len);
187                 fault_in_pages_readable(iov->iov_base, this_len);
188                 len -= this_len;
189                 iov++;
190         }
191 }
192
193 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
194                                   struct pipe_buffer *buf)
195 {
196         struct page *page = buf->page;
197
198         /*
199          * If nobody else uses this page, and we don't already have a
200          * temporary page, let's keep track of it as a one-deep
201          * allocation cache. (Otherwise just release our reference to it)
202          */
203         if (page_count(page) == 1 && !pipe->tmp_page)
204                 pipe->tmp_page = page;
205         else
206                 page_cache_release(page);
207 }
208
209 /**
210  * generic_pipe_buf_map - virtually map a pipe buffer
211  * @pipe:       the pipe that the buffer belongs to
212  * @buf:        the buffer that should be mapped
213  * @atomic:     whether to use an atomic map
214  *
215  * Description:
216  *      This function returns a kernel virtual address mapping for the
217  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
218  *      and the caller has to be careful not to fault before calling
219  *      the unmap function.
220  *
221  *      Note that this function occupies KM_USER0 if @atomic != 0.
222  */
223 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
224                            struct pipe_buffer *buf, int atomic)
225 {
226         if (atomic) {
227                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
228                 return kmap_atomic(buf->page, KM_USER0);
229         }
230
231         return kmap(buf->page);
232 }
233
234 /**
235  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
236  * @pipe:       the pipe that the buffer belongs to
237  * @buf:        the buffer that should be unmapped
238  * @map_data:   the data that the mapping function returned
239  *
240  * Description:
241  *      This function undoes the mapping that ->map() provided.
242  */
243 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
244                             struct pipe_buffer *buf, void *map_data)
245 {
246         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
247                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
248                 kunmap_atomic(map_data, KM_USER0);
249         } else
250                 kunmap(buf->page);
251 }
252
253 /**
254  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
255  * @pipe:       the pipe that the buffer belongs to
256  * @buf:        the buffer to attempt to steal
257  *
258  * Description:
259  *      This function attempts to steal the &struct page attached to
260  *      @buf. If successful, this function returns 0 and returns with
261  *      the page locked. The caller may then reuse the page for whatever
262  *      he wishes; the typical use is insertion into a different file
263  *      page cache.
264  */
265 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
266                            struct pipe_buffer *buf)
267 {
268         struct page *page = buf->page;
269
270         /*
271          * A reference of one is golden, that means that the owner of this
272          * page is the only one holding a reference to it. lock the page
273          * and return OK.
274          */
275         if (page_count(page) == 1) {
276                 lock_page(page);
277                 return 0;
278         }
279
280         return 1;
281 }
282
283 /**
284  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
285  * @pipe:       the pipe that the buffer belongs to
286  * @buf:        the buffer to get a reference to
287  *
288  * Description:
289  *      This function grabs an extra reference to @buf. It's used in
290  *      in the tee() system call, when we duplicate the buffers in one
291  *      pipe into another.
292  */
293 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
294 {
295         page_cache_get(buf->page);
296 }
297
298 /**
299  * generic_pipe_buf_confirm - verify contents of the pipe buffer
300  * @info:       the pipe that the buffer belongs to
301  * @buf:        the buffer to confirm
302  *
303  * Description:
304  *      This function does nothing, because the generic pipe code uses
305  *      pages that are always good when inserted into the pipe.
306  */
307 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
308                              struct pipe_buffer *buf)
309 {
310         return 0;
311 }
312
313 /**
314  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
315  * @pipe:       the pipe that the buffer belongs to
316  * @buf:        the buffer to put a reference to
317  *
318  * Description:
319  *      This function releases a reference to @buf.
320  */
321 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
322                               struct pipe_buffer *buf)
323 {
324         page_cache_release(buf->page);
325 }
326
327 static const struct pipe_buf_operations anon_pipe_buf_ops = {
328         .can_merge = 1,
329         .map = generic_pipe_buf_map,
330         .unmap = generic_pipe_buf_unmap,
331         .confirm = generic_pipe_buf_confirm,
332         .release = anon_pipe_buf_release,
333         .steal = generic_pipe_buf_steal,
334         .get = generic_pipe_buf_get,
335 };
336
337 static ssize_t
338 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
339            unsigned long nr_segs, loff_t pos)
340 {
341         struct file *filp = iocb->ki_filp;
342         struct inode *inode = filp->f_path.dentry->d_inode;
343         struct pipe_inode_info *pipe;
344         int do_wakeup;
345         ssize_t ret;
346         struct iovec *iov = (struct iovec *)_iov;
347         size_t total_len;
348
349         total_len = iov_length(iov, nr_segs);
350         /* Null read succeeds. */
351         if (unlikely(total_len == 0))
352                 return 0;
353
354         do_wakeup = 0;
355         ret = 0;
356         mutex_lock(&inode->i_mutex);
357         pipe = inode->i_pipe;
358         for (;;) {
359                 int bufs = pipe->nrbufs;
360                 if (bufs) {
361                         int curbuf = pipe->curbuf;
362                         struct pipe_buffer *buf = pipe->bufs + curbuf;
363                         const struct pipe_buf_operations *ops = buf->ops;
364                         void *addr;
365                         size_t chars = buf->len;
366                         int error, atomic;
367
368                         if (chars > total_len)
369                                 chars = total_len;
370
371                         error = ops->confirm(pipe, buf);
372                         if (error) {
373                                 if (!ret)
374                                         error = ret;
375                                 break;
376                         }
377
378                         atomic = !iov_fault_in_pages_write(iov, chars);
379 redo:
380                         addr = ops->map(pipe, buf, atomic);
381                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
382                         ops->unmap(pipe, buf, addr);
383                         if (unlikely(error)) {
384                                 /*
385                                  * Just retry with the slow path if we failed.
386                                  */
387                                 if (atomic) {
388                                         atomic = 0;
389                                         goto redo;
390                                 }
391                                 if (!ret)
392                                         ret = error;
393                                 break;
394                         }
395                         ret += chars;
396                         buf->offset += chars;
397                         buf->len -= chars;
398                         if (!buf->len) {
399                                 buf->ops = NULL;
400                                 ops->release(pipe, buf);
401                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
402                                 pipe->curbuf = curbuf;
403                                 pipe->nrbufs = --bufs;
404                                 do_wakeup = 1;
405                         }
406                         total_len -= chars;
407                         if (!total_len)
408                                 break;  /* common path: read succeeded */
409                 }
410                 if (bufs)       /* More to do? */
411                         continue;
412                 if (!pipe->writers)
413                         break;
414                 if (!pipe->waiting_writers) {
415                         /* syscall merging: Usually we must not sleep
416                          * if O_NONBLOCK is set, or if we got some data.
417                          * But if a writer sleeps in kernel space, then
418                          * we can wait for that data without violating POSIX.
419                          */
420                         if (ret)
421                                 break;
422                         if (filp->f_flags & O_NONBLOCK) {
423                                 ret = -EAGAIN;
424                                 break;
425                         }
426                 }
427                 if (signal_pending(current)) {
428                         if (!ret)
429                                 ret = -ERESTARTSYS;
430                         break;
431                 }
432                 if (do_wakeup) {
433                         wake_up_interruptible_sync(&pipe->wait);
434                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
435                 }
436                 pipe_wait(pipe);
437         }
438         mutex_unlock(&inode->i_mutex);
439
440         /* Signal writers asynchronously that there is more room. */
441         if (do_wakeup) {
442                 wake_up_interruptible_sync(&pipe->wait);
443                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
444         }
445         if (ret > 0)
446                 file_accessed(filp);
447         return ret;
448 }
449
450 static ssize_t
451 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
452             unsigned long nr_segs, loff_t ppos)
453 {
454         struct file *filp = iocb->ki_filp;
455         struct inode *inode = filp->f_path.dentry->d_inode;
456         struct pipe_inode_info *pipe;
457         ssize_t ret;
458         int do_wakeup;
459         struct iovec *iov = (struct iovec *)_iov;
460         size_t total_len;
461         ssize_t chars;
462
463         total_len = iov_length(iov, nr_segs);
464         /* Null write succeeds. */
465         if (unlikely(total_len == 0))
466                 return 0;
467
468         do_wakeup = 0;
469         ret = 0;
470         mutex_lock(&inode->i_mutex);
471         pipe = inode->i_pipe;
472
473         if (!pipe->readers) {
474                 send_sig(SIGPIPE, current, 0);
475                 ret = -EPIPE;
476                 goto out;
477         }
478
479         /* We try to merge small writes */
480         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
481         if (pipe->nrbufs && chars != 0) {
482                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
483                                                         (pipe->buffers - 1);
484                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
485                 const struct pipe_buf_operations *ops = buf->ops;
486                 int offset = buf->offset + buf->len;
487
488                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
489                         int error, atomic = 1;
490                         void *addr;
491
492                         error = ops->confirm(pipe, buf);
493                         if (error)
494                                 goto out;
495
496                         iov_fault_in_pages_read(iov, chars);
497 redo1:
498                         addr = ops->map(pipe, buf, atomic);
499                         error = pipe_iov_copy_from_user(offset + addr, iov,
500                                                         chars, atomic);
501                         ops->unmap(pipe, buf, addr);
502                         ret = error;
503                         do_wakeup = 1;
504                         if (error) {
505                                 if (atomic) {
506                                         atomic = 0;
507                                         goto redo1;
508                                 }
509                                 goto out;
510                         }
511                         buf->len += chars;
512                         total_len -= chars;
513                         ret = chars;
514                         if (!total_len)
515                                 goto out;
516                 }
517         }
518
519         for (;;) {
520                 int bufs;
521
522                 if (!pipe->readers) {
523                         send_sig(SIGPIPE, current, 0);
524                         if (!ret)
525                                 ret = -EPIPE;
526                         break;
527                 }
528                 bufs = pipe->nrbufs;
529                 if (bufs < pipe->buffers) {
530                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
531                         struct pipe_buffer *buf = pipe->bufs + newbuf;
532                         struct page *page = pipe->tmp_page;
533                         char *src;
534                         int error, atomic = 1;
535
536                         if (!page) {
537                                 page = alloc_page(GFP_HIGHUSER);
538                                 if (unlikely(!page)) {
539                                         ret = ret ? : -ENOMEM;
540                                         break;
541                                 }
542                                 pipe->tmp_page = page;
543                         }
544                         /* Always wake up, even if the copy fails. Otherwise
545                          * we lock up (O_NONBLOCK-)readers that sleep due to
546                          * syscall merging.
547                          * FIXME! Is this really true?
548                          */
549                         do_wakeup = 1;
550                         chars = PAGE_SIZE;
551                         if (chars > total_len)
552                                 chars = total_len;
553
554                         iov_fault_in_pages_read(iov, chars);
555 redo2:
556                         if (atomic)
557                                 src = kmap_atomic(page, KM_USER0);
558                         else
559                                 src = kmap(page);
560
561                         error = pipe_iov_copy_from_user(src, iov, chars,
562                                                         atomic);
563                         if (atomic)
564                                 kunmap_atomic(src, KM_USER0);
565                         else
566                                 kunmap(page);
567
568                         if (unlikely(error)) {
569                                 if (atomic) {
570                                         atomic = 0;
571                                         goto redo2;
572                                 }
573                                 if (!ret)
574                                         ret = error;
575                                 break;
576                         }
577                         ret += chars;
578
579                         /* Insert it into the buffer array */
580                         buf->page = page;
581                         buf->ops = &anon_pipe_buf_ops;
582                         buf->offset = 0;
583                         buf->len = chars;
584                         pipe->nrbufs = ++bufs;
585                         pipe->tmp_page = NULL;
586
587                         total_len -= chars;
588                         if (!total_len)
589                                 break;
590                 }
591                 if (bufs < pipe->buffers)
592                         continue;
593                 if (filp->f_flags & O_NONBLOCK) {
594                         if (!ret)
595                                 ret = -EAGAIN;
596                         break;
597                 }
598                 if (signal_pending(current)) {
599                         if (!ret)
600                                 ret = -ERESTARTSYS;
601                         break;
602                 }
603                 if (do_wakeup) {
604                         wake_up_interruptible_sync(&pipe->wait);
605                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
606                         do_wakeup = 0;
607                 }
608                 pipe->waiting_writers++;
609                 pipe_wait(pipe);
610                 pipe->waiting_writers--;
611         }
612 out:
613         mutex_unlock(&inode->i_mutex);
614         if (do_wakeup) {
615                 wake_up_interruptible_sync(&pipe->wait);
616                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
617         }
618         if (ret > 0)
619                 file_update_time(filp);
620         return ret;
621 }
622
623 static ssize_t
624 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
625 {
626         return -EBADF;
627 }
628
629 static ssize_t
630 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
631            loff_t *ppos)
632 {
633         return -EBADF;
634 }
635
636 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
637 {
638         struct inode *inode = filp->f_path.dentry->d_inode;
639         struct pipe_inode_info *pipe;
640         int count, buf, nrbufs;
641
642         switch (cmd) {
643                 case FIONREAD:
644                         mutex_lock(&inode->i_mutex);
645                         pipe = inode->i_pipe;
646                         count = 0;
647                         buf = pipe->curbuf;
648                         nrbufs = pipe->nrbufs;
649                         while (--nrbufs >= 0) {
650                                 count += pipe->bufs[buf].len;
651                                 buf = (buf+1) & (pipe->buffers - 1);
652                         }
653                         mutex_unlock(&inode->i_mutex);
654
655                         return put_user(count, (int __user *)arg);
656                 default:
657                         return -EINVAL;
658         }
659 }
660
661 /* No kernel lock held - fine */
662 static unsigned int
663 pipe_poll(struct file *filp, poll_table *wait)
664 {
665         unsigned int mask;
666         struct inode *inode = filp->f_path.dentry->d_inode;
667         struct pipe_inode_info *pipe = inode->i_pipe;
668         int nrbufs;
669
670         poll_wait(filp, &pipe->wait, wait);
671
672         /* Reading only -- no need for acquiring the semaphore.  */
673         nrbufs = pipe->nrbufs;
674         mask = 0;
675         if (filp->f_mode & FMODE_READ) {
676                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
677                 if (!pipe->writers && filp->f_version != pipe->w_counter)
678                         mask |= POLLHUP;
679         }
680
681         if (filp->f_mode & FMODE_WRITE) {
682                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
683                 /*
684                  * Most Unices do not set POLLERR for FIFOs but on Linux they
685                  * behave exactly like pipes for poll().
686                  */
687                 if (!pipe->readers)
688                         mask |= POLLERR;
689         }
690
691         return mask;
692 }
693
694 static int
695 pipe_release(struct inode *inode, int decr, int decw)
696 {
697         struct pipe_inode_info *pipe;
698
699         mutex_lock(&inode->i_mutex);
700         pipe = inode->i_pipe;
701         pipe->readers -= decr;
702         pipe->writers -= decw;
703
704         if (!pipe->readers && !pipe->writers) {
705                 free_pipe_info(inode);
706         } else {
707                 wake_up_interruptible_sync(&pipe->wait);
708                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
709                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
710         }
711         mutex_unlock(&inode->i_mutex);
712
713         return 0;
714 }
715
716 static int
717 pipe_read_fasync(int fd, struct file *filp, int on)
718 {
719         struct inode *inode = filp->f_path.dentry->d_inode;
720         int retval;
721
722         mutex_lock(&inode->i_mutex);
723         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
724         mutex_unlock(&inode->i_mutex);
725
726         return retval;
727 }
728
729
730 static int
731 pipe_write_fasync(int fd, struct file *filp, int on)
732 {
733         struct inode *inode = filp->f_path.dentry->d_inode;
734         int retval;
735
736         mutex_lock(&inode->i_mutex);
737         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
738         mutex_unlock(&inode->i_mutex);
739
740         return retval;
741 }
742
743
744 static int
745 pipe_rdwr_fasync(int fd, struct file *filp, int on)
746 {
747         struct inode *inode = filp->f_path.dentry->d_inode;
748         struct pipe_inode_info *pipe = inode->i_pipe;
749         int retval;
750
751         mutex_lock(&inode->i_mutex);
752         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
753         if (retval >= 0) {
754                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
755                 if (retval < 0) /* this can happen only if on == T */
756                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
757         }
758         mutex_unlock(&inode->i_mutex);
759         return retval;
760 }
761
762
763 static int
764 pipe_read_release(struct inode *inode, struct file *filp)
765 {
766         return pipe_release(inode, 1, 0);
767 }
768
769 static int
770 pipe_write_release(struct inode *inode, struct file *filp)
771 {
772         return pipe_release(inode, 0, 1);
773 }
774
775 static int
776 pipe_rdwr_release(struct inode *inode, struct file *filp)
777 {
778         int decr, decw;
779
780         decr = (filp->f_mode & FMODE_READ) != 0;
781         decw = (filp->f_mode & FMODE_WRITE) != 0;
782         return pipe_release(inode, decr, decw);
783 }
784
785 static int
786 pipe_read_open(struct inode *inode, struct file *filp)
787 {
788         int ret = -ENOENT;
789
790         mutex_lock(&inode->i_mutex);
791
792         if (inode->i_pipe) {
793                 ret = 0;
794                 inode->i_pipe->readers++;
795         }
796
797         mutex_unlock(&inode->i_mutex);
798
799         return ret;
800 }
801
802 static int
803 pipe_write_open(struct inode *inode, struct file *filp)
804 {
805         int ret = -ENOENT;
806
807         mutex_lock(&inode->i_mutex);
808
809         if (inode->i_pipe) {
810                 ret = 0;
811                 inode->i_pipe->writers++;
812         }
813
814         mutex_unlock(&inode->i_mutex);
815
816         return ret;
817 }
818
819 static int
820 pipe_rdwr_open(struct inode *inode, struct file *filp)
821 {
822         int ret = -ENOENT;
823
824         mutex_lock(&inode->i_mutex);
825
826         if (inode->i_pipe) {
827                 ret = 0;
828                 if (filp->f_mode & FMODE_READ)
829                         inode->i_pipe->readers++;
830                 if (filp->f_mode & FMODE_WRITE)
831                         inode->i_pipe->writers++;
832         }
833
834         mutex_unlock(&inode->i_mutex);
835
836         return ret;
837 }
838
839 /*
840  * The file_operations structs are not static because they
841  * are also used in linux/fs/fifo.c to do operations on FIFOs.
842  *
843  * Pipes reuse fifos' file_operations structs.
844  */
845 const struct file_operations read_pipefifo_fops = {
846         .llseek         = no_llseek,
847         .read           = do_sync_read,
848         .aio_read       = pipe_read,
849         .write          = bad_pipe_w,
850         .poll           = pipe_poll,
851         .unlocked_ioctl = pipe_ioctl,
852         .open           = pipe_read_open,
853         .release        = pipe_read_release,
854         .fasync         = pipe_read_fasync,
855 };
856
857 const struct file_operations write_pipefifo_fops = {
858         .llseek         = no_llseek,
859         .read           = bad_pipe_r,
860         .write          = do_sync_write,
861         .aio_write      = pipe_write,
862         .poll           = pipe_poll,
863         .unlocked_ioctl = pipe_ioctl,
864         .open           = pipe_write_open,
865         .release        = pipe_write_release,
866         .fasync         = pipe_write_fasync,
867 };
868
869 const struct file_operations rdwr_pipefifo_fops = {
870         .llseek         = no_llseek,
871         .read           = do_sync_read,
872         .aio_read       = pipe_read,
873         .write          = do_sync_write,
874         .aio_write      = pipe_write,
875         .poll           = pipe_poll,
876         .unlocked_ioctl = pipe_ioctl,
877         .open           = pipe_rdwr_open,
878         .release        = pipe_rdwr_release,
879         .fasync         = pipe_rdwr_fasync,
880 };
881
882 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
883 {
884         struct pipe_inode_info *pipe;
885
886         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
887         if (pipe) {
888                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
889                 if (pipe->bufs) {
890                         init_waitqueue_head(&pipe->wait);
891                         pipe->r_counter = pipe->w_counter = 1;
892                         pipe->inode = inode;
893                         pipe->buffers = PIPE_DEF_BUFFERS;
894                         return pipe;
895                 }
896                 kfree(pipe);
897         }
898
899         return NULL;
900 }
901
902 void __free_pipe_info(struct pipe_inode_info *pipe)
903 {
904         int i;
905
906         for (i = 0; i < pipe->buffers; i++) {
907                 struct pipe_buffer *buf = pipe->bufs + i;
908                 if (buf->ops)
909                         buf->ops->release(pipe, buf);
910         }
911         if (pipe->tmp_page)
912                 __free_page(pipe->tmp_page);
913         kfree(pipe->bufs);
914         kfree(pipe);
915 }
916
917 void free_pipe_info(struct inode *inode)
918 {
919         __free_pipe_info(inode->i_pipe);
920         inode->i_pipe = NULL;
921 }
922
923 static struct vfsmount *pipe_mnt __read_mostly;
924
925 /*
926  * pipefs_dname() is called from d_path().
927  */
928 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
929 {
930         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
931                                 dentry->d_inode->i_ino);
932 }
933
934 static const struct dentry_operations pipefs_dentry_operations = {
935         .d_dname        = pipefs_dname,
936 };
937
938 static struct inode * get_pipe_inode(void)
939 {
940         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
941         struct pipe_inode_info *pipe;
942
943         if (!inode)
944                 goto fail_inode;
945
946         pipe = alloc_pipe_info(inode);
947         if (!pipe)
948                 goto fail_iput;
949         inode->i_pipe = pipe;
950
951         pipe->readers = pipe->writers = 1;
952         inode->i_fop = &rdwr_pipefifo_fops;
953
954         /*
955          * Mark the inode dirty from the very beginning,
956          * that way it will never be moved to the dirty
957          * list because "mark_inode_dirty()" will think
958          * that it already _is_ on the dirty list.
959          */
960         inode->i_state = I_DIRTY;
961         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
962         inode->i_uid = current_fsuid();
963         inode->i_gid = current_fsgid();
964         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
965
966         return inode;
967
968 fail_iput:
969         iput(inode);
970
971 fail_inode:
972         return NULL;
973 }
974
975 struct file *create_write_pipe(int flags)
976 {
977         int err;
978         struct inode *inode;
979         struct file *f;
980         struct path path;
981         struct qstr name = { .name = "" };
982
983         err = -ENFILE;
984         inode = get_pipe_inode();
985         if (!inode)
986                 goto err;
987
988         err = -ENOMEM;
989         path.dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
990         if (!path.dentry)
991                 goto err_inode;
992         path.mnt = mntget(pipe_mnt);
993
994         path.dentry->d_op = &pipefs_dentry_operations;
995         d_instantiate(path.dentry, inode);
996
997         err = -ENFILE;
998         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
999         if (!f)
1000                 goto err_dentry;
1001         f->f_mapping = inode->i_mapping;
1002
1003         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
1004         f->f_version = 0;
1005
1006         return f;
1007
1008  err_dentry:
1009         free_pipe_info(inode);
1010         path_put(&path);
1011         return ERR_PTR(err);
1012
1013  err_inode:
1014         free_pipe_info(inode);
1015         iput(inode);
1016  err:
1017         return ERR_PTR(err);
1018 }
1019
1020 void free_write_pipe(struct file *f)
1021 {
1022         free_pipe_info(f->f_dentry->d_inode);
1023         path_put(&f->f_path);
1024         put_filp(f);
1025 }
1026
1027 struct file *create_read_pipe(struct file *wrf, int flags)
1028 {
1029         /* Grab pipe from the writer */
1030         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1031                                     &read_pipefifo_fops);
1032         if (!f)
1033                 return ERR_PTR(-ENFILE);
1034
1035         path_get(&wrf->f_path);
1036         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1037
1038         return f;
1039 }
1040
1041 int do_pipe_flags(int *fd, int flags)
1042 {
1043         struct file *fw, *fr;
1044         int error;
1045         int fdw, fdr;
1046
1047         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1048                 return -EINVAL;
1049
1050         fw = create_write_pipe(flags);
1051         if (IS_ERR(fw))
1052                 return PTR_ERR(fw);
1053         fr = create_read_pipe(fw, flags);
1054         error = PTR_ERR(fr);
1055         if (IS_ERR(fr))
1056                 goto err_write_pipe;
1057
1058         error = get_unused_fd_flags(flags);
1059         if (error < 0)
1060                 goto err_read_pipe;
1061         fdr = error;
1062
1063         error = get_unused_fd_flags(flags);
1064         if (error < 0)
1065                 goto err_fdr;
1066         fdw = error;
1067
1068         audit_fd_pair(fdr, fdw);
1069         fd_install(fdr, fr);
1070         fd_install(fdw, fw);
1071         fd[0] = fdr;
1072         fd[1] = fdw;
1073
1074         return 0;
1075
1076  err_fdr:
1077         put_unused_fd(fdr);
1078  err_read_pipe:
1079         path_put(&fr->f_path);
1080         put_filp(fr);
1081  err_write_pipe:
1082         free_write_pipe(fw);
1083         return error;
1084 }
1085
1086 /*
1087  * sys_pipe() is the normal C calling standard for creating
1088  * a pipe. It's not the way Unix traditionally does this, though.
1089  */
1090 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1091 {
1092         int fd[2];
1093         int error;
1094
1095         error = do_pipe_flags(fd, flags);
1096         if (!error) {
1097                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1098                         sys_close(fd[0]);
1099                         sys_close(fd[1]);
1100                         error = -EFAULT;
1101                 }
1102         }
1103         return error;
1104 }
1105
1106 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1107 {
1108         return sys_pipe2(fildes, 0);
1109 }
1110
1111 /*
1112  * Allocate a new array of pipe buffers and copy the info over. Returns the
1113  * pipe size if successful, or return -ERROR on error.
1114  */
1115 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1116 {
1117         struct pipe_buffer *bufs;
1118
1119         /*
1120          * Must be a power-of-2 currently
1121          */
1122         if (!is_power_of_2(arg))
1123                 return -EINVAL;
1124
1125         /*
1126          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1127          * expect a lot of shrink+grow operations, just free and allocate
1128          * again like we would do for growing. If the pipe currently
1129          * contains more buffers than arg, then return busy.
1130          */
1131         if (arg < pipe->nrbufs)
1132                 return -EBUSY;
1133
1134         bufs = kcalloc(arg, sizeof(struct pipe_buffer), GFP_KERNEL);
1135         if (unlikely(!bufs))
1136                 return -ENOMEM;
1137
1138         /*
1139          * The pipe array wraps around, so just start the new one at zero
1140          * and adjust the indexes.
1141          */
1142         if (pipe->nrbufs) {
1143                 const unsigned int tail = pipe->nrbufs & (pipe->buffers - 1);
1144                 const unsigned int head = pipe->nrbufs - tail;
1145
1146                 if (head)
1147                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1148                 if (tail)
1149                         memcpy(bufs + head, pipe->bufs + pipe->curbuf, tail * sizeof(struct pipe_buffer));
1150         }
1151
1152         pipe->curbuf = 0;
1153         kfree(pipe->bufs);
1154         pipe->bufs = bufs;
1155         pipe->buffers = arg;
1156         return arg;
1157 }
1158
1159 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1160 {
1161         struct pipe_inode_info *pipe;
1162         long ret;
1163
1164         pipe = file->f_path.dentry->d_inode->i_pipe;
1165         if (!pipe)
1166                 return -EBADF;
1167
1168         mutex_lock(&pipe->inode->i_mutex);
1169
1170         switch (cmd) {
1171         case F_SETPIPE_SZ:
1172                 if (!capable(CAP_SYS_ADMIN) && arg > pipe_max_pages) {
1173                         ret = -EINVAL;
1174                         goto out;
1175                 }
1176                 /*
1177                  * The pipe needs to be at least 2 pages large to
1178                  * guarantee POSIX behaviour.
1179                  */
1180                 if (arg < 2) {
1181                         ret = -EINVAL;
1182                         goto out;
1183                 }
1184                 ret = pipe_set_size(pipe, arg);
1185                 break;
1186         case F_GETPIPE_SZ:
1187                 ret = pipe->buffers;
1188                 break;
1189         default:
1190                 ret = -EINVAL;
1191                 break;
1192         }
1193
1194 out:
1195         mutex_unlock(&pipe->inode->i_mutex);
1196         return ret;
1197 }
1198
1199 /*
1200  * pipefs should _never_ be mounted by userland - too much of security hassle,
1201  * no real gain from having the whole whorehouse mounted. So we don't need
1202  * any operations on the root directory. However, we need a non-trivial
1203  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1204  */
1205 static int pipefs_get_sb(struct file_system_type *fs_type,
1206                          int flags, const char *dev_name, void *data,
1207                          struct vfsmount *mnt)
1208 {
1209         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1210 }
1211
1212 static struct file_system_type pipe_fs_type = {
1213         .name           = "pipefs",
1214         .get_sb         = pipefs_get_sb,
1215         .kill_sb        = kill_anon_super,
1216 };
1217
1218 static int __init init_pipe_fs(void)
1219 {
1220         int err = register_filesystem(&pipe_fs_type);
1221
1222         if (!err) {
1223                 pipe_mnt = kern_mount(&pipe_fs_type);
1224                 if (IS_ERR(pipe_mnt)) {
1225                         err = PTR_ERR(pipe_mnt);
1226                         unregister_filesystem(&pipe_fs_type);
1227                 }
1228         }
1229         return err;
1230 }
1231
1232 static void __exit exit_pipe_fs(void)
1233 {
1234         unregister_filesystem(&pipe_fs_type);
1235         mntput(pipe_mnt);
1236 }
1237
1238 fs_initcall(init_pipe_fs);
1239 module_exit(exit_pipe_fs);