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/log2.h>
15 #include <linux/mount.h>
16 #include <linux/magic.h>
17 #include <linux/pipe_fs_i.h>
18 #include <linux/uio.h>
19 #include <linux/highmem.h>
20 #include <linux/pagemap.h>
21 #include <linux/audit.h>
22 #include <linux/syscalls.h>
23 #include <linux/fcntl.h>
25 #include <asm/uaccess.h>
26 #include <asm/ioctls.h>
31 * The max size that a non-root user is allowed to grow the pipe. Can
32 * be set by root in /proc/sys/fs/pipe-max-size
34 unsigned int pipe_max_size = 1048576;
37 * Minimum pipe size, as required by POSIX
39 unsigned int pipe_min_size = PAGE_SIZE;
42 * We use a start+len construction, which provides full use of the
44 * -- Florian Coosmann (FGC)
46 * Reads with count = 0 should always return 0.
47 * -- Julian Bradfield 1999-06-07.
49 * FIFOs and Pipes now generate SIGIO for both readers and writers.
50 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
52 * pipe_read & write cleanup
53 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
56 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
59 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
62 void pipe_lock(struct pipe_inode_info *pipe)
65 * pipe_lock() nests non-pipe inode locks (for writing to a file)
67 pipe_lock_nested(pipe, I_MUTEX_PARENT);
69 EXPORT_SYMBOL(pipe_lock);
71 void pipe_unlock(struct pipe_inode_info *pipe)
74 mutex_unlock(&pipe->inode->i_mutex);
76 EXPORT_SYMBOL(pipe_unlock);
78 void pipe_double_lock(struct pipe_inode_info *pipe1,
79 struct pipe_inode_info *pipe2)
81 BUG_ON(pipe1 == pipe2);
84 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
85 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
87 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
88 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
92 /* Drop the inode semaphore and wait for a pipe event, atomically */
93 void pipe_wait(struct pipe_inode_info *pipe)
98 * Pipes are system-local resources, so sleeping on them
99 * is considered a noninteractive wait:
101 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
104 finish_wait(&pipe->wait, &wait);
109 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
115 while (!iov->iov_len)
117 copy = min_t(unsigned long, len, iov->iov_len);
120 if (__copy_from_user_inatomic(to, iov->iov_base, copy))
123 if (copy_from_user(to, iov->iov_base, copy))
128 iov->iov_base += copy;
129 iov->iov_len -= copy;
135 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
141 while (!iov->iov_len)
143 copy = min_t(unsigned long, len, iov->iov_len);
146 if (__copy_to_user_inatomic(iov->iov_base, from, copy))
149 if (copy_to_user(iov->iov_base, from, copy))
154 iov->iov_base += copy;
155 iov->iov_len -= copy;
161 * Attempt to pre-fault in the user memory, so we can use atomic copies.
162 * Returns the number of bytes not faulted in.
164 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
166 while (!iov->iov_len)
170 unsigned long this_len;
172 this_len = min_t(unsigned long, len, iov->iov_len);
173 if (fault_in_pages_writeable(iov->iov_base, this_len))
184 * Pre-fault in the user memory, so we can use atomic copies.
186 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
188 while (!iov->iov_len)
192 unsigned long this_len;
194 this_len = min_t(unsigned long, len, iov->iov_len);
195 fault_in_pages_readable(iov->iov_base, this_len);
201 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
202 struct pipe_buffer *buf)
204 struct page *page = buf->page;
207 * If nobody else uses this page, and we don't already have a
208 * temporary page, let's keep track of it as a one-deep
209 * allocation cache. (Otherwise just release our reference to it)
211 if (page_count(page) == 1 && !pipe->tmp_page)
212 pipe->tmp_page = page;
214 page_cache_release(page);
218 * generic_pipe_buf_map - virtually map a pipe buffer
219 * @pipe: the pipe that the buffer belongs to
220 * @buf: the buffer that should be mapped
221 * @atomic: whether to use an atomic map
224 * This function returns a kernel virtual address mapping for the
225 * pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
226 * and the caller has to be careful not to fault before calling
227 * the unmap function.
229 * Note that this function calls kmap_atomic() if @atomic != 0.
231 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
232 struct pipe_buffer *buf, int atomic)
235 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
236 return kmap_atomic(buf->page);
239 return kmap(buf->page);
241 EXPORT_SYMBOL(generic_pipe_buf_map);
244 * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
245 * @pipe: the pipe that the buffer belongs to
246 * @buf: the buffer that should be unmapped
247 * @map_data: the data that the mapping function returned
250 * This function undoes the mapping that ->map() provided.
252 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
253 struct pipe_buffer *buf, void *map_data)
255 if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
256 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
257 kunmap_atomic(map_data);
261 EXPORT_SYMBOL(generic_pipe_buf_unmap);
264 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
265 * @pipe: the pipe that the buffer belongs to
266 * @buf: the buffer to attempt to steal
269 * This function attempts to steal the &struct page attached to
270 * @buf. If successful, this function returns 0 and returns with
271 * the page locked. The caller may then reuse the page for whatever
272 * he wishes; the typical use is insertion into a different file
275 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
276 struct pipe_buffer *buf)
278 struct page *page = buf->page;
281 * A reference of one is golden, that means that the owner of this
282 * page is the only one holding a reference to it. lock the page
285 if (page_count(page) == 1) {
292 EXPORT_SYMBOL(generic_pipe_buf_steal);
295 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
296 * @pipe: the pipe that the buffer belongs to
297 * @buf: the buffer to get a reference to
300 * This function grabs an extra reference to @buf. It's used in
301 * in the tee() system call, when we duplicate the buffers in one
304 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
306 page_cache_get(buf->page);
308 EXPORT_SYMBOL(generic_pipe_buf_get);
311 * generic_pipe_buf_confirm - verify contents of the pipe buffer
312 * @info: the pipe that the buffer belongs to
313 * @buf: the buffer to confirm
316 * This function does nothing, because the generic pipe code uses
317 * pages that are always good when inserted into the pipe.
319 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
320 struct pipe_buffer *buf)
324 EXPORT_SYMBOL(generic_pipe_buf_confirm);
327 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
328 * @pipe: the pipe that the buffer belongs to
329 * @buf: the buffer to put a reference to
332 * This function releases a reference to @buf.
334 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
335 struct pipe_buffer *buf)
337 page_cache_release(buf->page);
339 EXPORT_SYMBOL(generic_pipe_buf_release);
341 static const struct pipe_buf_operations anon_pipe_buf_ops = {
343 .map = generic_pipe_buf_map,
344 .unmap = generic_pipe_buf_unmap,
345 .confirm = generic_pipe_buf_confirm,
346 .release = anon_pipe_buf_release,
347 .steal = generic_pipe_buf_steal,
348 .get = generic_pipe_buf_get,
351 static const struct pipe_buf_operations packet_pipe_buf_ops = {
353 .map = generic_pipe_buf_map,
354 .unmap = generic_pipe_buf_unmap,
355 .confirm = generic_pipe_buf_confirm,
356 .release = anon_pipe_buf_release,
357 .steal = generic_pipe_buf_steal,
358 .get = generic_pipe_buf_get,
362 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
363 unsigned long nr_segs, loff_t pos)
365 struct file *filp = iocb->ki_filp;
366 struct inode *inode = file_inode(filp);
367 struct pipe_inode_info *pipe;
370 struct iovec *iov = (struct iovec *)_iov;
373 total_len = iov_length(iov, nr_segs);
374 /* Null read succeeds. */
375 if (unlikely(total_len == 0))
380 mutex_lock(&inode->i_mutex);
381 pipe = inode->i_pipe;
383 int bufs = pipe->nrbufs;
385 int curbuf = pipe->curbuf;
386 struct pipe_buffer *buf = pipe->bufs + curbuf;
387 const struct pipe_buf_operations *ops = buf->ops;
389 size_t chars = buf->len;
392 if (chars > total_len)
395 error = ops->confirm(pipe, buf);
402 atomic = !iov_fault_in_pages_write(iov, chars);
404 addr = ops->map(pipe, buf, atomic);
405 error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
406 ops->unmap(pipe, buf, addr);
407 if (unlikely(error)) {
409 * Just retry with the slow path if we failed.
420 buf->offset += chars;
423 /* Was it a packet buffer? Clean up and exit */
424 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
431 ops->release(pipe, buf);
432 curbuf = (curbuf + 1) & (pipe->buffers - 1);
433 pipe->curbuf = curbuf;
434 pipe->nrbufs = --bufs;
439 break; /* common path: read succeeded */
441 if (bufs) /* More to do? */
445 if (!pipe->waiting_writers) {
446 /* syscall merging: Usually we must not sleep
447 * if O_NONBLOCK is set, or if we got some data.
448 * But if a writer sleeps in kernel space, then
449 * we can wait for that data without violating POSIX.
453 if (filp->f_flags & O_NONBLOCK) {
458 if (signal_pending(current)) {
464 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
465 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
469 mutex_unlock(&inode->i_mutex);
471 /* Signal writers asynchronously that there is more room. */
473 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
474 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
481 static inline int is_packetized(struct file *file)
483 return (file->f_flags & O_DIRECT) != 0;
487 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
488 unsigned long nr_segs, loff_t ppos)
490 struct file *filp = iocb->ki_filp;
491 struct inode *inode = file_inode(filp);
492 struct pipe_inode_info *pipe;
495 struct iovec *iov = (struct iovec *)_iov;
499 total_len = iov_length(iov, nr_segs);
500 /* Null write succeeds. */
501 if (unlikely(total_len == 0))
506 mutex_lock(&inode->i_mutex);
507 pipe = inode->i_pipe;
509 if (!pipe->readers) {
510 send_sig(SIGPIPE, current, 0);
515 /* We try to merge small writes */
516 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
517 if (pipe->nrbufs && chars != 0) {
518 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
520 struct pipe_buffer *buf = pipe->bufs + lastbuf;
521 const struct pipe_buf_operations *ops = buf->ops;
522 int offset = buf->offset + buf->len;
524 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
525 int error, atomic = 1;
528 error = ops->confirm(pipe, buf);
532 iov_fault_in_pages_read(iov, chars);
534 addr = ops->map(pipe, buf, atomic);
535 error = pipe_iov_copy_from_user(offset + addr, iov,
537 ops->unmap(pipe, buf, addr);
558 if (!pipe->readers) {
559 send_sig(SIGPIPE, current, 0);
565 if (bufs < pipe->buffers) {
566 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
567 struct pipe_buffer *buf = pipe->bufs + newbuf;
568 struct page *page = pipe->tmp_page;
570 int error, atomic = 1;
573 page = alloc_page(GFP_HIGHUSER);
574 if (unlikely(!page)) {
575 ret = ret ? : -ENOMEM;
578 pipe->tmp_page = page;
580 /* Always wake up, even if the copy fails. Otherwise
581 * we lock up (O_NONBLOCK-)readers that sleep due to
583 * FIXME! Is this really true?
587 if (chars > total_len)
590 iov_fault_in_pages_read(iov, chars);
593 src = kmap_atomic(page);
597 error = pipe_iov_copy_from_user(src, iov, chars,
604 if (unlikely(error)) {
615 /* Insert it into the buffer array */
617 buf->ops = &anon_pipe_buf_ops;
621 if (is_packetized(filp)) {
622 buf->ops = &packet_pipe_buf_ops;
623 buf->flags = PIPE_BUF_FLAG_PACKET;
625 pipe->nrbufs = ++bufs;
626 pipe->tmp_page = NULL;
632 if (bufs < pipe->buffers)
634 if (filp->f_flags & O_NONBLOCK) {
639 if (signal_pending(current)) {
645 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
646 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
649 pipe->waiting_writers++;
651 pipe->waiting_writers--;
654 mutex_unlock(&inode->i_mutex);
656 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
657 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
660 int err = file_update_time(filp);
667 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
669 struct inode *inode = file_inode(filp);
670 struct pipe_inode_info *pipe;
671 int count, buf, nrbufs;
675 mutex_lock(&inode->i_mutex);
676 pipe = inode->i_pipe;
679 nrbufs = pipe->nrbufs;
680 while (--nrbufs >= 0) {
681 count += pipe->bufs[buf].len;
682 buf = (buf+1) & (pipe->buffers - 1);
684 mutex_unlock(&inode->i_mutex);
686 return put_user(count, (int __user *)arg);
692 /* No kernel lock held - fine */
694 pipe_poll(struct file *filp, poll_table *wait)
697 struct inode *inode = file_inode(filp);
698 struct pipe_inode_info *pipe = inode->i_pipe;
701 poll_wait(filp, &pipe->wait, wait);
703 /* Reading only -- no need for acquiring the semaphore. */
704 nrbufs = pipe->nrbufs;
706 if (filp->f_mode & FMODE_READ) {
707 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
708 if (!pipe->writers && filp->f_version != pipe->w_counter)
712 if (filp->f_mode & FMODE_WRITE) {
713 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
715 * Most Unices do not set POLLERR for FIFOs but on Linux they
716 * behave exactly like pipes for poll().
726 pipe_release(struct inode *inode, struct file *file)
728 struct pipe_inode_info *pipe;
730 mutex_lock(&inode->i_mutex);
731 pipe = inode->i_pipe;
732 if (file->f_mode & FMODE_READ)
734 if (file->f_mode & FMODE_WRITE)
737 if (!pipe->readers && !pipe->writers) {
738 free_pipe_info(inode);
740 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
741 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
742 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
744 mutex_unlock(&inode->i_mutex);
750 pipe_fasync(int fd, struct file *filp, int on)
752 struct inode *inode = file_inode(filp);
753 struct pipe_inode_info *pipe = inode->i_pipe;
756 mutex_lock(&inode->i_mutex);
757 if (filp->f_mode & FMODE_READ)
758 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
759 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
760 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
761 if (retval < 0 && (filp->f_mode & FMODE_READ))
762 /* this can happen only if on == T */
763 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
765 mutex_unlock(&inode->i_mutex);
769 struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
771 struct pipe_inode_info *pipe;
773 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
775 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
777 init_waitqueue_head(&pipe->wait);
778 pipe->r_counter = pipe->w_counter = 1;
780 pipe->buffers = PIPE_DEF_BUFFERS;
789 void __free_pipe_info(struct pipe_inode_info *pipe)
793 for (i = 0; i < pipe->buffers; i++) {
794 struct pipe_buffer *buf = pipe->bufs + i;
796 buf->ops->release(pipe, buf);
799 __free_page(pipe->tmp_page);
804 void free_pipe_info(struct inode *inode)
806 __free_pipe_info(inode->i_pipe);
807 inode->i_pipe = NULL;
810 static struct vfsmount *pipe_mnt __read_mostly;
813 * pipefs_dname() is called from d_path().
815 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
817 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
818 dentry->d_inode->i_ino);
821 static const struct dentry_operations pipefs_dentry_operations = {
822 .d_dname = pipefs_dname,
825 static struct inode * get_pipe_inode(void)
827 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
828 struct pipe_inode_info *pipe;
833 inode->i_ino = get_next_ino();
835 pipe = alloc_pipe_info(inode);
838 inode->i_pipe = pipe;
840 pipe->readers = pipe->writers = 1;
841 inode->i_fop = &pipefifo_fops;
844 * Mark the inode dirty from the very beginning,
845 * that way it will never be moved to the dirty
846 * list because "mark_inode_dirty()" will think
847 * that it already _is_ on the dirty list.
849 inode->i_state = I_DIRTY;
850 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
851 inode->i_uid = current_fsuid();
852 inode->i_gid = current_fsgid();
853 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
864 int create_pipe_files(struct file **res, int flags)
867 struct inode *inode = get_pipe_inode();
870 static struct qstr name = { .name = "" };
876 path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
879 path.mnt = mntget(pipe_mnt);
881 d_instantiate(path.dentry, inode);
884 f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
888 f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
890 res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
895 res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
902 free_pipe_info(inode);
907 free_pipe_info(inode);
912 static int __do_pipe_flags(int *fd, struct file **files, int flags)
917 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
920 error = create_pipe_files(files, flags);
924 error = get_unused_fd_flags(flags);
929 error = get_unused_fd_flags(flags);
934 audit_fd_pair(fdr, fdw);
947 int do_pipe_flags(int *fd, int flags)
949 struct file *files[2];
950 int error = __do_pipe_flags(fd, files, flags);
952 fd_install(fd[0], files[0]);
953 fd_install(fd[1], files[1]);
959 * sys_pipe() is the normal C calling standard for creating
960 * a pipe. It's not the way Unix traditionally does this, though.
962 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
964 struct file *files[2];
968 error = __do_pipe_flags(fd, files, flags);
970 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
973 put_unused_fd(fd[0]);
974 put_unused_fd(fd[1]);
977 fd_install(fd[0], files[0]);
978 fd_install(fd[1], files[1]);
984 SYSCALL_DEFINE1(pipe, int __user *, fildes)
986 return sys_pipe2(fildes, 0);
989 static int wait_for_partner(struct inode* inode, unsigned int *cnt)
993 while (cur == *cnt) {
994 pipe_wait(inode->i_pipe);
995 if (signal_pending(current))
998 return cur == *cnt ? -ERESTARTSYS : 0;
1001 static void wake_up_partner(struct inode* inode)
1003 wake_up_interruptible(&inode->i_pipe->wait);
1006 static int fifo_open(struct inode *inode, struct file *filp)
1008 struct pipe_inode_info *pipe;
1009 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1012 mutex_lock(&inode->i_mutex);
1013 pipe = inode->i_pipe;
1016 pipe = alloc_pipe_info(inode);
1019 inode->i_pipe = pipe;
1021 filp->f_version = 0;
1023 /* We can only do regular read/write on fifos */
1024 filp->f_mode &= (FMODE_READ | FMODE_WRITE);
1026 switch (filp->f_mode) {
1030 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1031 * opened, even when there is no process writing the FIFO.
1034 if (pipe->readers++ == 0)
1035 wake_up_partner(inode);
1037 if (!is_pipe && !pipe->writers) {
1038 if ((filp->f_flags & O_NONBLOCK)) {
1039 /* suppress POLLHUP until we have
1041 filp->f_version = pipe->w_counter;
1043 if (wait_for_partner(inode, &pipe->w_counter))
1052 * POSIX.1 says that O_NONBLOCK means return -1 with
1053 * errno=ENXIO when there is no process reading the FIFO.
1056 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1060 if (!pipe->writers++)
1061 wake_up_partner(inode);
1063 if (!is_pipe && !pipe->readers) {
1064 if (wait_for_partner(inode, &pipe->r_counter))
1069 case FMODE_READ | FMODE_WRITE:
1072 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1073 * This implementation will NEVER block on a O_RDWR open, since
1074 * the process can at least talk to itself.
1081 if (pipe->readers == 1 || pipe->writers == 1)
1082 wake_up_partner(inode);
1091 mutex_unlock(&inode->i_mutex);
1095 if (!--pipe->readers)
1096 wake_up_interruptible(&pipe->wait);
1101 if (!--pipe->writers)
1102 wake_up_interruptible(&pipe->wait);
1107 if (!pipe->readers && !pipe->writers)
1108 free_pipe_info(inode);
1111 mutex_unlock(&inode->i_mutex);
1115 const struct file_operations pipefifo_fops = {
1117 .llseek = no_llseek,
1118 .read = do_sync_read,
1119 .aio_read = pipe_read,
1120 .write = do_sync_write,
1121 .aio_write = pipe_write,
1123 .unlocked_ioctl = pipe_ioctl,
1124 .release = pipe_release,
1125 .fasync = pipe_fasync,
1129 * Allocate a new array of pipe buffers and copy the info over. Returns the
1130 * pipe size if successful, or return -ERROR on error.
1132 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1134 struct pipe_buffer *bufs;
1137 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1138 * expect a lot of shrink+grow operations, just free and allocate
1139 * again like we would do for growing. If the pipe currently
1140 * contains more buffers than arg, then return busy.
1142 if (nr_pages < pipe->nrbufs)
1145 bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1146 if (unlikely(!bufs))
1150 * The pipe array wraps around, so just start the new one at zero
1151 * and adjust the indexes.
1157 tail = pipe->curbuf + pipe->nrbufs;
1158 if (tail < pipe->buffers)
1161 tail &= (pipe->buffers - 1);
1163 head = pipe->nrbufs - tail;
1165 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1167 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1173 pipe->buffers = nr_pages;
1174 return nr_pages * PAGE_SIZE;
1178 * Currently we rely on the pipe array holding a power-of-2 number
1181 static inline unsigned int round_pipe_size(unsigned int size)
1183 unsigned long nr_pages;
1185 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1186 return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1190 * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1191 * will return an error.
1193 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1194 size_t *lenp, loff_t *ppos)
1198 ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1199 if (ret < 0 || !write)
1202 pipe_max_size = round_pipe_size(pipe_max_size);
1207 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1208 * location, so checking ->i_pipe is not enough to verify that this is a
1211 struct pipe_inode_info *get_pipe_info(struct file *file)
1213 struct inode *i = file_inode(file);
1215 return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1218 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1220 struct pipe_inode_info *pipe;
1223 pipe = get_pipe_info(file);
1227 mutex_lock(&pipe->inode->i_mutex);
1230 case F_SETPIPE_SZ: {
1231 unsigned int size, nr_pages;
1233 size = round_pipe_size(arg);
1234 nr_pages = size >> PAGE_SHIFT;
1240 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1244 ret = pipe_set_size(pipe, nr_pages);
1248 ret = pipe->buffers * PAGE_SIZE;
1256 mutex_unlock(&pipe->inode->i_mutex);
1260 static const struct super_operations pipefs_ops = {
1261 .destroy_inode = free_inode_nonrcu,
1262 .statfs = simple_statfs,
1266 * pipefs should _never_ be mounted by userland - too much of security hassle,
1267 * no real gain from having the whole whorehouse mounted. So we don't need
1268 * any operations on the root directory. However, we need a non-trivial
1269 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1271 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1272 int flags, const char *dev_name, void *data)
1274 return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1275 &pipefs_dentry_operations, PIPEFS_MAGIC);
1278 static struct file_system_type pipe_fs_type = {
1280 .mount = pipefs_mount,
1281 .kill_sb = kill_anon_super,
1284 static int __init init_pipe_fs(void)
1286 int err = register_filesystem(&pipe_fs_type);
1289 pipe_mnt = kern_mount(&pipe_fs_type);
1290 if (IS_ERR(pipe_mnt)) {
1291 err = PTR_ERR(pipe_mnt);
1292 unregister_filesystem(&pipe_fs_type);
1298 fs_initcall(init_pipe_fs);