arm64: alternative: Provide if/else/endif assembler macros
[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                         int offset;
405
406                         if (chars > total_len)
407                                 chars = total_len;
408
409                         error = ops->confirm(pipe, buf);
410                         if (error) {
411                                 if (!ret)
412                                         ret = error;
413                                 break;
414                         }
415
416                         atomic = !iov_fault_in_pages_write(iov, chars);
417                         remaining = chars;
418                         offset = buf->offset;
419 redo:
420                         addr = ops->map(pipe, buf, atomic);
421                         error = pipe_iov_copy_to_user(iov, addr, &offset,
422                                                       &remaining, atomic);
423                         ops->unmap(pipe, buf, addr);
424                         if (unlikely(error)) {
425                                 /*
426                                  * Just retry with the slow path if we failed.
427                                  */
428                                 if (atomic) {
429                                         atomic = 0;
430                                         goto redo;
431                                 }
432                                 if (!ret)
433                                         ret = error;
434                                 break;
435                         }
436                         ret += chars;
437                         buf->offset += chars;
438                         buf->len -= chars;
439
440                         /* Was it a packet buffer? Clean up and exit */
441                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
442                                 total_len = chars;
443                                 buf->len = 0;
444                         }
445
446                         if (!buf->len) {
447                                 buf->ops = NULL;
448                                 ops->release(pipe, buf);
449                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
450                                 pipe->curbuf = curbuf;
451                                 pipe->nrbufs = --bufs;
452                                 do_wakeup = 1;
453                         }
454                         total_len -= chars;
455                         if (!total_len)
456                                 break;  /* common path: read succeeded */
457                 }
458                 if (bufs)       /* More to do? */
459                         continue;
460                 if (!pipe->writers)
461                         break;
462                 if (!pipe->waiting_writers) {
463                         /* syscall merging: Usually we must not sleep
464                          * if O_NONBLOCK is set, or if we got some data.
465                          * But if a writer sleeps in kernel space, then
466                          * we can wait for that data without violating POSIX.
467                          */
468                         if (ret)
469                                 break;
470                         if (filp->f_flags & O_NONBLOCK) {
471                                 ret = -EAGAIN;
472                                 break;
473                         }
474                 }
475                 if (signal_pending(current)) {
476                         if (!ret)
477                                 ret = -ERESTARTSYS;
478                         break;
479                 }
480                 if (do_wakeup) {
481                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
482                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
483                 }
484                 pipe_wait(pipe);
485         }
486         __pipe_unlock(pipe);
487
488         /* Signal writers asynchronously that there is more room. */
489         if (do_wakeup) {
490                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
491                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
492         }
493         if (ret > 0)
494                 file_accessed(filp);
495         return ret;
496 }
497
498 static inline int is_packetized(struct file *file)
499 {
500         return (file->f_flags & O_DIRECT) != 0;
501 }
502
503 static ssize_t
504 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
505             unsigned long nr_segs, loff_t ppos)
506 {
507         struct file *filp = iocb->ki_filp;
508         struct pipe_inode_info *pipe = filp->private_data;
509         ssize_t ret;
510         int do_wakeup;
511         struct iovec *iov = (struct iovec *)_iov;
512         size_t total_len;
513         ssize_t chars;
514
515         total_len = iov_length(iov, nr_segs);
516         /* Null write succeeds. */
517         if (unlikely(total_len == 0))
518                 return 0;
519
520         do_wakeup = 0;
521         ret = 0;
522         __pipe_lock(pipe);
523
524         if (!pipe->readers) {
525                 send_sig(SIGPIPE, current, 0);
526                 ret = -EPIPE;
527                 goto out;
528         }
529
530         /* We try to merge small writes */
531         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
532         if (pipe->nrbufs && chars != 0) {
533                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
534                                                         (pipe->buffers - 1);
535                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
536                 const struct pipe_buf_operations *ops = buf->ops;
537                 int offset = buf->offset + buf->len;
538
539                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
540                         int error, atomic = 1;
541                         void *addr;
542                         size_t remaining = chars;
543
544                         error = ops->confirm(pipe, buf);
545                         if (error)
546                                 goto out;
547
548                         iov_fault_in_pages_read(iov, chars);
549 redo1:
550                         addr = ops->map(pipe, buf, atomic);
551                         error = pipe_iov_copy_from_user(addr, &offset, iov,
552                                                         &remaining, atomic);
553                         ops->unmap(pipe, buf, addr);
554                         ret = error;
555                         do_wakeup = 1;
556                         if (error) {
557                                 if (atomic) {
558                                         atomic = 0;
559                                         goto redo1;
560                                 }
561                                 goto out;
562                         }
563                         buf->len += chars;
564                         total_len -= chars;
565                         ret = chars;
566                         if (!total_len)
567                                 goto out;
568                 }
569         }
570
571         for (;;) {
572                 int bufs;
573
574                 if (!pipe->readers) {
575                         send_sig(SIGPIPE, current, 0);
576                         if (!ret)
577                                 ret = -EPIPE;
578                         break;
579                 }
580                 bufs = pipe->nrbufs;
581                 if (bufs < pipe->buffers) {
582                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
583                         struct pipe_buffer *buf = pipe->bufs + newbuf;
584                         struct page *page = pipe->tmp_page;
585                         char *src;
586                         int error, atomic = 1;
587                         int offset = 0;
588                         size_t remaining;
589
590                         if (!page) {
591                                 page = alloc_page(GFP_HIGHUSER);
592                                 if (unlikely(!page)) {
593                                         ret = ret ? : -ENOMEM;
594                                         break;
595                                 }
596                                 pipe->tmp_page = page;
597                         }
598                         /* Always wake up, even if the copy fails. Otherwise
599                          * we lock up (O_NONBLOCK-)readers that sleep due to
600                          * syscall merging.
601                          * FIXME! Is this really true?
602                          */
603                         do_wakeup = 1;
604                         chars = PAGE_SIZE;
605                         if (chars > total_len)
606                                 chars = total_len;
607
608                         iov_fault_in_pages_read(iov, chars);
609                         remaining = chars;
610 redo2:
611                         if (atomic)
612                                 src = kmap_atomic(page);
613                         else
614                                 src = kmap(page);
615
616                         error = pipe_iov_copy_from_user(src, &offset, iov,
617                                                         &remaining, atomic);
618                         if (atomic)
619                                 kunmap_atomic(src);
620                         else
621                                 kunmap(page);
622
623                         if (unlikely(error)) {
624                                 if (atomic) {
625                                         atomic = 0;
626                                         goto redo2;
627                                 }
628                                 if (!ret)
629                                         ret = error;
630                                 break;
631                         }
632                         ret += chars;
633
634                         /* Insert it into the buffer array */
635                         buf->page = page;
636                         buf->ops = &anon_pipe_buf_ops;
637                         buf->offset = 0;
638                         buf->len = chars;
639                         buf->flags = 0;
640                         if (is_packetized(filp)) {
641                                 buf->ops = &packet_pipe_buf_ops;
642                                 buf->flags = PIPE_BUF_FLAG_PACKET;
643                         }
644                         pipe->nrbufs = ++bufs;
645                         pipe->tmp_page = NULL;
646
647                         total_len -= chars;
648                         if (!total_len)
649                                 break;
650                 }
651                 if (bufs < pipe->buffers)
652                         continue;
653                 if (filp->f_flags & O_NONBLOCK) {
654                         if (!ret)
655                                 ret = -EAGAIN;
656                         break;
657                 }
658                 if (signal_pending(current)) {
659                         if (!ret)
660                                 ret = -ERESTARTSYS;
661                         break;
662                 }
663                 if (do_wakeup) {
664                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
665                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
666                         do_wakeup = 0;
667                 }
668                 pipe->waiting_writers++;
669                 pipe_wait(pipe);
670                 pipe->waiting_writers--;
671         }
672 out:
673         __pipe_unlock(pipe);
674         if (do_wakeup) {
675                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
676                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
677         }
678         if (ret > 0) {
679                 int err = file_update_time(filp);
680                 if (err)
681                         ret = err;
682         }
683         return ret;
684 }
685
686 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
687 {
688         struct pipe_inode_info *pipe = filp->private_data;
689         int count, buf, nrbufs;
690
691         switch (cmd) {
692                 case FIONREAD:
693                         __pipe_lock(pipe);
694                         count = 0;
695                         buf = pipe->curbuf;
696                         nrbufs = pipe->nrbufs;
697                         while (--nrbufs >= 0) {
698                                 count += pipe->bufs[buf].len;
699                                 buf = (buf+1) & (pipe->buffers - 1);
700                         }
701                         __pipe_unlock(pipe);
702
703                         return put_user(count, (int __user *)arg);
704                 default:
705                         return -ENOIOCTLCMD;
706         }
707 }
708
709 /* No kernel lock held - fine */
710 static unsigned int
711 pipe_poll(struct file *filp, poll_table *wait)
712 {
713         unsigned int mask;
714         struct pipe_inode_info *pipe = filp->private_data;
715         int nrbufs;
716
717         poll_wait(filp, &pipe->wait, wait);
718
719         /* Reading only -- no need for acquiring the semaphore.  */
720         nrbufs = pipe->nrbufs;
721         mask = 0;
722         if (filp->f_mode & FMODE_READ) {
723                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
724                 if (!pipe->writers && filp->f_version != pipe->w_counter)
725                         mask |= POLLHUP;
726         }
727
728         if (filp->f_mode & FMODE_WRITE) {
729                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
730                 /*
731                  * Most Unices do not set POLLERR for FIFOs but on Linux they
732                  * behave exactly like pipes for poll().
733                  */
734                 if (!pipe->readers)
735                         mask |= POLLERR;
736         }
737
738         return mask;
739 }
740
741 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
742 {
743         int kill = 0;
744
745         spin_lock(&inode->i_lock);
746         if (!--pipe->files) {
747                 inode->i_pipe = NULL;
748                 kill = 1;
749         }
750         spin_unlock(&inode->i_lock);
751
752         if (kill)
753                 free_pipe_info(pipe);
754 }
755
756 static int
757 pipe_release(struct inode *inode, struct file *file)
758 {
759         struct pipe_inode_info *pipe = file->private_data;
760
761         __pipe_lock(pipe);
762         if (file->f_mode & FMODE_READ)
763                 pipe->readers--;
764         if (file->f_mode & FMODE_WRITE)
765                 pipe->writers--;
766
767         if (pipe->readers || pipe->writers) {
768                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
769                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
770                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
771         }
772         __pipe_unlock(pipe);
773
774         put_pipe_info(inode, pipe);
775         return 0;
776 }
777
778 static int
779 pipe_fasync(int fd, struct file *filp, int on)
780 {
781         struct pipe_inode_info *pipe = filp->private_data;
782         int retval = 0;
783
784         __pipe_lock(pipe);
785         if (filp->f_mode & FMODE_READ)
786                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
787         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
788                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
789                 if (retval < 0 && (filp->f_mode & FMODE_READ))
790                         /* this can happen only if on == T */
791                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
792         }
793         __pipe_unlock(pipe);
794         return retval;
795 }
796
797 struct pipe_inode_info *alloc_pipe_info(void)
798 {
799         struct pipe_inode_info *pipe;
800
801         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
802         if (pipe) {
803                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
804                 if (pipe->bufs) {
805                         init_waitqueue_head(&pipe->wait);
806                         pipe->r_counter = pipe->w_counter = 1;
807                         pipe->buffers = PIPE_DEF_BUFFERS;
808                         mutex_init(&pipe->mutex);
809                         return pipe;
810                 }
811                 kfree(pipe);
812         }
813
814         return NULL;
815 }
816
817 void free_pipe_info(struct pipe_inode_info *pipe)
818 {
819         int i;
820
821         for (i = 0; i < pipe->buffers; i++) {
822                 struct pipe_buffer *buf = pipe->bufs + i;
823                 if (buf->ops)
824                         buf->ops->release(pipe, buf);
825         }
826         if (pipe->tmp_page)
827                 __free_page(pipe->tmp_page);
828         kfree(pipe->bufs);
829         kfree(pipe);
830 }
831
832 static struct vfsmount *pipe_mnt __read_mostly;
833
834 /*
835  * pipefs_dname() is called from d_path().
836  */
837 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
838 {
839         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
840                                 dentry->d_inode->i_ino);
841 }
842
843 static const struct dentry_operations pipefs_dentry_operations = {
844         .d_dname        = pipefs_dname,
845 };
846
847 static struct inode * get_pipe_inode(void)
848 {
849         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
850         struct pipe_inode_info *pipe;
851
852         if (!inode)
853                 goto fail_inode;
854
855         inode->i_ino = get_next_ino();
856
857         pipe = alloc_pipe_info();
858         if (!pipe)
859                 goto fail_iput;
860
861         inode->i_pipe = pipe;
862         pipe->files = 2;
863         pipe->readers = pipe->writers = 1;
864         inode->i_fop = &pipefifo_fops;
865
866         /*
867          * Mark the inode dirty from the very beginning,
868          * that way it will never be moved to the dirty
869          * list because "mark_inode_dirty()" will think
870          * that it already _is_ on the dirty list.
871          */
872         inode->i_state = I_DIRTY;
873         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
874         inode->i_uid = current_fsuid();
875         inode->i_gid = current_fsgid();
876         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
877
878         return inode;
879
880 fail_iput:
881         iput(inode);
882
883 fail_inode:
884         return NULL;
885 }
886
887 int create_pipe_files(struct file **res, int flags)
888 {
889         int err;
890         struct inode *inode = get_pipe_inode();
891         struct file *f;
892         struct path path;
893         static struct qstr name = { .name = "" };
894
895         if (!inode)
896                 return -ENFILE;
897
898         err = -ENOMEM;
899         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
900         if (!path.dentry)
901                 goto err_inode;
902         path.mnt = mntget(pipe_mnt);
903
904         d_instantiate(path.dentry, inode);
905
906         err = -ENFILE;
907         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
908         if (IS_ERR(f))
909                 goto err_dentry;
910
911         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
912         f->private_data = inode->i_pipe;
913
914         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
915         if (IS_ERR(res[0]))
916                 goto err_file;
917
918         path_get(&path);
919         res[0]->private_data = inode->i_pipe;
920         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
921         res[1] = f;
922         return 0;
923
924 err_file:
925         put_filp(f);
926 err_dentry:
927         free_pipe_info(inode->i_pipe);
928         path_put(&path);
929         return err;
930
931 err_inode:
932         free_pipe_info(inode->i_pipe);
933         iput(inode);
934         return err;
935 }
936
937 static int __do_pipe_flags(int *fd, struct file **files, int flags)
938 {
939         int error;
940         int fdw, fdr;
941
942         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
943                 return -EINVAL;
944
945         error = create_pipe_files(files, flags);
946         if (error)
947                 return error;
948
949         error = get_unused_fd_flags(flags);
950         if (error < 0)
951                 goto err_read_pipe;
952         fdr = error;
953
954         error = get_unused_fd_flags(flags);
955         if (error < 0)
956                 goto err_fdr;
957         fdw = error;
958
959         audit_fd_pair(fdr, fdw);
960         fd[0] = fdr;
961         fd[1] = fdw;
962         return 0;
963
964  err_fdr:
965         put_unused_fd(fdr);
966  err_read_pipe:
967         fput(files[0]);
968         fput(files[1]);
969         return error;
970 }
971
972 int do_pipe_flags(int *fd, int flags)
973 {
974         struct file *files[2];
975         int error = __do_pipe_flags(fd, files, flags);
976         if (!error) {
977                 fd_install(fd[0], files[0]);
978                 fd_install(fd[1], files[1]);
979         }
980         return error;
981 }
982
983 /*
984  * sys_pipe() is the normal C calling standard for creating
985  * a pipe. It's not the way Unix traditionally does this, though.
986  */
987 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
988 {
989         struct file *files[2];
990         int fd[2];
991         int error;
992
993         error = __do_pipe_flags(fd, files, flags);
994         if (!error) {
995                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
996                         fput(files[0]);
997                         fput(files[1]);
998                         put_unused_fd(fd[0]);
999                         put_unused_fd(fd[1]);
1000                         error = -EFAULT;
1001                 } else {
1002                         fd_install(fd[0], files[0]);
1003                         fd_install(fd[1], files[1]);
1004                 }
1005         }
1006         return error;
1007 }
1008
1009 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1010 {
1011         return sys_pipe2(fildes, 0);
1012 }
1013
1014 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1015 {
1016         int cur = *cnt; 
1017
1018         while (cur == *cnt) {
1019                 pipe_wait(pipe);
1020                 if (signal_pending(current))
1021                         break;
1022         }
1023         return cur == *cnt ? -ERESTARTSYS : 0;
1024 }
1025
1026 static void wake_up_partner(struct pipe_inode_info *pipe)
1027 {
1028         wake_up_interruptible(&pipe->wait);
1029 }
1030
1031 static int fifo_open(struct inode *inode, struct file *filp)
1032 {
1033         struct pipe_inode_info *pipe;
1034         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1035         int ret;
1036
1037         filp->f_version = 0;
1038
1039         spin_lock(&inode->i_lock);
1040         if (inode->i_pipe) {
1041                 pipe = inode->i_pipe;
1042                 pipe->files++;
1043                 spin_unlock(&inode->i_lock);
1044         } else {
1045                 spin_unlock(&inode->i_lock);
1046                 pipe = alloc_pipe_info();
1047                 if (!pipe)
1048                         return -ENOMEM;
1049                 pipe->files = 1;
1050                 spin_lock(&inode->i_lock);
1051                 if (unlikely(inode->i_pipe)) {
1052                         inode->i_pipe->files++;
1053                         spin_unlock(&inode->i_lock);
1054                         free_pipe_info(pipe);
1055                         pipe = inode->i_pipe;
1056                 } else {
1057                         inode->i_pipe = pipe;
1058                         spin_unlock(&inode->i_lock);
1059                 }
1060         }
1061         filp->private_data = pipe;
1062         /* OK, we have a pipe and it's pinned down */
1063
1064         __pipe_lock(pipe);
1065
1066         /* We can only do regular read/write on fifos */
1067         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
1068
1069         switch (filp->f_mode) {
1070         case FMODE_READ:
1071         /*
1072          *  O_RDONLY
1073          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1074          *  opened, even when there is no process writing the FIFO.
1075          */
1076                 pipe->r_counter++;
1077                 if (pipe->readers++ == 0)
1078                         wake_up_partner(pipe);
1079
1080                 if (!is_pipe && !pipe->writers) {
1081                         if ((filp->f_flags & O_NONBLOCK)) {
1082                                 /* suppress POLLHUP until we have
1083                                  * seen a writer */
1084                                 filp->f_version = pipe->w_counter;
1085                         } else {
1086                                 if (wait_for_partner(pipe, &pipe->w_counter))
1087                                         goto err_rd;
1088                         }
1089                 }
1090                 break;
1091         
1092         case FMODE_WRITE:
1093         /*
1094          *  O_WRONLY
1095          *  POSIX.1 says that O_NONBLOCK means return -1 with
1096          *  errno=ENXIO when there is no process reading the FIFO.
1097          */
1098                 ret = -ENXIO;
1099                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1100                         goto err;
1101
1102                 pipe->w_counter++;
1103                 if (!pipe->writers++)
1104                         wake_up_partner(pipe);
1105
1106                 if (!is_pipe && !pipe->readers) {
1107                         if (wait_for_partner(pipe, &pipe->r_counter))
1108                                 goto err_wr;
1109                 }
1110                 break;
1111         
1112         case FMODE_READ | FMODE_WRITE:
1113         /*
1114          *  O_RDWR
1115          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1116          *  This implementation will NEVER block on a O_RDWR open, since
1117          *  the process can at least talk to itself.
1118          */
1119
1120                 pipe->readers++;
1121                 pipe->writers++;
1122                 pipe->r_counter++;
1123                 pipe->w_counter++;
1124                 if (pipe->readers == 1 || pipe->writers == 1)
1125                         wake_up_partner(pipe);
1126                 break;
1127
1128         default:
1129                 ret = -EINVAL;
1130                 goto err;
1131         }
1132
1133         /* Ok! */
1134         __pipe_unlock(pipe);
1135         return 0;
1136
1137 err_rd:
1138         if (!--pipe->readers)
1139                 wake_up_interruptible(&pipe->wait);
1140         ret = -ERESTARTSYS;
1141         goto err;
1142
1143 err_wr:
1144         if (!--pipe->writers)
1145                 wake_up_interruptible(&pipe->wait);
1146         ret = -ERESTARTSYS;
1147         goto err;
1148
1149 err:
1150         __pipe_unlock(pipe);
1151
1152         put_pipe_info(inode, pipe);
1153         return ret;
1154 }
1155
1156 const struct file_operations pipefifo_fops = {
1157         .open           = fifo_open,
1158         .llseek         = no_llseek,
1159         .read           = do_sync_read,
1160         .aio_read       = pipe_read,
1161         .write          = do_sync_write,
1162         .aio_write      = pipe_write,
1163         .poll           = pipe_poll,
1164         .unlocked_ioctl = pipe_ioctl,
1165         .release        = pipe_release,
1166         .fasync         = pipe_fasync,
1167 };
1168
1169 /*
1170  * Allocate a new array of pipe buffers and copy the info over. Returns the
1171  * pipe size if successful, or return -ERROR on error.
1172  */
1173 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1174 {
1175         struct pipe_buffer *bufs;
1176
1177         /*
1178          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1179          * expect a lot of shrink+grow operations, just free and allocate
1180          * again like we would do for growing. If the pipe currently
1181          * contains more buffers than arg, then return busy.
1182          */
1183         if (nr_pages < pipe->nrbufs)
1184                 return -EBUSY;
1185
1186         bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1187         if (unlikely(!bufs))
1188                 return -ENOMEM;
1189
1190         /*
1191          * The pipe array wraps around, so just start the new one at zero
1192          * and adjust the indexes.
1193          */
1194         if (pipe->nrbufs) {
1195                 unsigned int tail;
1196                 unsigned int head;
1197
1198                 tail = pipe->curbuf + pipe->nrbufs;
1199                 if (tail < pipe->buffers)
1200                         tail = 0;
1201                 else
1202                         tail &= (pipe->buffers - 1);
1203
1204                 head = pipe->nrbufs - tail;
1205                 if (head)
1206                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1207                 if (tail)
1208                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1209         }
1210
1211         pipe->curbuf = 0;
1212         kfree(pipe->bufs);
1213         pipe->bufs = bufs;
1214         pipe->buffers = nr_pages;
1215         return nr_pages * PAGE_SIZE;
1216 }
1217
1218 /*
1219  * Currently we rely on the pipe array holding a power-of-2 number
1220  * of pages.
1221  */
1222 static inline unsigned int round_pipe_size(unsigned int size)
1223 {
1224         unsigned long nr_pages;
1225
1226         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1227         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1228 }
1229
1230 /*
1231  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1232  * will return an error.
1233  */
1234 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1235                  size_t *lenp, loff_t *ppos)
1236 {
1237         int ret;
1238
1239         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1240         if (ret < 0 || !write)
1241                 return ret;
1242
1243         pipe_max_size = round_pipe_size(pipe_max_size);
1244         return ret;
1245 }
1246
1247 /*
1248  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1249  * location, so checking ->i_pipe is not enough to verify that this is a
1250  * pipe.
1251  */
1252 struct pipe_inode_info *get_pipe_info(struct file *file)
1253 {
1254         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1255 }
1256
1257 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1258 {
1259         struct pipe_inode_info *pipe;
1260         long ret;
1261
1262         pipe = get_pipe_info(file);
1263         if (!pipe)
1264                 return -EBADF;
1265
1266         __pipe_lock(pipe);
1267
1268         switch (cmd) {
1269         case F_SETPIPE_SZ: {
1270                 unsigned int size, nr_pages;
1271
1272                 size = round_pipe_size(arg);
1273                 nr_pages = size >> PAGE_SHIFT;
1274
1275                 ret = -EINVAL;
1276                 if (!nr_pages)
1277                         goto out;
1278
1279                 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1280                         ret = -EPERM;
1281                         goto out;
1282                 }
1283                 ret = pipe_set_size(pipe, nr_pages);
1284                 break;
1285                 }
1286         case F_GETPIPE_SZ:
1287                 ret = pipe->buffers * PAGE_SIZE;
1288                 break;
1289         default:
1290                 ret = -EINVAL;
1291                 break;
1292         }
1293
1294 out:
1295         __pipe_unlock(pipe);
1296         return ret;
1297 }
1298
1299 static const struct super_operations pipefs_ops = {
1300         .destroy_inode = free_inode_nonrcu,
1301         .statfs = simple_statfs,
1302 };
1303
1304 /*
1305  * pipefs should _never_ be mounted by userland - too much of security hassle,
1306  * no real gain from having the whole whorehouse mounted. So we don't need
1307  * any operations on the root directory. However, we need a non-trivial
1308  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1309  */
1310 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1311                          int flags, const char *dev_name, void *data)
1312 {
1313         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1314                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1315 }
1316
1317 static struct file_system_type pipe_fs_type = {
1318         .name           = "pipefs",
1319         .mount          = pipefs_mount,
1320         .kill_sb        = kill_anon_super,
1321 };
1322
1323 static int __init init_pipe_fs(void)
1324 {
1325         int err = register_filesystem(&pipe_fs_type);
1326
1327         if (!err) {
1328                 pipe_mnt = kern_mount(&pipe_fs_type);
1329                 if (IS_ERR(pipe_mnt)) {
1330                         err = PTR_ERR(pipe_mnt);
1331                         unregister_filesystem(&pipe_fs_type);
1332                 }
1333         }
1334         return err;
1335 }
1336
1337 fs_initcall(init_pipe_fs);