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
79static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
80{
81 if (pipe->files)
82 mutex_lock_nested(&pipe->mutex, subclass);
83}
84
85void pipe_lock(struct pipe_inode_info *pipe)
86{
87 /*
88 * pipe_lock() nests non-pipe inode locks (for writing to a file)
89 */
90 pipe_lock_nested(pipe, I_MUTEX_PARENT);
91}
92EXPORT_SYMBOL(pipe_lock);
93
94void pipe_unlock(struct pipe_inode_info *pipe)
95{
96 if (pipe->files)
97 mutex_unlock(&pipe->mutex);
98}
99EXPORT_SYMBOL(pipe_unlock);
100
101static inline void __pipe_lock(struct pipe_inode_info *pipe)
102{
103 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
104}
105
106static inline void __pipe_unlock(struct pipe_inode_info *pipe)
107{
108 mutex_unlock(&pipe->mutex);
109}
110
111void pipe_double_lock(struct pipe_inode_info *pipe1,
112 struct pipe_inode_info *pipe2)
113{
114 BUG_ON(pipe1 == pipe2);
115
116 if (pipe1 < pipe2) {
117 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
118 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
119 } else {
120 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
121 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
122 }
123}
124
125static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
126 struct pipe_buffer *buf)
127{
128 struct page *page = buf->page;
129
130 /*
131 * If nobody else uses this page, and we don't already have a
132 * temporary page, let's keep track of it as a one-deep
133 * allocation cache. (Otherwise just release our reference to it)
134 */
135 if (page_count(page) == 1 && !pipe->tmp_page)
136 pipe->tmp_page = page;
137 else
138 put_page(page);
139}
140
141static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
142 struct pipe_buffer *buf)
143{
144 struct page *page = buf->page;
145
146 if (page_count(page) != 1)
147 return false;
148 memcg_kmem_uncharge_page(page, 0);
149 __SetPageLocked(page);
150 return true;
151}
152
153/**
154 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
155 * @pipe: the pipe that the buffer belongs to
156 * @buf: the buffer to attempt to steal
157 *
158 * Description:
159 * This function attempts to steal the &struct page attached to
160 * @buf. If successful, this function returns 0 and returns with
161 * the page locked. The caller may then reuse the page for whatever
162 * he wishes; the typical use is insertion into a different file
163 * page cache.
164 */
165bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
166 struct pipe_buffer *buf)
167{
168 struct page *page = buf->page;
169
170 /*
171 * A reference of one is golden, that means that the owner of this
172 * page is the only one holding a reference to it. lock the page
173 * and return OK.
174 */
175 if (page_count(page) == 1) {
176 lock_page(page);
177 return true;
178 }
179 return false;
180}
181EXPORT_SYMBOL(generic_pipe_buf_try_steal);
182
183/**
184 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
185 * @pipe: the pipe that the buffer belongs to
186 * @buf: the buffer to get a reference to
187 *
188 * Description:
189 * This function grabs an extra reference to @buf. It's used in
190 * the tee() system call, when we duplicate the buffers in one
191 * pipe into another.
192 */
193bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
194{
195 return try_get_page(buf->page);
196}
197EXPORT_SYMBOL(generic_pipe_buf_get);
198
199/**
200 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
201 * @pipe: the pipe that the buffer belongs to
202 * @buf: the buffer to put a reference to
203 *
204 * Description:
205 * This function releases a reference to @buf.
206 */
207void generic_pipe_buf_release(struct pipe_inode_info *pipe,
208 struct pipe_buffer *buf)
209{
210 put_page(buf->page);
211}
212EXPORT_SYMBOL(generic_pipe_buf_release);
213
214static const struct pipe_buf_operations anon_pipe_buf_ops = {
215 .release = anon_pipe_buf_release,
216 .try_steal = anon_pipe_buf_try_steal,
217 .get = generic_pipe_buf_get,
218};
219
220/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
221static inline bool pipe_readable(const struct pipe_inode_info *pipe)
222{
223 unsigned int head = READ_ONCE(pipe->head);
224 unsigned int tail = READ_ONCE(pipe->tail);
225 unsigned int writers = READ_ONCE(pipe->writers);
226
227 return !pipe_empty(head, tail) || !writers;
228}
229
230static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
231 struct pipe_buffer *buf,
232 unsigned int tail)
233{
234 pipe_buf_release(pipe, buf);
235
236 /*
237 * If the pipe has a watch_queue, we need additional protection
238 * by the spinlock because notifications get posted with only
239 * this spinlock, no mutex
240 */
241 if (pipe_has_watch_queue(pipe)) {
242 spin_lock_irq(&pipe->rd_wait.lock);
243#ifdef CONFIG_WATCH_QUEUE
244 if (buf->flags & PIPE_BUF_FLAG_LOSS)
245 pipe->note_loss = true;
246#endif
247 pipe->tail = ++tail;
248 spin_unlock_irq(&pipe->rd_wait.lock);
249 return tail;
250 }
251
252 /*
253 * Without a watch_queue, we can simply increment the tail
254 * without the spinlock - the mutex is enough.
255 */
256 pipe->tail = ++tail;
257 return tail;
258}
259
260static ssize_t
261pipe_read(struct kiocb *iocb, struct iov_iter *to)
262{
263 size_t total_len = iov_iter_count(to);
264 struct file *filp = iocb->ki_filp;
265 struct pipe_inode_info *pipe = filp->private_data;
266 bool was_full, wake_next_reader = false;
267 ssize_t ret;
268
269 /* Null read succeeds. */
270 if (unlikely(total_len == 0))
271 return 0;
272
273 ret = 0;
274 __pipe_lock(pipe);
275
276 /*
277 * We only wake up writers if the pipe was full when we started
278 * reading in order to avoid unnecessary wakeups.
279 *
280 * But when we do wake up writers, we do so using a sync wakeup
281 * (WF_SYNC), because we want them to get going and generate more
282 * data for us.
283 */
284 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
285 for (;;) {
286 /* Read ->head with a barrier vs post_one_notification() */
287 unsigned int head = smp_load_acquire(&pipe->head);
288 unsigned int tail = pipe->tail;
289 unsigned int mask = pipe->ring_size - 1;
290
291#ifdef CONFIG_WATCH_QUEUE
292 if (pipe->note_loss) {
293 struct watch_notification n;
294
295 if (total_len < 8) {
296 if (ret == 0)
297 ret = -ENOBUFS;
298 break;
299 }
300
301 n.type = WATCH_TYPE_META;
302 n.subtype = WATCH_META_LOSS_NOTIFICATION;
303 n.info = watch_sizeof(n);
304 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
305 if (ret == 0)
306 ret = -EFAULT;
307 break;
308 }
309 ret += sizeof(n);
310 total_len -= sizeof(n);
311 pipe->note_loss = false;
312 }
313#endif
314
315 if (!pipe_empty(head, tail)) {
316 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
317 size_t chars = buf->len;
318 size_t written;
319 int error;
320
321 if (chars > total_len) {
322 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
323 if (ret == 0)
324 ret = -ENOBUFS;
325 break;
326 }
327 chars = total_len;
328 }
329
330 error = pipe_buf_confirm(pipe, buf);
331 if (error) {
332 if (!ret)
333 ret = error;
334 break;
335 }
336
337 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
338 if (unlikely(written < chars)) {
339 if (!ret)
340 ret = -EFAULT;
341 break;
342 }
343 ret += chars;
344 buf->offset += chars;
345 buf->len -= chars;
346
347 /* Was it a packet buffer? Clean up and exit */
348 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
349 total_len = chars;
350 buf->len = 0;
351 }
352
353 if (!buf->len)
354 tail = pipe_update_tail(pipe, buf, tail);
355 total_len -= chars;
356 if (!total_len)
357 break; /* common path: read succeeded */
358 if (!pipe_empty(head, tail)) /* More to do? */
359 continue;
360 }
361
362 if (!pipe->writers)
363 break;
364 if (ret)
365 break;
366 if ((filp->f_flags & O_NONBLOCK) ||
367 (iocb->ki_flags & IOCB_NOWAIT)) {
368 ret = -EAGAIN;
369 break;
370 }
371 __pipe_unlock(pipe);
372
373 /*
374 * We only get here if we didn't actually read anything.
375 *
376 * However, we could have seen (and removed) a zero-sized
377 * pipe buffer, and might have made space in the buffers
378 * that way.
379 *
380 * You can't make zero-sized pipe buffers by doing an empty
381 * write (not even in packet mode), but they can happen if
382 * the writer gets an EFAULT when trying to fill a buffer
383 * that already got allocated and inserted in the buffer
384 * array.
385 *
386 * So we still need to wake up any pending writers in the
387 * _very_ unlikely case that the pipe was full, but we got
388 * no data.
389 */
390 if (unlikely(was_full))
391 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
392 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
393
394 /*
395 * But because we didn't read anything, at this point we can
396 * just return directly with -ERESTARTSYS if we're interrupted,
397 * since we've done any required wakeups and there's no need
398 * to mark anything accessed. And we've dropped the lock.
399 */
400 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
401 return -ERESTARTSYS;
402
403 __pipe_lock(pipe);
404 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
405 wake_next_reader = true;
406 }
407 if (pipe_empty(pipe->head, pipe->tail))
408 wake_next_reader = false;
409 __pipe_unlock(pipe);
410
411 if (was_full)
412 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
413 if (wake_next_reader)
414 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
415 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
416 if (ret > 0)
417 file_accessed(filp);
418 return ret;
419}
420
421static inline int is_packetized(struct file *file)
422{
423 return (file->f_flags & O_DIRECT) != 0;
424}
425
426/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
427static inline bool pipe_writable(const struct pipe_inode_info *pipe)
428{
429 unsigned int head = READ_ONCE(pipe->head);
430 unsigned int tail = READ_ONCE(pipe->tail);
431 unsigned int max_usage = READ_ONCE(pipe->max_usage);
432
433 return !pipe_full(head, tail, max_usage) ||
434 !READ_ONCE(pipe->readers);
435}
436
437static ssize_t
438pipe_write(struct kiocb *iocb, struct iov_iter *from)
439{
440 struct file *filp = iocb->ki_filp;
441 struct pipe_inode_info *pipe = filp->private_data;
442 unsigned int head;
443 ssize_t ret = 0;
444 size_t total_len = iov_iter_count(from);
445 ssize_t chars;
446 bool was_empty = false;
447 bool wake_next_writer = false;
448
449 /*
450 * Reject writing to watch queue pipes before the point where we lock
451 * the pipe.
452 * Otherwise, lockdep would be unhappy if the caller already has another
453 * pipe locked.
454 * If we had to support locking a normal pipe and a notification pipe at
455 * the same time, we could set up lockdep annotations for that, but
456 * since we don't actually need that, it's simpler to just bail here.
457 */
458 if (pipe_has_watch_queue(pipe))
459 return -EXDEV;
460
461 /* Null write succeeds. */
462 if (unlikely(total_len == 0))
463 return 0;
464
465 __pipe_lock(pipe);
466
467 if (!pipe->readers) {
468 send_sig(SIGPIPE, current, 0);
469 ret = -EPIPE;
470 goto out;
471 }
472
473 /*
474 * If it wasn't empty we try to merge new data into
475 * the last buffer.
476 *
477 * That naturally merges small writes, but it also
478 * page-aligns the rest of the writes for large writes
479 * spanning multiple pages.
480 */
481 head = pipe->head;
482 was_empty = pipe_empty(head, pipe->tail);
483 chars = total_len & (PAGE_SIZE-1);
484 if (chars && !was_empty) {
485 unsigned int mask = pipe->ring_size - 1;
486 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
487 int offset = buf->offset + buf->len;
488
489 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
490 offset + chars <= PAGE_SIZE) {
491 ret = pipe_buf_confirm(pipe, buf);
492 if (ret)
493 goto out;
494
495 ret = copy_page_from_iter(buf->page, offset, chars, from);
496 if (unlikely(ret < chars)) {
497 ret = -EFAULT;
498 goto out;
499 }
500
501 buf->len += ret;
502 if (!iov_iter_count(from))
503 goto out;
504 }
505 }
506
507 for (;;) {
508 if (!pipe->readers) {
509 send_sig(SIGPIPE, current, 0);
510 if (!ret)
511 ret = -EPIPE;
512 break;
513 }
514
515 head = pipe->head;
516 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
517 unsigned int mask = pipe->ring_size - 1;
518 struct pipe_buffer *buf;
519 struct page *page = pipe->tmp_page;
520 int copied;
521
522 if (!page) {
523 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
524 if (unlikely(!page)) {
525 ret = ret ? : -ENOMEM;
526 break;
527 }
528 pipe->tmp_page = page;
529 }
530
531 /* Allocate a slot in the ring in advance and attach an
532 * empty buffer. If we fault or otherwise fail to use
533 * it, either the reader will consume it or it'll still
534 * be there for the next write.
535 */
536 pipe->head = head + 1;
537
538 /* Insert it into the buffer array */
539 buf = &pipe->bufs[head & mask];
540 buf->page = page;
541 buf->ops = &anon_pipe_buf_ops;
542 buf->offset = 0;
543 buf->len = 0;
544 if (is_packetized(filp))
545 buf->flags = PIPE_BUF_FLAG_PACKET;
546 else
547 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
548 pipe->tmp_page = NULL;
549
550 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
551 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
552 if (!ret)
553 ret = -EFAULT;
554 break;
555 }
556 ret += copied;
557 buf->len = copied;
558
559 if (!iov_iter_count(from))
560 break;
561 }
562
563 if (!pipe_full(head, pipe->tail, pipe->max_usage))
564 continue;
565
566 /* Wait for buffer space to become available. */
567 if ((filp->f_flags & O_NONBLOCK) ||
568 (iocb->ki_flags & IOCB_NOWAIT)) {
569 if (!ret)
570 ret = -EAGAIN;
571 break;
572 }
573 if (signal_pending(current)) {
574 if (!ret)
575 ret = -ERESTARTSYS;
576 break;
577 }
578
579 /*
580 * We're going to release the pipe lock and wait for more
581 * space. We wake up any readers if necessary, and then
582 * after waiting we need to re-check whether the pipe
583 * become empty while we dropped the lock.
584 */
585 __pipe_unlock(pipe);
586 if (was_empty)
587 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
588 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
589 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
590 __pipe_lock(pipe);
591 was_empty = pipe_empty(pipe->head, pipe->tail);
592 wake_next_writer = true;
593 }
594out:
595 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
596 wake_next_writer = false;
597 __pipe_unlock(pipe);
598
599 /*
600 * If we do do a wakeup event, we do a 'sync' wakeup, because we
601 * want the reader to start processing things asap, rather than
602 * leave the data pending.
603 *
604 * This is particularly important for small writes, because of
605 * how (for example) the GNU make jobserver uses small writes to
606 * wake up pending jobs
607 *
608 * Epoll nonsensically wants a wakeup whether the pipe
609 * was already empty or not.
610 */
611 if (was_empty || pipe->poll_usage)
612 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
613 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
614 if (wake_next_writer)
615 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
616 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
617 int err = file_update_time(filp);
618 if (err)
619 ret = err;
620 sb_end_write(file_inode(filp)->i_sb);
621 }
622 return ret;
623}
624
625static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
626{
627 struct pipe_inode_info *pipe = filp->private_data;
628 unsigned int count, head, tail, mask;
629
630 switch (cmd) {
631 case FIONREAD:
632 __pipe_lock(pipe);
633 count = 0;
634 head = pipe->head;
635 tail = pipe->tail;
636 mask = pipe->ring_size - 1;
637
638 while (tail != head) {
639 count += pipe->bufs[tail & mask].len;
640 tail++;
641 }
642 __pipe_unlock(pipe);
643
644 return put_user(count, (int __user *)arg);
645
646#ifdef CONFIG_WATCH_QUEUE
647 case IOC_WATCH_QUEUE_SET_SIZE: {
648 int ret;
649 __pipe_lock(pipe);
650 ret = watch_queue_set_size(pipe, arg);
651 __pipe_unlock(pipe);
652 return ret;
653 }
654
655 case IOC_WATCH_QUEUE_SET_FILTER:
656 return watch_queue_set_filter(
657 pipe, (struct watch_notification_filter __user *)arg);
658#endif
659
660 default:
661 return -ENOIOCTLCMD;
662 }
663}
664
665/* No kernel lock held - fine */
666static __poll_t
667pipe_poll(struct file *filp, poll_table *wait)
668{
669 __poll_t mask;
670 struct pipe_inode_info *pipe = filp->private_data;
671 unsigned int head, tail;
672
673 /* Epoll has some historical nasty semantics, this enables them */
674 WRITE_ONCE(pipe->poll_usage, true);
675
676 /*
677 * Reading pipe state only -- no need for acquiring the semaphore.
678 *
679 * But because this is racy, the code has to add the
680 * entry to the poll table _first_ ..
681 */
682 if (filp->f_mode & FMODE_READ)
683 poll_wait(filp, &pipe->rd_wait, wait);
684 if (filp->f_mode & FMODE_WRITE)
685 poll_wait(filp, &pipe->wr_wait, wait);
686
687 /*
688 * .. and only then can you do the racy tests. That way,
689 * if something changes and you got it wrong, the poll
690 * table entry will wake you up and fix it.
691 */
692 head = READ_ONCE(pipe->head);
693 tail = READ_ONCE(pipe->tail);
694
695 mask = 0;
696 if (filp->f_mode & FMODE_READ) {
697 if (!pipe_empty(head, tail))
698 mask |= EPOLLIN | EPOLLRDNORM;
699 if (!pipe->writers && filp->f_version != pipe->w_counter)
700 mask |= EPOLLHUP;
701 }
702
703 if (filp->f_mode & FMODE_WRITE) {
704 if (!pipe_full(head, tail, pipe->max_usage))
705 mask |= EPOLLOUT | EPOLLWRNORM;
706 /*
707 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
708 * behave exactly like pipes for poll().
709 */
710 if (!pipe->readers)
711 mask |= EPOLLERR;
712 }
713
714 return mask;
715}
716
717static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
718{
719 int kill = 0;
720
721 spin_lock(&inode->i_lock);
722 if (!--pipe->files) {
723 inode->i_pipe = NULL;
724 kill = 1;
725 }
726 spin_unlock(&inode->i_lock);
727
728 if (kill)
729 free_pipe_info(pipe);
730}
731
732static int
733pipe_release(struct inode *inode, struct file *file)
734{
735 struct pipe_inode_info *pipe = file->private_data;
736
737 __pipe_lock(pipe);
738 if (file->f_mode & FMODE_READ)
739 pipe->readers--;
740 if (file->f_mode & FMODE_WRITE)
741 pipe->writers--;
742
743 /* Was that the last reader or writer, but not the other side? */
744 if (!pipe->readers != !pipe->writers) {
745 wake_up_interruptible_all(&pipe->rd_wait);
746 wake_up_interruptible_all(&pipe->wr_wait);
747 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
748 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
749 }
750 __pipe_unlock(pipe);
751
752 put_pipe_info(inode, pipe);
753 return 0;
754}
755
756static int
757pipe_fasync(int fd, struct file *filp, int on)
758{
759 struct pipe_inode_info *pipe = filp->private_data;
760 int retval = 0;
761
762 __pipe_lock(pipe);
763 if (filp->f_mode & FMODE_READ)
764 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
765 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
766 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
767 if (retval < 0 && (filp->f_mode & FMODE_READ))
768 /* this can happen only if on == T */
769 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
770 }
771 __pipe_unlock(pipe);
772 return retval;
773}
774
775unsigned long account_pipe_buffers(struct user_struct *user,
776 unsigned long old, unsigned long new)
777{
778 return atomic_long_add_return(new - old, &user->pipe_bufs);
779}
780
781bool too_many_pipe_buffers_soft(unsigned long user_bufs)
782{
783 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
784
785 return soft_limit && user_bufs > soft_limit;
786}
787
788bool too_many_pipe_buffers_hard(unsigned long user_bufs)
789{
790 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
791
792 return hard_limit && user_bufs > hard_limit;
793}
794
795bool pipe_is_unprivileged_user(void)
796{
797 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
798}
799
800struct pipe_inode_info *alloc_pipe_info(void)
801{
802 struct pipe_inode_info *pipe;
803 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
804 struct user_struct *user = get_current_user();
805 unsigned long user_bufs;
806 unsigned int max_size = READ_ONCE(pipe_max_size);
807
808 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
809 if (pipe == NULL)
810 goto out_free_uid;
811
812 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
813 pipe_bufs = max_size >> PAGE_SHIFT;
814
815 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
816
817 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
818 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
819 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
820 }
821
822 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
823 goto out_revert_acct;
824
825 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
826 GFP_KERNEL_ACCOUNT);
827
828 if (pipe->bufs) {
829 init_waitqueue_head(&pipe->rd_wait);
830 init_waitqueue_head(&pipe->wr_wait);
831 pipe->r_counter = pipe->w_counter = 1;
832 pipe->max_usage = pipe_bufs;
833 pipe->ring_size = pipe_bufs;
834 pipe->nr_accounted = pipe_bufs;
835 pipe->user = user;
836 mutex_init(&pipe->mutex);
837 return pipe;
838 }
839
840out_revert_acct:
841 (void) account_pipe_buffers(user, pipe_bufs, 0);
842 kfree(pipe);
843out_free_uid:
844 free_uid(user);
845 return NULL;
846}
847
848void free_pipe_info(struct pipe_inode_info *pipe)
849{
850 unsigned int i;
851
852#ifdef CONFIG_WATCH_QUEUE
853 if (pipe->watch_queue)
854 watch_queue_clear(pipe->watch_queue);
855#endif
856
857 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
858 free_uid(pipe->user);
859 for (i = 0; i < pipe->ring_size; i++) {
860 struct pipe_buffer *buf = pipe->bufs + i;
861 if (buf->ops)
862 pipe_buf_release(pipe, buf);
863 }
864#ifdef CONFIG_WATCH_QUEUE
865 if (pipe->watch_queue)
866 put_watch_queue(pipe->watch_queue);
867#endif
868 if (pipe->tmp_page)
869 __free_page(pipe->tmp_page);
870 kfree(pipe->bufs);
871 kfree(pipe);
872}
873
874static struct vfsmount *pipe_mnt __ro_after_init;
875
876/*
877 * pipefs_dname() is called from d_path().
878 */
879static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
880{
881 return dynamic_dname(buffer, buflen, "pipe:[%lu]",
882 d_inode(dentry)->i_ino);
883}
884
885static const struct dentry_operations pipefs_dentry_operations = {
886 .d_dname = pipefs_dname,
887};
888
889static struct inode * get_pipe_inode(void)
890{
891 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
892 struct pipe_inode_info *pipe;
893
894 if (!inode)
895 goto fail_inode;
896
897 inode->i_ino = get_next_ino();
898
899 pipe = alloc_pipe_info();
900 if (!pipe)
901 goto fail_iput;
902
903 inode->i_pipe = pipe;
904 pipe->files = 2;
905 pipe->readers = pipe->writers = 1;
906 inode->i_fop = &pipefifo_fops;
907
908 /*
909 * Mark the inode dirty from the very beginning,
910 * that way it will never be moved to the dirty
911 * list because "mark_inode_dirty()" will think
912 * that it already _is_ on the dirty list.
913 */
914 inode->i_state = I_DIRTY;
915 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
916 inode->i_uid = current_fsuid();
917 inode->i_gid = current_fsgid();
918 simple_inode_init_ts(inode);
919
920 return inode;
921
922fail_iput:
923 iput(inode);
924
925fail_inode:
926 return NULL;
927}
928
929int create_pipe_files(struct file **res, int flags)
930{
931 struct inode *inode = get_pipe_inode();
932 struct file *f;
933 int error;
934
935 if (!inode)
936 return -ENFILE;
937
938 if (flags & O_NOTIFICATION_PIPE) {
939 error = watch_queue_init(inode->i_pipe);
940 if (error) {
941 free_pipe_info(inode->i_pipe);
942 iput(inode);
943 return error;
944 }
945 }
946
947 f = alloc_file_pseudo(inode, pipe_mnt, "",
948 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
949 &pipefifo_fops);
950 if (IS_ERR(f)) {
951 free_pipe_info(inode->i_pipe);
952 iput(inode);
953 return PTR_ERR(f);
954 }
955
956 f->private_data = inode->i_pipe;
957
958 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
959 &pipefifo_fops);
960 if (IS_ERR(res[0])) {
961 put_pipe_info(inode, inode->i_pipe);
962 fput(f);
963 return PTR_ERR(res[0]);
964 }
965 res[0]->private_data = inode->i_pipe;
966 res[1] = f;
967 stream_open(inode, res[0]);
968 stream_open(inode, res[1]);
969 return 0;
970}
971
972static int __do_pipe_flags(int *fd, struct file **files, int flags)
973{
974 int error;
975 int fdw, fdr;
976
977 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
978 return -EINVAL;
979
980 error = create_pipe_files(files, flags);
981 if (error)
982 return error;
983
984 error = get_unused_fd_flags(flags);
985 if (error < 0)
986 goto err_read_pipe;
987 fdr = error;
988
989 error = get_unused_fd_flags(flags);
990 if (error < 0)
991 goto err_fdr;
992 fdw = error;
993
994 audit_fd_pair(fdr, fdw);
995 fd[0] = fdr;
996 fd[1] = fdw;
997 /* pipe groks IOCB_NOWAIT */
998 files[0]->f_mode |= FMODE_NOWAIT;
999 files[1]->f_mode |= FMODE_NOWAIT;
1000 return 0;
1001
1002 err_fdr:
1003 put_unused_fd(fdr);
1004 err_read_pipe:
1005 fput(files[0]);
1006 fput(files[1]);
1007 return error;
1008}
1009
1010int do_pipe_flags(int *fd, int flags)
1011{
1012 struct file *files[2];
1013 int error = __do_pipe_flags(fd, files, flags);
1014 if (!error) {
1015 fd_install(fd[0], files[0]);
1016 fd_install(fd[1], files[1]);
1017 }
1018 return error;
1019}
1020
1021/*
1022 * sys_pipe() is the normal C calling standard for creating
1023 * a pipe. It's not the way Unix traditionally does this, though.
1024 */
1025static int do_pipe2(int __user *fildes, int flags)
1026{
1027 struct file *files[2];
1028 int fd[2];
1029 int error;
1030
1031 error = __do_pipe_flags(fd, files, flags);
1032 if (!error) {
1033 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1034 fput(files[0]);
1035 fput(files[1]);
1036 put_unused_fd(fd[0]);
1037 put_unused_fd(fd[1]);
1038 error = -EFAULT;
1039 } else {
1040 fd_install(fd[0], files[0]);
1041 fd_install(fd[1], files[1]);
1042 }
1043 }
1044 return error;
1045}
1046
1047SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1048{
1049 return do_pipe2(fildes, flags);
1050}
1051
1052SYSCALL_DEFINE1(pipe, int __user *, fildes)
1053{
1054 return do_pipe2(fildes, 0);
1055}
1056
1057/*
1058 * This is the stupid "wait for pipe to be readable or writable"
1059 * model.
1060 *
1061 * See pipe_read/write() for the proper kind of exclusive wait,
1062 * but that requires that we wake up any other readers/writers
1063 * if we then do not end up reading everything (ie the whole
1064 * "wake_next_reader/writer" logic in pipe_read/write()).
1065 */
1066void pipe_wait_readable(struct pipe_inode_info *pipe)
1067{
1068 pipe_unlock(pipe);
1069 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1070 pipe_lock(pipe);
1071}
1072
1073void pipe_wait_writable(struct pipe_inode_info *pipe)
1074{
1075 pipe_unlock(pipe);
1076 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1077 pipe_lock(pipe);
1078}
1079
1080/*
1081 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1082 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1083 * race with the count check and waitqueue prep.
1084 *
1085 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1086 * then check the condition you're waiting for, and only then sleep. But
1087 * because of the pipe lock, we can check the condition before being on
1088 * the wait queue.
1089 *
1090 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1091 */
1092static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1093{
1094 DEFINE_WAIT(rdwait);
1095 int cur = *cnt;
1096
1097 while (cur == *cnt) {
1098 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1099 pipe_unlock(pipe);
1100 schedule();
1101 finish_wait(&pipe->rd_wait, &rdwait);
1102 pipe_lock(pipe);
1103 if (signal_pending(current))
1104 break;
1105 }
1106 return cur == *cnt ? -ERESTARTSYS : 0;
1107}
1108
1109static void wake_up_partner(struct pipe_inode_info *pipe)
1110{
1111 wake_up_interruptible_all(&pipe->rd_wait);
1112}
1113
1114static int fifo_open(struct inode *inode, struct file *filp)
1115{
1116 struct pipe_inode_info *pipe;
1117 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1118 int ret;
1119
1120 filp->f_version = 0;
1121
1122 spin_lock(&inode->i_lock);
1123 if (inode->i_pipe) {
1124 pipe = inode->i_pipe;
1125 pipe->files++;
1126 spin_unlock(&inode->i_lock);
1127 } else {
1128 spin_unlock(&inode->i_lock);
1129 pipe = alloc_pipe_info();
1130 if (!pipe)
1131 return -ENOMEM;
1132 pipe->files = 1;
1133 spin_lock(&inode->i_lock);
1134 if (unlikely(inode->i_pipe)) {
1135 inode->i_pipe->files++;
1136 spin_unlock(&inode->i_lock);
1137 free_pipe_info(pipe);
1138 pipe = inode->i_pipe;
1139 } else {
1140 inode->i_pipe = pipe;
1141 spin_unlock(&inode->i_lock);
1142 }
1143 }
1144 filp->private_data = pipe;
1145 /* OK, we have a pipe and it's pinned down */
1146
1147 __pipe_lock(pipe);
1148
1149 /* We can only do regular read/write on fifos */
1150 stream_open(inode, filp);
1151
1152 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1153 case FMODE_READ:
1154 /*
1155 * O_RDONLY
1156 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1157 * opened, even when there is no process writing the FIFO.
1158 */
1159 pipe->r_counter++;
1160 if (pipe->readers++ == 0)
1161 wake_up_partner(pipe);
1162
1163 if (!is_pipe && !pipe->writers) {
1164 if ((filp->f_flags & O_NONBLOCK)) {
1165 /* suppress EPOLLHUP until we have
1166 * seen a writer */
1167 filp->f_version = pipe->w_counter;
1168 } else {
1169 if (wait_for_partner(pipe, &pipe->w_counter))
1170 goto err_rd;
1171 }
1172 }
1173 break;
1174
1175 case FMODE_WRITE:
1176 /*
1177 * O_WRONLY
1178 * POSIX.1 says that O_NONBLOCK means return -1 with
1179 * errno=ENXIO when there is no process reading the FIFO.
1180 */
1181 ret = -ENXIO;
1182 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1183 goto err;
1184
1185 pipe->w_counter++;
1186 if (!pipe->writers++)
1187 wake_up_partner(pipe);
1188
1189 if (!is_pipe && !pipe->readers) {
1190 if (wait_for_partner(pipe, &pipe->r_counter))
1191 goto err_wr;
1192 }
1193 break;
1194
1195 case FMODE_READ | FMODE_WRITE:
1196 /*
1197 * O_RDWR
1198 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1199 * This implementation will NEVER block on a O_RDWR open, since
1200 * the process can at least talk to itself.
1201 */
1202
1203 pipe->readers++;
1204 pipe->writers++;
1205 pipe->r_counter++;
1206 pipe->w_counter++;
1207 if (pipe->readers == 1 || pipe->writers == 1)
1208 wake_up_partner(pipe);
1209 break;
1210
1211 default:
1212 ret = -EINVAL;
1213 goto err;
1214 }
1215
1216 /* Ok! */
1217 __pipe_unlock(pipe);
1218 return 0;
1219
1220err_rd:
1221 if (!--pipe->readers)
1222 wake_up_interruptible(&pipe->wr_wait);
1223 ret = -ERESTARTSYS;
1224 goto err;
1225
1226err_wr:
1227 if (!--pipe->writers)
1228 wake_up_interruptible_all(&pipe->rd_wait);
1229 ret = -ERESTARTSYS;
1230 goto err;
1231
1232err:
1233 __pipe_unlock(pipe);
1234
1235 put_pipe_info(inode, pipe);
1236 return ret;
1237}
1238
1239const struct file_operations pipefifo_fops = {
1240 .open = fifo_open,
1241 .llseek = no_llseek,
1242 .read_iter = pipe_read,
1243 .write_iter = pipe_write,
1244 .poll = pipe_poll,
1245 .unlocked_ioctl = pipe_ioctl,
1246 .release = pipe_release,
1247 .fasync = pipe_fasync,
1248 .splice_write = iter_file_splice_write,
1249};
1250
1251/*
1252 * Currently we rely on the pipe array holding a power-of-2 number
1253 * of pages. Returns 0 on error.
1254 */
1255unsigned int round_pipe_size(unsigned int size)
1256{
1257 if (size > (1U << 31))
1258 return 0;
1259
1260 /* Minimum pipe size, as required by POSIX */
1261 if (size < PAGE_SIZE)
1262 return PAGE_SIZE;
1263
1264 return roundup_pow_of_two(size);
1265}
1266
1267/*
1268 * Resize the pipe ring to a number of slots.
1269 *
1270 * Note the pipe can be reduced in capacity, but only if the current
1271 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1272 * returned instead.
1273 */
1274int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1275{
1276 struct pipe_buffer *bufs;
1277 unsigned int head, tail, mask, n;
1278
1279 bufs = kcalloc(nr_slots, sizeof(*bufs),
1280 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1281 if (unlikely(!bufs))
1282 return -ENOMEM;
1283
1284 spin_lock_irq(&pipe->rd_wait.lock);
1285 mask = pipe->ring_size - 1;
1286 head = pipe->head;
1287 tail = pipe->tail;
1288
1289 n = pipe_occupancy(head, tail);
1290 if (nr_slots < n) {
1291 spin_unlock_irq(&pipe->rd_wait.lock);
1292 kfree(bufs);
1293 return -EBUSY;
1294 }
1295
1296 /*
1297 * The pipe array wraps around, so just start the new one at zero
1298 * and adjust the indices.
1299 */
1300 if (n > 0) {
1301 unsigned int h = head & mask;
1302 unsigned int t = tail & mask;
1303 if (h > t) {
1304 memcpy(bufs, pipe->bufs + t,
1305 n * sizeof(struct pipe_buffer));
1306 } else {
1307 unsigned int tsize = pipe->ring_size - t;
1308 if (h > 0)
1309 memcpy(bufs + tsize, pipe->bufs,
1310 h * sizeof(struct pipe_buffer));
1311 memcpy(bufs, pipe->bufs + t,
1312 tsize * sizeof(struct pipe_buffer));
1313 }
1314 }
1315
1316 head = n;
1317 tail = 0;
1318
1319 kfree(pipe->bufs);
1320 pipe->bufs = bufs;
1321 pipe->ring_size = nr_slots;
1322 if (pipe->max_usage > nr_slots)
1323 pipe->max_usage = nr_slots;
1324 pipe->tail = tail;
1325 pipe->head = head;
1326
1327 if (!pipe_has_watch_queue(pipe)) {
1328 pipe->max_usage = nr_slots;
1329 pipe->nr_accounted = nr_slots;
1330 }
1331
1332 spin_unlock_irq(&pipe->rd_wait.lock);
1333
1334 /* This might have made more room for writers */
1335 wake_up_interruptible(&pipe->wr_wait);
1336 return 0;
1337}
1338
1339/*
1340 * Allocate a new array of pipe buffers and copy the info over. Returns the
1341 * pipe size if successful, or return -ERROR on error.
1342 */
1343static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1344{
1345 unsigned long user_bufs;
1346 unsigned int nr_slots, size;
1347 long ret = 0;
1348
1349 if (pipe_has_watch_queue(pipe))
1350 return -EBUSY;
1351
1352 size = round_pipe_size(arg);
1353 nr_slots = size >> PAGE_SHIFT;
1354
1355 if (!nr_slots)
1356 return -EINVAL;
1357
1358 /*
1359 * If trying to increase the pipe capacity, check that an
1360 * unprivileged user is not trying to exceed various limits
1361 * (soft limit check here, hard limit check just below).
1362 * Decreasing the pipe capacity is always permitted, even
1363 * if the user is currently over a limit.
1364 */
1365 if (nr_slots > pipe->max_usage &&
1366 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1367 return -EPERM;
1368
1369 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1370
1371 if (nr_slots > pipe->max_usage &&
1372 (too_many_pipe_buffers_hard(user_bufs) ||
1373 too_many_pipe_buffers_soft(user_bufs)) &&
1374 pipe_is_unprivileged_user()) {
1375 ret = -EPERM;
1376 goto out_revert_acct;
1377 }
1378
1379 ret = pipe_resize_ring(pipe, nr_slots);
1380 if (ret < 0)
1381 goto out_revert_acct;
1382
1383 return pipe->max_usage * PAGE_SIZE;
1384
1385out_revert_acct:
1386 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1387 return ret;
1388}
1389
1390/*
1391 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1392 * not enough to verify that this is a pipe.
1393 */
1394struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1395{
1396 struct pipe_inode_info *pipe = file->private_data;
1397
1398 if (file->f_op != &pipefifo_fops || !pipe)
1399 return NULL;
1400 if (for_splice && pipe_has_watch_queue(pipe))
1401 return NULL;
1402 return pipe;
1403}
1404
1405long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1406{
1407 struct pipe_inode_info *pipe;
1408 long ret;
1409
1410 pipe = get_pipe_info(file, false);
1411 if (!pipe)
1412 return -EBADF;
1413
1414 __pipe_lock(pipe);
1415
1416 switch (cmd) {
1417 case F_SETPIPE_SZ:
1418 ret = pipe_set_size(pipe, arg);
1419 break;
1420 case F_GETPIPE_SZ:
1421 ret = pipe->max_usage * PAGE_SIZE;
1422 break;
1423 default:
1424 ret = -EINVAL;
1425 break;
1426 }
1427
1428 __pipe_unlock(pipe);
1429 return ret;
1430}
1431
1432static const struct super_operations pipefs_ops = {
1433 .destroy_inode = free_inode_nonrcu,
1434 .statfs = simple_statfs,
1435};
1436
1437/*
1438 * pipefs should _never_ be mounted by userland - too much of security hassle,
1439 * no real gain from having the whole whorehouse mounted. So we don't need
1440 * any operations on the root directory. However, we need a non-trivial
1441 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1442 */
1443
1444static int pipefs_init_fs_context(struct fs_context *fc)
1445{
1446 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1447 if (!ctx)
1448 return -ENOMEM;
1449 ctx->ops = &pipefs_ops;
1450 ctx->dops = &pipefs_dentry_operations;
1451 return 0;
1452}
1453
1454static struct file_system_type pipe_fs_type = {
1455 .name = "pipefs",
1456 .init_fs_context = pipefs_init_fs_context,
1457 .kill_sb = kill_anon_super,
1458};
1459
1460#ifdef CONFIG_SYSCTL
1461static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1462 unsigned int *valp,
1463 int write, void *data)
1464{
1465 if (write) {
1466 unsigned int val;
1467
1468 val = round_pipe_size(*lvalp);
1469 if (val == 0)
1470 return -EINVAL;
1471
1472 *valp = val;
1473 } else {
1474 unsigned int val = *valp;
1475 *lvalp = (unsigned long) val;
1476 }
1477
1478 return 0;
1479}
1480
1481static int proc_dopipe_max_size(struct ctl_table *table, int write,
1482 void *buffer, size_t *lenp, loff_t *ppos)
1483{
1484 return do_proc_douintvec(table, write, buffer, lenp, ppos,
1485 do_proc_dopipe_max_size_conv, NULL);
1486}
1487
1488static struct ctl_table fs_pipe_sysctls[] = {
1489 {
1490 .procname = "pipe-max-size",
1491 .data = &pipe_max_size,
1492 .maxlen = sizeof(pipe_max_size),
1493 .mode = 0644,
1494 .proc_handler = proc_dopipe_max_size,
1495 },
1496 {
1497 .procname = "pipe-user-pages-hard",
1498 .data = &pipe_user_pages_hard,
1499 .maxlen = sizeof(pipe_user_pages_hard),
1500 .mode = 0644,
1501 .proc_handler = proc_doulongvec_minmax,
1502 },
1503 {
1504 .procname = "pipe-user-pages-soft",
1505 .data = &pipe_user_pages_soft,
1506 .maxlen = sizeof(pipe_user_pages_soft),
1507 .mode = 0644,
1508 .proc_handler = proc_doulongvec_minmax,
1509 },
1510};
1511#endif
1512
1513static int __init init_pipe_fs(void)
1514{
1515 int err = register_filesystem(&pipe_fs_type);
1516
1517 if (!err) {
1518 pipe_mnt = kern_mount(&pipe_fs_type);
1519 if (IS_ERR(pipe_mnt)) {
1520 err = PTR_ERR(pipe_mnt);
1521 unregister_filesystem(&pipe_fs_type);
1522 }
1523 }
1524#ifdef CONFIG_SYSCTL
1525 register_sysctl_init("fs", fs_pipe_sysctls);
1526#endif
1527 return err;
1528}
1529
1530fs_initcall(init_pipe_fs);
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_version != 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
949 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
950 &pipefifo_fops);
951 if (IS_ERR(res[0])) {
952 put_pipe_info(inode, inode->i_pipe);
953 fput(f);
954 return PTR_ERR(res[0]);
955 }
956 res[0]->private_data = inode->i_pipe;
957 res[1] = f;
958 stream_open(inode, res[0]);
959 stream_open(inode, res[1]);
960 return 0;
961}
962
963static int __do_pipe_flags(int *fd, struct file **files, int flags)
964{
965 int error;
966 int fdw, fdr;
967
968 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
969 return -EINVAL;
970
971 error = create_pipe_files(files, flags);
972 if (error)
973 return error;
974
975 error = get_unused_fd_flags(flags);
976 if (error < 0)
977 goto err_read_pipe;
978 fdr = error;
979
980 error = get_unused_fd_flags(flags);
981 if (error < 0)
982 goto err_fdr;
983 fdw = error;
984
985 audit_fd_pair(fdr, fdw);
986 fd[0] = fdr;
987 fd[1] = fdw;
988 /* pipe groks IOCB_NOWAIT */
989 files[0]->f_mode |= FMODE_NOWAIT;
990 files[1]->f_mode |= FMODE_NOWAIT;
991 return 0;
992
993 err_fdr:
994 put_unused_fd(fdr);
995 err_read_pipe:
996 fput(files[0]);
997 fput(files[1]);
998 return error;
999}
1000
1001int do_pipe_flags(int *fd, int flags)
1002{
1003 struct file *files[2];
1004 int error = __do_pipe_flags(fd, files, flags);
1005 if (!error) {
1006 fd_install(fd[0], files[0]);
1007 fd_install(fd[1], files[1]);
1008 }
1009 return error;
1010}
1011
1012/*
1013 * sys_pipe() is the normal C calling standard for creating
1014 * a pipe. It's not the way Unix traditionally does this, though.
1015 */
1016static int do_pipe2(int __user *fildes, int flags)
1017{
1018 struct file *files[2];
1019 int fd[2];
1020 int error;
1021
1022 error = __do_pipe_flags(fd, files, flags);
1023 if (!error) {
1024 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1025 fput(files[0]);
1026 fput(files[1]);
1027 put_unused_fd(fd[0]);
1028 put_unused_fd(fd[1]);
1029 error = -EFAULT;
1030 } else {
1031 fd_install(fd[0], files[0]);
1032 fd_install(fd[1], files[1]);
1033 }
1034 }
1035 return error;
1036}
1037
1038SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1039{
1040 return do_pipe2(fildes, flags);
1041}
1042
1043SYSCALL_DEFINE1(pipe, int __user *, fildes)
1044{
1045 return do_pipe2(fildes, 0);
1046}
1047
1048/*
1049 * This is the stupid "wait for pipe to be readable or writable"
1050 * model.
1051 *
1052 * See pipe_read/write() for the proper kind of exclusive wait,
1053 * but that requires that we wake up any other readers/writers
1054 * if we then do not end up reading everything (ie the whole
1055 * "wake_next_reader/writer" logic in pipe_read/write()).
1056 */
1057void pipe_wait_readable(struct pipe_inode_info *pipe)
1058{
1059 pipe_unlock(pipe);
1060 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1061 pipe_lock(pipe);
1062}
1063
1064void pipe_wait_writable(struct pipe_inode_info *pipe)
1065{
1066 pipe_unlock(pipe);
1067 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1068 pipe_lock(pipe);
1069}
1070
1071/*
1072 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1073 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1074 * race with the count check and waitqueue prep.
1075 *
1076 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1077 * then check the condition you're waiting for, and only then sleep. But
1078 * because of the pipe lock, we can check the condition before being on
1079 * the wait queue.
1080 *
1081 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1082 */
1083static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1084{
1085 DEFINE_WAIT(rdwait);
1086 int cur = *cnt;
1087
1088 while (cur == *cnt) {
1089 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1090 pipe_unlock(pipe);
1091 schedule();
1092 finish_wait(&pipe->rd_wait, &rdwait);
1093 pipe_lock(pipe);
1094 if (signal_pending(current))
1095 break;
1096 }
1097 return cur == *cnt ? -ERESTARTSYS : 0;
1098}
1099
1100static void wake_up_partner(struct pipe_inode_info *pipe)
1101{
1102 wake_up_interruptible_all(&pipe->rd_wait);
1103}
1104
1105static int fifo_open(struct inode *inode, struct file *filp)
1106{
1107 struct pipe_inode_info *pipe;
1108 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1109 int ret;
1110
1111 filp->f_version = 0;
1112
1113 spin_lock(&inode->i_lock);
1114 if (inode->i_pipe) {
1115 pipe = inode->i_pipe;
1116 pipe->files++;
1117 spin_unlock(&inode->i_lock);
1118 } else {
1119 spin_unlock(&inode->i_lock);
1120 pipe = alloc_pipe_info();
1121 if (!pipe)
1122 return -ENOMEM;
1123 pipe->files = 1;
1124 spin_lock(&inode->i_lock);
1125 if (unlikely(inode->i_pipe)) {
1126 inode->i_pipe->files++;
1127 spin_unlock(&inode->i_lock);
1128 free_pipe_info(pipe);
1129 pipe = inode->i_pipe;
1130 } else {
1131 inode->i_pipe = pipe;
1132 spin_unlock(&inode->i_lock);
1133 }
1134 }
1135 filp->private_data = pipe;
1136 /* OK, we have a pipe and it's pinned down */
1137
1138 mutex_lock(&pipe->mutex);
1139
1140 /* We can only do regular read/write on fifos */
1141 stream_open(inode, filp);
1142
1143 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1144 case FMODE_READ:
1145 /*
1146 * O_RDONLY
1147 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1148 * opened, even when there is no process writing the FIFO.
1149 */
1150 pipe->r_counter++;
1151 if (pipe->readers++ == 0)
1152 wake_up_partner(pipe);
1153
1154 if (!is_pipe && !pipe->writers) {
1155 if ((filp->f_flags & O_NONBLOCK)) {
1156 /* suppress EPOLLHUP until we have
1157 * seen a writer */
1158 filp->f_version = pipe->w_counter;
1159 } else {
1160 if (wait_for_partner(pipe, &pipe->w_counter))
1161 goto err_rd;
1162 }
1163 }
1164 break;
1165
1166 case FMODE_WRITE:
1167 /*
1168 * O_WRONLY
1169 * POSIX.1 says that O_NONBLOCK means return -1 with
1170 * errno=ENXIO when there is no process reading the FIFO.
1171 */
1172 ret = -ENXIO;
1173 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1174 goto err;
1175
1176 pipe->w_counter++;
1177 if (!pipe->writers++)
1178 wake_up_partner(pipe);
1179
1180 if (!is_pipe && !pipe->readers) {
1181 if (wait_for_partner(pipe, &pipe->r_counter))
1182 goto err_wr;
1183 }
1184 break;
1185
1186 case FMODE_READ | FMODE_WRITE:
1187 /*
1188 * O_RDWR
1189 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1190 * This implementation will NEVER block on a O_RDWR open, since
1191 * the process can at least talk to itself.
1192 */
1193
1194 pipe->readers++;
1195 pipe->writers++;
1196 pipe->r_counter++;
1197 pipe->w_counter++;
1198 if (pipe->readers == 1 || pipe->writers == 1)
1199 wake_up_partner(pipe);
1200 break;
1201
1202 default:
1203 ret = -EINVAL;
1204 goto err;
1205 }
1206
1207 /* Ok! */
1208 mutex_unlock(&pipe->mutex);
1209 return 0;
1210
1211err_rd:
1212 if (!--pipe->readers)
1213 wake_up_interruptible(&pipe->wr_wait);
1214 ret = -ERESTARTSYS;
1215 goto err;
1216
1217err_wr:
1218 if (!--pipe->writers)
1219 wake_up_interruptible_all(&pipe->rd_wait);
1220 ret = -ERESTARTSYS;
1221 goto err;
1222
1223err:
1224 mutex_unlock(&pipe->mutex);
1225
1226 put_pipe_info(inode, pipe);
1227 return ret;
1228}
1229
1230const struct file_operations pipefifo_fops = {
1231 .open = fifo_open,
1232 .llseek = no_llseek,
1233 .read_iter = pipe_read,
1234 .write_iter = pipe_write,
1235 .poll = pipe_poll,
1236 .unlocked_ioctl = pipe_ioctl,
1237 .release = pipe_release,
1238 .fasync = pipe_fasync,
1239 .splice_write = iter_file_splice_write,
1240};
1241
1242/*
1243 * Currently we rely on the pipe array holding a power-of-2 number
1244 * of pages. Returns 0 on error.
1245 */
1246unsigned int round_pipe_size(unsigned int size)
1247{
1248 if (size > (1U << 31))
1249 return 0;
1250
1251 /* Minimum pipe size, as required by POSIX */
1252 if (size < PAGE_SIZE)
1253 return PAGE_SIZE;
1254
1255 return roundup_pow_of_two(size);
1256}
1257
1258/*
1259 * Resize the pipe ring to a number of slots.
1260 *
1261 * Note the pipe can be reduced in capacity, but only if the current
1262 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1263 * returned instead.
1264 */
1265int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1266{
1267 struct pipe_buffer *bufs;
1268 unsigned int head, tail, mask, n;
1269
1270 bufs = kcalloc(nr_slots, sizeof(*bufs),
1271 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1272 if (unlikely(!bufs))
1273 return -ENOMEM;
1274
1275 spin_lock_irq(&pipe->rd_wait.lock);
1276 mask = pipe->ring_size - 1;
1277 head = pipe->head;
1278 tail = pipe->tail;
1279
1280 n = pipe_occupancy(head, tail);
1281 if (nr_slots < n) {
1282 spin_unlock_irq(&pipe->rd_wait.lock);
1283 kfree(bufs);
1284 return -EBUSY;
1285 }
1286
1287 /*
1288 * The pipe array wraps around, so just start the new one at zero
1289 * and adjust the indices.
1290 */
1291 if (n > 0) {
1292 unsigned int h = head & mask;
1293 unsigned int t = tail & mask;
1294 if (h > t) {
1295 memcpy(bufs, pipe->bufs + t,
1296 n * sizeof(struct pipe_buffer));
1297 } else {
1298 unsigned int tsize = pipe->ring_size - t;
1299 if (h > 0)
1300 memcpy(bufs + tsize, pipe->bufs,
1301 h * sizeof(struct pipe_buffer));
1302 memcpy(bufs, pipe->bufs + t,
1303 tsize * sizeof(struct pipe_buffer));
1304 }
1305 }
1306
1307 head = n;
1308 tail = 0;
1309
1310 kfree(pipe->bufs);
1311 pipe->bufs = bufs;
1312 pipe->ring_size = nr_slots;
1313 if (pipe->max_usage > nr_slots)
1314 pipe->max_usage = nr_slots;
1315 pipe->tail = tail;
1316 pipe->head = head;
1317
1318 if (!pipe_has_watch_queue(pipe)) {
1319 pipe->max_usage = nr_slots;
1320 pipe->nr_accounted = nr_slots;
1321 }
1322
1323 spin_unlock_irq(&pipe->rd_wait.lock);
1324
1325 /* This might have made more room for writers */
1326 wake_up_interruptible(&pipe->wr_wait);
1327 return 0;
1328}
1329
1330/*
1331 * Allocate a new array of pipe buffers and copy the info over. Returns the
1332 * pipe size if successful, or return -ERROR on error.
1333 */
1334static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1335{
1336 unsigned long user_bufs;
1337 unsigned int nr_slots, size;
1338 long ret = 0;
1339
1340 if (pipe_has_watch_queue(pipe))
1341 return -EBUSY;
1342
1343 size = round_pipe_size(arg);
1344 nr_slots = size >> PAGE_SHIFT;
1345
1346 if (!nr_slots)
1347 return -EINVAL;
1348
1349 /*
1350 * If trying to increase the pipe capacity, check that an
1351 * unprivileged user is not trying to exceed various limits
1352 * (soft limit check here, hard limit check just below).
1353 * Decreasing the pipe capacity is always permitted, even
1354 * if the user is currently over a limit.
1355 */
1356 if (nr_slots > pipe->max_usage &&
1357 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1358 return -EPERM;
1359
1360 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1361
1362 if (nr_slots > pipe->max_usage &&
1363 (too_many_pipe_buffers_hard(user_bufs) ||
1364 too_many_pipe_buffers_soft(user_bufs)) &&
1365 pipe_is_unprivileged_user()) {
1366 ret = -EPERM;
1367 goto out_revert_acct;
1368 }
1369
1370 ret = pipe_resize_ring(pipe, nr_slots);
1371 if (ret < 0)
1372 goto out_revert_acct;
1373
1374 return pipe->max_usage * PAGE_SIZE;
1375
1376out_revert_acct:
1377 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1378 return ret;
1379}
1380
1381/*
1382 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1383 * not enough to verify that this is a pipe.
1384 */
1385struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1386{
1387 struct pipe_inode_info *pipe = file->private_data;
1388
1389 if (file->f_op != &pipefifo_fops || !pipe)
1390 return NULL;
1391 if (for_splice && pipe_has_watch_queue(pipe))
1392 return NULL;
1393 return pipe;
1394}
1395
1396long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1397{
1398 struct pipe_inode_info *pipe;
1399 long ret;
1400
1401 pipe = get_pipe_info(file, false);
1402 if (!pipe)
1403 return -EBADF;
1404
1405 mutex_lock(&pipe->mutex);
1406
1407 switch (cmd) {
1408 case F_SETPIPE_SZ:
1409 ret = pipe_set_size(pipe, arg);
1410 break;
1411 case F_GETPIPE_SZ:
1412 ret = pipe->max_usage * PAGE_SIZE;
1413 break;
1414 default:
1415 ret = -EINVAL;
1416 break;
1417 }
1418
1419 mutex_unlock(&pipe->mutex);
1420 return ret;
1421}
1422
1423static const struct super_operations pipefs_ops = {
1424 .destroy_inode = free_inode_nonrcu,
1425 .statfs = simple_statfs,
1426};
1427
1428/*
1429 * pipefs should _never_ be mounted by userland - too much of security hassle,
1430 * no real gain from having the whole whorehouse mounted. So we don't need
1431 * any operations on the root directory. However, we need a non-trivial
1432 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1433 */
1434
1435static int pipefs_init_fs_context(struct fs_context *fc)
1436{
1437 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1438 if (!ctx)
1439 return -ENOMEM;
1440 ctx->ops = &pipefs_ops;
1441 ctx->dops = &pipefs_dentry_operations;
1442 return 0;
1443}
1444
1445static struct file_system_type pipe_fs_type = {
1446 .name = "pipefs",
1447 .init_fs_context = pipefs_init_fs_context,
1448 .kill_sb = kill_anon_super,
1449};
1450
1451#ifdef CONFIG_SYSCTL
1452static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1453 unsigned int *valp,
1454 int write, void *data)
1455{
1456 if (write) {
1457 unsigned int val;
1458
1459 val = round_pipe_size(*lvalp);
1460 if (val == 0)
1461 return -EINVAL;
1462
1463 *valp = val;
1464 } else {
1465 unsigned int val = *valp;
1466 *lvalp = (unsigned long) val;
1467 }
1468
1469 return 0;
1470}
1471
1472static int proc_dopipe_max_size(struct ctl_table *table, int write,
1473 void *buffer, size_t *lenp, loff_t *ppos)
1474{
1475 return do_proc_douintvec(table, write, buffer, lenp, ppos,
1476 do_proc_dopipe_max_size_conv, NULL);
1477}
1478
1479static struct ctl_table fs_pipe_sysctls[] = {
1480 {
1481 .procname = "pipe-max-size",
1482 .data = &pipe_max_size,
1483 .maxlen = sizeof(pipe_max_size),
1484 .mode = 0644,
1485 .proc_handler = proc_dopipe_max_size,
1486 },
1487 {
1488 .procname = "pipe-user-pages-hard",
1489 .data = &pipe_user_pages_hard,
1490 .maxlen = sizeof(pipe_user_pages_hard),
1491 .mode = 0644,
1492 .proc_handler = proc_doulongvec_minmax,
1493 },
1494 {
1495 .procname = "pipe-user-pages-soft",
1496 .data = &pipe_user_pages_soft,
1497 .maxlen = sizeof(pipe_user_pages_soft),
1498 .mode = 0644,
1499 .proc_handler = proc_doulongvec_minmax,
1500 },
1501};
1502#endif
1503
1504static int __init init_pipe_fs(void)
1505{
1506 int err = register_filesystem(&pipe_fs_type);
1507
1508 if (!err) {
1509 pipe_mnt = kern_mount(&pipe_fs_type);
1510 if (IS_ERR(pipe_mnt)) {
1511 err = PTR_ERR(pipe_mnt);
1512 unregister_filesystem(&pipe_fs_type);
1513 }
1514 }
1515#ifdef CONFIG_SYSCTL
1516 register_sysctl_init("fs", fs_pipe_sysctls);
1517#endif
1518 return err;
1519}
1520
1521fs_initcall(init_pipe_fs);