pipe: iovec: Fix memory corruption when retrying atomic copy as non-atomic
[linux-3.10.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 #include <linux/aio.h>
25
26 #include <asm/uaccess.h>
27 #include <asm/ioctls.h>
28
29 #include "internal.h"
30
31 /*
32  * The max size that a non-root user is allowed to grow the pipe. Can
33  * be set by root in /proc/sys/fs/pipe-max-size
34  */
35 unsigned int pipe_max_size = 1048576;
36
37 /*
38  * Minimum pipe size, as required by POSIX
39  */
40 unsigned int pipe_min_size = PAGE_SIZE;
41
42 /*
43  * We use a start+len construction, which provides full use of the 
44  * allocated memory.
45  * -- Florian Coosmann (FGC)
46  * 
47  * Reads with count = 0 should always return 0.
48  * -- Julian Bradfield 1999-06-07.
49  *
50  * FIFOs and Pipes now generate SIGIO for both readers and writers.
51  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
52  *
53  * pipe_read & write cleanup
54  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
55  */
56
57 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
58 {
59         if (pipe->files)
60                 mutex_lock_nested(&pipe->mutex, subclass);
61 }
62
63 void pipe_lock(struct pipe_inode_info *pipe)
64 {
65         /*
66          * pipe_lock() nests non-pipe inode locks (for writing to a file)
67          */
68         pipe_lock_nested(pipe, I_MUTEX_PARENT);
69 }
70 EXPORT_SYMBOL(pipe_lock);
71
72 void pipe_unlock(struct pipe_inode_info *pipe)
73 {
74         if (pipe->files)
75                 mutex_unlock(&pipe->mutex);
76 }
77 EXPORT_SYMBOL(pipe_unlock);
78
79 static inline void __pipe_lock(struct pipe_inode_info *pipe)
80 {
81         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
82 }
83
84 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
85 {
86         mutex_unlock(&pipe->mutex);
87 }
88
89 void pipe_double_lock(struct pipe_inode_info *pipe1,
90                       struct pipe_inode_info *pipe2)
91 {
92         BUG_ON(pipe1 == pipe2);
93
94         if (pipe1 < pipe2) {
95                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
96                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
97         } else {
98                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
99                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
100         }
101 }
102
103 /* Drop the inode semaphore and wait for a pipe event, atomically */
104 void pipe_wait(struct pipe_inode_info *pipe)
105 {
106         DEFINE_WAIT(wait);
107
108         /*
109          * Pipes are system-local resources, so sleeping on them
110          * is considered a noninteractive wait:
111          */
112         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
113         pipe_unlock(pipe);
114         schedule();
115         finish_wait(&pipe->wait, &wait);
116         pipe_lock(pipe);
117 }
118
119 static int
120 pipe_iov_copy_from_user(void *addr, int *offset, struct iovec *iov,
121                         size_t *remaining, int atomic)
122 {
123         unsigned long copy;
124
125         while (*remaining > 0) {
126                 while (!iov->iov_len)
127                         iov++;
128                 copy = min_t(unsigned long, *remaining, iov->iov_len);
129
130                 if (atomic) {
131                         if (__copy_from_user_inatomic(addr + *offset,
132                                                       iov->iov_base, copy))
133                                 return -EFAULT;
134                 } else {
135                         if (copy_from_user(addr + *offset,
136                                            iov->iov_base, copy))
137                                 return -EFAULT;
138                 }
139                 *offset += copy;
140                 *remaining -= copy;
141                 iov->iov_base += copy;
142                 iov->iov_len -= copy;
143         }
144         return 0;
145 }
146
147 static int
148 pipe_iov_copy_to_user(struct iovec *iov, void *addr, int *offset,
149                       size_t *remaining, int atomic)
150 {
151         unsigned long copy;
152
153         while (*remaining > 0) {
154                 while (!iov->iov_len)
155                         iov++;
156                 copy = min_t(unsigned long, *remaining, iov->iov_len);
157
158                 if (atomic) {
159                         if (__copy_to_user_inatomic(iov->iov_base,
160                                                     addr + *offset, copy))
161                                 return -EFAULT;
162                 } else {
163                         if (copy_to_user(iov->iov_base,
164                                          addr + *offset, copy))
165                                 return -EFAULT;
166                 }
167                 *offset += copy;
168                 *remaining -= copy;
169                 iov->iov_base += copy;
170                 iov->iov_len -= copy;
171         }
172         return 0;
173 }
174
175 /*
176  * Attempt to pre-fault in the user memory, so we can use atomic copies.
177  * Returns the number of bytes not faulted in.
178  */
179 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
180 {
181         while (!iov->iov_len)
182                 iov++;
183
184         while (len > 0) {
185                 unsigned long this_len;
186
187                 this_len = min_t(unsigned long, len, iov->iov_len);
188                 if (fault_in_pages_writeable(iov->iov_base, this_len))
189                         break;
190
191                 len -= this_len;
192                 iov++;
193         }
194
195         return len;
196 }
197
198 /*
199  * Pre-fault in the user memory, so we can use atomic copies.
200  */
201 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
202 {
203         while (!iov->iov_len)
204                 iov++;
205
206         while (len > 0) {
207                 unsigned long this_len;
208
209                 this_len = min_t(unsigned long, len, iov->iov_len);
210                 fault_in_pages_readable(iov->iov_base, this_len);
211                 len -= this_len;
212                 iov++;
213         }
214 }
215
216 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
217                                   struct pipe_buffer *buf)
218 {
219         struct page *page = buf->page;
220
221         /*
222          * If nobody else uses this page, and we don't already have a
223          * temporary page, let's keep track of it as a one-deep
224          * allocation cache. (Otherwise just release our reference to it)
225          */
226         if (page_count(page) == 1 && !pipe->tmp_page)
227                 pipe->tmp_page = page;
228         else
229                 page_cache_release(page);
230 }
231
232 /**
233  * generic_pipe_buf_map - virtually map a pipe buffer
234  * @pipe:       the pipe that the buffer belongs to
235  * @buf:        the buffer that should be mapped
236  * @atomic:     whether to use an atomic map
237  *
238  * Description:
239  *      This function returns a kernel virtual address mapping for the
240  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
241  *      and the caller has to be careful not to fault before calling
242  *      the unmap function.
243  *
244  *      Note that this function calls kmap_atomic() if @atomic != 0.
245  */
246 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
247                            struct pipe_buffer *buf, int atomic)
248 {
249         if (atomic) {
250                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
251                 return kmap_atomic(buf->page);
252         }
253
254         return kmap(buf->page);
255 }
256 EXPORT_SYMBOL(generic_pipe_buf_map);
257
258 /**
259  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
260  * @pipe:       the pipe that the buffer belongs to
261  * @buf:        the buffer that should be unmapped
262  * @map_data:   the data that the mapping function returned
263  *
264  * Description:
265  *      This function undoes the mapping that ->map() provided.
266  */
267 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
268                             struct pipe_buffer *buf, void *map_data)
269 {
270         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
271                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
272                 kunmap_atomic(map_data);
273         } else
274                 kunmap(buf->page);
275 }
276 EXPORT_SYMBOL(generic_pipe_buf_unmap);
277
278 /**
279  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
280  * @pipe:       the pipe that the buffer belongs to
281  * @buf:        the buffer to attempt to steal
282  *
283  * Description:
284  *      This function attempts to steal the &struct page attached to
285  *      @buf. If successful, this function returns 0 and returns with
286  *      the page locked. The caller may then reuse the page for whatever
287  *      he wishes; the typical use is insertion into a different file
288  *      page cache.
289  */
290 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
291                            struct pipe_buffer *buf)
292 {
293         struct page *page = buf->page;
294
295         /*
296          * A reference of one is golden, that means that the owner of this
297          * page is the only one holding a reference to it. lock the page
298          * and return OK.
299          */
300         if (page_count(page) == 1) {
301                 lock_page(page);
302                 return 0;
303         }
304
305         return 1;
306 }
307 EXPORT_SYMBOL(generic_pipe_buf_steal);
308
309 /**
310  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
311  * @pipe:       the pipe that the buffer belongs to
312  * @buf:        the buffer to get a reference to
313  *
314  * Description:
315  *      This function grabs an extra reference to @buf. It's used in
316  *      in the tee() system call, when we duplicate the buffers in one
317  *      pipe into another.
318  */
319 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
320 {
321         page_cache_get(buf->page);
322 }
323 EXPORT_SYMBOL(generic_pipe_buf_get);
324
325 /**
326  * generic_pipe_buf_confirm - verify contents of the pipe buffer
327  * @info:       the pipe that the buffer belongs to
328  * @buf:        the buffer to confirm
329  *
330  * Description:
331  *      This function does nothing, because the generic pipe code uses
332  *      pages that are always good when inserted into the pipe.
333  */
334 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
335                              struct pipe_buffer *buf)
336 {
337         return 0;
338 }
339 EXPORT_SYMBOL(generic_pipe_buf_confirm);
340
341 /**
342  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
343  * @pipe:       the pipe that the buffer belongs to
344  * @buf:        the buffer to put a reference to
345  *
346  * Description:
347  *      This function releases a reference to @buf.
348  */
349 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
350                               struct pipe_buffer *buf)
351 {
352         page_cache_release(buf->page);
353 }
354 EXPORT_SYMBOL(generic_pipe_buf_release);
355
356 static const struct pipe_buf_operations anon_pipe_buf_ops = {
357         .can_merge = 1,
358         .map = generic_pipe_buf_map,
359         .unmap = generic_pipe_buf_unmap,
360         .confirm = generic_pipe_buf_confirm,
361         .release = anon_pipe_buf_release,
362         .steal = generic_pipe_buf_steal,
363         .get = generic_pipe_buf_get,
364 };
365
366 static const struct pipe_buf_operations packet_pipe_buf_ops = {
367         .can_merge = 0,
368         .map = generic_pipe_buf_map,
369         .unmap = generic_pipe_buf_unmap,
370         .confirm = generic_pipe_buf_confirm,
371         .release = anon_pipe_buf_release,
372         .steal = generic_pipe_buf_steal,
373         .get = generic_pipe_buf_get,
374 };
375
376 static ssize_t
377 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
378            unsigned long nr_segs, loff_t pos)
379 {
380         struct file *filp = iocb->ki_filp;
381         struct pipe_inode_info *pipe = filp->private_data;
382         int do_wakeup;
383         ssize_t ret;
384         struct iovec *iov = (struct iovec *)_iov;
385         size_t total_len;
386
387         total_len = iov_length(iov, nr_segs);
388         /* Null read succeeds. */
389         if (unlikely(total_len == 0))
390                 return 0;
391
392         do_wakeup = 0;
393         ret = 0;
394         __pipe_lock(pipe);
395         for (;;) {
396                 int bufs = pipe->nrbufs;
397                 if (bufs) {
398                         int curbuf = pipe->curbuf;
399                         struct pipe_buffer *buf = pipe->bufs + curbuf;
400                         const struct pipe_buf_operations *ops = buf->ops;
401                         void *addr;
402                         size_t chars = buf->len, remaining;
403                         int error, atomic;
404
405                         if (chars > total_len)
406                                 chars = total_len;
407
408                         error = ops->confirm(pipe, buf);
409                         if (error) {
410                                 if (!ret)
411                                         ret = error;
412                                 break;
413                         }
414
415                         atomic = !iov_fault_in_pages_write(iov, chars);
416                         remaining = chars;
417 redo:
418                         addr = ops->map(pipe, buf, atomic);
419                         error = pipe_iov_copy_to_user(iov, addr, &buf->offset,
420                                                       &remaining, atomic);
421                         ops->unmap(pipe, buf, addr);
422                         if (unlikely(error)) {
423                                 /*
424                                  * Just retry with the slow path if we failed.
425                                  */
426                                 if (atomic) {
427                                         atomic = 0;
428                                         goto redo;
429                                 }
430                                 if (!ret)
431                                         ret = error;
432                                 break;
433                         }
434                         ret += chars;
435                         buf->len -= chars;
436
437                         /* Was it a packet buffer? Clean up and exit */
438                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
439                                 total_len = chars;
440                                 buf->len = 0;
441                         }
442
443                         if (!buf->len) {
444                                 buf->ops = NULL;
445                                 ops->release(pipe, buf);
446                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
447                                 pipe->curbuf = curbuf;
448                                 pipe->nrbufs = --bufs;
449                                 do_wakeup = 1;
450                         }
451                         total_len -= chars;
452                         if (!total_len)
453                                 break;  /* common path: read succeeded */
454                 }
455                 if (bufs)       /* More to do? */
456                         continue;
457                 if (!pipe->writers)
458                         break;
459                 if (!pipe->waiting_writers) {
460                         /* syscall merging: Usually we must not sleep
461                          * if O_NONBLOCK is set, or if we got some data.
462                          * But if a writer sleeps in kernel space, then
463                          * we can wait for that data without violating POSIX.
464                          */
465                         if (ret)
466                                 break;
467                         if (filp->f_flags & O_NONBLOCK) {
468                                 ret = -EAGAIN;
469                                 break;
470                         }
471                 }
472                 if (signal_pending(current)) {
473                         if (!ret)
474                                 ret = -ERESTARTSYS;
475                         break;
476                 }
477                 if (do_wakeup) {
478                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
479                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
480                 }
481                 pipe_wait(pipe);
482         }
483         __pipe_unlock(pipe);
484
485         /* Signal writers asynchronously that there is more room. */
486         if (do_wakeup) {
487                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
488                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
489         }
490         if (ret > 0)
491                 file_accessed(filp);
492         return ret;
493 }
494
495 static inline int is_packetized(struct file *file)
496 {
497         return (file->f_flags & O_DIRECT) != 0;
498 }
499
500 static ssize_t
501 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
502             unsigned long nr_segs, loff_t ppos)
503 {
504         struct file *filp = iocb->ki_filp;
505         struct pipe_inode_info *pipe = filp->private_data;
506         ssize_t ret;
507         int do_wakeup;
508         struct iovec *iov = (struct iovec *)_iov;
509         size_t total_len;
510         ssize_t chars;
511
512         total_len = iov_length(iov, nr_segs);
513         /* Null write succeeds. */
514         if (unlikely(total_len == 0))
515                 return 0;
516
517         do_wakeup = 0;
518         ret = 0;
519         __pipe_lock(pipe);
520
521         if (!pipe->readers) {
522                 send_sig(SIGPIPE, current, 0);
523                 ret = -EPIPE;
524                 goto out;
525         }
526
527         /* We try to merge small writes */
528         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
529         if (pipe->nrbufs && chars != 0) {
530                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
531                                                         (pipe->buffers - 1);
532                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
533                 const struct pipe_buf_operations *ops = buf->ops;
534                 int offset = buf->offset + buf->len;
535
536                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
537                         int error, atomic = 1;
538                         void *addr;
539                         size_t remaining = chars;
540
541                         error = ops->confirm(pipe, buf);
542                         if (error)
543                                 goto out;
544
545                         iov_fault_in_pages_read(iov, chars);
546 redo1:
547                         addr = ops->map(pipe, buf, atomic);
548                         error = pipe_iov_copy_from_user(addr, &offset, iov,
549                                                         &remaining, atomic);
550                         ops->unmap(pipe, buf, addr);
551                         ret = error;
552                         do_wakeup = 1;
553                         if (error) {
554                                 if (atomic) {
555                                         atomic = 0;
556                                         goto redo1;
557                                 }
558                                 goto out;
559                         }
560                         buf->len += chars;
561                         total_len -= chars;
562                         ret = chars;
563                         if (!total_len)
564                                 goto out;
565                 }
566         }
567
568         for (;;) {
569                 int bufs;
570
571                 if (!pipe->readers) {
572                         send_sig(SIGPIPE, current, 0);
573                         if (!ret)
574                                 ret = -EPIPE;
575                         break;
576                 }
577                 bufs = pipe->nrbufs;
578                 if (bufs < pipe->buffers) {
579                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
580                         struct pipe_buffer *buf = pipe->bufs + newbuf;
581                         struct page *page = pipe->tmp_page;
582                         char *src;
583                         int error, atomic = 1;
584                         int offset = 0;
585                         size_t remaining;
586
587                         if (!page) {
588                                 page = alloc_page(GFP_HIGHUSER);
589                                 if (unlikely(!page)) {
590                                         ret = ret ? : -ENOMEM;
591                                         break;
592                                 }
593                                 pipe->tmp_page = page;
594                         }
595                         /* Always wake up, even if the copy fails. Otherwise
596                          * we lock up (O_NONBLOCK-)readers that sleep due to
597                          * syscall merging.
598                          * FIXME! Is this really true?
599                          */
600                         do_wakeup = 1;
601                         chars = PAGE_SIZE;
602                         if (chars > total_len)
603                                 chars = total_len;
604
605                         iov_fault_in_pages_read(iov, chars);
606                         remaining = chars;
607 redo2:
608                         if (atomic)
609                                 src = kmap_atomic(page);
610                         else
611                                 src = kmap(page);
612
613                         error = pipe_iov_copy_from_user(src, &offset, iov,
614                                                         &remaining, atomic);
615                         if (atomic)
616                                 kunmap_atomic(src);
617                         else
618                                 kunmap(page);
619
620                         if (unlikely(error)) {
621                                 if (atomic) {
622                                         atomic = 0;
623                                         goto redo2;
624                                 }
625                                 if (!ret)
626                                         ret = error;
627                                 break;
628                         }
629                         ret += chars;
630
631                         /* Insert it into the buffer array */
632                         buf->page = page;
633                         buf->ops = &anon_pipe_buf_ops;
634                         buf->offset = 0;
635                         buf->len = chars;
636                         buf->flags = 0;
637                         if (is_packetized(filp)) {
638                                 buf->ops = &packet_pipe_buf_ops;
639                                 buf->flags = PIPE_BUF_FLAG_PACKET;
640                         }
641                         pipe->nrbufs = ++bufs;
642                         pipe->tmp_page = NULL;
643
644                         total_len -= chars;
645                         if (!total_len)
646                                 break;
647                 }
648                 if (bufs < pipe->buffers)
649                         continue;
650                 if (filp->f_flags & O_NONBLOCK) {
651                         if (!ret)
652                                 ret = -EAGAIN;
653                         break;
654                 }
655                 if (signal_pending(current)) {
656                         if (!ret)
657                                 ret = -ERESTARTSYS;
658                         break;
659                 }
660                 if (do_wakeup) {
661                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
662                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
663                         do_wakeup = 0;
664                 }
665                 pipe->waiting_writers++;
666                 pipe_wait(pipe);
667                 pipe->waiting_writers--;
668         }
669 out:
670         __pipe_unlock(pipe);
671         if (do_wakeup) {
672                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
673                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
674         }
675         if (ret > 0) {
676                 int err = file_update_time(filp);
677                 if (err)
678                         ret = err;
679         }
680         return ret;
681 }
682
683 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
684 {
685         struct pipe_inode_info *pipe = filp->private_data;
686         int count, buf, nrbufs;
687
688         switch (cmd) {
689                 case FIONREAD:
690                         __pipe_lock(pipe);
691                         count = 0;
692                         buf = pipe->curbuf;
693                         nrbufs = pipe->nrbufs;
694                         while (--nrbufs >= 0) {
695                                 count += pipe->bufs[buf].len;
696                                 buf = (buf+1) & (pipe->buffers - 1);
697                         }
698                         __pipe_unlock(pipe);
699
700                         return put_user(count, (int __user *)arg);
701                 default:
702                         return -ENOIOCTLCMD;
703         }
704 }
705
706 /* No kernel lock held - fine */
707 static unsigned int
708 pipe_poll(struct file *filp, poll_table *wait)
709 {
710         unsigned int mask;
711         struct pipe_inode_info *pipe = filp->private_data;
712         int nrbufs;
713
714         poll_wait(filp, &pipe->wait, wait);
715
716         /* Reading only -- no need for acquiring the semaphore.  */
717         nrbufs = pipe->nrbufs;
718         mask = 0;
719         if (filp->f_mode & FMODE_READ) {
720                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
721                 if (!pipe->writers && filp->f_version != pipe->w_counter)
722                         mask |= POLLHUP;
723         }
724
725         if (filp->f_mode & FMODE_WRITE) {
726                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
727                 /*
728                  * Most Unices do not set POLLERR for FIFOs but on Linux they
729                  * behave exactly like pipes for poll().
730                  */
731                 if (!pipe->readers)
732                         mask |= POLLERR;
733         }
734
735         return mask;
736 }
737
738 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
739 {
740         int kill = 0;
741
742         spin_lock(&inode->i_lock);
743         if (!--pipe->files) {
744                 inode->i_pipe = NULL;
745                 kill = 1;
746         }
747         spin_unlock(&inode->i_lock);
748
749         if (kill)
750                 free_pipe_info(pipe);
751 }
752
753 static int
754 pipe_release(struct inode *inode, struct file *file)
755 {
756         struct pipe_inode_info *pipe = file->private_data;
757
758         __pipe_lock(pipe);
759         if (file->f_mode & FMODE_READ)
760                 pipe->readers--;
761         if (file->f_mode & FMODE_WRITE)
762                 pipe->writers--;
763
764         if (pipe->readers || pipe->writers) {
765                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
766                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
767                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
768         }
769         __pipe_unlock(pipe);
770
771         put_pipe_info(inode, pipe);
772         return 0;
773 }
774
775 static int
776 pipe_fasync(int fd, struct file *filp, int on)
777 {
778         struct pipe_inode_info *pipe = filp->private_data;
779         int retval = 0;
780
781         __pipe_lock(pipe);
782         if (filp->f_mode & FMODE_READ)
783                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
784         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
785                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
786                 if (retval < 0 && (filp->f_mode & FMODE_READ))
787                         /* this can happen only if on == T */
788                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
789         }
790         __pipe_unlock(pipe);
791         return retval;
792 }
793
794 struct pipe_inode_info *alloc_pipe_info(void)
795 {
796         struct pipe_inode_info *pipe;
797
798         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
799         if (pipe) {
800                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
801                 if (pipe->bufs) {
802                         init_waitqueue_head(&pipe->wait);
803                         pipe->r_counter = pipe->w_counter = 1;
804                         pipe->buffers = PIPE_DEF_BUFFERS;
805                         mutex_init(&pipe->mutex);
806                         return pipe;
807                 }
808                 kfree(pipe);
809         }
810
811         return NULL;
812 }
813
814 void free_pipe_info(struct pipe_inode_info *pipe)
815 {
816         int i;
817
818         for (i = 0; i < pipe->buffers; i++) {
819                 struct pipe_buffer *buf = pipe->bufs + i;
820                 if (buf->ops)
821                         buf->ops->release(pipe, buf);
822         }
823         if (pipe->tmp_page)
824                 __free_page(pipe->tmp_page);
825         kfree(pipe->bufs);
826         kfree(pipe);
827 }
828
829 static struct vfsmount *pipe_mnt __read_mostly;
830
831 /*
832  * pipefs_dname() is called from d_path().
833  */
834 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
835 {
836         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
837                                 dentry->d_inode->i_ino);
838 }
839
840 static const struct dentry_operations pipefs_dentry_operations = {
841         .d_dname        = pipefs_dname,
842 };
843
844 static struct inode * get_pipe_inode(void)
845 {
846         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
847         struct pipe_inode_info *pipe;
848
849         if (!inode)
850                 goto fail_inode;
851
852         inode->i_ino = get_next_ino();
853
854         pipe = alloc_pipe_info();
855         if (!pipe)
856                 goto fail_iput;
857
858         inode->i_pipe = pipe;
859         pipe->files = 2;
860         pipe->readers = pipe->writers = 1;
861         inode->i_fop = &pipefifo_fops;
862
863         /*
864          * Mark the inode dirty from the very beginning,
865          * that way it will never be moved to the dirty
866          * list because "mark_inode_dirty()" will think
867          * that it already _is_ on the dirty list.
868          */
869         inode->i_state = I_DIRTY;
870         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
871         inode->i_uid = current_fsuid();
872         inode->i_gid = current_fsgid();
873         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
874
875         return inode;
876
877 fail_iput:
878         iput(inode);
879
880 fail_inode:
881         return NULL;
882 }
883
884 int create_pipe_files(struct file **res, int flags)
885 {
886         int err;
887         struct inode *inode = get_pipe_inode();
888         struct file *f;
889         struct path path;
890         static struct qstr name = { .name = "" };
891
892         if (!inode)
893                 return -ENFILE;
894
895         err = -ENOMEM;
896         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
897         if (!path.dentry)
898                 goto err_inode;
899         path.mnt = mntget(pipe_mnt);
900
901         d_instantiate(path.dentry, inode);
902
903         err = -ENFILE;
904         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
905         if (IS_ERR(f))
906                 goto err_dentry;
907
908         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
909         f->private_data = inode->i_pipe;
910
911         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
912         if (IS_ERR(res[0]))
913                 goto err_file;
914
915         path_get(&path);
916         res[0]->private_data = inode->i_pipe;
917         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
918         res[1] = f;
919         return 0;
920
921 err_file:
922         put_filp(f);
923 err_dentry:
924         free_pipe_info(inode->i_pipe);
925         path_put(&path);
926         return err;
927
928 err_inode:
929         free_pipe_info(inode->i_pipe);
930         iput(inode);
931         return err;
932 }
933
934 static int __do_pipe_flags(int *fd, struct file **files, int flags)
935 {
936         int error;
937         int fdw, fdr;
938
939         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
940                 return -EINVAL;
941
942         error = create_pipe_files(files, flags);
943         if (error)
944                 return error;
945
946         error = get_unused_fd_flags(flags);
947         if (error < 0)
948                 goto err_read_pipe;
949         fdr = error;
950
951         error = get_unused_fd_flags(flags);
952         if (error < 0)
953                 goto err_fdr;
954         fdw = error;
955
956         audit_fd_pair(fdr, fdw);
957         fd[0] = fdr;
958         fd[1] = fdw;
959         return 0;
960
961  err_fdr:
962         put_unused_fd(fdr);
963  err_read_pipe:
964         fput(files[0]);
965         fput(files[1]);
966         return error;
967 }
968
969 int do_pipe_flags(int *fd, int flags)
970 {
971         struct file *files[2];
972         int error = __do_pipe_flags(fd, files, flags);
973         if (!error) {
974                 fd_install(fd[0], files[0]);
975                 fd_install(fd[1], files[1]);
976         }
977         return error;
978 }
979
980 /*
981  * sys_pipe() is the normal C calling standard for creating
982  * a pipe. It's not the way Unix traditionally does this, though.
983  */
984 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
985 {
986         struct file *files[2];
987         int fd[2];
988         int error;
989
990         error = __do_pipe_flags(fd, files, flags);
991         if (!error) {
992                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
993                         fput(files[0]);
994                         fput(files[1]);
995                         put_unused_fd(fd[0]);
996                         put_unused_fd(fd[1]);
997                         error = -EFAULT;
998                 } else {
999                         fd_install(fd[0], files[0]);
1000                         fd_install(fd[1], files[1]);
1001                 }
1002         }
1003         return error;
1004 }
1005
1006 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1007 {
1008         return sys_pipe2(fildes, 0);
1009 }
1010
1011 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1012 {
1013         int cur = *cnt; 
1014
1015         while (cur == *cnt) {
1016                 pipe_wait(pipe);
1017                 if (signal_pending(current))
1018                         break;
1019         }
1020         return cur == *cnt ? -ERESTARTSYS : 0;
1021 }
1022
1023 static void wake_up_partner(struct pipe_inode_info *pipe)
1024 {
1025         wake_up_interruptible(&pipe->wait);
1026 }
1027
1028 static int fifo_open(struct inode *inode, struct file *filp)
1029 {
1030         struct pipe_inode_info *pipe;
1031         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1032         int ret;
1033
1034         filp->f_version = 0;
1035
1036         spin_lock(&inode->i_lock);
1037         if (inode->i_pipe) {
1038                 pipe = inode->i_pipe;
1039                 pipe->files++;
1040                 spin_unlock(&inode->i_lock);
1041         } else {
1042                 spin_unlock(&inode->i_lock);
1043                 pipe = alloc_pipe_info();
1044                 if (!pipe)
1045                         return -ENOMEM;
1046                 pipe->files = 1;
1047                 spin_lock(&inode->i_lock);
1048                 if (unlikely(inode->i_pipe)) {
1049                         inode->i_pipe->files++;
1050                         spin_unlock(&inode->i_lock);
1051                         free_pipe_info(pipe);
1052                         pipe = inode->i_pipe;
1053                 } else {
1054                         inode->i_pipe = pipe;
1055                         spin_unlock(&inode->i_lock);
1056                 }
1057         }
1058         filp->private_data = pipe;
1059         /* OK, we have a pipe and it's pinned down */
1060
1061         __pipe_lock(pipe);
1062
1063         /* We can only do regular read/write on fifos */
1064         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
1065
1066         switch (filp->f_mode) {
1067         case FMODE_READ:
1068         /*
1069          *  O_RDONLY
1070          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1071          *  opened, even when there is no process writing the FIFO.
1072          */
1073                 pipe->r_counter++;
1074                 if (pipe->readers++ == 0)
1075                         wake_up_partner(pipe);
1076
1077                 if (!is_pipe && !pipe->writers) {
1078                         if ((filp->f_flags & O_NONBLOCK)) {
1079                                 /* suppress POLLHUP until we have
1080                                  * seen a writer */
1081                                 filp->f_version = pipe->w_counter;
1082                         } else {
1083                                 if (wait_for_partner(pipe, &pipe->w_counter))
1084                                         goto err_rd;
1085                         }
1086                 }
1087                 break;
1088         
1089         case FMODE_WRITE:
1090         /*
1091          *  O_WRONLY
1092          *  POSIX.1 says that O_NONBLOCK means return -1 with
1093          *  errno=ENXIO when there is no process reading the FIFO.
1094          */
1095                 ret = -ENXIO;
1096                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1097                         goto err;
1098
1099                 pipe->w_counter++;
1100                 if (!pipe->writers++)
1101                         wake_up_partner(pipe);
1102
1103                 if (!is_pipe && !pipe->readers) {
1104                         if (wait_for_partner(pipe, &pipe->r_counter))
1105                                 goto err_wr;
1106                 }
1107                 break;
1108         
1109         case FMODE_READ | FMODE_WRITE:
1110         /*
1111          *  O_RDWR
1112          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1113          *  This implementation will NEVER block on a O_RDWR open, since
1114          *  the process can at least talk to itself.
1115          */
1116
1117                 pipe->readers++;
1118                 pipe->writers++;
1119                 pipe->r_counter++;
1120                 pipe->w_counter++;
1121                 if (pipe->readers == 1 || pipe->writers == 1)
1122                         wake_up_partner(pipe);
1123                 break;
1124
1125         default:
1126                 ret = -EINVAL;
1127                 goto err;
1128         }
1129
1130         /* Ok! */
1131         __pipe_unlock(pipe);
1132         return 0;
1133
1134 err_rd:
1135         if (!--pipe->readers)
1136                 wake_up_interruptible(&pipe->wait);
1137         ret = -ERESTARTSYS;
1138         goto err;
1139
1140 err_wr:
1141         if (!--pipe->writers)
1142                 wake_up_interruptible(&pipe->wait);
1143         ret = -ERESTARTSYS;
1144         goto err;
1145
1146 err:
1147         __pipe_unlock(pipe);
1148
1149         put_pipe_info(inode, pipe);
1150         return ret;
1151 }
1152
1153 const struct file_operations pipefifo_fops = {
1154         .open           = fifo_open,
1155         .llseek         = no_llseek,
1156         .read           = do_sync_read,
1157         .aio_read       = pipe_read,
1158         .write          = do_sync_write,
1159         .aio_write      = pipe_write,
1160         .poll           = pipe_poll,
1161         .unlocked_ioctl = pipe_ioctl,
1162         .release        = pipe_release,
1163         .fasync         = pipe_fasync,
1164 };
1165
1166 /*
1167  * Allocate a new array of pipe buffers and copy the info over. Returns the
1168  * pipe size if successful, or return -ERROR on error.
1169  */
1170 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1171 {
1172         struct pipe_buffer *bufs;
1173
1174         /*
1175          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1176          * expect a lot of shrink+grow operations, just free and allocate
1177          * again like we would do for growing. If the pipe currently
1178          * contains more buffers than arg, then return busy.
1179          */
1180         if (nr_pages < pipe->nrbufs)
1181                 return -EBUSY;
1182
1183         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1184         if (unlikely(!bufs))
1185                 return -ENOMEM;
1186
1187         /*
1188          * The pipe array wraps around, so just start the new one at zero
1189          * and adjust the indexes.
1190          */
1191         if (pipe->nrbufs) {
1192                 unsigned int tail;
1193                 unsigned int head;
1194
1195                 tail = pipe->curbuf + pipe->nrbufs;
1196                 if (tail < pipe->buffers)
1197                         tail = 0;
1198                 else
1199                         tail &= (pipe->buffers - 1);
1200
1201                 head = pipe->nrbufs - tail;
1202                 if (head)
1203                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1204                 if (tail)
1205                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1206         }
1207
1208         pipe->curbuf = 0;
1209         kfree(pipe->bufs);
1210         pipe->bufs = bufs;
1211         pipe->buffers = nr_pages;
1212         return nr_pages * PAGE_SIZE;
1213 }
1214
1215 /*
1216  * Currently we rely on the pipe array holding a power-of-2 number
1217  * of pages.
1218  */
1219 static inline unsigned int round_pipe_size(unsigned int size)
1220 {
1221         unsigned long nr_pages;
1222
1223         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1224         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1225 }
1226
1227 /*
1228  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1229  * will return an error.
1230  */
1231 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1232                  size_t *lenp, loff_t *ppos)
1233 {
1234         int ret;
1235
1236         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1237         if (ret < 0 || !write)
1238                 return ret;
1239
1240         pipe_max_size = round_pipe_size(pipe_max_size);
1241         return ret;
1242 }
1243
1244 /*
1245  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1246  * location, so checking ->i_pipe is not enough to verify that this is a
1247  * pipe.
1248  */
1249 struct pipe_inode_info *get_pipe_info(struct file *file)
1250 {
1251         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1252 }
1253
1254 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1255 {
1256         struct pipe_inode_info *pipe;
1257         long ret;
1258
1259         pipe = get_pipe_info(file);
1260         if (!pipe)
1261                 return -EBADF;
1262
1263         __pipe_lock(pipe);
1264
1265         switch (cmd) {
1266         case F_SETPIPE_SZ: {
1267                 unsigned int size, nr_pages;
1268
1269                 size = round_pipe_size(arg);
1270                 nr_pages = size >> PAGE_SHIFT;
1271
1272                 ret = -EINVAL;
1273                 if (!nr_pages)
1274                         goto out;
1275
1276                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1277                         ret = -EPERM;
1278                         goto out;
1279                 }
1280                 ret = pipe_set_size(pipe, nr_pages);
1281                 break;
1282                 }
1283         case F_GETPIPE_SZ:
1284                 ret = pipe->buffers * PAGE_SIZE;
1285                 break;
1286         default:
1287                 ret = -EINVAL;
1288                 break;
1289         }
1290
1291 out:
1292         __pipe_unlock(pipe);
1293         return ret;
1294 }
1295
1296 static const struct super_operations pipefs_ops = {
1297         .destroy_inode = free_inode_nonrcu,
1298         .statfs = simple_statfs,
1299 };
1300
1301 /*
1302  * pipefs should _never_ be mounted by userland - too much of security hassle,
1303  * no real gain from having the whole whorehouse mounted. So we don't need
1304  * any operations on the root directory. However, we need a non-trivial
1305  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1306  */
1307 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1308                          int flags, const char *dev_name, void *data)
1309 {
1310         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1311                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1312 }
1313
1314 static struct file_system_type pipe_fs_type = {
1315         .name           = "pipefs",
1316         .mount          = pipefs_mount,
1317         .kill_sb        = kill_anon_super,
1318 };
1319
1320 static int __init init_pipe_fs(void)
1321 {
1322         int err = register_filesystem(&pipe_fs_type);
1323
1324         if (!err) {
1325                 pipe_mnt = kern_mount(&pipe_fs_type);
1326                 if (IS_ERR(pipe_mnt)) {
1327                         err = PTR_ERR(pipe_mnt);
1328                         unregister_filesystem(&pipe_fs_type);
1329                 }
1330         }
1331         return err;
1332 }
1333
1334 fs_initcall(init_pipe_fs);