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/*
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
25#include <asm/uaccess.h>
26#include <asm/ioctls.h>
27
28/*
29 * The max size that a non-root user is allowed to grow the pipe. Can
30 * be set by root in /proc/sys/fs/pipe-max-size
31 */
32unsigned int pipe_max_size = 1048576;
33
34/*
35 * Minimum pipe size, as required by POSIX
36 */
37unsigned int pipe_min_size = PAGE_SIZE;
38
39/*
40 * We use a start+len construction, which provides full use of the
41 * allocated memory.
42 * -- Florian Coosmann (FGC)
43 *
44 * Reads with count = 0 should always return 0.
45 * -- Julian Bradfield 1999-06-07.
46 *
47 * FIFOs and Pipes now generate SIGIO for both readers and writers.
48 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
49 *
50 * pipe_read & write cleanup
51 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
52 */
53
54static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
55{
56 if (pipe->inode)
57 mutex_lock_nested(&pipe->inode->i_mutex, subclass);
58}
59
60void pipe_lock(struct pipe_inode_info *pipe)
61{
62 /*
63 * pipe_lock() nests non-pipe inode locks (for writing to a file)
64 */
65 pipe_lock_nested(pipe, I_MUTEX_PARENT);
66}
67EXPORT_SYMBOL(pipe_lock);
68
69void pipe_unlock(struct pipe_inode_info *pipe)
70{
71 if (pipe->inode)
72 mutex_unlock(&pipe->inode->i_mutex);
73}
74EXPORT_SYMBOL(pipe_unlock);
75
76void pipe_double_lock(struct pipe_inode_info *pipe1,
77 struct pipe_inode_info *pipe2)
78{
79 BUG_ON(pipe1 == pipe2);
80
81 if (pipe1 < pipe2) {
82 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
83 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
84 } else {
85 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
86 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
87 }
88}
89
90/* Drop the inode semaphore and wait for a pipe event, atomically */
91void pipe_wait(struct pipe_inode_info *pipe)
92{
93 DEFINE_WAIT(wait);
94
95 /*
96 * Pipes are system-local resources, so sleeping on them
97 * is considered a noninteractive wait:
98 */
99 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
100 pipe_unlock(pipe);
101 schedule();
102 finish_wait(&pipe->wait, &wait);
103 pipe_lock(pipe);
104}
105
106static int
107pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
108 int atomic)
109{
110 unsigned long copy;
111
112 while (len > 0) {
113 while (!iov->iov_len)
114 iov++;
115 copy = min_t(unsigned long, len, iov->iov_len);
116
117 if (atomic) {
118 if (__copy_from_user_inatomic(to, iov->iov_base, copy))
119 return -EFAULT;
120 } else {
121 if (copy_from_user(to, iov->iov_base, copy))
122 return -EFAULT;
123 }
124 to += copy;
125 len -= copy;
126 iov->iov_base += copy;
127 iov->iov_len -= copy;
128 }
129 return 0;
130}
131
132static int
133pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
134 int atomic)
135{
136 unsigned long copy;
137
138 while (len > 0) {
139 while (!iov->iov_len)
140 iov++;
141 copy = min_t(unsigned long, len, iov->iov_len);
142
143 if (atomic) {
144 if (__copy_to_user_inatomic(iov->iov_base, from, copy))
145 return -EFAULT;
146 } else {
147 if (copy_to_user(iov->iov_base, from, copy))
148 return -EFAULT;
149 }
150 from += copy;
151 len -= copy;
152 iov->iov_base += copy;
153 iov->iov_len -= copy;
154 }
155 return 0;
156}
157
158/*
159 * Attempt to pre-fault in the user memory, so we can use atomic copies.
160 * Returns the number of bytes not faulted in.
161 */
162static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
163{
164 while (!iov->iov_len)
165 iov++;
166
167 while (len > 0) {
168 unsigned long this_len;
169
170 this_len = min_t(unsigned long, len, iov->iov_len);
171 if (fault_in_pages_writeable(iov->iov_base, this_len))
172 break;
173
174 len -= this_len;
175 iov++;
176 }
177
178 return len;
179}
180
181/*
182 * Pre-fault in the user memory, so we can use atomic copies.
183 */
184static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
185{
186 while (!iov->iov_len)
187 iov++;
188
189 while (len > 0) {
190 unsigned long this_len;
191
192 this_len = min_t(unsigned long, len, iov->iov_len);
193 fault_in_pages_readable(iov->iov_base, this_len);
194 len -= this_len;
195 iov++;
196 }
197}
198
199static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
200 struct pipe_buffer *buf)
201{
202 struct page *page = buf->page;
203
204 /*
205 * If nobody else uses this page, and we don't already have a
206 * temporary page, let's keep track of it as a one-deep
207 * allocation cache. (Otherwise just release our reference to it)
208 */
209 if (page_count(page) == 1 && !pipe->tmp_page)
210 pipe->tmp_page = page;
211 else
212 page_cache_release(page);
213}
214
215/**
216 * generic_pipe_buf_map - virtually map a pipe buffer
217 * @pipe: the pipe that the buffer belongs to
218 * @buf: the buffer that should be mapped
219 * @atomic: whether to use an atomic map
220 *
221 * Description:
222 * This function returns a kernel virtual address mapping for the
223 * pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
224 * and the caller has to be careful not to fault before calling
225 * the unmap function.
226 *
227 * Note that this function occupies KM_USER0 if @atomic != 0.
228 */
229void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
230 struct pipe_buffer *buf, int atomic)
231{
232 if (atomic) {
233 buf->flags |= PIPE_BUF_FLAG_ATOMIC;
234 return kmap_atomic(buf->page);
235 }
236
237 return kmap(buf->page);
238}
239EXPORT_SYMBOL(generic_pipe_buf_map);
240
241/**
242 * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
243 * @pipe: the pipe that the buffer belongs to
244 * @buf: the buffer that should be unmapped
245 * @map_data: the data that the mapping function returned
246 *
247 * Description:
248 * This function undoes the mapping that ->map() provided.
249 */
250void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
251 struct pipe_buffer *buf, void *map_data)
252{
253 if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
254 buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
255 kunmap_atomic(map_data);
256 } else
257 kunmap(buf->page);
258}
259EXPORT_SYMBOL(generic_pipe_buf_unmap);
260
261/**
262 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
263 * @pipe: the pipe that the buffer belongs to
264 * @buf: the buffer to attempt to steal
265 *
266 * Description:
267 * This function attempts to steal the &struct page attached to
268 * @buf. If successful, this function returns 0 and returns with
269 * the page locked. The caller may then reuse the page for whatever
270 * he wishes; the typical use is insertion into a different file
271 * page cache.
272 */
273int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
274 struct pipe_buffer *buf)
275{
276 struct page *page = buf->page;
277
278 /*
279 * A reference of one is golden, that means that the owner of this
280 * page is the only one holding a reference to it. lock the page
281 * and return OK.
282 */
283 if (page_count(page) == 1) {
284 lock_page(page);
285 return 0;
286 }
287
288 return 1;
289}
290EXPORT_SYMBOL(generic_pipe_buf_steal);
291
292/**
293 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
294 * @pipe: the pipe that the buffer belongs to
295 * @buf: the buffer to get a reference to
296 *
297 * Description:
298 * This function grabs an extra reference to @buf. It's used in
299 * in the tee() system call, when we duplicate the buffers in one
300 * pipe into another.
301 */
302void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
303{
304 page_cache_get(buf->page);
305}
306EXPORT_SYMBOL(generic_pipe_buf_get);
307
308/**
309 * generic_pipe_buf_confirm - verify contents of the pipe buffer
310 * @info: the pipe that the buffer belongs to
311 * @buf: the buffer to confirm
312 *
313 * Description:
314 * This function does nothing, because the generic pipe code uses
315 * pages that are always good when inserted into the pipe.
316 */
317int generic_pipe_buf_confirm(struct pipe_inode_info *info,
318 struct pipe_buffer *buf)
319{
320 return 0;
321}
322EXPORT_SYMBOL(generic_pipe_buf_confirm);
323
324/**
325 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
326 * @pipe: the pipe that the buffer belongs to
327 * @buf: the buffer to put a reference to
328 *
329 * Description:
330 * This function releases a reference to @buf.
331 */
332void generic_pipe_buf_release(struct pipe_inode_info *pipe,
333 struct pipe_buffer *buf)
334{
335 page_cache_release(buf->page);
336}
337EXPORT_SYMBOL(generic_pipe_buf_release);
338
339static const struct pipe_buf_operations anon_pipe_buf_ops = {
340 .can_merge = 1,
341 .map = generic_pipe_buf_map,
342 .unmap = generic_pipe_buf_unmap,
343 .confirm = generic_pipe_buf_confirm,
344 .release = anon_pipe_buf_release,
345 .steal = generic_pipe_buf_steal,
346 .get = generic_pipe_buf_get,
347};
348
349static const struct pipe_buf_operations packet_pipe_buf_ops = {
350 .can_merge = 0,
351 .map = generic_pipe_buf_map,
352 .unmap = generic_pipe_buf_unmap,
353 .confirm = generic_pipe_buf_confirm,
354 .release = anon_pipe_buf_release,
355 .steal = generic_pipe_buf_steal,
356 .get = generic_pipe_buf_get,
357};
358
359static ssize_t
360pipe_read(struct kiocb *iocb, const struct iovec *_iov,
361 unsigned long nr_segs, loff_t pos)
362{
363 struct file *filp = iocb->ki_filp;
364 struct inode *inode = filp->f_path.dentry->d_inode;
365 struct pipe_inode_info *pipe;
366 int do_wakeup;
367 ssize_t ret;
368 struct iovec *iov = (struct iovec *)_iov;
369 size_t total_len;
370
371 total_len = iov_length(iov, nr_segs);
372 /* Null read succeeds. */
373 if (unlikely(total_len == 0))
374 return 0;
375
376 do_wakeup = 0;
377 ret = 0;
378 mutex_lock(&inode->i_mutex);
379 pipe = inode->i_pipe;
380 for (;;) {
381 int bufs = pipe->nrbufs;
382 if (bufs) {
383 int curbuf = pipe->curbuf;
384 struct pipe_buffer *buf = pipe->bufs + curbuf;
385 const struct pipe_buf_operations *ops = buf->ops;
386 void *addr;
387 size_t chars = buf->len;
388 int error, atomic;
389
390 if (chars > total_len)
391 chars = total_len;
392
393 error = ops->confirm(pipe, buf);
394 if (error) {
395 if (!ret)
396 ret = error;
397 break;
398 }
399
400 atomic = !iov_fault_in_pages_write(iov, chars);
401redo:
402 addr = ops->map(pipe, buf, atomic);
403 error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
404 ops->unmap(pipe, buf, addr);
405 if (unlikely(error)) {
406 /*
407 * Just retry with the slow path if we failed.
408 */
409 if (atomic) {
410 atomic = 0;
411 goto redo;
412 }
413 if (!ret)
414 ret = error;
415 break;
416 }
417 ret += chars;
418 buf->offset += chars;
419 buf->len -= chars;
420
421 /* Was it a packet buffer? Clean up and exit */
422 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
423 total_len = chars;
424 buf->len = 0;
425 }
426
427 if (!buf->len) {
428 buf->ops = NULL;
429 ops->release(pipe, buf);
430 curbuf = (curbuf + 1) & (pipe->buffers - 1);
431 pipe->curbuf = curbuf;
432 pipe->nrbufs = --bufs;
433 do_wakeup = 1;
434 }
435 total_len -= chars;
436 if (!total_len)
437 break; /* common path: read succeeded */
438 }
439 if (bufs) /* More to do? */
440 continue;
441 if (!pipe->writers)
442 break;
443 if (!pipe->waiting_writers) {
444 /* syscall merging: Usually we must not sleep
445 * if O_NONBLOCK is set, or if we got some data.
446 * But if a writer sleeps in kernel space, then
447 * we can wait for that data without violating POSIX.
448 */
449 if (ret)
450 break;
451 if (filp->f_flags & O_NONBLOCK) {
452 ret = -EAGAIN;
453 break;
454 }
455 }
456 if (signal_pending(current)) {
457 if (!ret)
458 ret = -ERESTARTSYS;
459 break;
460 }
461 if (do_wakeup) {
462 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
463 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
464 }
465 pipe_wait(pipe);
466 }
467 mutex_unlock(&inode->i_mutex);
468
469 /* Signal writers asynchronously that there is more room. */
470 if (do_wakeup) {
471 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
472 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
473 }
474 if (ret > 0)
475 file_accessed(filp);
476 return ret;
477}
478
479static inline int is_packetized(struct file *file)
480{
481 return (file->f_flags & O_DIRECT) != 0;
482}
483
484static ssize_t
485pipe_write(struct kiocb *iocb, const struct iovec *_iov,
486 unsigned long nr_segs, loff_t ppos)
487{
488 struct file *filp = iocb->ki_filp;
489 struct inode *inode = filp->f_path.dentry->d_inode;
490 struct pipe_inode_info *pipe;
491 ssize_t ret;
492 int do_wakeup;
493 struct iovec *iov = (struct iovec *)_iov;
494 size_t total_len;
495 ssize_t chars;
496
497 total_len = iov_length(iov, nr_segs);
498 /* Null write succeeds. */
499 if (unlikely(total_len == 0))
500 return 0;
501
502 do_wakeup = 0;
503 ret = 0;
504 mutex_lock(&inode->i_mutex);
505 pipe = inode->i_pipe;
506
507 if (!pipe->readers) {
508 send_sig(SIGPIPE, current, 0);
509 ret = -EPIPE;
510 goto out;
511 }
512
513 /* We try to merge small writes */
514 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
515 if (pipe->nrbufs && chars != 0) {
516 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
517 (pipe->buffers - 1);
518 struct pipe_buffer *buf = pipe->bufs + lastbuf;
519 const struct pipe_buf_operations *ops = buf->ops;
520 int offset = buf->offset + buf->len;
521
522 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
523 int error, atomic = 1;
524 void *addr;
525
526 error = ops->confirm(pipe, buf);
527 if (error)
528 goto out;
529
530 iov_fault_in_pages_read(iov, chars);
531redo1:
532 addr = ops->map(pipe, buf, atomic);
533 error = pipe_iov_copy_from_user(offset + addr, iov,
534 chars, atomic);
535 ops->unmap(pipe, buf, addr);
536 ret = error;
537 do_wakeup = 1;
538 if (error) {
539 if (atomic) {
540 atomic = 0;
541 goto redo1;
542 }
543 goto out;
544 }
545 buf->len += chars;
546 total_len -= chars;
547 ret = chars;
548 if (!total_len)
549 goto out;
550 }
551 }
552
553 for (;;) {
554 int bufs;
555
556 if (!pipe->readers) {
557 send_sig(SIGPIPE, current, 0);
558 if (!ret)
559 ret = -EPIPE;
560 break;
561 }
562 bufs = pipe->nrbufs;
563 if (bufs < pipe->buffers) {
564 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
565 struct pipe_buffer *buf = pipe->bufs + newbuf;
566 struct page *page = pipe->tmp_page;
567 char *src;
568 int error, atomic = 1;
569
570 if (!page) {
571 page = alloc_page(GFP_HIGHUSER);
572 if (unlikely(!page)) {
573 ret = ret ? : -ENOMEM;
574 break;
575 }
576 pipe->tmp_page = page;
577 }
578 /* Always wake up, even if the copy fails. Otherwise
579 * we lock up (O_NONBLOCK-)readers that sleep due to
580 * syscall merging.
581 * FIXME! Is this really true?
582 */
583 do_wakeup = 1;
584 chars = PAGE_SIZE;
585 if (chars > total_len)
586 chars = total_len;
587
588 iov_fault_in_pages_read(iov, chars);
589redo2:
590 if (atomic)
591 src = kmap_atomic(page);
592 else
593 src = kmap(page);
594
595 error = pipe_iov_copy_from_user(src, iov, chars,
596 atomic);
597 if (atomic)
598 kunmap_atomic(src);
599 else
600 kunmap(page);
601
602 if (unlikely(error)) {
603 if (atomic) {
604 atomic = 0;
605 goto redo2;
606 }
607 if (!ret)
608 ret = error;
609 break;
610 }
611 ret += chars;
612
613 /* Insert it into the buffer array */
614 buf->page = page;
615 buf->ops = &anon_pipe_buf_ops;
616 buf->offset = 0;
617 buf->len = chars;
618 buf->flags = 0;
619 if (is_packetized(filp)) {
620 buf->ops = &packet_pipe_buf_ops;
621 buf->flags = PIPE_BUF_FLAG_PACKET;
622 }
623 pipe->nrbufs = ++bufs;
624 pipe->tmp_page = NULL;
625
626 total_len -= chars;
627 if (!total_len)
628 break;
629 }
630 if (bufs < pipe->buffers)
631 continue;
632 if (filp->f_flags & O_NONBLOCK) {
633 if (!ret)
634 ret = -EAGAIN;
635 break;
636 }
637 if (signal_pending(current)) {
638 if (!ret)
639 ret = -ERESTARTSYS;
640 break;
641 }
642 if (do_wakeup) {
643 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
644 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
645 do_wakeup = 0;
646 }
647 pipe->waiting_writers++;
648 pipe_wait(pipe);
649 pipe->waiting_writers--;
650 }
651out:
652 mutex_unlock(&inode->i_mutex);
653 if (do_wakeup) {
654 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
655 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
656 }
657 if (ret > 0) {
658 int err = file_update_time(filp);
659 if (err)
660 ret = err;
661 }
662 return ret;
663}
664
665static ssize_t
666bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
667{
668 return -EBADF;
669}
670
671static ssize_t
672bad_pipe_w(struct file *filp, const char __user *buf, size_t count,
673 loff_t *ppos)
674{
675 return -EBADF;
676}
677
678static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
679{
680 struct inode *inode = filp->f_path.dentry->d_inode;
681 struct pipe_inode_info *pipe;
682 int count, buf, nrbufs;
683
684 switch (cmd) {
685 case FIONREAD:
686 mutex_lock(&inode->i_mutex);
687 pipe = inode->i_pipe;
688 count = 0;
689 buf = pipe->curbuf;
690 nrbufs = pipe->nrbufs;
691 while (--nrbufs >= 0) {
692 count += pipe->bufs[buf].len;
693 buf = (buf+1) & (pipe->buffers - 1);
694 }
695 mutex_unlock(&inode->i_mutex);
696
697 return put_user(count, (int __user *)arg);
698 default:
699 return -ENOIOCTLCMD;
700 }
701}
702
703/* No kernel lock held - fine */
704static unsigned int
705pipe_poll(struct file *filp, poll_table *wait)
706{
707 unsigned int mask;
708 struct inode *inode = filp->f_path.dentry->d_inode;
709 struct pipe_inode_info *pipe = inode->i_pipe;
710 int nrbufs;
711
712 poll_wait(filp, &pipe->wait, wait);
713
714 /* Reading only -- no need for acquiring the semaphore. */
715 nrbufs = pipe->nrbufs;
716 mask = 0;
717 if (filp->f_mode & FMODE_READ) {
718 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
719 if (!pipe->writers && filp->f_version != pipe->w_counter)
720 mask |= POLLHUP;
721 }
722
723 if (filp->f_mode & FMODE_WRITE) {
724 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
725 /*
726 * Most Unices do not set POLLERR for FIFOs but on Linux they
727 * behave exactly like pipes for poll().
728 */
729 if (!pipe->readers)
730 mask |= POLLERR;
731 }
732
733 return mask;
734}
735
736static int
737pipe_release(struct inode *inode, int decr, int decw)
738{
739 struct pipe_inode_info *pipe;
740
741 mutex_lock(&inode->i_mutex);
742 pipe = inode->i_pipe;
743 pipe->readers -= decr;
744 pipe->writers -= decw;
745
746 if (!pipe->readers && !pipe->writers) {
747 free_pipe_info(inode);
748 } else {
749 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
750 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
751 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
752 }
753 mutex_unlock(&inode->i_mutex);
754
755 return 0;
756}
757
758static int
759pipe_read_fasync(int fd, struct file *filp, int on)
760{
761 struct inode *inode = filp->f_path.dentry->d_inode;
762 int retval;
763
764 mutex_lock(&inode->i_mutex);
765 retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_readers);
766 mutex_unlock(&inode->i_mutex);
767
768 return retval;
769}
770
771
772static int
773pipe_write_fasync(int fd, struct file *filp, int on)
774{
775 struct inode *inode = filp->f_path.dentry->d_inode;
776 int retval;
777
778 mutex_lock(&inode->i_mutex);
779 retval = fasync_helper(fd, filp, on, &inode->i_pipe->fasync_writers);
780 mutex_unlock(&inode->i_mutex);
781
782 return retval;
783}
784
785
786static int
787pipe_rdwr_fasync(int fd, struct file *filp, int on)
788{
789 struct inode *inode = filp->f_path.dentry->d_inode;
790 struct pipe_inode_info *pipe = inode->i_pipe;
791 int retval;
792
793 mutex_lock(&inode->i_mutex);
794 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
795 if (retval >= 0) {
796 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
797 if (retval < 0) /* this can happen only if on == T */
798 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
799 }
800 mutex_unlock(&inode->i_mutex);
801 return retval;
802}
803
804
805static int
806pipe_read_release(struct inode *inode, struct file *filp)
807{
808 return pipe_release(inode, 1, 0);
809}
810
811static int
812pipe_write_release(struct inode *inode, struct file *filp)
813{
814 return pipe_release(inode, 0, 1);
815}
816
817static int
818pipe_rdwr_release(struct inode *inode, struct file *filp)
819{
820 int decr, decw;
821
822 decr = (filp->f_mode & FMODE_READ) != 0;
823 decw = (filp->f_mode & FMODE_WRITE) != 0;
824 return pipe_release(inode, decr, decw);
825}
826
827static int
828pipe_read_open(struct inode *inode, struct file *filp)
829{
830 int ret = -ENOENT;
831
832 mutex_lock(&inode->i_mutex);
833
834 if (inode->i_pipe) {
835 ret = 0;
836 inode->i_pipe->readers++;
837 }
838
839 mutex_unlock(&inode->i_mutex);
840
841 return ret;
842}
843
844static int
845pipe_write_open(struct inode *inode, struct file *filp)
846{
847 int ret = -ENOENT;
848
849 mutex_lock(&inode->i_mutex);
850
851 if (inode->i_pipe) {
852 ret = 0;
853 inode->i_pipe->writers++;
854 }
855
856 mutex_unlock(&inode->i_mutex);
857
858 return ret;
859}
860
861static int
862pipe_rdwr_open(struct inode *inode, struct file *filp)
863{
864 int ret = -ENOENT;
865
866 mutex_lock(&inode->i_mutex);
867
868 if (inode->i_pipe) {
869 ret = 0;
870 if (filp->f_mode & FMODE_READ)
871 inode->i_pipe->readers++;
872 if (filp->f_mode & FMODE_WRITE)
873 inode->i_pipe->writers++;
874 }
875
876 mutex_unlock(&inode->i_mutex);
877
878 return ret;
879}
880
881/*
882 * The file_operations structs are not static because they
883 * are also used in linux/fs/fifo.c to do operations on FIFOs.
884 *
885 * Pipes reuse fifos' file_operations structs.
886 */
887const struct file_operations read_pipefifo_fops = {
888 .llseek = no_llseek,
889 .read = do_sync_read,
890 .aio_read = pipe_read,
891 .write = bad_pipe_w,
892 .poll = pipe_poll,
893 .unlocked_ioctl = pipe_ioctl,
894 .open = pipe_read_open,
895 .release = pipe_read_release,
896 .fasync = pipe_read_fasync,
897};
898
899const struct file_operations write_pipefifo_fops = {
900 .llseek = no_llseek,
901 .read = bad_pipe_r,
902 .write = do_sync_write,
903 .aio_write = pipe_write,
904 .poll = pipe_poll,
905 .unlocked_ioctl = pipe_ioctl,
906 .open = pipe_write_open,
907 .release = pipe_write_release,
908 .fasync = pipe_write_fasync,
909};
910
911const struct file_operations rdwr_pipefifo_fops = {
912 .llseek = no_llseek,
913 .read = do_sync_read,
914 .aio_read = pipe_read,
915 .write = do_sync_write,
916 .aio_write = pipe_write,
917 .poll = pipe_poll,
918 .unlocked_ioctl = pipe_ioctl,
919 .open = pipe_rdwr_open,
920 .release = pipe_rdwr_release,
921 .fasync = pipe_rdwr_fasync,
922};
923
924struct pipe_inode_info * alloc_pipe_info(struct inode *inode)
925{
926 struct pipe_inode_info *pipe;
927
928 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
929 if (pipe) {
930 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
931 if (pipe->bufs) {
932 init_waitqueue_head(&pipe->wait);
933 pipe->r_counter = pipe->w_counter = 1;
934 pipe->inode = inode;
935 pipe->buffers = PIPE_DEF_BUFFERS;
936 return pipe;
937 }
938 kfree(pipe);
939 }
940
941 return NULL;
942}
943
944void __free_pipe_info(struct pipe_inode_info *pipe)
945{
946 int i;
947
948 for (i = 0; i < pipe->buffers; i++) {
949 struct pipe_buffer *buf = pipe->bufs + i;
950 if (buf->ops)
951 buf->ops->release(pipe, buf);
952 }
953 if (pipe->tmp_page)
954 __free_page(pipe->tmp_page);
955 kfree(pipe->bufs);
956 kfree(pipe);
957}
958
959void free_pipe_info(struct inode *inode)
960{
961 __free_pipe_info(inode->i_pipe);
962 inode->i_pipe = NULL;
963}
964
965static struct vfsmount *pipe_mnt __read_mostly;
966
967/*
968 * pipefs_dname() is called from d_path().
969 */
970static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
971{
972 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
973 dentry->d_inode->i_ino);
974}
975
976static const struct dentry_operations pipefs_dentry_operations = {
977 .d_dname = pipefs_dname,
978};
979
980static struct inode * get_pipe_inode(void)
981{
982 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
983 struct pipe_inode_info *pipe;
984
985 if (!inode)
986 goto fail_inode;
987
988 inode->i_ino = get_next_ino();
989
990 pipe = alloc_pipe_info(inode);
991 if (!pipe)
992 goto fail_iput;
993 inode->i_pipe = pipe;
994
995 pipe->readers = pipe->writers = 1;
996 inode->i_fop = &rdwr_pipefifo_fops;
997
998 /*
999 * Mark the inode dirty from the very beginning,
1000 * that way it will never be moved to the dirty
1001 * list because "mark_inode_dirty()" will think
1002 * that it already _is_ on the dirty list.
1003 */
1004 inode->i_state = I_DIRTY;
1005 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
1006 inode->i_uid = current_fsuid();
1007 inode->i_gid = current_fsgid();
1008 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1009
1010 return inode;
1011
1012fail_iput:
1013 iput(inode);
1014
1015fail_inode:
1016 return NULL;
1017}
1018
1019struct file *create_write_pipe(int flags)
1020{
1021 int err;
1022 struct inode *inode;
1023 struct file *f;
1024 struct path path;
1025 struct qstr name = { .name = "" };
1026
1027 err = -ENFILE;
1028 inode = get_pipe_inode();
1029 if (!inode)
1030 goto err;
1031
1032 err = -ENOMEM;
1033 path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
1034 if (!path.dentry)
1035 goto err_inode;
1036 path.mnt = mntget(pipe_mnt);
1037
1038 d_instantiate(path.dentry, inode);
1039
1040 err = -ENFILE;
1041 f = alloc_file(&path, FMODE_WRITE, &write_pipefifo_fops);
1042 if (!f)
1043 goto err_dentry;
1044 f->f_mapping = inode->i_mapping;
1045
1046 f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
1047 f->f_version = 0;
1048
1049 return f;
1050
1051 err_dentry:
1052 free_pipe_info(inode);
1053 path_put(&path);
1054 return ERR_PTR(err);
1055
1056 err_inode:
1057 free_pipe_info(inode);
1058 iput(inode);
1059 err:
1060 return ERR_PTR(err);
1061}
1062
1063void free_write_pipe(struct file *f)
1064{
1065 free_pipe_info(f->f_dentry->d_inode);
1066 path_put(&f->f_path);
1067 put_filp(f);
1068}
1069
1070struct file *create_read_pipe(struct file *wrf, int flags)
1071{
1072 /* Grab pipe from the writer */
1073 struct file *f = alloc_file(&wrf->f_path, FMODE_READ,
1074 &read_pipefifo_fops);
1075 if (!f)
1076 return ERR_PTR(-ENFILE);
1077
1078 path_get(&wrf->f_path);
1079 f->f_flags = O_RDONLY | (flags & O_NONBLOCK);
1080
1081 return f;
1082}
1083
1084int do_pipe_flags(int *fd, int flags)
1085{
1086 struct file *fw, *fr;
1087 int error;
1088 int fdw, fdr;
1089
1090 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
1091 return -EINVAL;
1092
1093 fw = create_write_pipe(flags);
1094 if (IS_ERR(fw))
1095 return PTR_ERR(fw);
1096 fr = create_read_pipe(fw, flags);
1097 error = PTR_ERR(fr);
1098 if (IS_ERR(fr))
1099 goto err_write_pipe;
1100
1101 error = get_unused_fd_flags(flags);
1102 if (error < 0)
1103 goto err_read_pipe;
1104 fdr = error;
1105
1106 error = get_unused_fd_flags(flags);
1107 if (error < 0)
1108 goto err_fdr;
1109 fdw = error;
1110
1111 audit_fd_pair(fdr, fdw);
1112 fd_install(fdr, fr);
1113 fd_install(fdw, fw);
1114 fd[0] = fdr;
1115 fd[1] = fdw;
1116
1117 return 0;
1118
1119 err_fdr:
1120 put_unused_fd(fdr);
1121 err_read_pipe:
1122 path_put(&fr->f_path);
1123 put_filp(fr);
1124 err_write_pipe:
1125 free_write_pipe(fw);
1126 return error;
1127}
1128
1129/*
1130 * sys_pipe() is the normal C calling standard for creating
1131 * a pipe. It's not the way Unix traditionally does this, though.
1132 */
1133SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1134{
1135 int fd[2];
1136 int error;
1137
1138 error = do_pipe_flags(fd, flags);
1139 if (!error) {
1140 if (copy_to_user(fildes, fd, sizeof(fd))) {
1141 sys_close(fd[0]);
1142 sys_close(fd[1]);
1143 error = -EFAULT;
1144 }
1145 }
1146 return error;
1147}
1148
1149SYSCALL_DEFINE1(pipe, int __user *, fildes)
1150{
1151 return sys_pipe2(fildes, 0);
1152}
1153
1154/*
1155 * Allocate a new array of pipe buffers and copy the info over. Returns the
1156 * pipe size if successful, or return -ERROR on error.
1157 */
1158static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1159{
1160 struct pipe_buffer *bufs;
1161
1162 /*
1163 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1164 * expect a lot of shrink+grow operations, just free and allocate
1165 * again like we would do for growing. If the pipe currently
1166 * contains more buffers than arg, then return busy.
1167 */
1168 if (nr_pages < pipe->nrbufs)
1169 return -EBUSY;
1170
1171 bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1172 if (unlikely(!bufs))
1173 return -ENOMEM;
1174
1175 /*
1176 * The pipe array wraps around, so just start the new one at zero
1177 * and adjust the indexes.
1178 */
1179 if (pipe->nrbufs) {
1180 unsigned int tail;
1181 unsigned int head;
1182
1183 tail = pipe->curbuf + pipe->nrbufs;
1184 if (tail < pipe->buffers)
1185 tail = 0;
1186 else
1187 tail &= (pipe->buffers - 1);
1188
1189 head = pipe->nrbufs - tail;
1190 if (head)
1191 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1192 if (tail)
1193 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1194 }
1195
1196 pipe->curbuf = 0;
1197 kfree(pipe->bufs);
1198 pipe->bufs = bufs;
1199 pipe->buffers = nr_pages;
1200 return nr_pages * PAGE_SIZE;
1201}
1202
1203/*
1204 * Currently we rely on the pipe array holding a power-of-2 number
1205 * of pages.
1206 */
1207static inline unsigned int round_pipe_size(unsigned int size)
1208{
1209 unsigned long nr_pages;
1210
1211 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1212 return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1213}
1214
1215/*
1216 * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1217 * will return an error.
1218 */
1219int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1220 size_t *lenp, loff_t *ppos)
1221{
1222 int ret;
1223
1224 ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1225 if (ret < 0 || !write)
1226 return ret;
1227
1228 pipe_max_size = round_pipe_size(pipe_max_size);
1229 return ret;
1230}
1231
1232/*
1233 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1234 * location, so checking ->i_pipe is not enough to verify that this is a
1235 * pipe.
1236 */
1237struct pipe_inode_info *get_pipe_info(struct file *file)
1238{
1239 struct inode *i = file->f_path.dentry->d_inode;
1240
1241 return S_ISFIFO(i->i_mode) ? i->i_pipe : NULL;
1242}
1243
1244long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1245{
1246 struct pipe_inode_info *pipe;
1247 long ret;
1248
1249 pipe = get_pipe_info(file);
1250 if (!pipe)
1251 return -EBADF;
1252
1253 mutex_lock(&pipe->inode->i_mutex);
1254
1255 switch (cmd) {
1256 case F_SETPIPE_SZ: {
1257 unsigned int size, nr_pages;
1258
1259 size = round_pipe_size(arg);
1260 nr_pages = size >> PAGE_SHIFT;
1261
1262 ret = -EINVAL;
1263 if (!nr_pages)
1264 goto out;
1265
1266 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1267 ret = -EPERM;
1268 goto out;
1269 }
1270 ret = pipe_set_size(pipe, nr_pages);
1271 break;
1272 }
1273 case F_GETPIPE_SZ:
1274 ret = pipe->buffers * PAGE_SIZE;
1275 break;
1276 default:
1277 ret = -EINVAL;
1278 break;
1279 }
1280
1281out:
1282 mutex_unlock(&pipe->inode->i_mutex);
1283 return ret;
1284}
1285
1286static const struct super_operations pipefs_ops = {
1287 .destroy_inode = free_inode_nonrcu,
1288 .statfs = simple_statfs,
1289};
1290
1291/*
1292 * pipefs should _never_ be mounted by userland - too much of security hassle,
1293 * no real gain from having the whole whorehouse mounted. So we don't need
1294 * any operations on the root directory. However, we need a non-trivial
1295 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1296 */
1297static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1298 int flags, const char *dev_name, void *data)
1299{
1300 return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1301 &pipefs_dentry_operations, PIPEFS_MAGIC);
1302}
1303
1304static struct file_system_type pipe_fs_type = {
1305 .name = "pipefs",
1306 .mount = pipefs_mount,
1307 .kill_sb = kill_anon_super,
1308};
1309
1310static int __init init_pipe_fs(void)
1311{
1312 int err = register_filesystem(&pipe_fs_type);
1313
1314 if (!err) {
1315 pipe_mnt = kern_mount(&pipe_fs_type);
1316 if (IS_ERR(pipe_mnt)) {
1317 err = PTR_ERR(pipe_mnt);
1318 unregister_filesystem(&pipe_fs_type);
1319 }
1320 }
1321 return err;
1322}
1323
1324fs_initcall(init_pipe_fs);