splice: fix racy pipe->buffers uses
[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/magic.h>
17 #include <linux/pipe_fs_i.h>
18 #include <linux/uio.h>
19 #include <linux/highmem.h>
20 #include <linux/pagemap.h>
21 #include <linux/audit.h>
22 #include <linux/syscalls.h>
23 #include <linux/fcntl.h>
24
25 #include <asm/uaccess.h>
26 #include <asm/ioctls.h>
27
28 /*
29  * The max size that a non-root user is allowed to grow the pipe. Can
30  * be set by root in /proc/sys/fs/pipe-max-size
31  */
32 unsigned int pipe_max_size = 1048576;
33
34 /*
35  * Minimum pipe size, as required by POSIX
36  */
37 unsigned int pipe_min_size = PAGE_SIZE;
38
39 /*
40  * We use a start+len construction, which provides full use of the 
41  * allocated memory.
42  * -- Florian Coosmann (FGC)
43  * 
44  * Reads with count = 0 should always return 0.
45  * -- Julian Bradfield 1999-06-07.
46  *
47  * FIFOs and Pipes now generate SIGIO for both readers and writers.
48  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
49  *
50  * pipe_read & write cleanup
51  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
52  */
53
54 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
55 {
56         if (pipe->inode)
57                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
58 }
59
60 void pipe_lock(struct pipe_inode_info *pipe)
61 {
62         /*
63          * pipe_lock() nests non-pipe inode locks (for writing to a file)
64          */
65         pipe_lock_nested(pipe, I_MUTEX_PARENT);
66 }
67 EXPORT_SYMBOL(pipe_lock);
68
69 void pipe_unlock(struct pipe_inode_info *pipe)
70 {
71         if (pipe->inode)
72                 mutex_unlock(&pipe->inode->i_mutex);
73 }
74 EXPORT_SYMBOL(pipe_unlock);
75
76 void pipe_double_lock(struct pipe_inode_info *pipe1,
77                       struct pipe_inode_info *pipe2)
78 {
79         BUG_ON(pipe1 == pipe2);
80
81         if (pipe1 < pipe2) {
82                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
83                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
84         } else {
85                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
86                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
87         }
88 }
89
90 /* Drop the inode semaphore and wait for a pipe event, atomically */
91 void pipe_wait(struct pipe_inode_info *pipe)
92 {
93         DEFINE_WAIT(wait);
94
95         /*
96          * Pipes are system-local resources, so sleeping on them
97          * is considered a noninteractive wait:
98          */
99         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
100         pipe_unlock(pipe);
101         schedule();
102         finish_wait(&pipe->wait, &wait);
103         pipe_lock(pipe);
104 }
105
106 static int
107 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
108                         int atomic)
109 {
110         unsigned long copy;
111
112         while (len > 0) {
113                 while (!iov->iov_len)
114                         iov++;
115                 copy = min_t(unsigned long, len, iov->iov_len);
116
117                 if (atomic) {
118                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
119                                 return -EFAULT;
120                 } else {
121                         if (copy_from_user(to, iov->iov_base, copy))
122                                 return -EFAULT;
123                 }
124                 to += copy;
125                 len -= copy;
126                 iov->iov_base += copy;
127                 iov->iov_len -= copy;
128         }
129         return 0;
130 }
131
132 static int
133 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
134                       int atomic)
135 {
136         unsigned long copy;
137
138         while (len > 0) {
139                 while (!iov->iov_len)
140                         iov++;
141                 copy = min_t(unsigned long, len, iov->iov_len);
142
143                 if (atomic) {
144                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
145                                 return -EFAULT;
146                 } else {
147                         if (copy_to_user(iov->iov_base, from, copy))
148                                 return -EFAULT;
149                 }
150                 from += copy;
151                 len -= copy;
152                 iov->iov_base += copy;
153                 iov->iov_len -= copy;
154         }
155         return 0;
156 }
157
158 /*
159  * Attempt to pre-fault in the user memory, so we can use atomic copies.
160  * Returns the number of bytes not faulted in.
161  */
162 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
163 {
164         while (!iov->iov_len)
165                 iov++;
166
167         while (len > 0) {
168                 unsigned long this_len;
169
170                 this_len = min_t(unsigned long, len, iov->iov_len);
171                 if (fault_in_pages_writeable(iov->iov_base, this_len))
172                         break;
173
174                 len -= this_len;
175                 iov++;
176         }
177
178         return len;
179 }
180
181 /*
182  * Pre-fault in the user memory, so we can use atomic copies.
183  */
184 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
185 {
186         while (!iov->iov_len)
187                 iov++;
188
189         while (len > 0) {
190                 unsigned long this_len;
191
192                 this_len = min_t(unsigned long, len, iov->iov_len);
193                 fault_in_pages_readable(iov->iov_base, this_len);
194                 len -= this_len;
195                 iov++;
196         }
197 }
198
199 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
200                                   struct pipe_buffer *buf)
201 {
202         struct page *page = buf->page;
203
204         /*
205          * If nobody else uses this page, and we don't already have a
206          * temporary page, let's keep track of it as a one-deep
207          * allocation cache. (Otherwise just release our reference to it)
208          */
209         if (page_count(page) == 1 && !pipe->tmp_page)
210                 pipe->tmp_page = page;
211         else
212                 page_cache_release(page);
213 }
214
215 /**
216  * generic_pipe_buf_map - virtually map a pipe buffer
217  * @pipe:       the pipe that the buffer belongs to
218  * @buf:        the buffer that should be mapped
219  * @atomic:     whether to use an atomic map
220  *
221  * Description:
222  *      This function returns a kernel virtual address mapping for the
223  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
224  *      and the caller has to be careful not to fault before calling
225  *      the unmap function.
226  *
227  *      Note that this function occupies KM_USER0 if @atomic != 0.
228  */
229 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
230                            struct pipe_buffer *buf, int atomic)
231 {
232         if (atomic) {
233                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
234                 return kmap_atomic(buf->page);
235         }
236
237         return kmap(buf->page);
238 }
239 EXPORT_SYMBOL(generic_pipe_buf_map);
240
241 /**
242  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
243  * @pipe:       the pipe that the buffer belongs to
244  * @buf:        the buffer that should be unmapped
245  * @map_data:   the data that the mapping function returned
246  *
247  * Description:
248  *      This function undoes the mapping that ->map() provided.
249  */
250 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
251                             struct pipe_buffer *buf, void *map_data)
252 {
253         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
254                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
255                 kunmap_atomic(map_data);
256         } else
257                 kunmap(buf->page);
258 }
259 EXPORT_SYMBOL(generic_pipe_buf_unmap);
260
261 /**
262  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
263  * @pipe:       the pipe that the buffer belongs to
264  * @buf:        the buffer to attempt to steal
265  *
266  * Description:
267  *      This function attempts to steal the &struct page attached to
268  *      @buf. If successful, this function returns 0 and returns with
269  *      the page locked. The caller may then reuse the page for whatever
270  *      he wishes; the typical use is insertion into a different file
271  *      page cache.
272  */
273 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
274                            struct pipe_buffer *buf)
275 {
276         struct page *page = buf->page;
277
278         /*
279          * A reference of one is golden, that means that the owner of this
280          * page is the only one holding a reference to it. lock the page
281          * and return OK.
282          */
283         if (page_count(page) == 1) {
284                 lock_page(page);
285                 return 0;
286         }
287
288         return 1;
289 }
290 EXPORT_SYMBOL(generic_pipe_buf_steal);
291
292 /**
293  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
294  * @pipe:       the pipe that the buffer belongs to
295  * @buf:        the buffer to get a reference to
296  *
297  * Description:
298  *      This function grabs an extra reference to @buf. It's used in
299  *      in the tee() system call, when we duplicate the buffers in one
300  *      pipe into another.
301  */
302 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
303 {
304         page_cache_get(buf->page);
305 }
306 EXPORT_SYMBOL(generic_pipe_buf_get);
307
308 /**
309  * generic_pipe_buf_confirm - verify contents of the pipe buffer
310  * @info:       the pipe that the buffer belongs to
311  * @buf:        the buffer to confirm
312  *
313  * Description:
314  *      This function does nothing, because the generic pipe code uses
315  *      pages that are always good when inserted into the pipe.
316  */
317 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
318                              struct pipe_buffer *buf)
319 {
320         return 0;
321 }
322 EXPORT_SYMBOL(generic_pipe_buf_confirm);
323
324 /**
325  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
326  * @pipe:       the pipe that the buffer belongs to
327  * @buf:        the buffer to put a reference to
328  *
329  * Description:
330  *      This function releases a reference to @buf.
331  */
332 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
333                               struct pipe_buffer *buf)
334 {
335         page_cache_release(buf->page);
336 }
337 EXPORT_SYMBOL(generic_pipe_buf_release);
338
339 static const struct pipe_buf_operations anon_pipe_buf_ops = {
340         .can_merge = 1,
341         .map = generic_pipe_buf_map,
342         .unmap = generic_pipe_buf_unmap,
343         .confirm = generic_pipe_buf_confirm,
344         .release = anon_pipe_buf_release,
345         .steal = generic_pipe_buf_steal,
346         .get = generic_pipe_buf_get,
347 };
348
349 static const struct pipe_buf_operations packet_pipe_buf_ops = {
350         .can_merge = 0,
351         .map = generic_pipe_buf_map,
352         .unmap = generic_pipe_buf_unmap,
353         .confirm = generic_pipe_buf_confirm,
354         .release = anon_pipe_buf_release,
355         .steal = generic_pipe_buf_steal,
356         .get = generic_pipe_buf_get,
357 };
358
359 static ssize_t
360 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
361            unsigned long nr_segs, loff_t pos)
362 {
363         struct file *filp = iocb->ki_filp;
364         struct inode *inode = filp->f_path.dentry->d_inode;
365         struct pipe_inode_info *pipe;
366         int do_wakeup;
367         ssize_t ret;
368         struct iovec *iov = (struct iovec *)_iov;
369         size_t total_len;
370
371         total_len = iov_length(iov, nr_segs);
372         /* Null read succeeds. */
373         if (unlikely(total_len == 0))
374                 return 0;
375
376         do_wakeup = 0;
377         ret = 0;
378         mutex_lock(&inode->i_mutex);
379         pipe = inode->i_pipe;
380         for (;;) {
381                 int bufs = pipe->nrbufs;
382                 if (bufs) {
383                         int curbuf = pipe->curbuf;
384                         struct pipe_buffer *buf = pipe->bufs + curbuf;
385                         const struct pipe_buf_operations *ops = buf->ops;
386                         void *addr;
387                         size_t chars = buf->len;
388                         int error, atomic;
389
390                         if (chars > total_len)
391                                 chars = total_len;
392
393                         error = ops->confirm(pipe, buf);
394                         if (error) {
395                                 if (!ret)
396                                         ret = error;
397                                 break;
398                         }
399
400                         atomic = !iov_fault_in_pages_write(iov, chars);
401 redo:
402                         addr = ops->map(pipe, buf, atomic);
403                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
404                         ops->unmap(pipe, buf, addr);
405                         if (unlikely(error)) {
406                                 /*
407                                  * Just retry with the slow path if we failed.
408                                  */
409                                 if (atomic) {
410                                         atomic = 0;
411                                         goto redo;
412                                 }
413                                 if (!ret)
414                                         ret = error;
415                                 break;
416                         }
417                         ret += chars;
418                         buf->offset += chars;
419                         buf->len -= chars;
420
421                         /* Was it a packet buffer? Clean up and exit */
422                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
423                                 total_len = chars;
424                                 buf->len = 0;
425                         }
426
427                         if (!buf->len) {
428                                 buf->ops = NULL;
429                                 ops->release(pipe, buf);
430                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
431                                 pipe->curbuf = curbuf;
432                                 pipe->nrbufs = --bufs;
433                                 do_wakeup = 1;
434                         }
435                         total_len -= chars;
436                         if (!total_len)
437                                 break;  /* common path: read succeeded */
438                 }
439                 if (bufs)       /* More to do? */
440                         continue;
441                 if (!pipe->writers)
442                         break;
443                 if (!pipe->waiting_writers) {
444                         /* syscall merging: Usually we must not sleep
445                          * if O_NONBLOCK is set, or if we got some data.
446                          * But if a writer sleeps in kernel space, then
447                          * we can wait for that data without violating POSIX.
448                          */
449                         if (ret)
450                                 break;
451                         if (filp->f_flags & O_NONBLOCK) {
452                                 ret = -EAGAIN;
453                                 break;
454                         }
455                 }
456                 if (signal_pending(current)) {
457                         if (!ret)
458                                 ret = -ERESTARTSYS;
459                         break;
460                 }
461                 if (do_wakeup) {
462                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
463                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
464                 }
465                 pipe_wait(pipe);
466         }
467         mutex_unlock(&inode->i_mutex);
468
469         /* Signal writers asynchronously that there is more room. */
470         if (do_wakeup) {
471                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
472                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
473         }
474         if (ret > 0)
475                 file_accessed(filp);
476         return ret;
477 }
478
479 static inline int is_packetized(struct file *file)
480 {
481         return (file->f_flags & O_DIRECT) != 0;
482 }
483
484 static ssize_t
485 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
486             unsigned long nr_segs, loff_t ppos)
487 {
488         struct file *filp = iocb->ki_filp;
489         struct inode *inode = filp->f_path.dentry->d_inode;
490         struct pipe_inode_info *pipe;
491         ssize_t ret;
492         int do_wakeup;
493         struct iovec *iov = (struct iovec *)_iov;
494         size_t total_len;
495         ssize_t chars;
496
497         total_len = iov_length(iov, nr_segs);
498         /* Null write succeeds. */
499         if (unlikely(total_len == 0))
500                 return 0;
501
502         do_wakeup = 0;
503         ret = 0;
504         mutex_lock(&inode->i_mutex);
505         pipe = inode->i_pipe;
506
507         if (!pipe->readers) {
508                 send_sig(SIGPIPE, current, 0);
509                 ret = -EPIPE;
510                 goto out;
511         }
512
513         /* We try to merge small writes */
514         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
515         if (pipe->nrbufs && chars != 0) {
516                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
517                                                         (pipe->buffers - 1);
518                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
519                 const struct pipe_buf_operations *ops = buf->ops;
520                 int offset = buf->offset + buf->len;
521
522                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
523                         int error, atomic = 1;
524                         void *addr;
525
526                         error = ops->confirm(pipe, buf);
527                         if (error)
528                                 goto out;
529
530                         iov_fault_in_pages_read(iov, chars);
531 redo1:
532                         addr = ops->map(pipe, buf, atomic);
533                         error = pipe_iov_copy_from_user(offset + addr, iov,
534                                                         chars, atomic);
535                         ops->unmap(pipe, buf, addr);
536                         ret = error;
537                         do_wakeup = 1;
538                         if (error) {
539                                 if (atomic) {
540                                         atomic = 0;
541                                         goto redo1;
542                                 }
543                                 goto out;
544                         }
545                         buf->len += chars;
546                         total_len -= chars;
547                         ret = chars;
548                         if (!total_len)
549                                 goto out;
550                 }
551         }
552
553         for (;;) {
554                 int bufs;
555
556                 if (!pipe->readers) {
557                         send_sig(SIGPIPE, current, 0);
558                         if (!ret)
559                                 ret = -EPIPE;
560                         break;
561                 }
562                 bufs = pipe->nrbufs;
563                 if (bufs < pipe->buffers) {
564                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
565                         struct pipe_buffer *buf = pipe->bufs + newbuf;
566                         struct page *page = pipe->tmp_page;
567                         char *src;
568                         int error, atomic = 1;
569
570                         if (!page) {
571                                 page = alloc_page(GFP_HIGHUSER);
572                                 if (unlikely(!page)) {
573                                         ret = ret ? : -ENOMEM;
574                                         break;
575                                 }
576                                 pipe->tmp_page = page;
577                         }
578                         /* Always wake up, even if the copy fails. Otherwise
579                          * we lock up (O_NONBLOCK-)readers that sleep due to
580                          * syscall merging.
581                          * FIXME! Is this really true?
582                          */
583                         do_wakeup = 1;
584                         chars = PAGE_SIZE;
585                         if (chars > total_len)
586                                 chars = total_len;
587
588                         iov_fault_in_pages_read(iov, chars);
589 redo2:
590                         if (atomic)
591                                 src = kmap_atomic(page);
592                         else
593                                 src = kmap(page);
594
595                         error = pipe_iov_copy_from_user(src, iov, chars,
596                                                         atomic);
597                         if (atomic)
598                                 kunmap_atomic(src);
599                         else
600                                 kunmap(page);
601
602                         if (unlikely(error)) {
603                                 if (atomic) {
604                                         atomic = 0;
605                                         goto redo2;
606                                 }
607                                 if (!ret)
608                                         ret = error;
609                                 break;
610                         }
611                         ret += chars;
612
613                         /* Insert it into the buffer array */
614                         buf->page = page;
615                         buf->ops = &anon_pipe_buf_ops;
616                         buf->offset = 0;
617                         buf->len = chars;
618                         buf->flags = 0;
619                         if (is_packetized(filp)) {
620                                 buf->ops = &packet_pipe_buf_ops;
621                                 buf->flags = PIPE_BUF_FLAG_PACKET;
622                         }
623                         pipe->nrbufs = ++bufs;
624                         pipe->tmp_page = NULL;
625
626                         total_len -= chars;
627                         if (!total_len)
628                                 break;
629                 }
630                 if (bufs < pipe->buffers)
631                         continue;
632                 if (filp->f_flags & O_NONBLOCK) {
633                         if (!ret)
634                                 ret = -EAGAIN;
635                         break;
636                 }
637                 if (signal_pending(current)) {
638                         if (!ret)
639                                 ret = -ERESTARTSYS;
640                         break;
641                 }
642                 if (do_wakeup) {
643                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
644                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
645                         do_wakeup = 0;
646                 }
647                 pipe->waiting_writers++;
648                 pipe_wait(pipe);
649                 pipe->waiting_writers--;
650         }
651 out:
652         mutex_unlock(&inode->i_mutex);
653         if (do_wakeup) {
654                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
655                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
656         }
657         if (ret > 0)
658                 file_update_time(filp);
659         return ret;
660 }
661
662 static ssize_t
663 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
664 {
665         return -EBADF;
666 }
667
668 static ssize_t
669 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
670            loff_t *ppos)
671 {
672         return -EBADF;
673 }
674
675 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
676 {
677         struct inode *inode = filp->f_path.dentry->d_inode;
678         struct pipe_inode_info *pipe;
679         int count, buf, nrbufs;
680
681         switch (cmd) {
682                 case FIONREAD:
683                         mutex_lock(&inode->i_mutex);
684                         pipe = inode->i_pipe;
685                         count = 0;
686                         buf = pipe->curbuf;
687                         nrbufs = pipe->nrbufs;
688                         while (--nrbufs >= 0) {
689                                 count += pipe->bufs[buf].len;
690                                 buf = (buf+1) & (pipe->buffers - 1);
691                         }
692                         mutex_unlock(&inode->i_mutex);
693
694                         return put_user(count, (int __user *)arg);
695                 default:
696                         return -EINVAL;
697         }
698 }
699
700 /* No kernel lock held - fine */
701 static unsigned int
702 pipe_poll(struct file *filp, poll_table *wait)
703 {
704         unsigned int mask;
705         struct inode *inode = filp->f_path.dentry->d_inode;
706         struct pipe_inode_info *pipe = inode->i_pipe;
707         int nrbufs;
708
709         poll_wait(filp, &pipe->wait, wait);
710
711         /* Reading only -- no need for acquiring the semaphore.  */
712         nrbufs = pipe->nrbufs;
713         mask = 0;
714         if (filp->f_mode & FMODE_READ) {
715                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
716                 if (!pipe->writers && filp->f_version != pipe->w_counter)
717                         mask |= POLLHUP;
718         }
719
720         if (filp->f_mode & FMODE_WRITE) {
721                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
722                 /*
723                  * Most Unices do not set POLLERR for FIFOs but on Linux they
724                  * behave exactly like pipes for poll().
725                  */
726                 if (!pipe->readers)
727                         mask |= POLLERR;
728         }
729
730         return mask;
731 }
732
733 static int
734 pipe_release(struct inode *inode, int decr, int decw)
735 {
736         struct pipe_inode_info *pipe;
737
738         mutex_lock(&inode->i_mutex);
739         pipe = inode->i_pipe;
740         pipe->readers -= decr;
741         pipe->writers -= decw;
742
743         if (!pipe->readers && !pipe->writers) {
744                 free_pipe_info(inode);
745         } else {
746                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
747                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
748                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
749         }
750         mutex_unlock(&inode->i_mutex);
751
752         return 0;
753 }
754
755 static int
756 pipe_read_fasync(int fd, struct file *filp, int on)
757 {
758         struct inode *inode = filp->f_path.dentry->d_inode;
759         int retval;
760
761         mutex_lock(&inode->i_mutex);
762         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
763         mutex_unlock(&inode->i_mutex);
764
765         return retval;
766 }
767
768
769 static int
770 pipe_write_fasync(int fd, struct file *filp, int on)
771 {
772         struct inode *inode = filp->f_path.dentry->d_inode;
773         int retval;
774
775         mutex_lock(&inode->i_mutex);
776         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
777         mutex_unlock(&inode->i_mutex);
778
779         return retval;
780 }
781
782
783 static int
784 pipe_rdwr_fasync(int fd, struct file *filp, int on)
785 {
786         struct inode *inode = filp->f_path.dentry->d_inode;
787         struct pipe_inode_info *pipe = inode->i_pipe;
788         int retval;
789
790         mutex_lock(&inode->i_mutex);
791         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
792         if (retval >= 0) {
793                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
794                 if (retval < 0) /* this can happen only if on == T */
795                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
796         }
797         mutex_unlock(&inode->i_mutex);
798         return retval;
799 }
800
801
802 static int
803 pipe_read_release(struct inode *inode, struct file *filp)
804 {
805         return pipe_release(inode, 1, 0);
806 }
807
808 static int
809 pipe_write_release(struct inode *inode, struct file *filp)
810 {
811         return pipe_release(inode, 0, 1);
812 }
813
814 static int
815 pipe_rdwr_release(struct inode *inode, struct file *filp)
816 {
817         int decr, decw;
818
819         decr = (filp->f_mode & FMODE_READ) != 0;
820         decw = (filp->f_mode & FMODE_WRITE) != 0;
821         return pipe_release(inode, decr, decw);
822 }
823
824 static int
825 pipe_read_open(struct inode *inode, struct file *filp)
826 {
827         int ret = -ENOENT;
828
829         mutex_lock(&inode->i_mutex);
830
831         if (inode->i_pipe) {
832                 ret = 0;
833                 inode->i_pipe->readers++;
834         }
835
836         mutex_unlock(&inode->i_mutex);
837
838         return ret;
839 }
840
841 static int
842 pipe_write_open(struct inode *inode, struct file *filp)
843 {
844         int ret = -ENOENT;
845
846         mutex_lock(&inode->i_mutex);
847
848         if (inode->i_pipe) {
849                 ret = 0;
850                 inode->i_pipe->writers++;
851         }
852
853         mutex_unlock(&inode->i_mutex);
854
855         return ret;
856 }
857
858 static int
859 pipe_rdwr_open(struct inode *inode, struct file *filp)
860 {
861         int ret = -ENOENT;
862
863         mutex_lock(&inode->i_mutex);
864
865         if (inode->i_pipe) {
866                 ret = 0;
867                 if (filp->f_mode & FMODE_READ)
868                         inode->i_pipe->readers++;
869                 if (filp->f_mode & FMODE_WRITE)
870                         inode->i_pipe->writers++;
871         }
872
873         mutex_unlock(&inode->i_mutex);
874
875         return ret;
876 }
877
878 /*
879  * The file_operations structs are not static because they
880  * are also used in linux/fs/fifo.c to do operations on FIFOs.
881  *
882  * Pipes reuse fifos' file_operations structs.
883  */
884 const struct file_operations read_pipefifo_fops = {
885         .llseek         = no_llseek,
886         .read           = do_sync_read,
887         .aio_read       = pipe_read,
888         .write          = bad_pipe_w,
889         .poll           = pipe_poll,
890         .unlocked_ioctl = pipe_ioctl,
891         .open           = pipe_read_open,
892         .release        = pipe_read_release,
893         .fasync         = pipe_read_fasync,
894 };
895
896 const struct file_operations write_pipefifo_fops = {
897         .llseek         = no_llseek,
898         .read           = bad_pipe_r,
899         .write          = do_sync_write,
900         .aio_write      = pipe_write,
901         .poll           = pipe_poll,
902         .unlocked_ioctl = pipe_ioctl,
903         .open           = pipe_write_open,
904         .release        = pipe_write_release,
905         .fasync         = pipe_write_fasync,
906 };
907
908 const struct file_operations rdwr_pipefifo_fops = {
909         .llseek         = no_llseek,
910         .read           = do_sync_read,
911         .aio_read       = pipe_read,
912         .write          = do_sync_write,
913         .aio_write      = pipe_write,
914         .poll           = pipe_poll,
915         .unlocked_ioctl = pipe_ioctl,
916         .open           = pipe_rdwr_open,
917         .release        = pipe_rdwr_release,
918         .fasync         = pipe_rdwr_fasync,
919 };
920
921 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
922 {
923         struct pipe_inode_info *pipe;
924
925         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
926         if (pipe) {
927                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
928                 if (pipe->bufs) {
929                         init_waitqueue_head(&pipe->wait);
930                         pipe->r_counter = pipe->w_counter = 1;
931                         pipe->inode = inode;
932                         pipe->buffers = PIPE_DEF_BUFFERS;
933                         return pipe;
934                 }
935                 kfree(pipe);
936         }
937
938         return NULL;
939 }
940
941 void __free_pipe_info(struct pipe_inode_info *pipe)
942 {
943         int i;
944
945         for (i = 0; i < pipe->buffers; i++) {
946                 struct pipe_buffer *buf = pipe->bufs + i;
947                 if (buf->ops)
948                         buf->ops->release(pipe, buf);
949         }
950         if (pipe->tmp_page)
951                 __free_page(pipe->tmp_page);
952         kfree(pipe->bufs);
953         kfree(pipe);
954 }
955
956 void free_pipe_info(struct inode *inode)
957 {
958         __free_pipe_info(inode->i_pipe);
959         inode->i_pipe = NULL;
960 }
961
962 static struct vfsmount *pipe_mnt __read_mostly;
963
964 /*
965  * pipefs_dname() is called from d_path().
966  */
967 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
968 {
969         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
970                                 dentry->d_inode->i_ino);
971 }
972
973 static const struct dentry_operations pipefs_dentry_operations = {
974         .d_dname        = pipefs_dname,
975 };
976
977 static struct inode * get_pipe_inode(void)
978 {
979         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
980         struct pipe_inode_info *pipe;
981
982         if (!inode)
983                 goto fail_inode;
984
985         inode->i_ino = get_next_ino();
986
987         pipe = alloc_pipe_info(inode);
988         if (!pipe)
989                 goto fail_iput;
990         inode->i_pipe = pipe;
991
992         pipe->readers = pipe->writers = 1;
993         inode->i_fop = &rdwr_pipefifo_fops;
994
995         /*
996          * Mark the inode dirty from the very beginning,
997          * that way it will never be moved to the dirty
998          * list because "mark_inode_dirty()" will think
999          * that it already _is_ on the dirty list.
1000          */
1001         inode->i_state = I_DIRTY;
1002         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1003         inode->i_uid = current_fsuid();
1004         inode->i_gid = current_fsgid();
1005         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1006
1007         return inode;
1008
1009 fail_iput:
1010         iput(inode);
1011
1012 fail_inode:
1013         return NULL;
1014 }
1015
1016 struct file *create_write_pipe(int flags)
1017 {
1018         int err;
1019         struct inode *inode;
1020         struct file *f;
1021         struct path path;
1022         struct qstr name = { .name = "" };
1023
1024         err = -ENFILE;
1025         inode = get_pipe_inode();
1026         if (!inode)
1027                 goto err;
1028
1029         err = -ENOMEM;
1030         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1031         if (!path.dentry)
1032                 goto err_inode;
1033         path.mnt = mntget(pipe_mnt);
1034
1035         d_instantiate(path.dentry, inode);
1036
1037         err = -ENFILE;
1038         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1039         if (!f)
1040                 goto err_dentry;
1041         f->f_mapping = inode->i_mapping;
1042
1043         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
1044         f->f_version = 0;
1045
1046         return f;
1047
1048  err_dentry:
1049         free_pipe_info(inode);
1050         path_put(&path);
1051         return ERR_PTR(err);
1052
1053  err_inode:
1054         free_pipe_info(inode);
1055         iput(inode);
1056  err:
1057         return ERR_PTR(err);
1058 }
1059
1060 void free_write_pipe(struct file *f)
1061 {
1062         free_pipe_info(f->f_dentry->d_inode);
1063         path_put(&f->f_path);
1064         put_filp(f);
1065 }
1066
1067 struct file *create_read_pipe(struct file *wrf, int flags)
1068 {
1069         /* Grab pipe from the writer */
1070         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1071                                     &read_pipefifo_fops);
1072         if (!f)
1073                 return ERR_PTR(-ENFILE);
1074
1075         path_get(&wrf->f_path);
1076         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1077
1078         return f;
1079 }
1080
1081 int do_pipe_flags(int *fd, int flags)
1082 {
1083         struct file *fw, *fr;
1084         int error;
1085         int fdw, fdr;
1086
1087         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
1088                 return -EINVAL;
1089
1090         fw = create_write_pipe(flags);
1091         if (IS_ERR(fw))
1092                 return PTR_ERR(fw);
1093         fr = create_read_pipe(fw, flags);
1094         error = PTR_ERR(fr);
1095         if (IS_ERR(fr))
1096                 goto err_write_pipe;
1097
1098         error = get_unused_fd_flags(flags);
1099         if (error < 0)
1100                 goto err_read_pipe;
1101         fdr = error;
1102
1103         error = get_unused_fd_flags(flags);
1104         if (error < 0)
1105                 goto err_fdr;
1106         fdw = error;
1107
1108         audit_fd_pair(fdr, fdw);
1109         fd_install(fdr, fr);
1110         fd_install(fdw, fw);
1111         fd[0] = fdr;
1112         fd[1] = fdw;
1113
1114         return 0;
1115
1116  err_fdr:
1117         put_unused_fd(fdr);
1118  err_read_pipe:
1119         path_put(&fr->f_path);
1120         put_filp(fr);
1121  err_write_pipe:
1122         free_write_pipe(fw);
1123         return error;
1124 }
1125
1126 /*
1127  * sys_pipe() is the normal C calling standard for creating
1128  * a pipe. It's not the way Unix traditionally does this, though.
1129  */
1130 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1131 {
1132         int fd[2];
1133         int error;
1134
1135         error = do_pipe_flags(fd, flags);
1136         if (!error) {
1137                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1138                         sys_close(fd[0]);
1139                         sys_close(fd[1]);
1140                         error = -EFAULT;
1141                 }
1142         }
1143         return error;
1144 }
1145
1146 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1147 {
1148         return sys_pipe2(fildes, 0);
1149 }
1150
1151 /*
1152  * Allocate a new array of pipe buffers and copy the info over. Returns the
1153  * pipe size if successful, or return -ERROR on error.
1154  */
1155 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1156 {
1157         struct pipe_buffer *bufs;
1158
1159         /*
1160          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1161          * expect a lot of shrink+grow operations, just free and allocate
1162          * again like we would do for growing. If the pipe currently
1163          * contains more buffers than arg, then return busy.
1164          */
1165         if (nr_pages < pipe->nrbufs)
1166                 return -EBUSY;
1167
1168         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1169         if (unlikely(!bufs))
1170                 return -ENOMEM;
1171
1172         /*
1173          * The pipe array wraps around, so just start the new one at zero
1174          * and adjust the indexes.
1175          */
1176         if (pipe->nrbufs) {
1177                 unsigned int tail;
1178                 unsigned int head;
1179
1180                 tail = pipe->curbuf + pipe->nrbufs;
1181                 if (tail < pipe->buffers)
1182                         tail = 0;
1183                 else
1184                         tail &= (pipe->buffers - 1);
1185
1186                 head = pipe->nrbufs - tail;
1187                 if (head)
1188                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1189                 if (tail)
1190                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1191         }
1192
1193         pipe->curbuf = 0;
1194         kfree(pipe->bufs);
1195         pipe->bufs = bufs;
1196         pipe->buffers = nr_pages;
1197         return nr_pages * PAGE_SIZE;
1198 }
1199
1200 /*
1201  * Currently we rely on the pipe array holding a power-of-2 number
1202  * of pages.
1203  */
1204 static inline unsigned int round_pipe_size(unsigned int size)
1205 {
1206         unsigned long nr_pages;
1207
1208         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1209         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1210 }
1211
1212 /*
1213  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1214  * will return an error.
1215  */
1216 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1217                  size_t *lenp, loff_t *ppos)
1218 {
1219         int ret;
1220
1221         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1222         if (ret < 0 || !write)
1223                 return ret;
1224
1225         pipe_max_size = round_pipe_size(pipe_max_size);
1226         return ret;
1227 }
1228
1229 /*
1230  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1231  * location, so checking ->i_pipe is not enough to verify that this is a
1232  * pipe.
1233  */
1234 struct pipe_inode_info *get_pipe_info(struct file *file)
1235 {
1236         struct inode *i = file->f_path.dentry->d_inode;
1237
1238         return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1239 }
1240
1241 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1242 {
1243         struct pipe_inode_info *pipe;
1244         long ret;
1245
1246         pipe = get_pipe_info(file);
1247         if (!pipe)
1248                 return -EBADF;
1249
1250         mutex_lock(&pipe->inode->i_mutex);
1251
1252         switch (cmd) {
1253         case F_SETPIPE_SZ: {
1254                 unsigned int size, nr_pages;
1255
1256                 size = round_pipe_size(arg);
1257                 nr_pages = size >> PAGE_SHIFT;
1258
1259                 ret = -EINVAL;
1260                 if (!nr_pages)
1261                         goto out;
1262
1263                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1264                         ret = -EPERM;
1265                         goto out;
1266                 }
1267                 ret = pipe_set_size(pipe, nr_pages);
1268                 break;
1269                 }
1270         case F_GETPIPE_SZ:
1271                 ret = pipe->buffers * PAGE_SIZE;
1272                 break;
1273         default:
1274                 ret = -EINVAL;
1275                 break;
1276         }
1277
1278 out:
1279         mutex_unlock(&pipe->inode->i_mutex);
1280         return ret;
1281 }
1282
1283 static const struct super_operations pipefs_ops = {
1284         .destroy_inode = free_inode_nonrcu,
1285         .statfs = simple_statfs,
1286 };
1287
1288 /*
1289  * pipefs should _never_ be mounted by userland - too much of security hassle,
1290  * no real gain from having the whole whorehouse mounted. So we don't need
1291  * any operations on the root directory. However, we need a non-trivial
1292  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1293  */
1294 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1295                          int flags, const char *dev_name, void *data)
1296 {
1297         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1298                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1299 }
1300
1301 static struct file_system_type pipe_fs_type = {
1302         .name           = "pipefs",
1303         .mount          = pipefs_mount,
1304         .kill_sb        = kill_anon_super,
1305 };
1306
1307 static int __init init_pipe_fs(void)
1308 {
1309         int err = register_filesystem(&pipe_fs_type);
1310
1311         if (!err) {
1312                 pipe_mnt = kern_mount(&pipe_fs_type);
1313                 if (IS_ERR(pipe_mnt)) {
1314                         err = PTR_ERR(pipe_mnt);
1315                         unregister_filesystem(&pipe_fs_type);
1316                 }
1317         }
1318         return err;
1319 }
1320
1321 fs_initcall(init_pipe_fs);