mm: export generic_pipe_buf_*() to modules
[linux-2.6.git] / fs / pipe.c
1 /*
2  *  linux/fs/pipe.c
3  *
4  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
5  */
6
7 #include <linux/mm.h>
8 #include <linux/file.h>
9 #include <linux/poll.h>
10 #include <linux/slab.h>
11 #include <linux/module.h>
12 #include <linux/init.h>
13 #include <linux/fs.h>
14 #include <linux/log2.h>
15 #include <linux/mount.h>
16 #include <linux/pipe_fs_i.h>
17 #include <linux/uio.h>
18 #include <linux/highmem.h>
19 #include <linux/pagemap.h>
20 #include <linux/audit.h>
21 #include <linux/syscalls.h>
22 #include <linux/fcntl.h>
23
24 #include <asm/uaccess.h>
25 #include <asm/ioctls.h>
26
27 /*
28  * The max size that a non-root user is allowed to grow the pipe. Can
29  * be set by root in /proc/sys/fs/pipe-max-pages
30  */
31 unsigned int pipe_max_pages = PIPE_DEF_BUFFERS * 16;
32
33 /*
34  * We use a start+len construction, which provides full use of the 
35  * allocated memory.
36  * -- Florian Coosmann (FGC)
37  * 
38  * Reads with count = 0 should always return 0.
39  * -- Julian Bradfield 1999-06-07.
40  *
41  * FIFOs and Pipes now generate SIGIO for both readers and writers.
42  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
43  *
44  * pipe_read & write cleanup
45  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
46  */
47
48 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
49 {
50         if (pipe->inode)
51                 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
52 }
53
54 void pipe_lock(struct pipe_inode_info *pipe)
55 {
56         /*
57          * pipe_lock() nests non-pipe inode locks (for writing to a file)
58          */
59         pipe_lock_nested(pipe, I_MUTEX_PARENT);
60 }
61 EXPORT_SYMBOL(pipe_lock);
62
63 void pipe_unlock(struct pipe_inode_info *pipe)
64 {
65         if (pipe->inode)
66                 mutex_unlock(&pipe->inode->i_mutex);
67 }
68 EXPORT_SYMBOL(pipe_unlock);
69
70 void pipe_double_lock(struct pipe_inode_info *pipe1,
71                       struct pipe_inode_info *pipe2)
72 {
73         BUG_ON(pipe1 == pipe2);
74
75         if (pipe1 < pipe2) {
76                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
77                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
78         } else {
79                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
80                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
81         }
82 }
83
84 /* Drop the inode semaphore and wait for a pipe event, atomically */
85 void pipe_wait(struct pipe_inode_info *pipe)
86 {
87         DEFINE_WAIT(wait);
88
89         /*
90          * Pipes are system-local resources, so sleeping on them
91          * is considered a noninteractive wait:
92          */
93         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
94         pipe_unlock(pipe);
95         schedule();
96         finish_wait(&pipe->wait, &wait);
97         pipe_lock(pipe);
98 }
99
100 static int
101 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
102                         int atomic)
103 {
104         unsigned long copy;
105
106         while (len > 0) {
107                 while (!iov->iov_len)
108                         iov++;
109                 copy = min_t(unsigned long, len, iov->iov_len);
110
111                 if (atomic) {
112                         if (__copy_from_user_inatomic(to, iov->iov_base, copy))
113                                 return -EFAULT;
114                 } else {
115                         if (copy_from_user(to, iov->iov_base, copy))
116                                 return -EFAULT;
117                 }
118                 to += copy;
119                 len -= copy;
120                 iov->iov_base += copy;
121                 iov->iov_len -= copy;
122         }
123         return 0;
124 }
125
126 static int
127 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
128                       int atomic)
129 {
130         unsigned long copy;
131
132         while (len > 0) {
133                 while (!iov->iov_len)
134                         iov++;
135                 copy = min_t(unsigned long, len, iov->iov_len);
136
137                 if (atomic) {
138                         if (__copy_to_user_inatomic(iov->iov_base, from, copy))
139                                 return -EFAULT;
140                 } else {
141                         if (copy_to_user(iov->iov_base, from, copy))
142                                 return -EFAULT;
143                 }
144                 from += copy;
145                 len -= copy;
146                 iov->iov_base += copy;
147                 iov->iov_len -= copy;
148         }
149         return 0;
150 }
151
152 /*
153  * Attempt to pre-fault in the user memory, so we can use atomic copies.
154  * Returns the number of bytes not faulted in.
155  */
156 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
157 {
158         while (!iov->iov_len)
159                 iov++;
160
161         while (len > 0) {
162                 unsigned long this_len;
163
164                 this_len = min_t(unsigned long, len, iov->iov_len);
165                 if (fault_in_pages_writeable(iov->iov_base, this_len))
166                         break;
167
168                 len -= this_len;
169                 iov++;
170         }
171
172         return len;
173 }
174
175 /*
176  * Pre-fault in the user memory, so we can use atomic copies.
177  */
178 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
179 {
180         while (!iov->iov_len)
181                 iov++;
182
183         while (len > 0) {
184                 unsigned long this_len;
185
186                 this_len = min_t(unsigned long, len, iov->iov_len);
187                 fault_in_pages_readable(iov->iov_base, this_len);
188                 len -= this_len;
189                 iov++;
190         }
191 }
192
193 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
194                                   struct pipe_buffer *buf)
195 {
196         struct page *page = buf->page;
197
198         /*
199          * If nobody else uses this page, and we don't already have a
200          * temporary page, let's keep track of it as a one-deep
201          * allocation cache. (Otherwise just release our reference to it)
202          */
203         if (page_count(page) == 1 && !pipe->tmp_page)
204                 pipe->tmp_page = page;
205         else
206                 page_cache_release(page);
207 }
208
209 /**
210  * generic_pipe_buf_map - virtually map a pipe buffer
211  * @pipe:       the pipe that the buffer belongs to
212  * @buf:        the buffer that should be mapped
213  * @atomic:     whether to use an atomic map
214  *
215  * Description:
216  *      This function returns a kernel virtual address mapping for the
217  *      pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
218  *      and the caller has to be careful not to fault before calling
219  *      the unmap function.
220  *
221  *      Note that this function occupies KM_USER0 if @atomic != 0.
222  */
223 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
224                            struct pipe_buffer *buf, int atomic)
225 {
226         if (atomic) {
227                 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
228                 return kmap_atomic(buf->page, KM_USER0);
229         }
230
231         return kmap(buf->page);
232 }
233 EXPORT_SYMBOL(generic_pipe_buf_map);
234
235 /**
236  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
237  * @pipe:       the pipe that the buffer belongs to
238  * @buf:        the buffer that should be unmapped
239  * @map_data:   the data that the mapping function returned
240  *
241  * Description:
242  *      This function undoes the mapping that ->map() provided.
243  */
244 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
245                             struct pipe_buffer *buf, void *map_data)
246 {
247         if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
248                 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
249                 kunmap_atomic(map_data, KM_USER0);
250         } else
251                 kunmap(buf->page);
252 }
253 EXPORT_SYMBOL(generic_pipe_buf_unmap);
254
255 /**
256  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
257  * @pipe:       the pipe that the buffer belongs to
258  * @buf:        the buffer to attempt to steal
259  *
260  * Description:
261  *      This function attempts to steal the &struct page attached to
262  *      @buf. If successful, this function returns 0 and returns with
263  *      the page locked. The caller may then reuse the page for whatever
264  *      he wishes; the typical use is insertion into a different file
265  *      page cache.
266  */
267 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
268                            struct pipe_buffer *buf)
269 {
270         struct page *page = buf->page;
271
272         /*
273          * A reference of one is golden, that means that the owner of this
274          * page is the only one holding a reference to it. lock the page
275          * and return OK.
276          */
277         if (page_count(page) == 1) {
278                 lock_page(page);
279                 return 0;
280         }
281
282         return 1;
283 }
284 EXPORT_SYMBOL(generic_pipe_buf_steal);
285
286 /**
287  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
288  * @pipe:       the pipe that the buffer belongs to
289  * @buf:        the buffer to get a reference to
290  *
291  * Description:
292  *      This function grabs an extra reference to @buf. It's used in
293  *      in the tee() system call, when we duplicate the buffers in one
294  *      pipe into another.
295  */
296 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
297 {
298         page_cache_get(buf->page);
299 }
300 EXPORT_SYMBOL(generic_pipe_buf_get);
301
302 /**
303  * generic_pipe_buf_confirm - verify contents of the pipe buffer
304  * @info:       the pipe that the buffer belongs to
305  * @buf:        the buffer to confirm
306  *
307  * Description:
308  *      This function does nothing, because the generic pipe code uses
309  *      pages that are always good when inserted into the pipe.
310  */
311 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
312                              struct pipe_buffer *buf)
313 {
314         return 0;
315 }
316 EXPORT_SYMBOL(generic_pipe_buf_confirm);
317
318 /**
319  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
320  * @pipe:       the pipe that the buffer belongs to
321  * @buf:        the buffer to put a reference to
322  *
323  * Description:
324  *      This function releases a reference to @buf.
325  */
326 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
327                               struct pipe_buffer *buf)
328 {
329         page_cache_release(buf->page);
330 }
331 EXPORT_SYMBOL(generic_pipe_buf_release);
332
333 static const struct pipe_buf_operations anon_pipe_buf_ops = {
334         .can_merge = 1,
335         .map = generic_pipe_buf_map,
336         .unmap = generic_pipe_buf_unmap,
337         .confirm = generic_pipe_buf_confirm,
338         .release = anon_pipe_buf_release,
339         .steal = generic_pipe_buf_steal,
340         .get = generic_pipe_buf_get,
341 };
342
343 static ssize_t
344 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
345            unsigned long nr_segs, loff_t pos)
346 {
347         struct file *filp = iocb->ki_filp;
348         struct inode *inode = filp->f_path.dentry->d_inode;
349         struct pipe_inode_info *pipe;
350         int do_wakeup;
351         ssize_t ret;
352         struct iovec *iov = (struct iovec *)_iov;
353         size_t total_len;
354
355         total_len = iov_length(iov, nr_segs);
356         /* Null read succeeds. */
357         if (unlikely(total_len == 0))
358                 return 0;
359
360         do_wakeup = 0;
361         ret = 0;
362         mutex_lock(&inode->i_mutex);
363         pipe = inode->i_pipe;
364         for (;;) {
365                 int bufs = pipe->nrbufs;
366                 if (bufs) {
367                         int curbuf = pipe->curbuf;
368                         struct pipe_buffer *buf = pipe->bufs + curbuf;
369                         const struct pipe_buf_operations *ops = buf->ops;
370                         void *addr;
371                         size_t chars = buf->len;
372                         int error, atomic;
373
374                         if (chars > total_len)
375                                 chars = total_len;
376
377                         error = ops->confirm(pipe, buf);
378                         if (error) {
379                                 if (!ret)
380                                         error = ret;
381                                 break;
382                         }
383
384                         atomic = !iov_fault_in_pages_write(iov, chars);
385 redo:
386                         addr = ops->map(pipe, buf, atomic);
387                         error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
388                         ops->unmap(pipe, buf, addr);
389                         if (unlikely(error)) {
390                                 /*
391                                  * Just retry with the slow path if we failed.
392                                  */
393                                 if (atomic) {
394                                         atomic = 0;
395                                         goto redo;
396                                 }
397                                 if (!ret)
398                                         ret = error;
399                                 break;
400                         }
401                         ret += chars;
402                         buf->offset += chars;
403                         buf->len -= chars;
404                         if (!buf->len) {
405                                 buf->ops = NULL;
406                                 ops->release(pipe, buf);
407                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
408                                 pipe->curbuf = curbuf;
409                                 pipe->nrbufs = --bufs;
410                                 do_wakeup = 1;
411                         }
412                         total_len -= chars;
413                         if (!total_len)
414                                 break;  /* common path: read succeeded */
415                 }
416                 if (bufs)       /* More to do? */
417                         continue;
418                 if (!pipe->writers)
419                         break;
420                 if (!pipe->waiting_writers) {
421                         /* syscall merging: Usually we must not sleep
422                          * if O_NONBLOCK is set, or if we got some data.
423                          * But if a writer sleeps in kernel space, then
424                          * we can wait for that data without violating POSIX.
425                          */
426                         if (ret)
427                                 break;
428                         if (filp->f_flags & O_NONBLOCK) {
429                                 ret = -EAGAIN;
430                                 break;
431                         }
432                 }
433                 if (signal_pending(current)) {
434                         if (!ret)
435                                 ret = -ERESTARTSYS;
436                         break;
437                 }
438                 if (do_wakeup) {
439                         wake_up_interruptible_sync(&pipe->wait);
440                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
441                 }
442                 pipe_wait(pipe);
443         }
444         mutex_unlock(&inode->i_mutex);
445
446         /* Signal writers asynchronously that there is more room. */
447         if (do_wakeup) {
448                 wake_up_interruptible_sync(&pipe->wait);
449                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
450         }
451         if (ret > 0)
452                 file_accessed(filp);
453         return ret;
454 }
455
456 static ssize_t
457 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
458             unsigned long nr_segs, loff_t ppos)
459 {
460         struct file *filp = iocb->ki_filp;
461         struct inode *inode = filp->f_path.dentry->d_inode;
462         struct pipe_inode_info *pipe;
463         ssize_t ret;
464         int do_wakeup;
465         struct iovec *iov = (struct iovec *)_iov;
466         size_t total_len;
467         ssize_t chars;
468
469         total_len = iov_length(iov, nr_segs);
470         /* Null write succeeds. */
471         if (unlikely(total_len == 0))
472                 return 0;
473
474         do_wakeup = 0;
475         ret = 0;
476         mutex_lock(&inode->i_mutex);
477         pipe = inode->i_pipe;
478
479         if (!pipe->readers) {
480                 send_sig(SIGPIPE, current, 0);
481                 ret = -EPIPE;
482                 goto out;
483         }
484
485         /* We try to merge small writes */
486         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
487         if (pipe->nrbufs && chars != 0) {
488                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
489                                                         (pipe->buffers - 1);
490                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
491                 const struct pipe_buf_operations *ops = buf->ops;
492                 int offset = buf->offset + buf->len;
493
494                 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
495                         int error, atomic = 1;
496                         void *addr;
497
498                         error = ops->confirm(pipe, buf);
499                         if (error)
500                                 goto out;
501
502                         iov_fault_in_pages_read(iov, chars);
503 redo1:
504                         addr = ops->map(pipe, buf, atomic);
505                         error = pipe_iov_copy_from_user(offset + addr, iov,
506                                                         chars, atomic);
507                         ops->unmap(pipe, buf, addr);
508                         ret = error;
509                         do_wakeup = 1;
510                         if (error) {
511                                 if (atomic) {
512                                         atomic = 0;
513                                         goto redo1;
514                                 }
515                                 goto out;
516                         }
517                         buf->len += chars;
518                         total_len -= chars;
519                         ret = chars;
520                         if (!total_len)
521                                 goto out;
522                 }
523         }
524
525         for (;;) {
526                 int bufs;
527
528                 if (!pipe->readers) {
529                         send_sig(SIGPIPE, current, 0);
530                         if (!ret)
531                                 ret = -EPIPE;
532                         break;
533                 }
534                 bufs = pipe->nrbufs;
535                 if (bufs < pipe->buffers) {
536                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
537                         struct pipe_buffer *buf = pipe->bufs + newbuf;
538                         struct page *page = pipe->tmp_page;
539                         char *src;
540                         int error, atomic = 1;
541
542                         if (!page) {
543                                 page = alloc_page(GFP_HIGHUSER);
544                                 if (unlikely(!page)) {
545                                         ret = ret ? : -ENOMEM;
546                                         break;
547                                 }
548                                 pipe->tmp_page = page;
549                         }
550                         /* Always wake up, even if the copy fails. Otherwise
551                          * we lock up (O_NONBLOCK-)readers that sleep due to
552                          * syscall merging.
553                          * FIXME! Is this really true?
554                          */
555                         do_wakeup = 1;
556                         chars = PAGE_SIZE;
557                         if (chars > total_len)
558                                 chars = total_len;
559
560                         iov_fault_in_pages_read(iov, chars);
561 redo2:
562                         if (atomic)
563                                 src = kmap_atomic(page, KM_USER0);
564                         else
565                                 src = kmap(page);
566
567                         error = pipe_iov_copy_from_user(src, iov, chars,
568                                                         atomic);
569                         if (atomic)
570                                 kunmap_atomic(src, KM_USER0);
571                         else
572                                 kunmap(page);
573
574                         if (unlikely(error)) {
575                                 if (atomic) {
576                                         atomic = 0;
577                                         goto redo2;
578                                 }
579                                 if (!ret)
580                                         ret = error;
581                                 break;
582                         }
583                         ret += chars;
584
585                         /* Insert it into the buffer array */
586                         buf->page = page;
587                         buf->ops = &anon_pipe_buf_ops;
588                         buf->offset = 0;
589                         buf->len = chars;
590                         pipe->nrbufs = ++bufs;
591                         pipe->tmp_page = NULL;
592
593                         total_len -= chars;
594                         if (!total_len)
595                                 break;
596                 }
597                 if (bufs < pipe->buffers)
598                         continue;
599                 if (filp->f_flags & O_NONBLOCK) {
600                         if (!ret)
601                                 ret = -EAGAIN;
602                         break;
603                 }
604                 if (signal_pending(current)) {
605                         if (!ret)
606                                 ret = -ERESTARTSYS;
607                         break;
608                 }
609                 if (do_wakeup) {
610                         wake_up_interruptible_sync(&pipe->wait);
611                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
612                         do_wakeup = 0;
613                 }
614                 pipe->waiting_writers++;
615                 pipe_wait(pipe);
616                 pipe->waiting_writers--;
617         }
618 out:
619         mutex_unlock(&inode->i_mutex);
620         if (do_wakeup) {
621                 wake_up_interruptible_sync(&pipe->wait);
622                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
623         }
624         if (ret > 0)
625                 file_update_time(filp);
626         return ret;
627 }
628
629 static ssize_t
630 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
631 {
632         return -EBADF;
633 }
634
635 static ssize_t
636 bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
637            loff_t *ppos)
638 {
639         return -EBADF;
640 }
641
642 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
643 {
644         struct inode *inode = filp->f_path.dentry->d_inode;
645         struct pipe_inode_info *pipe;
646         int count, buf, nrbufs;
647
648         switch (cmd) {
649                 case FIONREAD:
650                         mutex_lock(&inode->i_mutex);
651                         pipe = inode->i_pipe;
652                         count = 0;
653                         buf = pipe->curbuf;
654                         nrbufs = pipe->nrbufs;
655                         while (--nrbufs >= 0) {
656                                 count += pipe->bufs[buf].len;
657                                 buf = (buf+1) & (pipe->buffers - 1);
658                         }
659                         mutex_unlock(&inode->i_mutex);
660
661                         return put_user(count, (int __user *)arg);
662                 default:
663                         return -EINVAL;
664         }
665 }
666
667 /* No kernel lock held - fine */
668 static unsigned int
669 pipe_poll(struct file *filp, poll_table *wait)
670 {
671         unsigned int mask;
672         struct inode *inode = filp->f_path.dentry->d_inode;
673         struct pipe_inode_info *pipe = inode->i_pipe;
674         int nrbufs;
675
676         poll_wait(filp, &pipe->wait, wait);
677
678         /* Reading only -- no need for acquiring the semaphore.  */
679         nrbufs = pipe->nrbufs;
680         mask = 0;
681         if (filp->f_mode & FMODE_READ) {
682                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
683                 if (!pipe->writers && filp->f_version != pipe->w_counter)
684                         mask |= POLLHUP;
685         }
686
687         if (filp->f_mode & FMODE_WRITE) {
688                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
689                 /*
690                  * Most Unices do not set POLLERR for FIFOs but on Linux they
691                  * behave exactly like pipes for poll().
692                  */
693                 if (!pipe->readers)
694                         mask |= POLLERR;
695         }
696
697         return mask;
698 }
699
700 static int
701 pipe_release(struct inode *inode, int decr, int decw)
702 {
703         struct pipe_inode_info *pipe;
704
705         mutex_lock(&inode->i_mutex);
706         pipe = inode->i_pipe;
707         pipe->readers -= decr;
708         pipe->writers -= decw;
709
710         if (!pipe->readers && !pipe->writers) {
711                 free_pipe_info(inode);
712         } else {
713                 wake_up_interruptible_sync(&pipe->wait);
714                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
715                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
716         }
717         mutex_unlock(&inode->i_mutex);
718
719         return 0;
720 }
721
722 static int
723 pipe_read_fasync(int fd, struct file *filp, int on)
724 {
725         struct inode *inode = filp->f_path.dentry->d_inode;
726         int retval;
727
728         mutex_lock(&inode->i_mutex);
729         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
730         mutex_unlock(&inode->i_mutex);
731
732         return retval;
733 }
734
735
736 static int
737 pipe_write_fasync(int fd, struct file *filp, int on)
738 {
739         struct inode *inode = filp->f_path.dentry->d_inode;
740         int retval;
741
742         mutex_lock(&inode->i_mutex);
743         retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
744         mutex_unlock(&inode->i_mutex);
745
746         return retval;
747 }
748
749
750 static int
751 pipe_rdwr_fasync(int fd, struct file *filp, int on)
752 {
753         struct inode *inode = filp->f_path.dentry->d_inode;
754         struct pipe_inode_info *pipe = inode->i_pipe;
755         int retval;
756
757         mutex_lock(&inode->i_mutex);
758         retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
759         if (retval >= 0) {
760                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
761                 if (retval < 0) /* this can happen only if on == T */
762                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
763         }
764         mutex_unlock(&inode->i_mutex);
765         return retval;
766 }
767
768
769 static int
770 pipe_read_release(struct inode *inode, struct file *filp)
771 {
772         return pipe_release(inode, 1, 0);
773 }
774
775 static int
776 pipe_write_release(struct inode *inode, struct file *filp)
777 {
778         return pipe_release(inode, 0, 1);
779 }
780
781 static int
782 pipe_rdwr_release(struct inode *inode, struct file *filp)
783 {
784         int decr, decw;
785
786         decr = (filp->f_mode & FMODE_READ) != 0;
787         decw = (filp->f_mode & FMODE_WRITE) != 0;
788         return pipe_release(inode, decr, decw);
789 }
790
791 static int
792 pipe_read_open(struct inode *inode, struct file *filp)
793 {
794         int ret = -ENOENT;
795
796         mutex_lock(&inode->i_mutex);
797
798         if (inode->i_pipe) {
799                 ret = 0;
800                 inode->i_pipe->readers++;
801         }
802
803         mutex_unlock(&inode->i_mutex);
804
805         return ret;
806 }
807
808 static int
809 pipe_write_open(struct inode *inode, struct file *filp)
810 {
811         int ret = -ENOENT;
812
813         mutex_lock(&inode->i_mutex);
814
815         if (inode->i_pipe) {
816                 ret = 0;
817                 inode->i_pipe->writers++;
818         }
819
820         mutex_unlock(&inode->i_mutex);
821
822         return ret;
823 }
824
825 static int
826 pipe_rdwr_open(struct inode *inode, struct file *filp)
827 {
828         int ret = -ENOENT;
829
830         mutex_lock(&inode->i_mutex);
831
832         if (inode->i_pipe) {
833                 ret = 0;
834                 if (filp->f_mode & FMODE_READ)
835                         inode->i_pipe->readers++;
836                 if (filp->f_mode & FMODE_WRITE)
837                         inode->i_pipe->writers++;
838         }
839
840         mutex_unlock(&inode->i_mutex);
841
842         return ret;
843 }
844
845 /*
846  * The file_operations structs are not static because they
847  * are also used in linux/fs/fifo.c to do operations on FIFOs.
848  *
849  * Pipes reuse fifos' file_operations structs.
850  */
851 const struct file_operations read_pipefifo_fops = {
852         .llseek         = no_llseek,
853         .read           = do_sync_read,
854         .aio_read       = pipe_read,
855         .write          = bad_pipe_w,
856         .poll           = pipe_poll,
857         .unlocked_ioctl = pipe_ioctl,
858         .open           = pipe_read_open,
859         .release        = pipe_read_release,
860         .fasync         = pipe_read_fasync,
861 };
862
863 const struct file_operations write_pipefifo_fops = {
864         .llseek         = no_llseek,
865         .read           = bad_pipe_r,
866         .write          = do_sync_write,
867         .aio_write      = pipe_write,
868         .poll           = pipe_poll,
869         .unlocked_ioctl = pipe_ioctl,
870         .open           = pipe_write_open,
871         .release        = pipe_write_release,
872         .fasync         = pipe_write_fasync,
873 };
874
875 const struct file_operations rdwr_pipefifo_fops = {
876         .llseek         = no_llseek,
877         .read           = do_sync_read,
878         .aio_read       = pipe_read,
879         .write          = do_sync_write,
880         .aio_write      = pipe_write,
881         .poll           = pipe_poll,
882         .unlocked_ioctl = pipe_ioctl,
883         .open           = pipe_rdwr_open,
884         .release        = pipe_rdwr_release,
885         .fasync         = pipe_rdwr_fasync,
886 };
887
888 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
889 {
890         struct pipe_inode_info *pipe;
891
892         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
893         if (pipe) {
894                 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
895                 if (pipe->bufs) {
896                         init_waitqueue_head(&pipe->wait);
897                         pipe->r_counter = pipe->w_counter = 1;
898                         pipe->inode = inode;
899                         pipe->buffers = PIPE_DEF_BUFFERS;
900                         return pipe;
901                 }
902                 kfree(pipe);
903         }
904
905         return NULL;
906 }
907
908 void __free_pipe_info(struct pipe_inode_info *pipe)
909 {
910         int i;
911
912         for (i = 0; i < pipe->buffers; i++) {
913                 struct pipe_buffer *buf = pipe->bufs + i;
914                 if (buf->ops)
915                         buf->ops->release(pipe, buf);
916         }
917         if (pipe->tmp_page)
918                 __free_page(pipe->tmp_page);
919         kfree(pipe->bufs);
920         kfree(pipe);
921 }
922
923 void free_pipe_info(struct inode *inode)
924 {
925         __free_pipe_info(inode->i_pipe);
926         inode->i_pipe = NULL;
927 }
928
929 static struct vfsmount *pipe_mnt __read_mostly;
930
931 /*
932  * pipefs_dname() is called from d_path().
933  */
934 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
935 {
936         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
937                                 dentry->d_inode->i_ino);
938 }
939
940 static const struct dentry_operations pipefs_dentry_operations = {
941         .d_dname        = pipefs_dname,
942 };
943
944 static struct inode * get_pipe_inode(void)
945 {
946         struct inode *inode = new_inode(pipe_mnt->mnt_sb);
947         struct pipe_inode_info *pipe;
948
949         if (!inode)
950                 goto fail_inode;
951
952         pipe = alloc_pipe_info(inode);
953         if (!pipe)
954                 goto fail_iput;
955         inode->i_pipe = pipe;
956
957         pipe->readers = pipe->writers = 1;
958         inode->i_fop = &rdwr_pipefifo_fops;
959
960         /*
961          * Mark the inode dirty from the very beginning,
962          * that way it will never be moved to the dirty
963          * list because "mark_inode_dirty()" will think
964          * that it already _is_ on the dirty list.
965          */
966         inode->i_state = I_DIRTY;
967         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
968         inode->i_uid = current_fsuid();
969         inode->i_gid = current_fsgid();
970         inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
971
972         return inode;
973
974 fail_iput:
975         iput(inode);
976
977 fail_inode:
978         return NULL;
979 }
980
981 struct file *create_write_pipe(int flags)
982 {
983         int err;
984         struct inode *inode;
985         struct file *f;
986         struct path path;
987         struct qstr name = { .name = "" };
988
989         err = -ENFILE;
990         inode = get_pipe_inode();
991         if (!inode)
992                 goto err;
993
994         err = -ENOMEM;
995         path.dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &name);
996         if (!path.dentry)
997                 goto err_inode;
998         path.mnt = mntget(pipe_mnt);
999
1000         path.dentry->d_op = &pipefs_dentry_operations;
1001         d_instantiate(path.dentry, inode);
1002
1003         err = -ENFILE;
1004         f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1005         if (!f)
1006                 goto err_dentry;
1007         f->f_mapping = inode->i_mapping;
1008
1009         f->f_flags = O_WRONLY | (flags & O_NONBLOCK);
1010         f->f_version = 0;
1011
1012         return f;
1013
1014  err_dentry:
1015         free_pipe_info(inode);
1016         path_put(&path);
1017         return ERR_PTR(err);
1018
1019  err_inode:
1020         free_pipe_info(inode);
1021         iput(inode);
1022  err:
1023         return ERR_PTR(err);
1024 }
1025
1026 void free_write_pipe(struct file *f)
1027 {
1028         free_pipe_info(f->f_dentry->d_inode);
1029         path_put(&f->f_path);
1030         put_filp(f);
1031 }
1032
1033 struct file *create_read_pipe(struct file *wrf, int flags)
1034 {
1035         /* Grab pipe from the writer */
1036         struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1037                                     &read_pipefifo_fops);
1038         if (!f)
1039                 return ERR_PTR(-ENFILE);
1040
1041         path_get(&wrf->f_path);
1042         f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1043
1044         return f;
1045 }
1046
1047 int do_pipe_flags(int *fd, int flags)
1048 {
1049         struct file *fw, *fr;
1050         int error;
1051         int fdw, fdr;
1052
1053         if (flags & ~(O_CLOEXEC | O_NONBLOCK))
1054                 return -EINVAL;
1055
1056         fw = create_write_pipe(flags);
1057         if (IS_ERR(fw))
1058                 return PTR_ERR(fw);
1059         fr = create_read_pipe(fw, flags);
1060         error = PTR_ERR(fr);
1061         if (IS_ERR(fr))
1062                 goto err_write_pipe;
1063
1064         error = get_unused_fd_flags(flags);
1065         if (error < 0)
1066                 goto err_read_pipe;
1067         fdr = error;
1068
1069         error = get_unused_fd_flags(flags);
1070         if (error < 0)
1071                 goto err_fdr;
1072         fdw = error;
1073
1074         audit_fd_pair(fdr, fdw);
1075         fd_install(fdr, fr);
1076         fd_install(fdw, fw);
1077         fd[0] = fdr;
1078         fd[1] = fdw;
1079
1080         return 0;
1081
1082  err_fdr:
1083         put_unused_fd(fdr);
1084  err_read_pipe:
1085         path_put(&fr->f_path);
1086         put_filp(fr);
1087  err_write_pipe:
1088         free_write_pipe(fw);
1089         return error;
1090 }
1091
1092 /*
1093  * sys_pipe() is the normal C calling standard for creating
1094  * a pipe. It's not the way Unix traditionally does this, though.
1095  */
1096 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1097 {
1098         int fd[2];
1099         int error;
1100
1101         error = do_pipe_flags(fd, flags);
1102         if (!error) {
1103                 if (copy_to_user(fildes, fd, sizeof(fd))) {
1104                         sys_close(fd[0]);
1105                         sys_close(fd[1]);
1106                         error = -EFAULT;
1107                 }
1108         }
1109         return error;
1110 }
1111
1112 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1113 {
1114         return sys_pipe2(fildes, 0);
1115 }
1116
1117 /*
1118  * Allocate a new array of pipe buffers and copy the info over. Returns the
1119  * pipe size if successful, or return -ERROR on error.
1120  */
1121 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1122 {
1123         struct pipe_buffer *bufs;
1124
1125         /*
1126          * Must be a power-of-2 currently
1127          */
1128         if (!is_power_of_2(arg))
1129                 return -EINVAL;
1130
1131         /*
1132          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1133          * expect a lot of shrink+grow operations, just free and allocate
1134          * again like we would do for growing. If the pipe currently
1135          * contains more buffers than arg, then return busy.
1136          */
1137         if (arg < pipe->nrbufs)
1138                 return -EBUSY;
1139
1140         bufs = kcalloc(arg, sizeof(struct pipe_buffer), GFP_KERNEL);
1141         if (unlikely(!bufs))
1142                 return -ENOMEM;
1143
1144         /*
1145          * The pipe array wraps around, so just start the new one at zero
1146          * and adjust the indexes.
1147          */
1148         if (pipe->nrbufs) {
1149                 const unsigned int tail = pipe->nrbufs & (pipe->buffers - 1);
1150                 const unsigned int head = pipe->nrbufs - tail;
1151
1152                 if (head)
1153                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1154                 if (tail)
1155                         memcpy(bufs + head, pipe->bufs + pipe->curbuf, tail * sizeof(struct pipe_buffer));
1156         }
1157
1158         pipe->curbuf = 0;
1159         kfree(pipe->bufs);
1160         pipe->bufs = bufs;
1161         pipe->buffers = arg;
1162         return arg;
1163 }
1164
1165 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1166 {
1167         struct pipe_inode_info *pipe;
1168         long ret;
1169
1170         pipe = file->f_path.dentry->d_inode->i_pipe;
1171         if (!pipe)
1172                 return -EBADF;
1173
1174         mutex_lock(&pipe->inode->i_mutex);
1175
1176         switch (cmd) {
1177         case F_SETPIPE_SZ:
1178                 if (!capable(CAP_SYS_ADMIN) && arg > pipe_max_pages)
1179                         return -EINVAL;
1180                 /*
1181                  * The pipe needs to be at least 2 pages large to
1182                  * guarantee POSIX behaviour.
1183                  */
1184                 if (arg < 2)
1185                         return -EINVAL;
1186                 ret = pipe_set_size(pipe, arg);
1187                 break;
1188         case F_GETPIPE_SZ:
1189                 ret = pipe->buffers;
1190                 break;
1191         default:
1192                 ret = -EINVAL;
1193                 break;
1194         }
1195
1196         mutex_unlock(&pipe->inode->i_mutex);
1197         return ret;
1198 }
1199
1200 /*
1201  * pipefs should _never_ be mounted by userland - too much of security hassle,
1202  * no real gain from having the whole whorehouse mounted. So we don't need
1203  * any operations on the root directory. However, we need a non-trivial
1204  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1205  */
1206 static int pipefs_get_sb(struct file_system_type *fs_type,
1207                          int flags, const char *dev_name, void *data,
1208                          struct vfsmount *mnt)
1209 {
1210         return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC, mnt);
1211 }
1212
1213 static struct file_system_type pipe_fs_type = {
1214         .name           = "pipefs",
1215         .get_sb         = pipefs_get_sb,
1216         .kill_sb        = kill_anon_super,
1217 };
1218
1219 static int __init init_pipe_fs(void)
1220 {
1221         int err = register_filesystem(&pipe_fs_type);
1222
1223         if (!err) {
1224                 pipe_mnt = kern_mount(&pipe_fs_type);
1225                 if (IS_ERR(pipe_mnt)) {
1226                         err = PTR_ERR(pipe_mnt);
1227                         unregister_filesystem(&pipe_fs_type);
1228                 }
1229         }
1230         return err;
1231 }
1232
1233 static void __exit exit_pipe_fs(void)
1234 {
1235         unregister_filesystem(&pipe_fs_type);
1236         mntput(pipe_mnt);
1237 }
1238
1239 fs_initcall(init_pipe_fs);
1240 module_exit(exit_pipe_fs);