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