4 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
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>
14 #include <linux/mount.h>
15 #include <linux/pipe_fs_i.h>
16 #include <linux/uio.h>
17 #include <linux/highmem.h>
18 #include <linux/pagemap.h>
20 #include <asm/uaccess.h>
21 #include <asm/ioctls.h>
24 * We use a start+len construction, which provides full use of the
26 * -- Florian Coosmann (FGC)
28 * Reads with count = 0 should always return 0.
29 * -- Julian Bradfield 1999-06-07.
31 * FIFOs and Pipes now generate SIGIO for both readers and writers.
32 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
34 * pipe_read & write cleanup
35 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
38 /* Drop the inode semaphore and wait for a pipe event, atomically */
39 void pipe_wait(struct inode
* inode
)
44 * Pipes are system-local resources, so sleeping on them
45 * is considered a noninteractive wait:
47 prepare_to_wait(PIPE_WAIT(*inode
), &wait
, TASK_INTERRUPTIBLE
|TASK_NONINTERACTIVE
);
48 mutex_unlock(PIPE_MUTEX(*inode
));
50 finish_wait(PIPE_WAIT(*inode
), &wait
);
51 mutex_lock(PIPE_MUTEX(*inode
));
55 pipe_iov_copy_from_user(void *to
, struct iovec
*iov
, unsigned long len
)
62 copy
= min_t(unsigned long, len
, iov
->iov_len
);
64 if (copy_from_user(to
, iov
->iov_base
, copy
))
68 iov
->iov_base
+= copy
;
75 pipe_iov_copy_to_user(struct iovec
*iov
, const void *from
, unsigned long len
)
82 copy
= min_t(unsigned long, len
, iov
->iov_len
);
84 if (copy_to_user(iov
->iov_base
, from
, copy
))
88 iov
->iov_base
+= copy
;
94 static void anon_pipe_buf_release(struct pipe_inode_info
*info
, struct pipe_buffer
*buf
)
96 struct page
*page
= buf
->page
;
99 * If nobody else uses this page, and we don't already have a
100 * temporary page, let's keep track of it as a one-deep
103 if (page_count(page
) == 1 && !info
->tmp_page
) {
104 info
->tmp_page
= page
;
109 * Otherwise just release our reference to it
111 page_cache_release(page
);
114 static void *anon_pipe_buf_map(struct file
*file
, struct pipe_inode_info
*info
, struct pipe_buffer
*buf
)
116 return kmap(buf
->page
);
119 static void anon_pipe_buf_unmap(struct pipe_inode_info
*info
, struct pipe_buffer
*buf
)
124 static struct pipe_buf_operations anon_pipe_buf_ops
= {
126 .map
= anon_pipe_buf_map
,
127 .unmap
= anon_pipe_buf_unmap
,
128 .release
= anon_pipe_buf_release
,
132 pipe_readv(struct file
*filp
, const struct iovec
*_iov
,
133 unsigned long nr_segs
, loff_t
*ppos
)
135 struct inode
*inode
= filp
->f_dentry
->d_inode
;
136 struct pipe_inode_info
*info
;
139 struct iovec
*iov
= (struct iovec
*)_iov
;
142 total_len
= iov_length(iov
, nr_segs
);
143 /* Null read succeeds. */
144 if (unlikely(total_len
== 0))
149 mutex_lock(PIPE_MUTEX(*inode
));
150 info
= inode
->i_pipe
;
152 int bufs
= info
->nrbufs
;
154 int curbuf
= info
->curbuf
;
155 struct pipe_buffer
*buf
= info
->bufs
+ curbuf
;
156 struct pipe_buf_operations
*ops
= buf
->ops
;
158 size_t chars
= buf
->len
;
161 if (chars
> total_len
)
164 addr
= ops
->map(filp
, info
, buf
);
170 error
= pipe_iov_copy_to_user(iov
, addr
+ buf
->offset
, chars
);
171 ops
->unmap(info
, buf
);
172 if (unlikely(error
)) {
173 if (!ret
) ret
= -EFAULT
;
177 buf
->offset
+= chars
;
181 ops
->release(info
, buf
);
182 curbuf
= (curbuf
+ 1) & (PIPE_BUFFERS
-1);
183 info
->curbuf
= curbuf
;
184 info
->nrbufs
= --bufs
;
189 break; /* common path: read succeeded */
191 if (bufs
) /* More to do? */
193 if (!PIPE_WRITERS(*inode
))
195 if (!PIPE_WAITING_WRITERS(*inode
)) {
196 /* syscall merging: Usually we must not sleep
197 * if O_NONBLOCK is set, or if we got some data.
198 * But if a writer sleeps in kernel space, then
199 * we can wait for that data without violating POSIX.
203 if (filp
->f_flags
& O_NONBLOCK
) {
208 if (signal_pending(current
)) {
209 if (!ret
) ret
= -ERESTARTSYS
;
213 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
214 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
218 mutex_unlock(PIPE_MUTEX(*inode
));
219 /* Signal writers asynchronously that there is more room. */
221 wake_up_interruptible(PIPE_WAIT(*inode
));
222 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
230 pipe_read(struct file
*filp
, char __user
*buf
, size_t count
, loff_t
*ppos
)
232 struct iovec iov
= { .iov_base
= buf
, .iov_len
= count
};
233 return pipe_readv(filp
, &iov
, 1, ppos
);
237 pipe_writev(struct file
*filp
, const struct iovec
*_iov
,
238 unsigned long nr_segs
, loff_t
*ppos
)
240 struct inode
*inode
= filp
->f_dentry
->d_inode
;
241 struct pipe_inode_info
*info
;
244 struct iovec
*iov
= (struct iovec
*)_iov
;
248 total_len
= iov_length(iov
, nr_segs
);
249 /* Null write succeeds. */
250 if (unlikely(total_len
== 0))
255 mutex_lock(PIPE_MUTEX(*inode
));
256 info
= inode
->i_pipe
;
258 if (!PIPE_READERS(*inode
)) {
259 send_sig(SIGPIPE
, current
, 0);
264 /* We try to merge small writes */
265 chars
= total_len
& (PAGE_SIZE
-1); /* size of the last buffer */
266 if (info
->nrbufs
&& chars
!= 0) {
267 int lastbuf
= (info
->curbuf
+ info
->nrbufs
- 1) & (PIPE_BUFFERS
-1);
268 struct pipe_buffer
*buf
= info
->bufs
+ lastbuf
;
269 struct pipe_buf_operations
*ops
= buf
->ops
;
270 int offset
= buf
->offset
+ buf
->len
;
271 if (ops
->can_merge
&& offset
+ chars
<= PAGE_SIZE
) {
275 addr
= ops
->map(filp
, info
, buf
);
277 error
= PTR_ERR(addr
);
280 error
= pipe_iov_copy_from_user(offset
+ addr
, iov
,
282 ops
->unmap(info
, buf
);
297 if (!PIPE_READERS(*inode
)) {
298 send_sig(SIGPIPE
, current
, 0);
299 if (!ret
) ret
= -EPIPE
;
303 if (bufs
< PIPE_BUFFERS
) {
304 int newbuf
= (info
->curbuf
+ bufs
) & (PIPE_BUFFERS
-1);
305 struct pipe_buffer
*buf
= info
->bufs
+ newbuf
;
306 struct page
*page
= info
->tmp_page
;
310 page
= alloc_page(GFP_HIGHUSER
);
311 if (unlikely(!page
)) {
312 ret
= ret
? : -ENOMEM
;
315 info
->tmp_page
= page
;
317 /* Always wakeup, even if the copy fails. Otherwise
318 * we lock up (O_NONBLOCK-)readers that sleep due to
320 * FIXME! Is this really true?
324 if (chars
> total_len
)
327 error
= pipe_iov_copy_from_user(kmap(page
), iov
, chars
);
329 if (unlikely(error
)) {
330 if (!ret
) ret
= -EFAULT
;
335 /* Insert it into the buffer array */
337 buf
->ops
= &anon_pipe_buf_ops
;
340 info
->nrbufs
= ++bufs
;
341 info
->tmp_page
= NULL
;
347 if (bufs
< PIPE_BUFFERS
)
349 if (filp
->f_flags
& O_NONBLOCK
) {
350 if (!ret
) ret
= -EAGAIN
;
353 if (signal_pending(current
)) {
354 if (!ret
) ret
= -ERESTARTSYS
;
358 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
359 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
362 PIPE_WAITING_WRITERS(*inode
)++;
364 PIPE_WAITING_WRITERS(*inode
)--;
367 mutex_unlock(PIPE_MUTEX(*inode
));
369 wake_up_interruptible(PIPE_WAIT(*inode
));
370 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
373 file_update_time(filp
);
378 pipe_write(struct file
*filp
, const char __user
*buf
,
379 size_t count
, loff_t
*ppos
)
381 struct iovec iov
= { .iov_base
= (void __user
*)buf
, .iov_len
= count
};
382 return pipe_writev(filp
, &iov
, 1, ppos
);
386 bad_pipe_r(struct file
*filp
, char __user
*buf
, size_t count
, loff_t
*ppos
)
392 bad_pipe_w(struct file
*filp
, const char __user
*buf
, size_t count
, loff_t
*ppos
)
398 pipe_ioctl(struct inode
*pino
, struct file
*filp
,
399 unsigned int cmd
, unsigned long arg
)
401 struct inode
*inode
= filp
->f_dentry
->d_inode
;
402 struct pipe_inode_info
*info
;
403 int count
, buf
, nrbufs
;
407 mutex_lock(PIPE_MUTEX(*inode
));
408 info
= inode
->i_pipe
;
411 nrbufs
= info
->nrbufs
;
412 while (--nrbufs
>= 0) {
413 count
+= info
->bufs
[buf
].len
;
414 buf
= (buf
+1) & (PIPE_BUFFERS
-1);
416 mutex_unlock(PIPE_MUTEX(*inode
));
417 return put_user(count
, (int __user
*)arg
);
423 /* No kernel lock held - fine */
425 pipe_poll(struct file
*filp
, poll_table
*wait
)
428 struct inode
*inode
= filp
->f_dentry
->d_inode
;
429 struct pipe_inode_info
*info
= inode
->i_pipe
;
432 poll_wait(filp
, PIPE_WAIT(*inode
), wait
);
434 /* Reading only -- no need for acquiring the semaphore. */
435 nrbufs
= info
->nrbufs
;
437 if (filp
->f_mode
& FMODE_READ
) {
438 mask
= (nrbufs
> 0) ? POLLIN
| POLLRDNORM
: 0;
439 if (!PIPE_WRITERS(*inode
) && filp
->f_version
!= PIPE_WCOUNTER(*inode
))
443 if (filp
->f_mode
& FMODE_WRITE
) {
444 mask
|= (nrbufs
< PIPE_BUFFERS
) ? POLLOUT
| POLLWRNORM
: 0;
446 * Most Unices do not set POLLERR for FIFOs but on Linux they
447 * behave exactly like pipes for poll().
449 if (!PIPE_READERS(*inode
))
457 pipe_release(struct inode
*inode
, int decr
, int decw
)
459 mutex_lock(PIPE_MUTEX(*inode
));
460 PIPE_READERS(*inode
) -= decr
;
461 PIPE_WRITERS(*inode
) -= decw
;
462 if (!PIPE_READERS(*inode
) && !PIPE_WRITERS(*inode
)) {
463 free_pipe_info(inode
);
465 wake_up_interruptible(PIPE_WAIT(*inode
));
466 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
467 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
469 mutex_unlock(PIPE_MUTEX(*inode
));
475 pipe_read_fasync(int fd
, struct file
*filp
, int on
)
477 struct inode
*inode
= filp
->f_dentry
->d_inode
;
480 mutex_lock(PIPE_MUTEX(*inode
));
481 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_READERS(*inode
));
482 mutex_unlock(PIPE_MUTEX(*inode
));
492 pipe_write_fasync(int fd
, struct file
*filp
, int on
)
494 struct inode
*inode
= filp
->f_dentry
->d_inode
;
497 mutex_lock(PIPE_MUTEX(*inode
));
498 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_WRITERS(*inode
));
499 mutex_unlock(PIPE_MUTEX(*inode
));
509 pipe_rdwr_fasync(int fd
, struct file
*filp
, int on
)
511 struct inode
*inode
= filp
->f_dentry
->d_inode
;
514 mutex_lock(PIPE_MUTEX(*inode
));
516 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_READERS(*inode
));
519 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_WRITERS(*inode
));
521 mutex_unlock(PIPE_MUTEX(*inode
));
531 pipe_read_release(struct inode
*inode
, struct file
*filp
)
533 pipe_read_fasync(-1, filp
, 0);
534 return pipe_release(inode
, 1, 0);
538 pipe_write_release(struct inode
*inode
, struct file
*filp
)
540 pipe_write_fasync(-1, filp
, 0);
541 return pipe_release(inode
, 0, 1);
545 pipe_rdwr_release(struct inode
*inode
, struct file
*filp
)
549 pipe_rdwr_fasync(-1, filp
, 0);
550 decr
= (filp
->f_mode
& FMODE_READ
) != 0;
551 decw
= (filp
->f_mode
& FMODE_WRITE
) != 0;
552 return pipe_release(inode
, decr
, decw
);
556 pipe_read_open(struct inode
*inode
, struct file
*filp
)
558 /* We could have perhaps used atomic_t, but this and friends
559 below are the only places. So it doesn't seem worthwhile. */
560 mutex_lock(PIPE_MUTEX(*inode
));
561 PIPE_READERS(*inode
)++;
562 mutex_unlock(PIPE_MUTEX(*inode
));
568 pipe_write_open(struct inode
*inode
, struct file
*filp
)
570 mutex_lock(PIPE_MUTEX(*inode
));
571 PIPE_WRITERS(*inode
)++;
572 mutex_unlock(PIPE_MUTEX(*inode
));
578 pipe_rdwr_open(struct inode
*inode
, struct file
*filp
)
580 mutex_lock(PIPE_MUTEX(*inode
));
581 if (filp
->f_mode
& FMODE_READ
)
582 PIPE_READERS(*inode
)++;
583 if (filp
->f_mode
& FMODE_WRITE
)
584 PIPE_WRITERS(*inode
)++;
585 mutex_unlock(PIPE_MUTEX(*inode
));
591 * The file_operations structs are not static because they
592 * are also used in linux/fs/fifo.c to do operations on FIFOs.
594 const struct file_operations read_fifo_fops
= {
601 .open
= pipe_read_open
,
602 .release
= pipe_read_release
,
603 .fasync
= pipe_read_fasync
,
606 const struct file_operations write_fifo_fops
= {
610 .writev
= pipe_writev
,
613 .open
= pipe_write_open
,
614 .release
= pipe_write_release
,
615 .fasync
= pipe_write_fasync
,
618 const struct file_operations rdwr_fifo_fops
= {
623 .writev
= pipe_writev
,
626 .open
= pipe_rdwr_open
,
627 .release
= pipe_rdwr_release
,
628 .fasync
= pipe_rdwr_fasync
,
631 static struct file_operations read_pipe_fops
= {
638 .open
= pipe_read_open
,
639 .release
= pipe_read_release
,
640 .fasync
= pipe_read_fasync
,
643 static struct file_operations write_pipe_fops
= {
647 .writev
= pipe_writev
,
650 .open
= pipe_write_open
,
651 .release
= pipe_write_release
,
652 .fasync
= pipe_write_fasync
,
655 static struct file_operations rdwr_pipe_fops
= {
660 .writev
= pipe_writev
,
663 .open
= pipe_rdwr_open
,
664 .release
= pipe_rdwr_release
,
665 .fasync
= pipe_rdwr_fasync
,
668 void free_pipe_info(struct inode
*inode
)
671 struct pipe_inode_info
*info
= inode
->i_pipe
;
673 inode
->i_pipe
= NULL
;
674 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
675 struct pipe_buffer
*buf
= info
->bufs
+ i
;
677 buf
->ops
->release(info
, buf
);
680 __free_page(info
->tmp_page
);
684 struct inode
* pipe_new(struct inode
* inode
)
686 struct pipe_inode_info
*info
;
688 info
= kzalloc(sizeof(struct pipe_inode_info
), GFP_KERNEL
);
691 inode
->i_pipe
= info
;
693 init_waitqueue_head(PIPE_WAIT(*inode
));
694 PIPE_RCOUNTER(*inode
) = PIPE_WCOUNTER(*inode
) = 1;
701 static struct vfsmount
*pipe_mnt __read_mostly
;
702 static int pipefs_delete_dentry(struct dentry
*dentry
)
706 static struct dentry_operations pipefs_dentry_operations
= {
707 .d_delete
= pipefs_delete_dentry
,
710 static struct inode
* get_pipe_inode(void)
712 struct inode
*inode
= new_inode(pipe_mnt
->mnt_sb
);
719 PIPE_READERS(*inode
) = PIPE_WRITERS(*inode
) = 1;
720 inode
->i_fop
= &rdwr_pipe_fops
;
723 * Mark the inode dirty from the very beginning,
724 * that way it will never be moved to the dirty
725 * list because "mark_inode_dirty()" will think
726 * that it already _is_ on the dirty list.
728 inode
->i_state
= I_DIRTY
;
729 inode
->i_mode
= S_IFIFO
| S_IRUSR
| S_IWUSR
;
730 inode
->i_uid
= current
->fsuid
;
731 inode
->i_gid
= current
->fsgid
;
732 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
733 inode
->i_blksize
= PAGE_SIZE
;
746 struct dentry
*dentry
;
747 struct inode
* inode
;
748 struct file
*f1
, *f2
;
753 f1
= get_empty_filp();
757 f2
= get_empty_filp();
761 inode
= get_pipe_inode();
765 error
= get_unused_fd();
767 goto close_f12_inode
;
770 error
= get_unused_fd();
772 goto close_f12_inode_i
;
776 sprintf(name
, "[%lu]", inode
->i_ino
);
778 this.len
= strlen(name
);
779 this.hash
= inode
->i_ino
; /* will go */
780 dentry
= d_alloc(pipe_mnt
->mnt_sb
->s_root
, &this);
782 goto close_f12_inode_i_j
;
783 dentry
->d_op
= &pipefs_dentry_operations
;
784 d_add(dentry
, inode
);
785 f1
->f_vfsmnt
= f2
->f_vfsmnt
= mntget(mntget(pipe_mnt
));
786 f1
->f_dentry
= f2
->f_dentry
= dget(dentry
);
787 f1
->f_mapping
= f2
->f_mapping
= inode
->i_mapping
;
790 f1
->f_pos
= f2
->f_pos
= 0;
791 f1
->f_flags
= O_RDONLY
;
792 f1
->f_op
= &read_pipe_fops
;
793 f1
->f_mode
= FMODE_READ
;
797 f2
->f_flags
= O_WRONLY
;
798 f2
->f_op
= &write_pipe_fops
;
799 f2
->f_mode
= FMODE_WRITE
;
813 free_pipe_info(inode
);
824 * pipefs should _never_ be mounted by userland - too much of security hassle,
825 * no real gain from having the whole whorehouse mounted. So we don't need
826 * any operations on the root directory. However, we need a non-trivial
827 * d_name - pipe: will go nicely and kill the special-casing in procfs.
830 static struct super_block
*pipefs_get_sb(struct file_system_type
*fs_type
,
831 int flags
, const char *dev_name
, void *data
)
833 return get_sb_pseudo(fs_type
, "pipe:", NULL
, PIPEFS_MAGIC
);
836 static struct file_system_type pipe_fs_type
= {
838 .get_sb
= pipefs_get_sb
,
839 .kill_sb
= kill_anon_super
,
842 static int __init
init_pipe_fs(void)
844 int err
= register_filesystem(&pipe_fs_type
);
846 pipe_mnt
= kern_mount(&pipe_fs_type
);
847 if (IS_ERR(pipe_mnt
)) {
848 err
= PTR_ERR(pipe_mnt
);
849 unregister_filesystem(&pipe_fs_type
);
855 static void __exit
exit_pipe_fs(void)
857 unregister_filesystem(&pipe_fs_type
);
861 fs_initcall(init_pipe_fs
);
862 module_exit(exit_pipe_fs
);