1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
9 #include <linux/file.h>
10 #include <linux/poll.h>
11 #include <linux/slab.h>
12 #include <linux/module.h>
13 #include <linux/init.h>
15 #include <linux/log2.h>
16 #include <linux/mount.h>
17 #include <linux/pseudo_fs.h>
18 #include <linux/magic.h>
19 #include <linux/pipe_fs_i.h>
20 #include <linux/uio.h>
21 #include <linux/highmem.h>
22 #include <linux/pagemap.h>
23 #include <linux/audit.h>
24 #include <linux/syscalls.h>
25 #include <linux/fcntl.h>
26 #include <linux/memcontrol.h>
27 #include <linux/watch_queue.h>
29 #include <linux/uaccess.h>
30 #include <asm/ioctls.h>
35 * The max size that a non-root user is allowed to grow the pipe. Can
36 * be set by root in /proc/sys/fs/pipe-max-size
38 unsigned int pipe_max_size = 1048576;
40 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
41 * matches default values.
43 unsigned long pipe_user_pages_hard;
44 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
47 * We use head and tail indices that aren't masked off, except at the point of
48 * dereference, but rather they're allowed to wrap naturally. This means there
49 * isn't a dead spot in the buffer, but the ring has to be a power of two and
51 * -- David Howells 2019-09-23.
53 * Reads with count = 0 should always return 0.
54 * -- Julian Bradfield 1999-06-07.
56 * FIFOs and Pipes now generate SIGIO for both readers and writers.
57 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
59 * pipe_read & write cleanup
60 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
63 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
66 mutex_lock_nested(&pipe->mutex, subclass);
69 void pipe_lock(struct pipe_inode_info *pipe)
72 * pipe_lock() nests non-pipe inode locks (for writing to a file)
74 pipe_lock_nested(pipe, I_MUTEX_PARENT);
76 EXPORT_SYMBOL(pipe_lock);
78 void pipe_unlock(struct pipe_inode_info *pipe)
81 mutex_unlock(&pipe->mutex);
83 EXPORT_SYMBOL(pipe_unlock);
85 static inline void __pipe_lock(struct pipe_inode_info *pipe)
87 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
90 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
92 mutex_unlock(&pipe->mutex);
95 void pipe_double_lock(struct pipe_inode_info *pipe1,
96 struct pipe_inode_info *pipe2)
98 BUG_ON(pipe1 == pipe2);
101 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
102 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
104 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
105 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
109 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
110 struct pipe_buffer *buf)
112 struct page *page = buf->page;
115 * If nobody else uses this page, and we don't already have a
116 * temporary page, let's keep track of it as a one-deep
117 * allocation cache. (Otherwise just release our reference to it)
119 if (page_count(page) == 1 && !pipe->tmp_page)
120 pipe->tmp_page = page;
125 static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
126 struct pipe_buffer *buf)
128 struct page *page = buf->page;
130 if (page_count(page) != 1)
132 memcg_kmem_uncharge_page(page, 0);
133 __SetPageLocked(page);
138 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
139 * @pipe: the pipe that the buffer belongs to
140 * @buf: the buffer to attempt to steal
143 * This function attempts to steal the &struct page attached to
144 * @buf. If successful, this function returns 0 and returns with
145 * the page locked. The caller may then reuse the page for whatever
146 * he wishes; the typical use is insertion into a different file
149 bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
150 struct pipe_buffer *buf)
152 struct page *page = buf->page;
155 * A reference of one is golden, that means that the owner of this
156 * page is the only one holding a reference to it. lock the page
159 if (page_count(page) == 1) {
165 EXPORT_SYMBOL(generic_pipe_buf_try_steal);
168 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
169 * @pipe: the pipe that the buffer belongs to
170 * @buf: the buffer to get a reference to
173 * This function grabs an extra reference to @buf. It's used in
174 * the tee() system call, when we duplicate the buffers in one
177 bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
179 return try_get_page(buf->page);
181 EXPORT_SYMBOL(generic_pipe_buf_get);
184 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
185 * @pipe: the pipe that the buffer belongs to
186 * @buf: the buffer to put a reference to
189 * This function releases a reference to @buf.
191 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
192 struct pipe_buffer *buf)
196 EXPORT_SYMBOL(generic_pipe_buf_release);
198 static const struct pipe_buf_operations anon_pipe_buf_ops = {
199 .release = anon_pipe_buf_release,
200 .try_steal = anon_pipe_buf_try_steal,
201 .get = generic_pipe_buf_get,
204 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
205 static inline bool pipe_readable(const struct pipe_inode_info *pipe)
207 unsigned int head = READ_ONCE(pipe->head);
208 unsigned int tail = READ_ONCE(pipe->tail);
209 unsigned int writers = READ_ONCE(pipe->writers);
211 return !pipe_empty(head, tail) || !writers;
215 pipe_read(struct kiocb *iocb, struct iov_iter *to)
217 size_t total_len = iov_iter_count(to);
218 struct file *filp = iocb->ki_filp;
219 struct pipe_inode_info *pipe = filp->private_data;
220 bool was_full, wake_next_reader = false;
223 /* Null read succeeds. */
224 if (unlikely(total_len == 0))
231 * We only wake up writers if the pipe was full when we started
232 * reading in order to avoid unnecessary wakeups.
234 * But when we do wake up writers, we do so using a sync wakeup
235 * (WF_SYNC), because we want them to get going and generate more
238 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
240 unsigned int head = pipe->head;
241 unsigned int tail = pipe->tail;
242 unsigned int mask = pipe->ring_size - 1;
244 #ifdef CONFIG_WATCH_QUEUE
245 if (pipe->note_loss) {
246 struct watch_notification n;
254 n.type = WATCH_TYPE_META;
255 n.subtype = WATCH_META_LOSS_NOTIFICATION;
256 n.info = watch_sizeof(n);
257 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
263 total_len -= sizeof(n);
264 pipe->note_loss = false;
268 if (!pipe_empty(head, tail)) {
269 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
270 size_t chars = buf->len;
274 if (chars > total_len) {
275 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
283 error = pipe_buf_confirm(pipe, buf);
290 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
291 if (unlikely(written < chars)) {
297 buf->offset += chars;
300 /* Was it a packet buffer? Clean up and exit */
301 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
307 pipe_buf_release(pipe, buf);
308 spin_lock_irq(&pipe->rd_wait.lock);
309 #ifdef CONFIG_WATCH_QUEUE
310 if (buf->flags & PIPE_BUF_FLAG_LOSS)
311 pipe->note_loss = true;
315 spin_unlock_irq(&pipe->rd_wait.lock);
319 break; /* common path: read succeeded */
320 if (!pipe_empty(head, tail)) /* More to do? */
328 if (filp->f_flags & O_NONBLOCK) {
335 * We only get here if we didn't actually read anything.
337 * However, we could have seen (and removed) a zero-sized
338 * pipe buffer, and might have made space in the buffers
341 * You can't make zero-sized pipe buffers by doing an empty
342 * write (not even in packet mode), but they can happen if
343 * the writer gets an EFAULT when trying to fill a buffer
344 * that already got allocated and inserted in the buffer
347 * So we still need to wake up any pending writers in the
348 * _very_ unlikely case that the pipe was full, but we got
351 if (unlikely(was_full)) {
352 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
353 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
357 * But because we didn't read anything, at this point we can
358 * just return directly with -ERESTARTSYS if we're interrupted,
359 * since we've done any required wakeups and there's no need
360 * to mark anything accessed. And we've dropped the lock.
362 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
366 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
367 wake_next_reader = true;
369 if (pipe_empty(pipe->head, pipe->tail))
370 wake_next_reader = false;
374 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
375 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
377 if (wake_next_reader)
378 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
384 static inline int is_packetized(struct file *file)
386 return (file->f_flags & O_DIRECT) != 0;
389 /* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
390 static inline bool pipe_writable(const struct pipe_inode_info *pipe)
392 unsigned int head = READ_ONCE(pipe->head);
393 unsigned int tail = READ_ONCE(pipe->tail);
394 unsigned int max_usage = READ_ONCE(pipe->max_usage);
396 return !pipe_full(head, tail, max_usage) ||
397 !READ_ONCE(pipe->readers);
401 pipe_write(struct kiocb *iocb, struct iov_iter *from)
403 struct file *filp = iocb->ki_filp;
404 struct pipe_inode_info *pipe = filp->private_data;
407 size_t total_len = iov_iter_count(from);
409 bool was_empty = false;
410 bool wake_next_writer = false;
412 /* Null write succeeds. */
413 if (unlikely(total_len == 0))
418 if (!pipe->readers) {
419 send_sig(SIGPIPE, current, 0);
424 #ifdef CONFIG_WATCH_QUEUE
425 if (pipe->watch_queue) {
432 * Only wake up if the pipe started out empty, since
433 * otherwise there should be no readers waiting.
435 * If it wasn't empty we try to merge new data into
438 * That naturally merges small writes, but it also
439 * page-aligs the rest of the writes for large writes
440 * spanning multiple pages.
443 was_empty = pipe_empty(head, pipe->tail);
444 chars = total_len & (PAGE_SIZE-1);
445 if (chars && !was_empty) {
446 unsigned int mask = pipe->ring_size - 1;
447 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
448 int offset = buf->offset + buf->len;
450 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
451 offset + chars <= PAGE_SIZE) {
452 ret = pipe_buf_confirm(pipe, buf);
456 ret = copy_page_from_iter(buf->page, offset, chars, from);
457 if (unlikely(ret < chars)) {
463 if (!iov_iter_count(from))
469 if (!pipe->readers) {
470 send_sig(SIGPIPE, current, 0);
477 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
478 unsigned int mask = pipe->ring_size - 1;
479 struct pipe_buffer *buf = &pipe->bufs[head & mask];
480 struct page *page = pipe->tmp_page;
484 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
485 if (unlikely(!page)) {
486 ret = ret ? : -ENOMEM;
489 pipe->tmp_page = page;
492 /* Allocate a slot in the ring in advance and attach an
493 * empty buffer. If we fault or otherwise fail to use
494 * it, either the reader will consume it or it'll still
495 * be there for the next write.
497 spin_lock_irq(&pipe->rd_wait.lock);
500 if (pipe_full(head, pipe->tail, pipe->max_usage)) {
501 spin_unlock_irq(&pipe->rd_wait.lock);
505 pipe->head = head + 1;
506 spin_unlock_irq(&pipe->rd_wait.lock);
508 /* Insert it into the buffer array */
509 buf = &pipe->bufs[head & mask];
511 buf->ops = &anon_pipe_buf_ops;
514 if (is_packetized(filp))
515 buf->flags = PIPE_BUF_FLAG_PACKET;
517 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
518 pipe->tmp_page = NULL;
520 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
521 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
530 if (!iov_iter_count(from))
534 if (!pipe_full(head, pipe->tail, pipe->max_usage))
537 /* Wait for buffer space to become available. */
538 if (filp->f_flags & O_NONBLOCK) {
543 if (signal_pending(current)) {
550 * We're going to release the pipe lock and wait for more
551 * space. We wake up any readers if necessary, and then
552 * after waiting we need to re-check whether the pipe
553 * become empty while we dropped the lock.
557 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
558 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
560 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
562 was_empty = pipe_empty(pipe->head, pipe->tail);
563 wake_next_writer = true;
566 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
567 wake_next_writer = false;
571 * If we do do a wakeup event, we do a 'sync' wakeup, because we
572 * want the reader to start processing things asap, rather than
573 * leave the data pending.
575 * This is particularly important for small writes, because of
576 * how (for example) the GNU make jobserver uses small writes to
577 * wake up pending jobs
580 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
581 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
583 if (wake_next_writer)
584 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
585 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
586 int err = file_update_time(filp);
589 sb_end_write(file_inode(filp)->i_sb);
594 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
596 struct pipe_inode_info *pipe = filp->private_data;
597 int count, head, tail, mask;
605 mask = pipe->ring_size - 1;
607 while (tail != head) {
608 count += pipe->bufs[tail & mask].len;
613 return put_user(count, (int __user *)arg);
615 #ifdef CONFIG_WATCH_QUEUE
616 case IOC_WATCH_QUEUE_SET_SIZE: {
619 ret = watch_queue_set_size(pipe, arg);
624 case IOC_WATCH_QUEUE_SET_FILTER:
625 return watch_queue_set_filter(
626 pipe, (struct watch_notification_filter __user *)arg);
634 /* No kernel lock held - fine */
636 pipe_poll(struct file *filp, poll_table *wait)
639 struct pipe_inode_info *pipe = filp->private_data;
640 unsigned int head, tail;
643 * Reading pipe state only -- no need for acquiring the semaphore.
645 * But because this is racy, the code has to add the
646 * entry to the poll table _first_ ..
648 if (filp->f_mode & FMODE_READ)
649 poll_wait(filp, &pipe->rd_wait, wait);
650 if (filp->f_mode & FMODE_WRITE)
651 poll_wait(filp, &pipe->wr_wait, wait);
654 * .. and only then can you do the racy tests. That way,
655 * if something changes and you got it wrong, the poll
656 * table entry will wake you up and fix it.
658 head = READ_ONCE(pipe->head);
659 tail = READ_ONCE(pipe->tail);
662 if (filp->f_mode & FMODE_READ) {
663 if (!pipe_empty(head, tail))
664 mask |= EPOLLIN | EPOLLRDNORM;
665 if (!pipe->writers && filp->f_version != pipe->w_counter)
669 if (filp->f_mode & FMODE_WRITE) {
670 if (!pipe_full(head, tail, pipe->max_usage))
671 mask |= EPOLLOUT | EPOLLWRNORM;
673 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
674 * behave exactly like pipes for poll().
683 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
687 spin_lock(&inode->i_lock);
688 if (!--pipe->files) {
689 inode->i_pipe = NULL;
692 spin_unlock(&inode->i_lock);
695 free_pipe_info(pipe);
699 pipe_release(struct inode *inode, struct file *file)
701 struct pipe_inode_info *pipe = file->private_data;
704 if (file->f_mode & FMODE_READ)
706 if (file->f_mode & FMODE_WRITE)
709 /* Was that the last reader or writer, but not the other side? */
710 if (!pipe->readers != !pipe->writers) {
711 wake_up_interruptible_all(&pipe->rd_wait);
712 wake_up_interruptible_all(&pipe->wr_wait);
713 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
714 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
718 put_pipe_info(inode, pipe);
723 pipe_fasync(int fd, struct file *filp, int on)
725 struct pipe_inode_info *pipe = filp->private_data;
729 if (filp->f_mode & FMODE_READ)
730 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
731 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
732 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
733 if (retval < 0 && (filp->f_mode & FMODE_READ))
734 /* this can happen only if on == T */
735 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
741 unsigned long account_pipe_buffers(struct user_struct *user,
742 unsigned long old, unsigned long new)
744 return atomic_long_add_return(new - old, &user->pipe_bufs);
747 bool too_many_pipe_buffers_soft(unsigned long user_bufs)
749 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
751 return soft_limit && user_bufs > soft_limit;
754 bool too_many_pipe_buffers_hard(unsigned long user_bufs)
756 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
758 return hard_limit && user_bufs > hard_limit;
761 bool pipe_is_unprivileged_user(void)
763 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
766 struct pipe_inode_info *alloc_pipe_info(void)
768 struct pipe_inode_info *pipe;
769 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
770 struct user_struct *user = get_current_user();
771 unsigned long user_bufs;
772 unsigned int max_size = READ_ONCE(pipe_max_size);
774 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
778 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
779 pipe_bufs = max_size >> PAGE_SHIFT;
781 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
783 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
784 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
788 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
789 goto out_revert_acct;
791 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
795 init_waitqueue_head(&pipe->rd_wait);
796 init_waitqueue_head(&pipe->wr_wait);
797 pipe->r_counter = pipe->w_counter = 1;
798 pipe->max_usage = pipe_bufs;
799 pipe->ring_size = pipe_bufs;
800 pipe->nr_accounted = pipe_bufs;
802 mutex_init(&pipe->mutex);
807 (void) account_pipe_buffers(user, pipe_bufs, 0);
814 void free_pipe_info(struct pipe_inode_info *pipe)
818 #ifdef CONFIG_WATCH_QUEUE
819 if (pipe->watch_queue) {
820 watch_queue_clear(pipe->watch_queue);
821 put_watch_queue(pipe->watch_queue);
825 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
826 free_uid(pipe->user);
827 for (i = 0; i < pipe->ring_size; i++) {
828 struct pipe_buffer *buf = pipe->bufs + i;
830 pipe_buf_release(pipe, buf);
833 __free_page(pipe->tmp_page);
838 static struct vfsmount *pipe_mnt __read_mostly;
841 * pipefs_dname() is called from d_path().
843 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
845 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
846 d_inode(dentry)->i_ino);
849 static const struct dentry_operations pipefs_dentry_operations = {
850 .d_dname = pipefs_dname,
853 static struct inode * get_pipe_inode(void)
855 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
856 struct pipe_inode_info *pipe;
861 inode->i_ino = get_next_ino();
863 pipe = alloc_pipe_info();
867 inode->i_pipe = pipe;
869 pipe->readers = pipe->writers = 1;
870 inode->i_fop = &pipefifo_fops;
873 * Mark the inode dirty from the very beginning,
874 * that way it will never be moved to the dirty
875 * list because "mark_inode_dirty()" will think
876 * that it already _is_ on the dirty list.
878 inode->i_state = I_DIRTY;
879 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
880 inode->i_uid = current_fsuid();
881 inode->i_gid = current_fsgid();
882 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
893 int create_pipe_files(struct file **res, int flags)
895 struct inode *inode = get_pipe_inode();
902 if (flags & O_NOTIFICATION_PIPE) {
903 error = watch_queue_init(inode->i_pipe);
905 free_pipe_info(inode->i_pipe);
911 f = alloc_file_pseudo(inode, pipe_mnt, "",
912 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
915 free_pipe_info(inode->i_pipe);
920 f->private_data = inode->i_pipe;
922 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
924 if (IS_ERR(res[0])) {
925 put_pipe_info(inode, inode->i_pipe);
927 return PTR_ERR(res[0]);
929 res[0]->private_data = inode->i_pipe;
931 stream_open(inode, res[0]);
932 stream_open(inode, res[1]);
936 static int __do_pipe_flags(int *fd, struct file **files, int flags)
941 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
944 error = create_pipe_files(files, flags);
948 error = get_unused_fd_flags(flags);
953 error = get_unused_fd_flags(flags);
958 audit_fd_pair(fdr, fdw);
971 int do_pipe_flags(int *fd, int flags)
973 struct file *files[2];
974 int error = __do_pipe_flags(fd, files, flags);
976 fd_install(fd[0], files[0]);
977 fd_install(fd[1], files[1]);
983 * sys_pipe() is the normal C calling standard for creating
984 * a pipe. It's not the way Unix traditionally does this, though.
986 static int do_pipe2(int __user *fildes, int flags)
988 struct file *files[2];
992 error = __do_pipe_flags(fd, files, flags);
994 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
997 put_unused_fd(fd[0]);
998 put_unused_fd(fd[1]);
1001 fd_install(fd[0], files[0]);
1002 fd_install(fd[1], files[1]);
1008 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1010 return do_pipe2(fildes, flags);
1013 SYSCALL_DEFINE1(pipe, int __user *, fildes)
1015 return do_pipe2(fildes, 0);
1019 * This is the stupid "wait for pipe to be readable or writable"
1022 * See pipe_read/write() for the proper kind of exclusive wait,
1023 * but that requires that we wake up any other readers/writers
1024 * if we then do not end up reading everything (ie the whole
1025 * "wake_next_reader/writer" logic in pipe_read/write()).
1027 void pipe_wait_readable(struct pipe_inode_info *pipe)
1030 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1034 void pipe_wait_writable(struct pipe_inode_info *pipe)
1037 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1042 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1043 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1044 * race with the count check and waitqueue prep.
1046 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1047 * then check the condition you're waiting for, and only then sleep. But
1048 * because of the pipe lock, we can check the condition before being on
1051 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1053 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1055 DEFINE_WAIT(rdwait);
1058 while (cur == *cnt) {
1059 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1062 finish_wait(&pipe->rd_wait, &rdwait);
1064 if (signal_pending(current))
1067 return cur == *cnt ? -ERESTARTSYS : 0;
1070 static void wake_up_partner(struct pipe_inode_info *pipe)
1072 wake_up_interruptible_all(&pipe->rd_wait);
1075 static int fifo_open(struct inode *inode, struct file *filp)
1077 struct pipe_inode_info *pipe;
1078 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1081 filp->f_version = 0;
1083 spin_lock(&inode->i_lock);
1084 if (inode->i_pipe) {
1085 pipe = inode->i_pipe;
1087 spin_unlock(&inode->i_lock);
1089 spin_unlock(&inode->i_lock);
1090 pipe = alloc_pipe_info();
1094 spin_lock(&inode->i_lock);
1095 if (unlikely(inode->i_pipe)) {
1096 inode->i_pipe->files++;
1097 spin_unlock(&inode->i_lock);
1098 free_pipe_info(pipe);
1099 pipe = inode->i_pipe;
1101 inode->i_pipe = pipe;
1102 spin_unlock(&inode->i_lock);
1105 filp->private_data = pipe;
1106 /* OK, we have a pipe and it's pinned down */
1110 /* We can only do regular read/write on fifos */
1111 stream_open(inode, filp);
1113 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1117 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1118 * opened, even when there is no process writing the FIFO.
1121 if (pipe->readers++ == 0)
1122 wake_up_partner(pipe);
1124 if (!is_pipe && !pipe->writers) {
1125 if ((filp->f_flags & O_NONBLOCK)) {
1126 /* suppress EPOLLHUP until we have
1128 filp->f_version = pipe->w_counter;
1130 if (wait_for_partner(pipe, &pipe->w_counter))
1139 * POSIX.1 says that O_NONBLOCK means return -1 with
1140 * errno=ENXIO when there is no process reading the FIFO.
1143 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1147 if (!pipe->writers++)
1148 wake_up_partner(pipe);
1150 if (!is_pipe && !pipe->readers) {
1151 if (wait_for_partner(pipe, &pipe->r_counter))
1156 case FMODE_READ | FMODE_WRITE:
1159 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1160 * This implementation will NEVER block on a O_RDWR open, since
1161 * the process can at least talk to itself.
1168 if (pipe->readers == 1 || pipe->writers == 1)
1169 wake_up_partner(pipe);
1178 __pipe_unlock(pipe);
1182 if (!--pipe->readers)
1183 wake_up_interruptible(&pipe->wr_wait);
1188 if (!--pipe->writers)
1189 wake_up_interruptible_all(&pipe->rd_wait);
1194 __pipe_unlock(pipe);
1196 put_pipe_info(inode, pipe);
1200 const struct file_operations pipefifo_fops = {
1202 .llseek = no_llseek,
1203 .read_iter = pipe_read,
1204 .write_iter = pipe_write,
1206 .unlocked_ioctl = pipe_ioctl,
1207 .release = pipe_release,
1208 .fasync = pipe_fasync,
1209 .splice_write = iter_file_splice_write,
1213 * Currently we rely on the pipe array holding a power-of-2 number
1214 * of pages. Returns 0 on error.
1216 unsigned int round_pipe_size(unsigned long size)
1218 if (size > (1U << 31))
1221 /* Minimum pipe size, as required by POSIX */
1222 if (size < PAGE_SIZE)
1225 return roundup_pow_of_two(size);
1229 * Resize the pipe ring to a number of slots.
1231 int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1233 struct pipe_buffer *bufs;
1234 unsigned int head, tail, mask, n;
1237 * We can shrink the pipe, if arg is greater than the ring occupancy.
1238 * Since we don't expect a lot of shrink+grow operations, just free and
1239 * allocate again like we would do for growing. If the pipe currently
1240 * contains more buffers than arg, then return busy.
1242 mask = pipe->ring_size - 1;
1245 n = pipe_occupancy(pipe->head, pipe->tail);
1249 bufs = kcalloc(nr_slots, sizeof(*bufs),
1250 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1251 if (unlikely(!bufs))
1255 * The pipe array wraps around, so just start the new one at zero
1256 * and adjust the indices.
1259 unsigned int h = head & mask;
1260 unsigned int t = tail & mask;
1262 memcpy(bufs, pipe->bufs + t,
1263 n * sizeof(struct pipe_buffer));
1265 unsigned int tsize = pipe->ring_size - t;
1267 memcpy(bufs + tsize, pipe->bufs,
1268 h * sizeof(struct pipe_buffer));
1269 memcpy(bufs, pipe->bufs + t,
1270 tsize * sizeof(struct pipe_buffer));
1279 pipe->ring_size = nr_slots;
1280 if (pipe->max_usage > nr_slots)
1281 pipe->max_usage = nr_slots;
1285 /* This might have made more room for writers */
1286 wake_up_interruptible(&pipe->wr_wait);
1291 * Allocate a new array of pipe buffers and copy the info over. Returns the
1292 * pipe size if successful, or return -ERROR on error.
1294 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1296 unsigned long user_bufs;
1297 unsigned int nr_slots, size;
1300 #ifdef CONFIG_WATCH_QUEUE
1301 if (pipe->watch_queue)
1305 size = round_pipe_size(arg);
1306 nr_slots = size >> PAGE_SHIFT;
1312 * If trying to increase the pipe capacity, check that an
1313 * unprivileged user is not trying to exceed various limits
1314 * (soft limit check here, hard limit check just below).
1315 * Decreasing the pipe capacity is always permitted, even
1316 * if the user is currently over a limit.
1318 if (nr_slots > pipe->max_usage &&
1319 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1322 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1324 if (nr_slots > pipe->max_usage &&
1325 (too_many_pipe_buffers_hard(user_bufs) ||
1326 too_many_pipe_buffers_soft(user_bufs)) &&
1327 pipe_is_unprivileged_user()) {
1329 goto out_revert_acct;
1332 ret = pipe_resize_ring(pipe, nr_slots);
1334 goto out_revert_acct;
1336 pipe->max_usage = nr_slots;
1337 pipe->nr_accounted = nr_slots;
1338 return pipe->max_usage * PAGE_SIZE;
1341 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1346 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1347 * not enough to verify that this is a pipe.
1349 struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1351 struct pipe_inode_info *pipe = file->private_data;
1353 if (file->f_op != &pipefifo_fops || !pipe)
1355 #ifdef CONFIG_WATCH_QUEUE
1356 if (for_splice && pipe->watch_queue)
1362 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1364 struct pipe_inode_info *pipe;
1367 pipe = get_pipe_info(file, false);
1375 ret = pipe_set_size(pipe, arg);
1378 ret = pipe->max_usage * PAGE_SIZE;
1385 __pipe_unlock(pipe);
1389 static const struct super_operations pipefs_ops = {
1390 .destroy_inode = free_inode_nonrcu,
1391 .statfs = simple_statfs,
1395 * pipefs should _never_ be mounted by userland - too much of security hassle,
1396 * no real gain from having the whole whorehouse mounted. So we don't need
1397 * any operations on the root directory. However, we need a non-trivial
1398 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1401 static int pipefs_init_fs_context(struct fs_context *fc)
1403 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1406 ctx->ops = &pipefs_ops;
1407 ctx->dops = &pipefs_dentry_operations;
1411 static struct file_system_type pipe_fs_type = {
1413 .init_fs_context = pipefs_init_fs_context,
1414 .kill_sb = kill_anon_super,
1417 static int __init init_pipe_fs(void)
1419 int err = register_filesystem(&pipe_fs_type);
1422 pipe_mnt = kern_mount(&pipe_fs_type);
1423 if (IS_ERR(pipe_mnt)) {
1424 err = PTR_ERR(pipe_mnt);
1425 unregister_filesystem(&pipe_fs_type);
1431 fs_initcall(init_pipe_fs);