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 int anon_pipe_buf_steal(struct pipe_inode_info
*info
,
125 struct pipe_buffer
*buf
)
130 static struct pipe_buf_operations anon_pipe_buf_ops
= {
132 .map
= anon_pipe_buf_map
,
133 .unmap
= anon_pipe_buf_unmap
,
134 .release
= anon_pipe_buf_release
,
135 .steal
= anon_pipe_buf_steal
,
139 pipe_readv(struct file
*filp
, const struct iovec
*_iov
,
140 unsigned long nr_segs
, loff_t
*ppos
)
142 struct inode
*inode
= filp
->f_dentry
->d_inode
;
143 struct pipe_inode_info
*info
;
146 struct iovec
*iov
= (struct iovec
*)_iov
;
149 total_len
= iov_length(iov
, nr_segs
);
150 /* Null read succeeds. */
151 if (unlikely(total_len
== 0))
156 mutex_lock(PIPE_MUTEX(*inode
));
157 info
= inode
->i_pipe
;
159 int bufs
= info
->nrbufs
;
161 int curbuf
= info
->curbuf
;
162 struct pipe_buffer
*buf
= info
->bufs
+ curbuf
;
163 struct pipe_buf_operations
*ops
= buf
->ops
;
165 size_t chars
= buf
->len
;
168 if (chars
> total_len
)
171 addr
= ops
->map(filp
, info
, buf
);
177 error
= pipe_iov_copy_to_user(iov
, addr
+ buf
->offset
, chars
);
178 ops
->unmap(info
, buf
);
179 if (unlikely(error
)) {
180 if (!ret
) ret
= -EFAULT
;
184 buf
->offset
+= chars
;
188 ops
->release(info
, buf
);
189 curbuf
= (curbuf
+ 1) & (PIPE_BUFFERS
-1);
190 info
->curbuf
= curbuf
;
191 info
->nrbufs
= --bufs
;
196 break; /* common path: read succeeded */
198 if (bufs
) /* More to do? */
200 if (!PIPE_WRITERS(*inode
))
202 if (!PIPE_WAITING_WRITERS(*inode
)) {
203 /* syscall merging: Usually we must not sleep
204 * if O_NONBLOCK is set, or if we got some data.
205 * But if a writer sleeps in kernel space, then
206 * we can wait for that data without violating POSIX.
210 if (filp
->f_flags
& O_NONBLOCK
) {
215 if (signal_pending(current
)) {
216 if (!ret
) ret
= -ERESTARTSYS
;
220 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
221 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
225 mutex_unlock(PIPE_MUTEX(*inode
));
226 /* Signal writers asynchronously that there is more room. */
228 wake_up_interruptible(PIPE_WAIT(*inode
));
229 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
237 pipe_read(struct file
*filp
, char __user
*buf
, size_t count
, loff_t
*ppos
)
239 struct iovec iov
= { .iov_base
= buf
, .iov_len
= count
};
240 return pipe_readv(filp
, &iov
, 1, ppos
);
244 pipe_writev(struct file
*filp
, const struct iovec
*_iov
,
245 unsigned long nr_segs
, loff_t
*ppos
)
247 struct inode
*inode
= filp
->f_dentry
->d_inode
;
248 struct pipe_inode_info
*info
;
251 struct iovec
*iov
= (struct iovec
*)_iov
;
255 total_len
= iov_length(iov
, nr_segs
);
256 /* Null write succeeds. */
257 if (unlikely(total_len
== 0))
262 mutex_lock(PIPE_MUTEX(*inode
));
263 info
= inode
->i_pipe
;
265 if (!PIPE_READERS(*inode
)) {
266 send_sig(SIGPIPE
, current
, 0);
271 /* We try to merge small writes */
272 chars
= total_len
& (PAGE_SIZE
-1); /* size of the last buffer */
273 if (info
->nrbufs
&& chars
!= 0) {
274 int lastbuf
= (info
->curbuf
+ info
->nrbufs
- 1) & (PIPE_BUFFERS
-1);
275 struct pipe_buffer
*buf
= info
->bufs
+ lastbuf
;
276 struct pipe_buf_operations
*ops
= buf
->ops
;
277 int offset
= buf
->offset
+ buf
->len
;
278 if (ops
->can_merge
&& offset
+ chars
<= PAGE_SIZE
) {
282 addr
= ops
->map(filp
, info
, buf
);
284 error
= PTR_ERR(addr
);
287 error
= pipe_iov_copy_from_user(offset
+ addr
, iov
,
289 ops
->unmap(info
, buf
);
304 if (!PIPE_READERS(*inode
)) {
305 send_sig(SIGPIPE
, current
, 0);
306 if (!ret
) ret
= -EPIPE
;
310 if (bufs
< PIPE_BUFFERS
) {
311 int newbuf
= (info
->curbuf
+ bufs
) & (PIPE_BUFFERS
-1);
312 struct pipe_buffer
*buf
= info
->bufs
+ newbuf
;
313 struct page
*page
= info
->tmp_page
;
317 page
= alloc_page(GFP_HIGHUSER
);
318 if (unlikely(!page
)) {
319 ret
= ret
? : -ENOMEM
;
322 info
->tmp_page
= page
;
324 /* Always wakeup, even if the copy fails. Otherwise
325 * we lock up (O_NONBLOCK-)readers that sleep due to
327 * FIXME! Is this really true?
331 if (chars
> total_len
)
334 error
= pipe_iov_copy_from_user(kmap(page
), iov
, chars
);
336 if (unlikely(error
)) {
337 if (!ret
) ret
= -EFAULT
;
342 /* Insert it into the buffer array */
344 buf
->ops
= &anon_pipe_buf_ops
;
347 info
->nrbufs
= ++bufs
;
348 info
->tmp_page
= NULL
;
354 if (bufs
< PIPE_BUFFERS
)
356 if (filp
->f_flags
& O_NONBLOCK
) {
357 if (!ret
) ret
= -EAGAIN
;
360 if (signal_pending(current
)) {
361 if (!ret
) ret
= -ERESTARTSYS
;
365 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
366 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
369 PIPE_WAITING_WRITERS(*inode
)++;
371 PIPE_WAITING_WRITERS(*inode
)--;
374 mutex_unlock(PIPE_MUTEX(*inode
));
376 wake_up_interruptible(PIPE_WAIT(*inode
));
377 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
380 file_update_time(filp
);
385 pipe_write(struct file
*filp
, const char __user
*buf
,
386 size_t count
, loff_t
*ppos
)
388 struct iovec iov
= { .iov_base
= (void __user
*)buf
, .iov_len
= count
};
389 return pipe_writev(filp
, &iov
, 1, ppos
);
393 bad_pipe_r(struct file
*filp
, char __user
*buf
, size_t count
, loff_t
*ppos
)
399 bad_pipe_w(struct file
*filp
, const char __user
*buf
, size_t count
, loff_t
*ppos
)
405 pipe_ioctl(struct inode
*pino
, struct file
*filp
,
406 unsigned int cmd
, unsigned long arg
)
408 struct inode
*inode
= filp
->f_dentry
->d_inode
;
409 struct pipe_inode_info
*info
;
410 int count
, buf
, nrbufs
;
414 mutex_lock(PIPE_MUTEX(*inode
));
415 info
= inode
->i_pipe
;
418 nrbufs
= info
->nrbufs
;
419 while (--nrbufs
>= 0) {
420 count
+= info
->bufs
[buf
].len
;
421 buf
= (buf
+1) & (PIPE_BUFFERS
-1);
423 mutex_unlock(PIPE_MUTEX(*inode
));
424 return put_user(count
, (int __user
*)arg
);
430 /* No kernel lock held - fine */
432 pipe_poll(struct file
*filp
, poll_table
*wait
)
435 struct inode
*inode
= filp
->f_dentry
->d_inode
;
436 struct pipe_inode_info
*info
= inode
->i_pipe
;
439 poll_wait(filp
, PIPE_WAIT(*inode
), wait
);
441 /* Reading only -- no need for acquiring the semaphore. */
442 nrbufs
= info
->nrbufs
;
444 if (filp
->f_mode
& FMODE_READ
) {
445 mask
= (nrbufs
> 0) ? POLLIN
| POLLRDNORM
: 0;
446 if (!PIPE_WRITERS(*inode
) && filp
->f_version
!= PIPE_WCOUNTER(*inode
))
450 if (filp
->f_mode
& FMODE_WRITE
) {
451 mask
|= (nrbufs
< PIPE_BUFFERS
) ? POLLOUT
| POLLWRNORM
: 0;
453 * Most Unices do not set POLLERR for FIFOs but on Linux they
454 * behave exactly like pipes for poll().
456 if (!PIPE_READERS(*inode
))
464 pipe_release(struct inode
*inode
, int decr
, int decw
)
466 mutex_lock(PIPE_MUTEX(*inode
));
467 PIPE_READERS(*inode
) -= decr
;
468 PIPE_WRITERS(*inode
) -= decw
;
469 if (!PIPE_READERS(*inode
) && !PIPE_WRITERS(*inode
)) {
470 free_pipe_info(inode
);
472 wake_up_interruptible(PIPE_WAIT(*inode
));
473 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
474 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
476 mutex_unlock(PIPE_MUTEX(*inode
));
482 pipe_read_fasync(int fd
, struct file
*filp
, int on
)
484 struct inode
*inode
= filp
->f_dentry
->d_inode
;
487 mutex_lock(PIPE_MUTEX(*inode
));
488 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_READERS(*inode
));
489 mutex_unlock(PIPE_MUTEX(*inode
));
499 pipe_write_fasync(int fd
, struct file
*filp
, int on
)
501 struct inode
*inode
= filp
->f_dentry
->d_inode
;
504 mutex_lock(PIPE_MUTEX(*inode
));
505 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_WRITERS(*inode
));
506 mutex_unlock(PIPE_MUTEX(*inode
));
516 pipe_rdwr_fasync(int fd
, struct file
*filp
, int on
)
518 struct inode
*inode
= filp
->f_dentry
->d_inode
;
521 mutex_lock(PIPE_MUTEX(*inode
));
523 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_READERS(*inode
));
526 retval
= fasync_helper(fd
, filp
, on
, PIPE_FASYNC_WRITERS(*inode
));
528 mutex_unlock(PIPE_MUTEX(*inode
));
538 pipe_read_release(struct inode
*inode
, struct file
*filp
)
540 pipe_read_fasync(-1, filp
, 0);
541 return pipe_release(inode
, 1, 0);
545 pipe_write_release(struct inode
*inode
, struct file
*filp
)
547 pipe_write_fasync(-1, filp
, 0);
548 return pipe_release(inode
, 0, 1);
552 pipe_rdwr_release(struct inode
*inode
, struct file
*filp
)
556 pipe_rdwr_fasync(-1, filp
, 0);
557 decr
= (filp
->f_mode
& FMODE_READ
) != 0;
558 decw
= (filp
->f_mode
& FMODE_WRITE
) != 0;
559 return pipe_release(inode
, decr
, decw
);
563 pipe_read_open(struct inode
*inode
, struct file
*filp
)
565 /* We could have perhaps used atomic_t, but this and friends
566 below are the only places. So it doesn't seem worthwhile. */
567 mutex_lock(PIPE_MUTEX(*inode
));
568 PIPE_READERS(*inode
)++;
569 mutex_unlock(PIPE_MUTEX(*inode
));
575 pipe_write_open(struct inode
*inode
, struct file
*filp
)
577 mutex_lock(PIPE_MUTEX(*inode
));
578 PIPE_WRITERS(*inode
)++;
579 mutex_unlock(PIPE_MUTEX(*inode
));
585 pipe_rdwr_open(struct inode
*inode
, struct file
*filp
)
587 mutex_lock(PIPE_MUTEX(*inode
));
588 if (filp
->f_mode
& FMODE_READ
)
589 PIPE_READERS(*inode
)++;
590 if (filp
->f_mode
& FMODE_WRITE
)
591 PIPE_WRITERS(*inode
)++;
592 mutex_unlock(PIPE_MUTEX(*inode
));
598 * The file_operations structs are not static because they
599 * are also used in linux/fs/fifo.c to do operations on FIFOs.
601 const struct file_operations read_fifo_fops
= {
608 .open
= pipe_read_open
,
609 .release
= pipe_read_release
,
610 .fasync
= pipe_read_fasync
,
613 const struct file_operations write_fifo_fops
= {
617 .writev
= pipe_writev
,
620 .open
= pipe_write_open
,
621 .release
= pipe_write_release
,
622 .fasync
= pipe_write_fasync
,
625 const struct file_operations rdwr_fifo_fops
= {
630 .writev
= pipe_writev
,
633 .open
= pipe_rdwr_open
,
634 .release
= pipe_rdwr_release
,
635 .fasync
= pipe_rdwr_fasync
,
638 static struct file_operations read_pipe_fops
= {
645 .open
= pipe_read_open
,
646 .release
= pipe_read_release
,
647 .fasync
= pipe_read_fasync
,
650 static struct file_operations write_pipe_fops
= {
654 .writev
= pipe_writev
,
657 .open
= pipe_write_open
,
658 .release
= pipe_write_release
,
659 .fasync
= pipe_write_fasync
,
662 static struct file_operations rdwr_pipe_fops
= {
667 .writev
= pipe_writev
,
670 .open
= pipe_rdwr_open
,
671 .release
= pipe_rdwr_release
,
672 .fasync
= pipe_rdwr_fasync
,
675 void free_pipe_info(struct inode
*inode
)
678 struct pipe_inode_info
*info
= inode
->i_pipe
;
680 inode
->i_pipe
= NULL
;
681 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
682 struct pipe_buffer
*buf
= info
->bufs
+ i
;
684 buf
->ops
->release(info
, buf
);
687 __free_page(info
->tmp_page
);
691 struct inode
* pipe_new(struct inode
* inode
)
693 struct pipe_inode_info
*info
;
695 info
= kzalloc(sizeof(struct pipe_inode_info
), GFP_KERNEL
);
698 inode
->i_pipe
= info
;
700 init_waitqueue_head(PIPE_WAIT(*inode
));
701 PIPE_RCOUNTER(*inode
) = PIPE_WCOUNTER(*inode
) = 1;
708 static struct vfsmount
*pipe_mnt __read_mostly
;
709 static int pipefs_delete_dentry(struct dentry
*dentry
)
713 static struct dentry_operations pipefs_dentry_operations
= {
714 .d_delete
= pipefs_delete_dentry
,
717 static struct inode
* get_pipe_inode(void)
719 struct inode
*inode
= new_inode(pipe_mnt
->mnt_sb
);
726 PIPE_READERS(*inode
) = PIPE_WRITERS(*inode
) = 1;
727 inode
->i_fop
= &rdwr_pipe_fops
;
730 * Mark the inode dirty from the very beginning,
731 * that way it will never be moved to the dirty
732 * list because "mark_inode_dirty()" will think
733 * that it already _is_ on the dirty list.
735 inode
->i_state
= I_DIRTY
;
736 inode
->i_mode
= S_IFIFO
| S_IRUSR
| S_IWUSR
;
737 inode
->i_uid
= current
->fsuid
;
738 inode
->i_gid
= current
->fsgid
;
739 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
740 inode
->i_blksize
= PAGE_SIZE
;
753 struct dentry
*dentry
;
754 struct inode
* inode
;
755 struct file
*f1
, *f2
;
760 f1
= get_empty_filp();
764 f2
= get_empty_filp();
768 inode
= get_pipe_inode();
772 error
= get_unused_fd();
774 goto close_f12_inode
;
777 error
= get_unused_fd();
779 goto close_f12_inode_i
;
783 sprintf(name
, "[%lu]", inode
->i_ino
);
785 this.len
= strlen(name
);
786 this.hash
= inode
->i_ino
; /* will go */
787 dentry
= d_alloc(pipe_mnt
->mnt_sb
->s_root
, &this);
789 goto close_f12_inode_i_j
;
790 dentry
->d_op
= &pipefs_dentry_operations
;
791 d_add(dentry
, inode
);
792 f1
->f_vfsmnt
= f2
->f_vfsmnt
= mntget(mntget(pipe_mnt
));
793 f1
->f_dentry
= f2
->f_dentry
= dget(dentry
);
794 f1
->f_mapping
= f2
->f_mapping
= inode
->i_mapping
;
797 f1
->f_pos
= f2
->f_pos
= 0;
798 f1
->f_flags
= O_RDONLY
;
799 f1
->f_op
= &read_pipe_fops
;
800 f1
->f_mode
= FMODE_READ
;
804 f2
->f_flags
= O_WRONLY
;
805 f2
->f_op
= &write_pipe_fops
;
806 f2
->f_mode
= FMODE_WRITE
;
820 free_pipe_info(inode
);
831 * pipefs should _never_ be mounted by userland - too much of security hassle,
832 * no real gain from having the whole whorehouse mounted. So we don't need
833 * any operations on the root directory. However, we need a non-trivial
834 * d_name - pipe: will go nicely and kill the special-casing in procfs.
837 static struct super_block
*pipefs_get_sb(struct file_system_type
*fs_type
,
838 int flags
, const char *dev_name
, void *data
)
840 return get_sb_pseudo(fs_type
, "pipe:", NULL
, PIPEFS_MAGIC
);
843 static struct file_system_type pipe_fs_type
= {
845 .get_sb
= pipefs_get_sb
,
846 .kill_sb
= kill_anon_super
,
849 static int __init
init_pipe_fs(void)
851 int err
= register_filesystem(&pipe_fs_type
);
853 pipe_mnt
= kern_mount(&pipe_fs_type
);
854 if (IS_ERR(pipe_mnt
)) {
855 err
= PTR_ERR(pipe_mnt
);
856 unregister_filesystem(&pipe_fs_type
);
862 static void __exit
exit_pipe_fs(void)
864 unregister_filesystem(&pipe_fs_type
);
868 fs_initcall(init_pipe_fs
);
869 module_exit(exit_pipe_fs
);