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