Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/pipe.c
4 *
5 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
6 */
7
8#include <linux/mm.h>
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>
14#include <linux/fs.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>
28#include <linux/sysctl.h>
29
30#include <linux/uaccess.h>
31#include <asm/ioctls.h>
32
33#include "internal.h"
34
35/*
36 * New pipe buffers will be restricted to this size while the user is exceeding
37 * their pipe buffer quota. The general pipe use case needs at least two
38 * buffers: one for data yet to be read, and one for new data. If this is less
39 * than two, then a write to a non-empty pipe may block even if the pipe is not
40 * full. This can occur with GNU make jobserver or similar uses of pipes as
41 * semaphores: multiple processes may be waiting to write tokens back to the
42 * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
43 *
44 * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
45 * own risk, namely: pipe writes to non-full pipes may block until the pipe is
46 * emptied.
47 */
48#define PIPE_MIN_DEF_BUFFERS 2
49
50/*
51 * The max size that a non-root user is allowed to grow the pipe. Can
52 * be set by root in /proc/sys/fs/pipe-max-size
53 */
54static unsigned int pipe_max_size = 1048576;
55
56/* Maximum allocatable pages per user. Hard limit is unset by default, soft
57 * matches default values.
58 */
59static unsigned long pipe_user_pages_hard;
60static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
61
62/*
63 * We use head and tail indices that aren't masked off, except at the point of
64 * dereference, but rather they're allowed to wrap naturally. This means there
65 * isn't a dead spot in the buffer, but the ring has to be a power of two and
66 * <= 2^31.
67 * -- David Howells 2019-09-23.
68 *
69 * Reads with count = 0 should always return 0.
70 * -- Julian Bradfield 1999-06-07.
71 *
72 * FIFOs and Pipes now generate SIGIO for both readers and writers.
73 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
74 *
75 * pipe_read & write cleanup
76 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
77 */
78
79#define cmp_int(l, r) ((l > r) - (l < r))
80
81#ifdef CONFIG_PROVE_LOCKING
82static int pipe_lock_cmp_fn(const struct lockdep_map *a,
83 const struct lockdep_map *b)
84{
85 return cmp_int((unsigned long) a, (unsigned long) b);
86}
87#endif
88
89void pipe_lock(struct pipe_inode_info *pipe)
90{
91 if (pipe->files)
92 mutex_lock(&pipe->mutex);
93}
94EXPORT_SYMBOL(pipe_lock);
95
96void pipe_unlock(struct pipe_inode_info *pipe)
97{
98 if (pipe->files)
99 mutex_unlock(&pipe->mutex);
100}
101EXPORT_SYMBOL(pipe_unlock);
102
103void pipe_double_lock(struct pipe_inode_info *pipe1,
104 struct pipe_inode_info *pipe2)
105{
106 BUG_ON(pipe1 == pipe2);
107
108 if (pipe1 > pipe2)
109 swap(pipe1, pipe2);
110
111 pipe_lock(pipe1);
112 pipe_lock(pipe2);
113}
114
115static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
116 struct pipe_buffer *buf)
117{
118 struct page *page = buf->page;
119
120 /*
121 * If nobody else uses this page, and we don't already have a
122 * temporary page, let's keep track of it as a one-deep
123 * allocation cache. (Otherwise just release our reference to it)
124 */
125 if (page_count(page) == 1 && !pipe->tmp_page)
126 pipe->tmp_page = page;
127 else
128 put_page(page);
129}
130
131static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
132 struct pipe_buffer *buf)
133{
134 struct page *page = buf->page;
135
136 if (page_count(page) != 1)
137 return false;
138 memcg_kmem_uncharge_page(page, 0);
139 __SetPageLocked(page);
140 return true;
141}
142
143/**
144 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
145 * @pipe: the pipe that the buffer belongs to
146 * @buf: the buffer to attempt to steal
147 *
148 * Description:
149 * This function attempts to steal the &struct page attached to
150 * @buf. If successful, this function returns 0 and returns with
151 * the page locked. The caller may then reuse the page for whatever
152 * he wishes; the typical use is insertion into a different file
153 * page cache.
154 */
155bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
156 struct pipe_buffer *buf)
157{
158 struct page *page = buf->page;
159
160 /*
161 * A reference of one is golden, that means that the owner of this
162 * page is the only one holding a reference to it. lock the page
163 * and return OK.
164 */
165 if (page_count(page) == 1) {
166 lock_page(page);
167 return true;
168 }
169 return false;
170}
171EXPORT_SYMBOL(generic_pipe_buf_try_steal);
172
173/**
174 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
175 * @pipe: the pipe that the buffer belongs to
176 * @buf: the buffer to get a reference to
177 *
178 * Description:
179 * This function grabs an extra reference to @buf. It's used in
180 * the tee() system call, when we duplicate the buffers in one
181 * pipe into another.
182 */
183bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
184{
185 return try_get_page(buf->page);
186}
187EXPORT_SYMBOL(generic_pipe_buf_get);
188
189/**
190 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
191 * @pipe: the pipe that the buffer belongs to
192 * @buf: the buffer to put a reference to
193 *
194 * Description:
195 * This function releases a reference to @buf.
196 */
197void generic_pipe_buf_release(struct pipe_inode_info *pipe,
198 struct pipe_buffer *buf)
199{
200 put_page(buf->page);
201}
202EXPORT_SYMBOL(generic_pipe_buf_release);
203
204static const struct pipe_buf_operations anon_pipe_buf_ops = {
205 .release = anon_pipe_buf_release,
206 .try_steal = anon_pipe_buf_try_steal,
207 .get = generic_pipe_buf_get,
208};
209
210/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
211static inline bool pipe_readable(const struct pipe_inode_info *pipe)
212{
213 unsigned int head = READ_ONCE(pipe->head);
214 unsigned int tail = READ_ONCE(pipe->tail);
215 unsigned int writers = READ_ONCE(pipe->writers);
216
217 return !pipe_empty(head, tail) || !writers;
218}
219
220static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
221 struct pipe_buffer *buf,
222 unsigned int tail)
223{
224 pipe_buf_release(pipe, buf);
225
226 /*
227 * If the pipe has a watch_queue, we need additional protection
228 * by the spinlock because notifications get posted with only
229 * this spinlock, no mutex
230 */
231 if (pipe_has_watch_queue(pipe)) {
232 spin_lock_irq(&pipe->rd_wait.lock);
233#ifdef CONFIG_WATCH_QUEUE
234 if (buf->flags & PIPE_BUF_FLAG_LOSS)
235 pipe->note_loss = true;
236#endif
237 pipe->tail = ++tail;
238 spin_unlock_irq(&pipe->rd_wait.lock);
239 return tail;
240 }
241
242 /*
243 * Without a watch_queue, we can simply increment the tail
244 * without the spinlock - the mutex is enough.
245 */
246 pipe->tail = ++tail;
247 return tail;
248}
249
250static ssize_t
251pipe_read(struct kiocb *iocb, struct iov_iter *to)
252{
253 size_t total_len = iov_iter_count(to);
254 struct file *filp = iocb->ki_filp;
255 struct pipe_inode_info *pipe = filp->private_data;
256 bool was_full, wake_next_reader = false;
257 ssize_t ret;
258
259 /* Null read succeeds. */
260 if (unlikely(total_len == 0))
261 return 0;
262
263 ret = 0;
264 mutex_lock(&pipe->mutex);
265
266 /*
267 * We only wake up writers if the pipe was full when we started
268 * reading in order to avoid unnecessary wakeups.
269 *
270 * But when we do wake up writers, we do so using a sync wakeup
271 * (WF_SYNC), because we want them to get going and generate more
272 * data for us.
273 */
274 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
275 for (;;) {
276 /* Read ->head with a barrier vs post_one_notification() */
277 unsigned int head = smp_load_acquire(&pipe->head);
278 unsigned int tail = pipe->tail;
279 unsigned int mask = pipe->ring_size - 1;
280
281#ifdef CONFIG_WATCH_QUEUE
282 if (pipe->note_loss) {
283 struct watch_notification n;
284
285 if (total_len < 8) {
286 if (ret == 0)
287 ret = -ENOBUFS;
288 break;
289 }
290
291 n.type = WATCH_TYPE_META;
292 n.subtype = WATCH_META_LOSS_NOTIFICATION;
293 n.info = watch_sizeof(n);
294 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
295 if (ret == 0)
296 ret = -EFAULT;
297 break;
298 }
299 ret += sizeof(n);
300 total_len -= sizeof(n);
301 pipe->note_loss = false;
302 }
303#endif
304
305 if (!pipe_empty(head, tail)) {
306 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
307 size_t chars = buf->len;
308 size_t written;
309 int error;
310
311 if (chars > total_len) {
312 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
313 if (ret == 0)
314 ret = -ENOBUFS;
315 break;
316 }
317 chars = total_len;
318 }
319
320 error = pipe_buf_confirm(pipe, buf);
321 if (error) {
322 if (!ret)
323 ret = error;
324 break;
325 }
326
327 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
328 if (unlikely(written < chars)) {
329 if (!ret)
330 ret = -EFAULT;
331 break;
332 }
333 ret += chars;
334 buf->offset += chars;
335 buf->len -= chars;
336
337 /* Was it a packet buffer? Clean up and exit */
338 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
339 total_len = chars;
340 buf->len = 0;
341 }
342
343 if (!buf->len)
344 tail = pipe_update_tail(pipe, buf, tail);
345 total_len -= chars;
346 if (!total_len)
347 break; /* common path: read succeeded */
348 if (!pipe_empty(head, tail)) /* More to do? */
349 continue;
350 }
351
352 if (!pipe->writers)
353 break;
354 if (ret)
355 break;
356 if ((filp->f_flags & O_NONBLOCK) ||
357 (iocb->ki_flags & IOCB_NOWAIT)) {
358 ret = -EAGAIN;
359 break;
360 }
361 mutex_unlock(&pipe->mutex);
362
363 /*
364 * We only get here if we didn't actually read anything.
365 *
366 * However, we could have seen (and removed) a zero-sized
367 * pipe buffer, and might have made space in the buffers
368 * that way.
369 *
370 * You can't make zero-sized pipe buffers by doing an empty
371 * write (not even in packet mode), but they can happen if
372 * the writer gets an EFAULT when trying to fill a buffer
373 * that already got allocated and inserted in the buffer
374 * array.
375 *
376 * So we still need to wake up any pending writers in the
377 * _very_ unlikely case that the pipe was full, but we got
378 * no data.
379 */
380 if (unlikely(was_full))
381 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
382 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
383
384 /*
385 * But because we didn't read anything, at this point we can
386 * just return directly with -ERESTARTSYS if we're interrupted,
387 * since we've done any required wakeups and there's no need
388 * to mark anything accessed. And we've dropped the lock.
389 */
390 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
391 return -ERESTARTSYS;
392
393 mutex_lock(&pipe->mutex);
394 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
395 wake_next_reader = true;
396 }
397 if (pipe_empty(pipe->head, pipe->tail))
398 wake_next_reader = false;
399 mutex_unlock(&pipe->mutex);
400
401 if (was_full)
402 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
403 if (wake_next_reader)
404 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
405 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
406 if (ret > 0)
407 file_accessed(filp);
408 return ret;
409}
410
411static inline int is_packetized(struct file *file)
412{
413 return (file->f_flags & O_DIRECT) != 0;
414}
415
416/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
417static inline bool pipe_writable(const struct pipe_inode_info *pipe)
418{
419 unsigned int head = READ_ONCE(pipe->head);
420 unsigned int tail = READ_ONCE(pipe->tail);
421 unsigned int max_usage = READ_ONCE(pipe->max_usage);
422
423 return !pipe_full(head, tail, max_usage) ||
424 !READ_ONCE(pipe->readers);
425}
426
427static ssize_t
428pipe_write(struct kiocb *iocb, struct iov_iter *from)
429{
430 struct file *filp = iocb->ki_filp;
431 struct pipe_inode_info *pipe = filp->private_data;
432 unsigned int head;
433 ssize_t ret = 0;
434 size_t total_len = iov_iter_count(from);
435 ssize_t chars;
436 bool was_empty = false;
437 bool wake_next_writer = false;
438
439 /*
440 * Reject writing to watch queue pipes before the point where we lock
441 * the pipe.
442 * Otherwise, lockdep would be unhappy if the caller already has another
443 * pipe locked.
444 * If we had to support locking a normal pipe and a notification pipe at
445 * the same time, we could set up lockdep annotations for that, but
446 * since we don't actually need that, it's simpler to just bail here.
447 */
448 if (pipe_has_watch_queue(pipe))
449 return -EXDEV;
450
451 /* Null write succeeds. */
452 if (unlikely(total_len == 0))
453 return 0;
454
455 mutex_lock(&pipe->mutex);
456
457 if (!pipe->readers) {
458 send_sig(SIGPIPE, current, 0);
459 ret = -EPIPE;
460 goto out;
461 }
462
463 /*
464 * If it wasn't empty we try to merge new data into
465 * the last buffer.
466 *
467 * That naturally merges small writes, but it also
468 * page-aligns the rest of the writes for large writes
469 * spanning multiple pages.
470 */
471 head = pipe->head;
472 was_empty = pipe_empty(head, pipe->tail);
473 chars = total_len & (PAGE_SIZE-1);
474 if (chars && !was_empty) {
475 unsigned int mask = pipe->ring_size - 1;
476 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
477 int offset = buf->offset + buf->len;
478
479 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
480 offset + chars <= PAGE_SIZE) {
481 ret = pipe_buf_confirm(pipe, buf);
482 if (ret)
483 goto out;
484
485 ret = copy_page_from_iter(buf->page, offset, chars, from);
486 if (unlikely(ret < chars)) {
487 ret = -EFAULT;
488 goto out;
489 }
490
491 buf->len += ret;
492 if (!iov_iter_count(from))
493 goto out;
494 }
495 }
496
497 for (;;) {
498 if (!pipe->readers) {
499 send_sig(SIGPIPE, current, 0);
500 if (!ret)
501 ret = -EPIPE;
502 break;
503 }
504
505 head = pipe->head;
506 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
507 unsigned int mask = pipe->ring_size - 1;
508 struct pipe_buffer *buf;
509 struct page *page = pipe->tmp_page;
510 int copied;
511
512 if (!page) {
513 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
514 if (unlikely(!page)) {
515 ret = ret ? : -ENOMEM;
516 break;
517 }
518 pipe->tmp_page = page;
519 }
520
521 /* Allocate a slot in the ring in advance and attach an
522 * empty buffer. If we fault or otherwise fail to use
523 * it, either the reader will consume it or it'll still
524 * be there for the next write.
525 */
526 pipe->head = head + 1;
527
528 /* Insert it into the buffer array */
529 buf = &pipe->bufs[head & mask];
530 buf->page = page;
531 buf->ops = &anon_pipe_buf_ops;
532 buf->offset = 0;
533 buf->len = 0;
534 if (is_packetized(filp))
535 buf->flags = PIPE_BUF_FLAG_PACKET;
536 else
537 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
538 pipe->tmp_page = NULL;
539
540 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
541 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
542 if (!ret)
543 ret = -EFAULT;
544 break;
545 }
546 ret += copied;
547 buf->len = copied;
548
549 if (!iov_iter_count(from))
550 break;
551 }
552
553 if (!pipe_full(head, pipe->tail, pipe->max_usage))
554 continue;
555
556 /* Wait for buffer space to become available. */
557 if ((filp->f_flags & O_NONBLOCK) ||
558 (iocb->ki_flags & IOCB_NOWAIT)) {
559 if (!ret)
560 ret = -EAGAIN;
561 break;
562 }
563 if (signal_pending(current)) {
564 if (!ret)
565 ret = -ERESTARTSYS;
566 break;
567 }
568
569 /*
570 * We're going to release the pipe lock and wait for more
571 * space. We wake up any readers if necessary, and then
572 * after waiting we need to re-check whether the pipe
573 * become empty while we dropped the lock.
574 */
575 mutex_unlock(&pipe->mutex);
576 if (was_empty)
577 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
578 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
579 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
580 mutex_lock(&pipe->mutex);
581 was_empty = pipe_empty(pipe->head, pipe->tail);
582 wake_next_writer = true;
583 }
584out:
585 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
586 wake_next_writer = false;
587 mutex_unlock(&pipe->mutex);
588
589 /*
590 * If we do do a wakeup event, we do a 'sync' wakeup, because we
591 * want the reader to start processing things asap, rather than
592 * leave the data pending.
593 *
594 * This is particularly important for small writes, because of
595 * how (for example) the GNU make jobserver uses small writes to
596 * wake up pending jobs
597 *
598 * Epoll nonsensically wants a wakeup whether the pipe
599 * was already empty or not.
600 */
601 if (was_empty || pipe->poll_usage)
602 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
603 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
604 if (wake_next_writer)
605 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
606 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
607 int err = file_update_time(filp);
608 if (err)
609 ret = err;
610 sb_end_write(file_inode(filp)->i_sb);
611 }
612 return ret;
613}
614
615static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
616{
617 struct pipe_inode_info *pipe = filp->private_data;
618 unsigned int count, head, tail, mask;
619
620 switch (cmd) {
621 case FIONREAD:
622 mutex_lock(&pipe->mutex);
623 count = 0;
624 head = pipe->head;
625 tail = pipe->tail;
626 mask = pipe->ring_size - 1;
627
628 while (tail != head) {
629 count += pipe->bufs[tail & mask].len;
630 tail++;
631 }
632 mutex_unlock(&pipe->mutex);
633
634 return put_user(count, (int __user *)arg);
635
636#ifdef CONFIG_WATCH_QUEUE
637 case IOC_WATCH_QUEUE_SET_SIZE: {
638 int ret;
639 mutex_lock(&pipe->mutex);
640 ret = watch_queue_set_size(pipe, arg);
641 mutex_unlock(&pipe->mutex);
642 return ret;
643 }
644
645 case IOC_WATCH_QUEUE_SET_FILTER:
646 return watch_queue_set_filter(
647 pipe, (struct watch_notification_filter __user *)arg);
648#endif
649
650 default:
651 return -ENOIOCTLCMD;
652 }
653}
654
655/* No kernel lock held - fine */
656static __poll_t
657pipe_poll(struct file *filp, poll_table *wait)
658{
659 __poll_t mask;
660 struct pipe_inode_info *pipe = filp->private_data;
661 unsigned int head, tail;
662
663 /* Epoll has some historical nasty semantics, this enables them */
664 WRITE_ONCE(pipe->poll_usage, true);
665
666 /*
667 * Reading pipe state only -- no need for acquiring the semaphore.
668 *
669 * But because this is racy, the code has to add the
670 * entry to the poll table _first_ ..
671 */
672 if (filp->f_mode & FMODE_READ)
673 poll_wait(filp, &pipe->rd_wait, wait);
674 if (filp->f_mode & FMODE_WRITE)
675 poll_wait(filp, &pipe->wr_wait, wait);
676
677 /*
678 * .. and only then can you do the racy tests. That way,
679 * if something changes and you got it wrong, the poll
680 * table entry will wake you up and fix it.
681 */
682 head = READ_ONCE(pipe->head);
683 tail = READ_ONCE(pipe->tail);
684
685 mask = 0;
686 if (filp->f_mode & FMODE_READ) {
687 if (!pipe_empty(head, tail))
688 mask |= EPOLLIN | EPOLLRDNORM;
689 if (!pipe->writers && filp->f_pipe != pipe->w_counter)
690 mask |= EPOLLHUP;
691 }
692
693 if (filp->f_mode & FMODE_WRITE) {
694 if (!pipe_full(head, tail, pipe->max_usage))
695 mask |= EPOLLOUT | EPOLLWRNORM;
696 /*
697 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
698 * behave exactly like pipes for poll().
699 */
700 if (!pipe->readers)
701 mask |= EPOLLERR;
702 }
703
704 return mask;
705}
706
707static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
708{
709 int kill = 0;
710
711 spin_lock(&inode->i_lock);
712 if (!--pipe->files) {
713 inode->i_pipe = NULL;
714 kill = 1;
715 }
716 spin_unlock(&inode->i_lock);
717
718 if (kill)
719 free_pipe_info(pipe);
720}
721
722static int
723pipe_release(struct inode *inode, struct file *file)
724{
725 struct pipe_inode_info *pipe = file->private_data;
726
727 mutex_lock(&pipe->mutex);
728 if (file->f_mode & FMODE_READ)
729 pipe->readers--;
730 if (file->f_mode & FMODE_WRITE)
731 pipe->writers--;
732
733 /* Was that the last reader or writer, but not the other side? */
734 if (!pipe->readers != !pipe->writers) {
735 wake_up_interruptible_all(&pipe->rd_wait);
736 wake_up_interruptible_all(&pipe->wr_wait);
737 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
738 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
739 }
740 mutex_unlock(&pipe->mutex);
741
742 put_pipe_info(inode, pipe);
743 return 0;
744}
745
746static int
747pipe_fasync(int fd, struct file *filp, int on)
748{
749 struct pipe_inode_info *pipe = filp->private_data;
750 int retval = 0;
751
752 mutex_lock(&pipe->mutex);
753 if (filp->f_mode & FMODE_READ)
754 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
755 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
756 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
757 if (retval < 0 && (filp->f_mode & FMODE_READ))
758 /* this can happen only if on == T */
759 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
760 }
761 mutex_unlock(&pipe->mutex);
762 return retval;
763}
764
765unsigned long account_pipe_buffers(struct user_struct *user,
766 unsigned long old, unsigned long new)
767{
768 return atomic_long_add_return(new - old, &user->pipe_bufs);
769}
770
771bool too_many_pipe_buffers_soft(unsigned long user_bufs)
772{
773 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
774
775 return soft_limit && user_bufs > soft_limit;
776}
777
778bool too_many_pipe_buffers_hard(unsigned long user_bufs)
779{
780 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
781
782 return hard_limit && user_bufs > hard_limit;
783}
784
785bool pipe_is_unprivileged_user(void)
786{
787 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
788}
789
790struct pipe_inode_info *alloc_pipe_info(void)
791{
792 struct pipe_inode_info *pipe;
793 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
794 struct user_struct *user = get_current_user();
795 unsigned long user_bufs;
796 unsigned int max_size = READ_ONCE(pipe_max_size);
797
798 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
799 if (pipe == NULL)
800 goto out_free_uid;
801
802 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
803 pipe_bufs = max_size >> PAGE_SHIFT;
804
805 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
806
807 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
808 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
809 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
810 }
811
812 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
813 goto out_revert_acct;
814
815 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
816 GFP_KERNEL_ACCOUNT);
817
818 if (pipe->bufs) {
819 init_waitqueue_head(&pipe->rd_wait);
820 init_waitqueue_head(&pipe->wr_wait);
821 pipe->r_counter = pipe->w_counter = 1;
822 pipe->max_usage = pipe_bufs;
823 pipe->ring_size = pipe_bufs;
824 pipe->nr_accounted = pipe_bufs;
825 pipe->user = user;
826 mutex_init(&pipe->mutex);
827 lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);
828 return pipe;
829 }
830
831out_revert_acct:
832 (void) account_pipe_buffers(user, pipe_bufs, 0);
833 kfree(pipe);
834out_free_uid:
835 free_uid(user);
836 return NULL;
837}
838
839void free_pipe_info(struct pipe_inode_info *pipe)
840{
841 unsigned int i;
842
843#ifdef CONFIG_WATCH_QUEUE
844 if (pipe->watch_queue)
845 watch_queue_clear(pipe->watch_queue);
846#endif
847
848 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
849 free_uid(pipe->user);
850 for (i = 0; i < pipe->ring_size; i++) {
851 struct pipe_buffer *buf = pipe->bufs + i;
852 if (buf->ops)
853 pipe_buf_release(pipe, buf);
854 }
855#ifdef CONFIG_WATCH_QUEUE
856 if (pipe->watch_queue)
857 put_watch_queue(pipe->watch_queue);
858#endif
859 if (pipe->tmp_page)
860 __free_page(pipe->tmp_page);
861 kfree(pipe->bufs);
862 kfree(pipe);
863}
864
865static struct vfsmount *pipe_mnt __ro_after_init;
866
867/*
868 * pipefs_dname() is called from d_path().
869 */
870static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
871{
872 return dynamic_dname(buffer, buflen, "pipe:[%lu]",
873 d_inode(dentry)->i_ino);
874}
875
876static const struct dentry_operations pipefs_dentry_operations = {
877 .d_dname = pipefs_dname,
878};
879
880static struct inode * get_pipe_inode(void)
881{
882 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
883 struct pipe_inode_info *pipe;
884
885 if (!inode)
886 goto fail_inode;
887
888 inode->i_ino = get_next_ino();
889
890 pipe = alloc_pipe_info();
891 if (!pipe)
892 goto fail_iput;
893
894 inode->i_pipe = pipe;
895 pipe->files = 2;
896 pipe->readers = pipe->writers = 1;
897 inode->i_fop = &pipefifo_fops;
898
899 /*
900 * Mark the inode dirty from the very beginning,
901 * that way it will never be moved to the dirty
902 * list because "mark_inode_dirty()" will think
903 * that it already _is_ on the dirty list.
904 */
905 inode->i_state = I_DIRTY;
906 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
907 inode->i_uid = current_fsuid();
908 inode->i_gid = current_fsgid();
909 simple_inode_init_ts(inode);
910
911 return inode;
912
913fail_iput:
914 iput(inode);
915
916fail_inode:
917 return NULL;
918}
919
920int create_pipe_files(struct file **res, int flags)
921{
922 struct inode *inode = get_pipe_inode();
923 struct file *f;
924 int error;
925
926 if (!inode)
927 return -ENFILE;
928
929 if (flags & O_NOTIFICATION_PIPE) {
930 error = watch_queue_init(inode->i_pipe);
931 if (error) {
932 free_pipe_info(inode->i_pipe);
933 iput(inode);
934 return error;
935 }
936 }
937
938 f = alloc_file_pseudo(inode, pipe_mnt, "",
939 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
940 &pipefifo_fops);
941 if (IS_ERR(f)) {
942 free_pipe_info(inode->i_pipe);
943 iput(inode);
944 return PTR_ERR(f);
945 }
946
947 f->private_data = inode->i_pipe;
948 f->f_pipe = 0;
949
950 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
951 &pipefifo_fops);
952 if (IS_ERR(res[0])) {
953 put_pipe_info(inode, inode->i_pipe);
954 fput(f);
955 return PTR_ERR(res[0]);
956 }
957 res[0]->private_data = inode->i_pipe;
958 res[0]->f_pipe = 0;
959 res[1] = f;
960 stream_open(inode, res[0]);
961 stream_open(inode, res[1]);
962 return 0;
963}
964
965static int __do_pipe_flags(int *fd, struct file **files, int flags)
966{
967 int error;
968 int fdw, fdr;
969
970 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
971 return -EINVAL;
972
973 error = create_pipe_files(files, flags);
974 if (error)
975 return error;
976
977 error = get_unused_fd_flags(flags);
978 if (error < 0)
979 goto err_read_pipe;
980 fdr = error;
981
982 error = get_unused_fd_flags(flags);
983 if (error < 0)
984 goto err_fdr;
985 fdw = error;
986
987 audit_fd_pair(fdr, fdw);
988 fd[0] = fdr;
989 fd[1] = fdw;
990 /* pipe groks IOCB_NOWAIT */
991 files[0]->f_mode |= FMODE_NOWAIT;
992 files[1]->f_mode |= FMODE_NOWAIT;
993 return 0;
994
995 err_fdr:
996 put_unused_fd(fdr);
997 err_read_pipe:
998 fput(files[0]);
999 fput(files[1]);
1000 return error;
1001}
1002
1003int do_pipe_flags(int *fd, int flags)
1004{
1005 struct file *files[2];
1006 int error = __do_pipe_flags(fd, files, flags);
1007 if (!error) {
1008 fd_install(fd[0], files[0]);
1009 fd_install(fd[1], files[1]);
1010 }
1011 return error;
1012}
1013
1014/*
1015 * sys_pipe() is the normal C calling standard for creating
1016 * a pipe. It's not the way Unix traditionally does this, though.
1017 */
1018static int do_pipe2(int __user *fildes, int flags)
1019{
1020 struct file *files[2];
1021 int fd[2];
1022 int error;
1023
1024 error = __do_pipe_flags(fd, files, flags);
1025 if (!error) {
1026 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1027 fput(files[0]);
1028 fput(files[1]);
1029 put_unused_fd(fd[0]);
1030 put_unused_fd(fd[1]);
1031 error = -EFAULT;
1032 } else {
1033 fd_install(fd[0], files[0]);
1034 fd_install(fd[1], files[1]);
1035 }
1036 }
1037 return error;
1038}
1039
1040SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1041{
1042 return do_pipe2(fildes, flags);
1043}
1044
1045SYSCALL_DEFINE1(pipe, int __user *, fildes)
1046{
1047 return do_pipe2(fildes, 0);
1048}
1049
1050/*
1051 * This is the stupid "wait for pipe to be readable or writable"
1052 * model.
1053 *
1054 * See pipe_read/write() for the proper kind of exclusive wait,
1055 * but that requires that we wake up any other readers/writers
1056 * if we then do not end up reading everything (ie the whole
1057 * "wake_next_reader/writer" logic in pipe_read/write()).
1058 */
1059void pipe_wait_readable(struct pipe_inode_info *pipe)
1060{
1061 pipe_unlock(pipe);
1062 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1063 pipe_lock(pipe);
1064}
1065
1066void pipe_wait_writable(struct pipe_inode_info *pipe)
1067{
1068 pipe_unlock(pipe);
1069 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1070 pipe_lock(pipe);
1071}
1072
1073/*
1074 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1075 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1076 * race with the count check and waitqueue prep.
1077 *
1078 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1079 * then check the condition you're waiting for, and only then sleep. But
1080 * because of the pipe lock, we can check the condition before being on
1081 * the wait queue.
1082 *
1083 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1084 */
1085static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1086{
1087 DEFINE_WAIT(rdwait);
1088 int cur = *cnt;
1089
1090 while (cur == *cnt) {
1091 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1092 pipe_unlock(pipe);
1093 schedule();
1094 finish_wait(&pipe->rd_wait, &rdwait);
1095 pipe_lock(pipe);
1096 if (signal_pending(current))
1097 break;
1098 }
1099 return cur == *cnt ? -ERESTARTSYS : 0;
1100}
1101
1102static void wake_up_partner(struct pipe_inode_info *pipe)
1103{
1104 wake_up_interruptible_all(&pipe->rd_wait);
1105}
1106
1107static int fifo_open(struct inode *inode, struct file *filp)
1108{
1109 struct pipe_inode_info *pipe;
1110 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1111 int ret;
1112
1113 filp->f_pipe = 0;
1114
1115 spin_lock(&inode->i_lock);
1116 if (inode->i_pipe) {
1117 pipe = inode->i_pipe;
1118 pipe->files++;
1119 spin_unlock(&inode->i_lock);
1120 } else {
1121 spin_unlock(&inode->i_lock);
1122 pipe = alloc_pipe_info();
1123 if (!pipe)
1124 return -ENOMEM;
1125 pipe->files = 1;
1126 spin_lock(&inode->i_lock);
1127 if (unlikely(inode->i_pipe)) {
1128 inode->i_pipe->files++;
1129 spin_unlock(&inode->i_lock);
1130 free_pipe_info(pipe);
1131 pipe = inode->i_pipe;
1132 } else {
1133 inode->i_pipe = pipe;
1134 spin_unlock(&inode->i_lock);
1135 }
1136 }
1137 filp->private_data = pipe;
1138 /* OK, we have a pipe and it's pinned down */
1139
1140 mutex_lock(&pipe->mutex);
1141
1142 /* We can only do regular read/write on fifos */
1143 stream_open(inode, filp);
1144
1145 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1146 case FMODE_READ:
1147 /*
1148 * O_RDONLY
1149 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1150 * opened, even when there is no process writing the FIFO.
1151 */
1152 pipe->r_counter++;
1153 if (pipe->readers++ == 0)
1154 wake_up_partner(pipe);
1155
1156 if (!is_pipe && !pipe->writers) {
1157 if ((filp->f_flags & O_NONBLOCK)) {
1158 /* suppress EPOLLHUP until we have
1159 * seen a writer */
1160 filp->f_pipe = pipe->w_counter;
1161 } else {
1162 if (wait_for_partner(pipe, &pipe->w_counter))
1163 goto err_rd;
1164 }
1165 }
1166 break;
1167
1168 case FMODE_WRITE:
1169 /*
1170 * O_WRONLY
1171 * POSIX.1 says that O_NONBLOCK means return -1 with
1172 * errno=ENXIO when there is no process reading the FIFO.
1173 */
1174 ret = -ENXIO;
1175 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1176 goto err;
1177
1178 pipe->w_counter++;
1179 if (!pipe->writers++)
1180 wake_up_partner(pipe);
1181
1182 if (!is_pipe && !pipe->readers) {
1183 if (wait_for_partner(pipe, &pipe->r_counter))
1184 goto err_wr;
1185 }
1186 break;
1187
1188 case FMODE_READ | FMODE_WRITE:
1189 /*
1190 * O_RDWR
1191 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1192 * This implementation will NEVER block on a O_RDWR open, since
1193 * the process can at least talk to itself.
1194 */
1195
1196 pipe->readers++;
1197 pipe->writers++;
1198 pipe->r_counter++;
1199 pipe->w_counter++;
1200 if (pipe->readers == 1 || pipe->writers == 1)
1201 wake_up_partner(pipe);
1202 break;
1203
1204 default:
1205 ret = -EINVAL;
1206 goto err;
1207 }
1208
1209 /* Ok! */
1210 mutex_unlock(&pipe->mutex);
1211 return 0;
1212
1213err_rd:
1214 if (!--pipe->readers)
1215 wake_up_interruptible(&pipe->wr_wait);
1216 ret = -ERESTARTSYS;
1217 goto err;
1218
1219err_wr:
1220 if (!--pipe->writers)
1221 wake_up_interruptible_all(&pipe->rd_wait);
1222 ret = -ERESTARTSYS;
1223 goto err;
1224
1225err:
1226 mutex_unlock(&pipe->mutex);
1227
1228 put_pipe_info(inode, pipe);
1229 return ret;
1230}
1231
1232const struct file_operations pipefifo_fops = {
1233 .open = fifo_open,
1234 .read_iter = pipe_read,
1235 .write_iter = pipe_write,
1236 .poll = pipe_poll,
1237 .unlocked_ioctl = pipe_ioctl,
1238 .release = pipe_release,
1239 .fasync = pipe_fasync,
1240 .splice_write = iter_file_splice_write,
1241};
1242
1243/*
1244 * Currently we rely on the pipe array holding a power-of-2 number
1245 * of pages. Returns 0 on error.
1246 */
1247unsigned int round_pipe_size(unsigned int size)
1248{
1249 if (size > (1U << 31))
1250 return 0;
1251
1252 /* Minimum pipe size, as required by POSIX */
1253 if (size < PAGE_SIZE)
1254 return PAGE_SIZE;
1255
1256 return roundup_pow_of_two(size);
1257}
1258
1259/*
1260 * Resize the pipe ring to a number of slots.
1261 *
1262 * Note the pipe can be reduced in capacity, but only if the current
1263 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1264 * returned instead.
1265 */
1266int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1267{
1268 struct pipe_buffer *bufs;
1269 unsigned int head, tail, mask, n;
1270
1271 bufs = kcalloc(nr_slots, sizeof(*bufs),
1272 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1273 if (unlikely(!bufs))
1274 return -ENOMEM;
1275
1276 spin_lock_irq(&pipe->rd_wait.lock);
1277 mask = pipe->ring_size - 1;
1278 head = pipe->head;
1279 tail = pipe->tail;
1280
1281 n = pipe_occupancy(head, tail);
1282 if (nr_slots < n) {
1283 spin_unlock_irq(&pipe->rd_wait.lock);
1284 kfree(bufs);
1285 return -EBUSY;
1286 }
1287
1288 /*
1289 * The pipe array wraps around, so just start the new one at zero
1290 * and adjust the indices.
1291 */
1292 if (n > 0) {
1293 unsigned int h = head & mask;
1294 unsigned int t = tail & mask;
1295 if (h > t) {
1296 memcpy(bufs, pipe->bufs + t,
1297 n * sizeof(struct pipe_buffer));
1298 } else {
1299 unsigned int tsize = pipe->ring_size - t;
1300 if (h > 0)
1301 memcpy(bufs + tsize, pipe->bufs,
1302 h * sizeof(struct pipe_buffer));
1303 memcpy(bufs, pipe->bufs + t,
1304 tsize * sizeof(struct pipe_buffer));
1305 }
1306 }
1307
1308 head = n;
1309 tail = 0;
1310
1311 kfree(pipe->bufs);
1312 pipe->bufs = bufs;
1313 pipe->ring_size = nr_slots;
1314 if (pipe->max_usage > nr_slots)
1315 pipe->max_usage = nr_slots;
1316 pipe->tail = tail;
1317 pipe->head = head;
1318
1319 if (!pipe_has_watch_queue(pipe)) {
1320 pipe->max_usage = nr_slots;
1321 pipe->nr_accounted = nr_slots;
1322 }
1323
1324 spin_unlock_irq(&pipe->rd_wait.lock);
1325
1326 /* This might have made more room for writers */
1327 wake_up_interruptible(&pipe->wr_wait);
1328 return 0;
1329}
1330
1331/*
1332 * Allocate a new array of pipe buffers and copy the info over. Returns the
1333 * pipe size if successful, or return -ERROR on error.
1334 */
1335static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1336{
1337 unsigned long user_bufs;
1338 unsigned int nr_slots, size;
1339 long ret = 0;
1340
1341 if (pipe_has_watch_queue(pipe))
1342 return -EBUSY;
1343
1344 size = round_pipe_size(arg);
1345 nr_slots = size >> PAGE_SHIFT;
1346
1347 if (!nr_slots)
1348 return -EINVAL;
1349
1350 /*
1351 * If trying to increase the pipe capacity, check that an
1352 * unprivileged user is not trying to exceed various limits
1353 * (soft limit check here, hard limit check just below).
1354 * Decreasing the pipe capacity is always permitted, even
1355 * if the user is currently over a limit.
1356 */
1357 if (nr_slots > pipe->max_usage &&
1358 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1359 return -EPERM;
1360
1361 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1362
1363 if (nr_slots > pipe->max_usage &&
1364 (too_many_pipe_buffers_hard(user_bufs) ||
1365 too_many_pipe_buffers_soft(user_bufs)) &&
1366 pipe_is_unprivileged_user()) {
1367 ret = -EPERM;
1368 goto out_revert_acct;
1369 }
1370
1371 ret = pipe_resize_ring(pipe, nr_slots);
1372 if (ret < 0)
1373 goto out_revert_acct;
1374
1375 return pipe->max_usage * PAGE_SIZE;
1376
1377out_revert_acct:
1378 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1379 return ret;
1380}
1381
1382/*
1383 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1384 * not enough to verify that this is a pipe.
1385 */
1386struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1387{
1388 struct pipe_inode_info *pipe = file->private_data;
1389
1390 if (file->f_op != &pipefifo_fops || !pipe)
1391 return NULL;
1392 if (for_splice && pipe_has_watch_queue(pipe))
1393 return NULL;
1394 return pipe;
1395}
1396
1397long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1398{
1399 struct pipe_inode_info *pipe;
1400 long ret;
1401
1402 pipe = get_pipe_info(file, false);
1403 if (!pipe)
1404 return -EBADF;
1405
1406 mutex_lock(&pipe->mutex);
1407
1408 switch (cmd) {
1409 case F_SETPIPE_SZ:
1410 ret = pipe_set_size(pipe, arg);
1411 break;
1412 case F_GETPIPE_SZ:
1413 ret = pipe->max_usage * PAGE_SIZE;
1414 break;
1415 default:
1416 ret = -EINVAL;
1417 break;
1418 }
1419
1420 mutex_unlock(&pipe->mutex);
1421 return ret;
1422}
1423
1424static const struct super_operations pipefs_ops = {
1425 .destroy_inode = free_inode_nonrcu,
1426 .statfs = simple_statfs,
1427};
1428
1429/*
1430 * pipefs should _never_ be mounted by userland - too much of security hassle,
1431 * no real gain from having the whole file system mounted. So we don't need
1432 * any operations on the root directory. However, we need a non-trivial
1433 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1434 */
1435
1436static int pipefs_init_fs_context(struct fs_context *fc)
1437{
1438 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1439 if (!ctx)
1440 return -ENOMEM;
1441 ctx->ops = &pipefs_ops;
1442 ctx->dops = &pipefs_dentry_operations;
1443 return 0;
1444}
1445
1446static struct file_system_type pipe_fs_type = {
1447 .name = "pipefs",
1448 .init_fs_context = pipefs_init_fs_context,
1449 .kill_sb = kill_anon_super,
1450};
1451
1452#ifdef CONFIG_SYSCTL
1453static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1454 unsigned int *valp,
1455 int write, void *data)
1456{
1457 if (write) {
1458 unsigned int val;
1459
1460 val = round_pipe_size(*lvalp);
1461 if (val == 0)
1462 return -EINVAL;
1463
1464 *valp = val;
1465 } else {
1466 unsigned int val = *valp;
1467 *lvalp = (unsigned long) val;
1468 }
1469
1470 return 0;
1471}
1472
1473static int proc_dopipe_max_size(const struct ctl_table *table, int write,
1474 void *buffer, size_t *lenp, loff_t *ppos)
1475{
1476 return do_proc_douintvec(table, write, buffer, lenp, ppos,
1477 do_proc_dopipe_max_size_conv, NULL);
1478}
1479
1480static struct ctl_table fs_pipe_sysctls[] = {
1481 {
1482 .procname = "pipe-max-size",
1483 .data = &pipe_max_size,
1484 .maxlen = sizeof(pipe_max_size),
1485 .mode = 0644,
1486 .proc_handler = proc_dopipe_max_size,
1487 },
1488 {
1489 .procname = "pipe-user-pages-hard",
1490 .data = &pipe_user_pages_hard,
1491 .maxlen = sizeof(pipe_user_pages_hard),
1492 .mode = 0644,
1493 .proc_handler = proc_doulongvec_minmax,
1494 },
1495 {
1496 .procname = "pipe-user-pages-soft",
1497 .data = &pipe_user_pages_soft,
1498 .maxlen = sizeof(pipe_user_pages_soft),
1499 .mode = 0644,
1500 .proc_handler = proc_doulongvec_minmax,
1501 },
1502};
1503#endif
1504
1505static int __init init_pipe_fs(void)
1506{
1507 int err = register_filesystem(&pipe_fs_type);
1508
1509 if (!err) {
1510 pipe_mnt = kern_mount(&pipe_fs_type);
1511 if (IS_ERR(pipe_mnt)) {
1512 err = PTR_ERR(pipe_mnt);
1513 unregister_filesystem(&pipe_fs_type);
1514 }
1515 }
1516#ifdef CONFIG_SYSCTL
1517 register_sysctl_init("fs", fs_pipe_sysctls);
1518#endif
1519 return err;
1520}
1521
1522fs_initcall(init_pipe_fs);
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/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>
24#include <linux/aio.h>
25
26#include <asm/uaccess.h>
27#include <asm/ioctls.h>
28
29#include "internal.h"
30
31/*
32 * The max size that a non-root user is allowed to grow the pipe. Can
33 * be set by root in /proc/sys/fs/pipe-max-size
34 */
35unsigned int pipe_max_size = 1048576;
36
37/*
38 * Minimum pipe size, as required by POSIX
39 */
40unsigned int pipe_min_size = PAGE_SIZE;
41
42/*
43 * We use a start+len construction, which provides full use of the
44 * allocated memory.
45 * -- Florian Coosmann (FGC)
46 *
47 * Reads with count = 0 should always return 0.
48 * -- Julian Bradfield 1999-06-07.
49 *
50 * FIFOs and Pipes now generate SIGIO for both readers and writers.
51 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
52 *
53 * pipe_read & write cleanup
54 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
55 */
56
57static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
58{
59 if (pipe->files)
60 mutex_lock_nested(&pipe->mutex, subclass);
61}
62
63void pipe_lock(struct pipe_inode_info *pipe)
64{
65 /*
66 * pipe_lock() nests non-pipe inode locks (for writing to a file)
67 */
68 pipe_lock_nested(pipe, I_MUTEX_PARENT);
69}
70EXPORT_SYMBOL(pipe_lock);
71
72void pipe_unlock(struct pipe_inode_info *pipe)
73{
74 if (pipe->files)
75 mutex_unlock(&pipe->mutex);
76}
77EXPORT_SYMBOL(pipe_unlock);
78
79static inline void __pipe_lock(struct pipe_inode_info *pipe)
80{
81 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
82}
83
84static inline void __pipe_unlock(struct pipe_inode_info *pipe)
85{
86 mutex_unlock(&pipe->mutex);
87}
88
89void pipe_double_lock(struct pipe_inode_info *pipe1,
90 struct pipe_inode_info *pipe2)
91{
92 BUG_ON(pipe1 == pipe2);
93
94 if (pipe1 < pipe2) {
95 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
96 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
97 } else {
98 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
99 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
100 }
101}
102
103/* Drop the inode semaphore and wait for a pipe event, atomically */
104void pipe_wait(struct pipe_inode_info *pipe)
105{
106 DEFINE_WAIT(wait);
107
108 /*
109 * Pipes are system-local resources, so sleeping on them
110 * is considered a noninteractive wait:
111 */
112 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
113 pipe_unlock(pipe);
114 schedule();
115 finish_wait(&pipe->wait, &wait);
116 pipe_lock(pipe);
117}
118
119static int
120pipe_iov_copy_from_user(void *to, struct iovec *iov, 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_from_user_inatomic(to, iov->iov_base, copy))
132 return -EFAULT;
133 } else {
134 if (copy_from_user(to, iov->iov_base, copy))
135 return -EFAULT;
136 }
137 to += copy;
138 len -= copy;
139 iov->iov_base += copy;
140 iov->iov_len -= copy;
141 }
142 return 0;
143}
144
145/*
146 * Pre-fault in the user memory, so we can use atomic copies.
147 */
148static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
149{
150 while (!iov->iov_len)
151 iov++;
152
153 while (len > 0) {
154 unsigned long this_len;
155
156 this_len = min_t(unsigned long, len, iov->iov_len);
157 fault_in_pages_readable(iov->iov_base, this_len);
158 len -= this_len;
159 iov++;
160 }
161}
162
163static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
164 struct pipe_buffer *buf)
165{
166 struct page *page = buf->page;
167
168 /*
169 * If nobody else uses this page, and we don't already have a
170 * temporary page, let's keep track of it as a one-deep
171 * allocation cache. (Otherwise just release our reference to it)
172 */
173 if (page_count(page) == 1 && !pipe->tmp_page)
174 pipe->tmp_page = page;
175 else
176 page_cache_release(page);
177}
178
179/**
180 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
181 * @pipe: the pipe that the buffer belongs to
182 * @buf: the buffer to attempt to steal
183 *
184 * Description:
185 * This function attempts to steal the &struct page attached to
186 * @buf. If successful, this function returns 0 and returns with
187 * the page locked. The caller may then reuse the page for whatever
188 * he wishes; the typical use is insertion into a different file
189 * page cache.
190 */
191int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
192 struct pipe_buffer *buf)
193{
194 struct page *page = buf->page;
195
196 /*
197 * A reference of one is golden, that means that the owner of this
198 * page is the only one holding a reference to it. lock the page
199 * and return OK.
200 */
201 if (page_count(page) == 1) {
202 lock_page(page);
203 return 0;
204 }
205
206 return 1;
207}
208EXPORT_SYMBOL(generic_pipe_buf_steal);
209
210/**
211 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
212 * @pipe: the pipe that the buffer belongs to
213 * @buf: the buffer to get a reference to
214 *
215 * Description:
216 * This function grabs an extra reference to @buf. It's used in
217 * in the tee() system call, when we duplicate the buffers in one
218 * pipe into another.
219 */
220void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
221{
222 page_cache_get(buf->page);
223}
224EXPORT_SYMBOL(generic_pipe_buf_get);
225
226/**
227 * generic_pipe_buf_confirm - verify contents of the pipe buffer
228 * @info: the pipe that the buffer belongs to
229 * @buf: the buffer to confirm
230 *
231 * Description:
232 * This function does nothing, because the generic pipe code uses
233 * pages that are always good when inserted into the pipe.
234 */
235int generic_pipe_buf_confirm(struct pipe_inode_info *info,
236 struct pipe_buffer *buf)
237{
238 return 0;
239}
240EXPORT_SYMBOL(generic_pipe_buf_confirm);
241
242/**
243 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
244 * @pipe: the pipe that the buffer belongs to
245 * @buf: the buffer to put a reference to
246 *
247 * Description:
248 * This function releases a reference to @buf.
249 */
250void generic_pipe_buf_release(struct pipe_inode_info *pipe,
251 struct pipe_buffer *buf)
252{
253 page_cache_release(buf->page);
254}
255EXPORT_SYMBOL(generic_pipe_buf_release);
256
257static const struct pipe_buf_operations anon_pipe_buf_ops = {
258 .can_merge = 1,
259 .confirm = generic_pipe_buf_confirm,
260 .release = anon_pipe_buf_release,
261 .steal = generic_pipe_buf_steal,
262 .get = generic_pipe_buf_get,
263};
264
265static const struct pipe_buf_operations packet_pipe_buf_ops = {
266 .can_merge = 0,
267 .confirm = generic_pipe_buf_confirm,
268 .release = anon_pipe_buf_release,
269 .steal = generic_pipe_buf_steal,
270 .get = generic_pipe_buf_get,
271};
272
273static ssize_t
274pipe_read(struct kiocb *iocb, const struct iovec *_iov,
275 unsigned long nr_segs, loff_t pos)
276{
277 struct file *filp = iocb->ki_filp;
278 struct pipe_inode_info *pipe = filp->private_data;
279 int do_wakeup;
280 ssize_t ret;
281 struct iovec *iov = (struct iovec *)_iov;
282 size_t total_len;
283 struct iov_iter iter;
284
285 total_len = iov_length(iov, nr_segs);
286 /* Null read succeeds. */
287 if (unlikely(total_len == 0))
288 return 0;
289
290 iov_iter_init(&iter, iov, nr_segs, total_len, 0);
291
292 do_wakeup = 0;
293 ret = 0;
294 __pipe_lock(pipe);
295 for (;;) {
296 int bufs = pipe->nrbufs;
297 if (bufs) {
298 int curbuf = pipe->curbuf;
299 struct pipe_buffer *buf = pipe->bufs + curbuf;
300 const struct pipe_buf_operations *ops = buf->ops;
301 size_t chars = buf->len;
302 size_t written;
303 int error;
304
305 if (chars > total_len)
306 chars = total_len;
307
308 error = ops->confirm(pipe, buf);
309 if (error) {
310 if (!ret)
311 ret = error;
312 break;
313 }
314
315 written = copy_page_to_iter(buf->page, buf->offset, chars, &iter);
316 if (unlikely(written < chars)) {
317 if (!ret)
318 ret = -EFAULT;
319 break;
320 }
321 ret += chars;
322 buf->offset += chars;
323 buf->len -= chars;
324
325 /* Was it a packet buffer? Clean up and exit */
326 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
327 total_len = chars;
328 buf->len = 0;
329 }
330
331 if (!buf->len) {
332 buf->ops = NULL;
333 ops->release(pipe, buf);
334 curbuf = (curbuf + 1) & (pipe->buffers - 1);
335 pipe->curbuf = curbuf;
336 pipe->nrbufs = --bufs;
337 do_wakeup = 1;
338 }
339 total_len -= chars;
340 if (!total_len)
341 break; /* common path: read succeeded */
342 }
343 if (bufs) /* More to do? */
344 continue;
345 if (!pipe->writers)
346 break;
347 if (!pipe->waiting_writers) {
348 /* syscall merging: Usually we must not sleep
349 * if O_NONBLOCK is set, or if we got some data.
350 * But if a writer sleeps in kernel space, then
351 * we can wait for that data without violating POSIX.
352 */
353 if (ret)
354 break;
355 if (filp->f_flags & O_NONBLOCK) {
356 ret = -EAGAIN;
357 break;
358 }
359 }
360 if (signal_pending(current)) {
361 if (!ret)
362 ret = -ERESTARTSYS;
363 break;
364 }
365 if (do_wakeup) {
366 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
367 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
368 }
369 pipe_wait(pipe);
370 }
371 __pipe_unlock(pipe);
372
373 /* Signal writers asynchronously that there is more room. */
374 if (do_wakeup) {
375 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
376 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
377 }
378 if (ret > 0)
379 file_accessed(filp);
380 return ret;
381}
382
383static inline int is_packetized(struct file *file)
384{
385 return (file->f_flags & O_DIRECT) != 0;
386}
387
388static ssize_t
389pipe_write(struct kiocb *iocb, const struct iovec *_iov,
390 unsigned long nr_segs, loff_t ppos)
391{
392 struct file *filp = iocb->ki_filp;
393 struct pipe_inode_info *pipe = filp->private_data;
394 ssize_t ret;
395 int do_wakeup;
396 struct iovec *iov = (struct iovec *)_iov;
397 size_t total_len;
398 ssize_t chars;
399
400 total_len = iov_length(iov, nr_segs);
401 /* Null write succeeds. */
402 if (unlikely(total_len == 0))
403 return 0;
404
405 do_wakeup = 0;
406 ret = 0;
407 __pipe_lock(pipe);
408
409 if (!pipe->readers) {
410 send_sig(SIGPIPE, current, 0);
411 ret = -EPIPE;
412 goto out;
413 }
414
415 /* We try to merge small writes */
416 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
417 if (pipe->nrbufs && chars != 0) {
418 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
419 (pipe->buffers - 1);
420 struct pipe_buffer *buf = pipe->bufs + lastbuf;
421 const struct pipe_buf_operations *ops = buf->ops;
422 int offset = buf->offset + buf->len;
423
424 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
425 int error, atomic = 1;
426 void *addr;
427
428 error = ops->confirm(pipe, buf);
429 if (error)
430 goto out;
431
432 iov_fault_in_pages_read(iov, chars);
433redo1:
434 if (atomic)
435 addr = kmap_atomic(buf->page);
436 else
437 addr = kmap(buf->page);
438 error = pipe_iov_copy_from_user(offset + addr, iov,
439 chars, atomic);
440 if (atomic)
441 kunmap_atomic(addr);
442 else
443 kunmap(buf->page);
444 ret = error;
445 do_wakeup = 1;
446 if (error) {
447 if (atomic) {
448 atomic = 0;
449 goto redo1;
450 }
451 goto out;
452 }
453 buf->len += chars;
454 total_len -= chars;
455 ret = chars;
456 if (!total_len)
457 goto out;
458 }
459 }
460
461 for (;;) {
462 int bufs;
463
464 if (!pipe->readers) {
465 send_sig(SIGPIPE, current, 0);
466 if (!ret)
467 ret = -EPIPE;
468 break;
469 }
470 bufs = pipe->nrbufs;
471 if (bufs < pipe->buffers) {
472 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
473 struct pipe_buffer *buf = pipe->bufs + newbuf;
474 struct page *page = pipe->tmp_page;
475 char *src;
476 int error, atomic = 1;
477
478 if (!page) {
479 page = alloc_page(GFP_HIGHUSER);
480 if (unlikely(!page)) {
481 ret = ret ? : -ENOMEM;
482 break;
483 }
484 pipe->tmp_page = page;
485 }
486 /* Always wake up, even if the copy fails. Otherwise
487 * we lock up (O_NONBLOCK-)readers that sleep due to
488 * syscall merging.
489 * FIXME! Is this really true?
490 */
491 do_wakeup = 1;
492 chars = PAGE_SIZE;
493 if (chars > total_len)
494 chars = total_len;
495
496 iov_fault_in_pages_read(iov, chars);
497redo2:
498 if (atomic)
499 src = kmap_atomic(page);
500 else
501 src = kmap(page);
502
503 error = pipe_iov_copy_from_user(src, iov, chars,
504 atomic);
505 if (atomic)
506 kunmap_atomic(src);
507 else
508 kunmap(page);
509
510 if (unlikely(error)) {
511 if (atomic) {
512 atomic = 0;
513 goto redo2;
514 }
515 if (!ret)
516 ret = error;
517 break;
518 }
519 ret += chars;
520
521 /* Insert it into the buffer array */
522 buf->page = page;
523 buf->ops = &anon_pipe_buf_ops;
524 buf->offset = 0;
525 buf->len = chars;
526 buf->flags = 0;
527 if (is_packetized(filp)) {
528 buf->ops = &packet_pipe_buf_ops;
529 buf->flags = PIPE_BUF_FLAG_PACKET;
530 }
531 pipe->nrbufs = ++bufs;
532 pipe->tmp_page = NULL;
533
534 total_len -= chars;
535 if (!total_len)
536 break;
537 }
538 if (bufs < pipe->buffers)
539 continue;
540 if (filp->f_flags & O_NONBLOCK) {
541 if (!ret)
542 ret = -EAGAIN;
543 break;
544 }
545 if (signal_pending(current)) {
546 if (!ret)
547 ret = -ERESTARTSYS;
548 break;
549 }
550 if (do_wakeup) {
551 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
552 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
553 do_wakeup = 0;
554 }
555 pipe->waiting_writers++;
556 pipe_wait(pipe);
557 pipe->waiting_writers--;
558 }
559out:
560 __pipe_unlock(pipe);
561 if (do_wakeup) {
562 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
563 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
564 }
565 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
566 int err = file_update_time(filp);
567 if (err)
568 ret = err;
569 sb_end_write(file_inode(filp)->i_sb);
570 }
571 return ret;
572}
573
574static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
575{
576 struct pipe_inode_info *pipe = filp->private_data;
577 int count, buf, nrbufs;
578
579 switch (cmd) {
580 case FIONREAD:
581 __pipe_lock(pipe);
582 count = 0;
583 buf = pipe->curbuf;
584 nrbufs = pipe->nrbufs;
585 while (--nrbufs >= 0) {
586 count += pipe->bufs[buf].len;
587 buf = (buf+1) & (pipe->buffers - 1);
588 }
589 __pipe_unlock(pipe);
590
591 return put_user(count, (int __user *)arg);
592 default:
593 return -ENOIOCTLCMD;
594 }
595}
596
597/* No kernel lock held - fine */
598static unsigned int
599pipe_poll(struct file *filp, poll_table *wait)
600{
601 unsigned int mask;
602 struct pipe_inode_info *pipe = filp->private_data;
603 int nrbufs;
604
605 poll_wait(filp, &pipe->wait, wait);
606
607 /* Reading only -- no need for acquiring the semaphore. */
608 nrbufs = pipe->nrbufs;
609 mask = 0;
610 if (filp->f_mode & FMODE_READ) {
611 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
612 if (!pipe->writers && filp->f_version != pipe->w_counter)
613 mask |= POLLHUP;
614 }
615
616 if (filp->f_mode & FMODE_WRITE) {
617 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
618 /*
619 * Most Unices do not set POLLERR for FIFOs but on Linux they
620 * behave exactly like pipes for poll().
621 */
622 if (!pipe->readers)
623 mask |= POLLERR;
624 }
625
626 return mask;
627}
628
629static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
630{
631 int kill = 0;
632
633 spin_lock(&inode->i_lock);
634 if (!--pipe->files) {
635 inode->i_pipe = NULL;
636 kill = 1;
637 }
638 spin_unlock(&inode->i_lock);
639
640 if (kill)
641 free_pipe_info(pipe);
642}
643
644static int
645pipe_release(struct inode *inode, struct file *file)
646{
647 struct pipe_inode_info *pipe = file->private_data;
648
649 __pipe_lock(pipe);
650 if (file->f_mode & FMODE_READ)
651 pipe->readers--;
652 if (file->f_mode & FMODE_WRITE)
653 pipe->writers--;
654
655 if (pipe->readers || pipe->writers) {
656 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
657 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
658 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
659 }
660 __pipe_unlock(pipe);
661
662 put_pipe_info(inode, pipe);
663 return 0;
664}
665
666static int
667pipe_fasync(int fd, struct file *filp, int on)
668{
669 struct pipe_inode_info *pipe = filp->private_data;
670 int retval = 0;
671
672 __pipe_lock(pipe);
673 if (filp->f_mode & FMODE_READ)
674 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
675 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
676 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
677 if (retval < 0 && (filp->f_mode & FMODE_READ))
678 /* this can happen only if on == T */
679 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
680 }
681 __pipe_unlock(pipe);
682 return retval;
683}
684
685struct pipe_inode_info *alloc_pipe_info(void)
686{
687 struct pipe_inode_info *pipe;
688
689 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
690 if (pipe) {
691 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
692 if (pipe->bufs) {
693 init_waitqueue_head(&pipe->wait);
694 pipe->r_counter = pipe->w_counter = 1;
695 pipe->buffers = PIPE_DEF_BUFFERS;
696 mutex_init(&pipe->mutex);
697 return pipe;
698 }
699 kfree(pipe);
700 }
701
702 return NULL;
703}
704
705void free_pipe_info(struct pipe_inode_info *pipe)
706{
707 int i;
708
709 for (i = 0; i < pipe->buffers; i++) {
710 struct pipe_buffer *buf = pipe->bufs + i;
711 if (buf->ops)
712 buf->ops->release(pipe, buf);
713 }
714 if (pipe->tmp_page)
715 __free_page(pipe->tmp_page);
716 kfree(pipe->bufs);
717 kfree(pipe);
718}
719
720static struct vfsmount *pipe_mnt __read_mostly;
721
722/*
723 * pipefs_dname() is called from d_path().
724 */
725static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
726{
727 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
728 dentry->d_inode->i_ino);
729}
730
731static const struct dentry_operations pipefs_dentry_operations = {
732 .d_dname = pipefs_dname,
733};
734
735static struct inode * get_pipe_inode(void)
736{
737 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
738 struct pipe_inode_info *pipe;
739
740 if (!inode)
741 goto fail_inode;
742
743 inode->i_ino = get_next_ino();
744
745 pipe = alloc_pipe_info();
746 if (!pipe)
747 goto fail_iput;
748
749 inode->i_pipe = pipe;
750 pipe->files = 2;
751 pipe->readers = pipe->writers = 1;
752 inode->i_fop = &pipefifo_fops;
753
754 /*
755 * Mark the inode dirty from the very beginning,
756 * that way it will never be moved to the dirty
757 * list because "mark_inode_dirty()" will think
758 * that it already _is_ on the dirty list.
759 */
760 inode->i_state = I_DIRTY;
761 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
762 inode->i_uid = current_fsuid();
763 inode->i_gid = current_fsgid();
764 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
765
766 return inode;
767
768fail_iput:
769 iput(inode);
770
771fail_inode:
772 return NULL;
773}
774
775int create_pipe_files(struct file **res, int flags)
776{
777 int err;
778 struct inode *inode = get_pipe_inode();
779 struct file *f;
780 struct path path;
781 static struct qstr name = { .name = "" };
782
783 if (!inode)
784 return -ENFILE;
785
786 err = -ENOMEM;
787 path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
788 if (!path.dentry)
789 goto err_inode;
790 path.mnt = mntget(pipe_mnt);
791
792 d_instantiate(path.dentry, inode);
793
794 err = -ENFILE;
795 f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
796 if (IS_ERR(f))
797 goto err_dentry;
798
799 f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
800 f->private_data = inode->i_pipe;
801
802 res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
803 if (IS_ERR(res[0]))
804 goto err_file;
805
806 path_get(&path);
807 res[0]->private_data = inode->i_pipe;
808 res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
809 res[1] = f;
810 return 0;
811
812err_file:
813 put_filp(f);
814err_dentry:
815 free_pipe_info(inode->i_pipe);
816 path_put(&path);
817 return err;
818
819err_inode:
820 free_pipe_info(inode->i_pipe);
821 iput(inode);
822 return err;
823}
824
825static int __do_pipe_flags(int *fd, struct file **files, int flags)
826{
827 int error;
828 int fdw, fdr;
829
830 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
831 return -EINVAL;
832
833 error = create_pipe_files(files, flags);
834 if (error)
835 return error;
836
837 error = get_unused_fd_flags(flags);
838 if (error < 0)
839 goto err_read_pipe;
840 fdr = error;
841
842 error = get_unused_fd_flags(flags);
843 if (error < 0)
844 goto err_fdr;
845 fdw = error;
846
847 audit_fd_pair(fdr, fdw);
848 fd[0] = fdr;
849 fd[1] = fdw;
850 return 0;
851
852 err_fdr:
853 put_unused_fd(fdr);
854 err_read_pipe:
855 fput(files[0]);
856 fput(files[1]);
857 return error;
858}
859
860int do_pipe_flags(int *fd, int flags)
861{
862 struct file *files[2];
863 int error = __do_pipe_flags(fd, files, flags);
864 if (!error) {
865 fd_install(fd[0], files[0]);
866 fd_install(fd[1], files[1]);
867 }
868 return error;
869}
870
871/*
872 * sys_pipe() is the normal C calling standard for creating
873 * a pipe. It's not the way Unix traditionally does this, though.
874 */
875SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
876{
877 struct file *files[2];
878 int fd[2];
879 int error;
880
881 error = __do_pipe_flags(fd, files, flags);
882 if (!error) {
883 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
884 fput(files[0]);
885 fput(files[1]);
886 put_unused_fd(fd[0]);
887 put_unused_fd(fd[1]);
888 error = -EFAULT;
889 } else {
890 fd_install(fd[0], files[0]);
891 fd_install(fd[1], files[1]);
892 }
893 }
894 return error;
895}
896
897SYSCALL_DEFINE1(pipe, int __user *, fildes)
898{
899 return sys_pipe2(fildes, 0);
900}
901
902static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
903{
904 int cur = *cnt;
905
906 while (cur == *cnt) {
907 pipe_wait(pipe);
908 if (signal_pending(current))
909 break;
910 }
911 return cur == *cnt ? -ERESTARTSYS : 0;
912}
913
914static void wake_up_partner(struct pipe_inode_info *pipe)
915{
916 wake_up_interruptible(&pipe->wait);
917}
918
919static int fifo_open(struct inode *inode, struct file *filp)
920{
921 struct pipe_inode_info *pipe;
922 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
923 int ret;
924
925 filp->f_version = 0;
926
927 spin_lock(&inode->i_lock);
928 if (inode->i_pipe) {
929 pipe = inode->i_pipe;
930 pipe->files++;
931 spin_unlock(&inode->i_lock);
932 } else {
933 spin_unlock(&inode->i_lock);
934 pipe = alloc_pipe_info();
935 if (!pipe)
936 return -ENOMEM;
937 pipe->files = 1;
938 spin_lock(&inode->i_lock);
939 if (unlikely(inode->i_pipe)) {
940 inode->i_pipe->files++;
941 spin_unlock(&inode->i_lock);
942 free_pipe_info(pipe);
943 pipe = inode->i_pipe;
944 } else {
945 inode->i_pipe = pipe;
946 spin_unlock(&inode->i_lock);
947 }
948 }
949 filp->private_data = pipe;
950 /* OK, we have a pipe and it's pinned down */
951
952 __pipe_lock(pipe);
953
954 /* We can only do regular read/write on fifos */
955 filp->f_mode &= (FMODE_READ | FMODE_WRITE);
956
957 switch (filp->f_mode) {
958 case FMODE_READ:
959 /*
960 * O_RDONLY
961 * POSIX.1 says that O_NONBLOCK means return with the FIFO
962 * opened, even when there is no process writing the FIFO.
963 */
964 pipe->r_counter++;
965 if (pipe->readers++ == 0)
966 wake_up_partner(pipe);
967
968 if (!is_pipe && !pipe->writers) {
969 if ((filp->f_flags & O_NONBLOCK)) {
970 /* suppress POLLHUP until we have
971 * seen a writer */
972 filp->f_version = pipe->w_counter;
973 } else {
974 if (wait_for_partner(pipe, &pipe->w_counter))
975 goto err_rd;
976 }
977 }
978 break;
979
980 case FMODE_WRITE:
981 /*
982 * O_WRONLY
983 * POSIX.1 says that O_NONBLOCK means return -1 with
984 * errno=ENXIO when there is no process reading the FIFO.
985 */
986 ret = -ENXIO;
987 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
988 goto err;
989
990 pipe->w_counter++;
991 if (!pipe->writers++)
992 wake_up_partner(pipe);
993
994 if (!is_pipe && !pipe->readers) {
995 if (wait_for_partner(pipe, &pipe->r_counter))
996 goto err_wr;
997 }
998 break;
999
1000 case FMODE_READ | FMODE_WRITE:
1001 /*
1002 * O_RDWR
1003 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1004 * This implementation will NEVER block on a O_RDWR open, since
1005 * the process can at least talk to itself.
1006 */
1007
1008 pipe->readers++;
1009 pipe->writers++;
1010 pipe->r_counter++;
1011 pipe->w_counter++;
1012 if (pipe->readers == 1 || pipe->writers == 1)
1013 wake_up_partner(pipe);
1014 break;
1015
1016 default:
1017 ret = -EINVAL;
1018 goto err;
1019 }
1020
1021 /* Ok! */
1022 __pipe_unlock(pipe);
1023 return 0;
1024
1025err_rd:
1026 if (!--pipe->readers)
1027 wake_up_interruptible(&pipe->wait);
1028 ret = -ERESTARTSYS;
1029 goto err;
1030
1031err_wr:
1032 if (!--pipe->writers)
1033 wake_up_interruptible(&pipe->wait);
1034 ret = -ERESTARTSYS;
1035 goto err;
1036
1037err:
1038 __pipe_unlock(pipe);
1039
1040 put_pipe_info(inode, pipe);
1041 return ret;
1042}
1043
1044const struct file_operations pipefifo_fops = {
1045 .open = fifo_open,
1046 .llseek = no_llseek,
1047 .read = do_sync_read,
1048 .aio_read = pipe_read,
1049 .write = do_sync_write,
1050 .aio_write = pipe_write,
1051 .poll = pipe_poll,
1052 .unlocked_ioctl = pipe_ioctl,
1053 .release = pipe_release,
1054 .fasync = pipe_fasync,
1055};
1056
1057/*
1058 * Allocate a new array of pipe buffers and copy the info over. Returns the
1059 * pipe size if successful, or return -ERROR on error.
1060 */
1061static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1062{
1063 struct pipe_buffer *bufs;
1064
1065 /*
1066 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1067 * expect a lot of shrink+grow operations, just free and allocate
1068 * again like we would do for growing. If the pipe currently
1069 * contains more buffers than arg, then return busy.
1070 */
1071 if (nr_pages < pipe->nrbufs)
1072 return -EBUSY;
1073
1074 bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1075 if (unlikely(!bufs))
1076 return -ENOMEM;
1077
1078 /*
1079 * The pipe array wraps around, so just start the new one at zero
1080 * and adjust the indexes.
1081 */
1082 if (pipe->nrbufs) {
1083 unsigned int tail;
1084 unsigned int head;
1085
1086 tail = pipe->curbuf + pipe->nrbufs;
1087 if (tail < pipe->buffers)
1088 tail = 0;
1089 else
1090 tail &= (pipe->buffers - 1);
1091
1092 head = pipe->nrbufs - tail;
1093 if (head)
1094 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1095 if (tail)
1096 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1097 }
1098
1099 pipe->curbuf = 0;
1100 kfree(pipe->bufs);
1101 pipe->bufs = bufs;
1102 pipe->buffers = nr_pages;
1103 return nr_pages * PAGE_SIZE;
1104}
1105
1106/*
1107 * Currently we rely on the pipe array holding a power-of-2 number
1108 * of pages.
1109 */
1110static inline unsigned int round_pipe_size(unsigned int size)
1111{
1112 unsigned long nr_pages;
1113
1114 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1115 return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1116}
1117
1118/*
1119 * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1120 * will return an error.
1121 */
1122int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1123 size_t *lenp, loff_t *ppos)
1124{
1125 int ret;
1126
1127 ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1128 if (ret < 0 || !write)
1129 return ret;
1130
1131 pipe_max_size = round_pipe_size(pipe_max_size);
1132 return ret;
1133}
1134
1135/*
1136 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1137 * location, so checking ->i_pipe is not enough to verify that this is a
1138 * pipe.
1139 */
1140struct pipe_inode_info *get_pipe_info(struct file *file)
1141{
1142 return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1143}
1144
1145long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1146{
1147 struct pipe_inode_info *pipe;
1148 long ret;
1149
1150 pipe = get_pipe_info(file);
1151 if (!pipe)
1152 return -EBADF;
1153
1154 __pipe_lock(pipe);
1155
1156 switch (cmd) {
1157 case F_SETPIPE_SZ: {
1158 unsigned int size, nr_pages;
1159
1160 size = round_pipe_size(arg);
1161 nr_pages = size >> PAGE_SHIFT;
1162
1163 ret = -EINVAL;
1164 if (!nr_pages)
1165 goto out;
1166
1167 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1168 ret = -EPERM;
1169 goto out;
1170 }
1171 ret = pipe_set_size(pipe, nr_pages);
1172 break;
1173 }
1174 case F_GETPIPE_SZ:
1175 ret = pipe->buffers * PAGE_SIZE;
1176 break;
1177 default:
1178 ret = -EINVAL;
1179 break;
1180 }
1181
1182out:
1183 __pipe_unlock(pipe);
1184 return ret;
1185}
1186
1187static const struct super_operations pipefs_ops = {
1188 .destroy_inode = free_inode_nonrcu,
1189 .statfs = simple_statfs,
1190};
1191
1192/*
1193 * pipefs should _never_ be mounted by userland - too much of security hassle,
1194 * no real gain from having the whole whorehouse mounted. So we don't need
1195 * any operations on the root directory. However, we need a non-trivial
1196 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1197 */
1198static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1199 int flags, const char *dev_name, void *data)
1200{
1201 return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1202 &pipefs_dentry_operations, PIPEFS_MAGIC);
1203}
1204
1205static struct file_system_type pipe_fs_type = {
1206 .name = "pipefs",
1207 .mount = pipefs_mount,
1208 .kill_sb = kill_anon_super,
1209};
1210
1211static int __init init_pipe_fs(void)
1212{
1213 int err = register_filesystem(&pipe_fs_type);
1214
1215 if (!err) {
1216 pipe_mnt = kern_mount(&pipe_fs_type);
1217 if (IS_ERR(pipe_mnt)) {
1218 err = PTR_ERR(pipe_mnt);
1219 unregister_filesystem(&pipe_fs_type);
1220 }
1221 }
1222 return err;
1223}
1224
1225fs_initcall(init_pipe_fs);