projects/linux·blob master · refs

1635 lines · 39.5 KiB

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#include <linux/sort.h>
30
31#include <linux/uaccess.h>
32#include <asm/ioctls.h>
33
34#include "internal.h"
35
36/*
37 * New pipe buffers will be restricted to this size while the user is exceeding
38 * their pipe buffer quota. The general pipe use case needs at least two
39 * buffers: one for data yet to be read, and one for new data. If this is less
40 * than two, then a write to a non-empty pipe may block even if the pipe is not
41 * full. This can occur with GNU make jobserver or similar uses of pipes as
42 * semaphores: multiple processes may be waiting to write tokens back to the
43 * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
44 *
45 * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
46 * own risk, namely: pipe writes to non-full pipes may block until the pipe is
47 * emptied.
48 */
49#define PIPE_MIN_DEF_BUFFERS 2
50
51/*
52 * The max size that a non-root user is allowed to grow the pipe. Can
53 * be set by root in /proc/sys/fs/pipe-max-size
54 */
55static unsigned int pipe_max_size = 1048576;
56
57/* Maximum allocatable pages per user. Hard limit is unset by default, soft
58 * matches default values.
59 */
60static unsigned long pipe_user_pages_hard;
61static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
62
63/*
64 * We use head and tail indices that aren't masked off, except at the point of
65 * dereference, but rather they're allowed to wrap naturally. This means there
66 * isn't a dead spot in the buffer, but the ring has to be a power of two and
67 * <= 2^31.
68 * -- David Howells 2019-09-23.
69 *
70 * Reads with count = 0 should always return 0.
71 * -- Julian Bradfield 1999-06-07.
72 *
73 * FIFOs and Pipes now generate SIGIO for both readers and writers.
74 * -- Jeremy Elson <[email protected]> 2001-08-16
75 *
76 * pipe_read & write cleanup
77 * -- Manfred Spraul <[email protected]> 2002-05-09
78 */
79
80#ifdef CONFIG_PROVE_LOCKING
81static int pipe_lock_cmp_fn(const struct lockdep_map *a,
82 const struct lockdep_map *b)
83{
84 return cmp_int((unsigned long) a, (unsigned long) b);
85}
86#endif
87
88void pipe_lock(struct pipe_inode_info *pipe)
89{
90 if (pipe->files)
91 mutex_lock(&pipe->mutex);
92}
93EXPORT_SYMBOL(pipe_lock);
94
95void pipe_unlock(struct pipe_inode_info *pipe)
96{
97 if (pipe->files)
98 mutex_unlock(&pipe->mutex);
99}
100EXPORT_SYMBOL(pipe_unlock);
101
102void pipe_double_lock(struct pipe_inode_info *pipe1,
103 struct pipe_inode_info *pipe2)
104{
105 BUG_ON(pipe1 == pipe2);
106
107 if (pipe1 > pipe2)
108 swap(pipe1, pipe2);
109
110 pipe_lock(pipe1);
111 pipe_lock(pipe2);
112}
113
114static struct page *anon_pipe_prealloc_pop(struct anon_pipe_prealloc *prealloc)
115{
116 if (!prealloc->count)
117 return NULL;
118
119 prealloc->count--;
120
121 return prealloc->pages[prealloc->count];
122}
123
124/* Push a page to the prealloc pool. Returns true if added, false if full. */
125static bool anon_pipe_prealloc_push(struct anon_pipe_prealloc *prealloc,
126 struct page *page)
127{
128 if (prealloc->count >= PIPE_PREALLOC_MAX)
129 return false;
130 prealloc->pages[prealloc->count++] = page;
131 return true;
132}
133
134/*
135 * Top up the pipe's own pool, then take pipe->mutex and return with it held.
136 * The shortfall is allocated outside the lock; the push and the caller's write
137 * then run under a single lock acquisition, avoiding a separate prefill
138 * lock/unlock cycle. anon_pipe_get_page() drains the pool instead of allocating
139 * under the lock.
140 */
141static void anon_pipe_prefill_and_lock(struct pipe_inode_info *pipe, size_t total_len)
142{
143 struct page *pages[PIPE_PREALLOC_MAX];
144 unsigned int want, have, need, n = 0;
145
146 want = min_t(unsigned int, DIV_ROUND_UP(total_len, PAGE_SIZE),
147 PIPE_PREALLOC_MAX);
148 /* Unlocked read; the pool is refilled under the lock below. */
149 have = min_t(unsigned int, READ_ONCE(pipe->prealloc.count), want);
150 need = want - have;
151
152 if (!need) {
153 mutex_lock(&pipe->mutex);
154 return;
155 }
156
157 while (n < need) {
158 struct page *page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
159
160 if (!page)
161 break;
162 pages[n++] = page;
163 }
164
165 mutex_lock(&pipe->mutex);
166 while (n && anon_pipe_prealloc_push(&pipe->prealloc, pages[n - 1]))
167 n--;
168
169 /*
170 * Just flush any extra page that got affected by the TOCTOU
171 * effect
172 */
173 while (n)
174 put_page(pages[--n]);
175}
176
177/*
178 * Called with pipe->mutex held. Trim the pool down to PIPE_PREALLOC_KEEP under
179 * the lock, drop it, then free the excess outside the critical section.
180 */
181static void anon_pipe_trim_and_unlock(struct pipe_inode_info *pipe)
182{
183 struct page *excess[PIPE_PREALLOC_MAX];
184 unsigned int nexcess = 0;
185
186 while (pipe->prealloc.count > PIPE_PREALLOC_KEEP)
187 excess[nexcess++] = anon_pipe_prealloc_pop(&pipe->prealloc);
188 mutex_unlock(&pipe->mutex);
189
190 while (nexcess)
191 put_page(excess[--nexcess]);
192}
193
194static struct page *anon_pipe_get_page(struct pipe_inode_info *pipe)
195{
196 struct page *page;
197
198 /* Drain the prealloc pool before allocating. Called with mutex held. */
199 page = anon_pipe_prealloc_pop(&pipe->prealloc);
200 if (page)
201 return page;
202
203 /* FWIW: This is called with pipe->mutex held */
204 return alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
205}
206
207static void anon_pipe_put_page(struct pipe_inode_info *pipe,
208 struct page *page)
209{
210 if (page_count(page) == 1 &&
211 anon_pipe_prealloc_push(&pipe->prealloc, page))
212 return;
213
214 put_page(page);
215}
216
217static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
218 struct pipe_buffer *buf)
219{
220 struct page *page = buf->page;
221
222 anon_pipe_put_page(pipe, page);
223}
224
225static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
226 struct pipe_buffer *buf)
227{
228 struct page *page = buf->page;
229
230 if (page_count(page) != 1)
231 return false;
232 memcg_kmem_uncharge_page(page, 0);
233 __SetPageLocked(page);
234 return true;
235}
236
237/**
238 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
239 * @pipe: the pipe that the buffer belongs to
240 * @buf: the buffer to attempt to steal
241 *
242 * Description:
243 * This function attempts to steal the &struct page attached to
244 * @buf. If successful, this function returns 0 and returns with
245 * the page locked. The caller may then reuse the page for whatever
246 * he wishes; the typical use is insertion into a different file
247 * page cache.
248 */
249bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
250 struct pipe_buffer *buf)
251{
252 struct page *page = buf->page;
253
254 /*
255 * A reference of one is golden, that means that the owner of this
256 * page is the only one holding a reference to it. lock the page
257 * and return OK.
258 */
259 if (page_count(page) == 1) {
260 lock_page(page);
261 return true;
262 }
263 return false;
264}
265EXPORT_SYMBOL(generic_pipe_buf_try_steal);
266
267/**
268 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
269 * @pipe: the pipe that the buffer belongs to
270 * @buf: the buffer to get a reference to
271 *
272 * Description:
273 * This function grabs an extra reference to @buf. It's used in
274 * the tee() system call, when we duplicate the buffers in one
275 * pipe into another.
276 */
277bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
278{
279 return try_get_page(buf->page);
280}
281EXPORT_SYMBOL(generic_pipe_buf_get);
282
283/**
284 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
285 * @pipe: the pipe that the buffer belongs to
286 * @buf: the buffer to put a reference to
287 *
288 * Description:
289 * This function releases a reference to @buf.
290 */
291void generic_pipe_buf_release(struct pipe_inode_info *pipe,
292 struct pipe_buffer *buf)
293{
294 put_page(buf->page);
295}
296EXPORT_SYMBOL(generic_pipe_buf_release);
297
298static const struct pipe_buf_operations anon_pipe_buf_ops = {
299 .release = anon_pipe_buf_release,
300 .try_steal = anon_pipe_buf_try_steal,
301 .get = generic_pipe_buf_get,
302};
303
304/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
305static inline bool pipe_readable(const struct pipe_inode_info *pipe)
306{
307 union pipe_index idx = { .head_tail = READ_ONCE(pipe->head_tail) };
308 unsigned int writers = READ_ONCE(pipe->writers);
309
310 return !pipe_empty(idx.head, idx.tail) || !writers;
311}
312
313static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
314 struct pipe_buffer *buf,
315 unsigned int tail)
316{
317 pipe_buf_release(pipe, buf);
318
319 /*
320 * If the pipe has a watch_queue, we need additional protection
321 * by the spinlock because notifications get posted with only
322 * this spinlock, no mutex
323 */
324 if (pipe_has_watch_queue(pipe)) {
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 pipe->tail = ++tail;
331 spin_unlock_irq(&pipe->rd_wait.lock);
332 return tail;
333 }
334
335 /*
336 * Without a watch_queue, we can simply increment the tail
337 * without the spinlock - the mutex is enough.
338 */
339 pipe->tail = ++tail;
340 return tail;
341}
342
343static ssize_t
344anon_pipe_read(struct kiocb *iocb, struct iov_iter *to)
345{
346 size_t total_len = iov_iter_count(to);
347 struct file *filp = iocb->ki_filp;
348 struct pipe_inode_info *pipe = filp->private_data;
349 bool wake_writer = false, wake_next_reader = false;
350 ssize_t ret;
351
352 /* Null read succeeds. */
353 if (unlikely(total_len == 0))
354 return 0;
355
356 ret = 0;
357 mutex_lock(&pipe->mutex);
358
359 /*
360 * We only wake up writers if the pipe was full when we started reading
361 * and it is no longer full after reading to avoid unnecessary wakeups.
362 *
363 * But when we do wake up writers, we do so using a sync wakeup
364 * (WF_SYNC), because we want them to get going and generate more
365 * data for us.
366 */
367 for (;;) {
368 /* Read ->head with a barrier vs post_one_notification() */
369 unsigned int head = smp_load_acquire(&pipe->head);
370 unsigned int tail = pipe->tail;
371
372#ifdef CONFIG_WATCH_QUEUE
373 if (pipe->note_loss) {
374 struct watch_notification n;
375
376 if (total_len < 8) {
377 if (ret == 0)
378 ret = -ENOBUFS;
379 break;
380 }
381
382 n.type = WATCH_TYPE_META;
383 n.subtype = WATCH_META_LOSS_NOTIFICATION;
384 n.info = watch_sizeof(n);
385 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
386 if (ret == 0)
387 ret = -EFAULT;
388 break;
389 }
390 ret += sizeof(n);
391 total_len -= sizeof(n);
392 pipe->note_loss = false;
393 }
394#endif
395
396 if (!pipe_empty(head, tail)) {
397 struct pipe_buffer *buf = pipe_buf(pipe, tail);
398 size_t chars = buf->len;
399 size_t written;
400 int error;
401
402 if (chars > total_len) {
403 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
404 if (ret == 0)
405 ret = -ENOBUFS;
406 break;
407 }
408 chars = total_len;
409 }
410
411 error = pipe_buf_confirm(pipe, buf);
412 if (error) {
413 if (!ret)
414 ret = error;
415 break;
416 }
417
418 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
419 if (unlikely(written < chars)) {
420 if (!ret)
421 ret = -EFAULT;
422 break;
423 }
424 ret += chars;
425 buf->offset += chars;
426 buf->len -= chars;
427
428 /* Was it a packet buffer? Clean up and exit */
429 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
430 total_len = chars;
431 buf->len = 0;
432 }
433
434 if (!buf->len) {
435 wake_writer |= pipe_full(head, tail, pipe->max_usage);
436 tail = pipe_update_tail(pipe, buf, tail);
437 }
438 total_len -= chars;
439 if (!total_len)
440 break; /* common path: read succeeded */
441 if (!pipe_empty(head, tail)) /* More to do? */
442 continue;
443 }
444
445 if (!pipe->writers)
446 break;
447 if (ret)
448 break;
449 if ((filp->f_flags & O_NONBLOCK) ||
450 (iocb->ki_flags & IOCB_NOWAIT)) {
451 ret = -EAGAIN;
452 break;
453 }
454 mutex_unlock(&pipe->mutex);
455 /*
456 * We only get here if we didn't actually read anything.
457 *
458 * But because we didn't read anything, at this point we can
459 * just return directly with -ERESTARTSYS if we're interrupted,
460 * since we've done any required wakeups and there's no need
461 * to mark anything accessed. And we've dropped the lock.
462 */
463 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
464 return -ERESTARTSYS;
465
466 wake_next_reader = true;
467 mutex_lock(&pipe->mutex);
468 }
469 if (pipe_is_empty(pipe))
470 wake_next_reader = false;
471 /* Consumed buffers may have refilled the pool; trim it and unlock. */
472 anon_pipe_trim_and_unlock(pipe);
473
474 if (wake_writer)
475 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
476 if (wake_next_reader)
477 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
478 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
479 return ret;
480}
481
482static ssize_t
483fifo_pipe_read(struct kiocb *iocb, struct iov_iter *to)
484{
485 int ret = anon_pipe_read(iocb, to);
486 if (ret > 0)
487 file_accessed(iocb->ki_filp);
488 return ret;
489}
490
491static inline int is_packetized(struct file *file)
492{
493 return (file->f_flags & O_DIRECT) != 0;
494}
495
496/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
497static inline bool pipe_writable(const struct pipe_inode_info *pipe)
498{
499 union pipe_index idx = { .head_tail = READ_ONCE(pipe->head_tail) };
500 unsigned int max_usage = READ_ONCE(pipe->max_usage);
501
502 return !pipe_full(idx.head, idx.tail, max_usage) ||
503 !READ_ONCE(pipe->readers);
504}
505
506static ssize_t
507anon_pipe_write(struct kiocb *iocb, struct iov_iter *from)
508{
509 struct file *filp = iocb->ki_filp;
510 struct pipe_inode_info *pipe = filp->private_data;
511 unsigned int head;
512 ssize_t ret = 0;
513 size_t total_len = iov_iter_count(from);
514 ssize_t chars;
515 bool was_empty = false;
516 bool wake_next_writer = false;
517
518 /*
519 * Reject writing to watch queue pipes before the point where we lock
520 * the pipe.
521 * Otherwise, lockdep would be unhappy if the caller already has another
522 * pipe locked.
523 * If we had to support locking a normal pipe and a notification pipe at
524 * the same time, we could set up lockdep annotations for that, but
525 * since we don't actually need that, it's simpler to just bail here.
526 */
527 if (pipe_has_watch_queue(pipe))
528 return -EXDEV;
529
530 /* Null write succeeds. */
531 if (unlikely(total_len == 0))
532 return 0;
533
534 anon_pipe_prefill_and_lock(pipe, total_len);
535
536 if (!pipe->readers) {
537 if ((iocb->ki_flags & IOCB_NOSIGNAL) == 0)
538 send_sig(SIGPIPE, current, 0);
539 ret = -EPIPE;
540 goto out;
541 }
542
543 /*
544 * If it wasn't empty we try to merge new data into
545 * the last buffer.
546 *
547 * That naturally merges small writes, but it also
548 * page-aligns the rest of the writes for large writes
549 * spanning multiple pages.
550 */
551 head = pipe->head;
552 was_empty = pipe_empty(head, pipe->tail);
553 chars = total_len & (PAGE_SIZE-1);
554 if (chars && !was_empty) {
555 struct pipe_buffer *buf = pipe_buf(pipe, head - 1);
556 int offset = buf->offset + buf->len;
557
558 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
559 offset + chars <= PAGE_SIZE) {
560 ret = pipe_buf_confirm(pipe, buf);
561 if (ret)
562 goto out;
563
564 ret = copy_page_from_iter(buf->page, offset, chars, from);
565 if (unlikely(ret < chars)) {
566 ret = -EFAULT;
567 goto out;
568 }
569
570 buf->len += ret;
571 if (!iov_iter_count(from))
572 goto out;
573 }
574 }
575
576 for (;;) {
577 if (!pipe->readers) {
578 if ((iocb->ki_flags & IOCB_NOSIGNAL) == 0)
579 send_sig(SIGPIPE, current, 0);
580 if (!ret)
581 ret = -EPIPE;
582 break;
583 }
584
585 head = pipe->head;
586 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
587 struct pipe_buffer *buf;
588 struct page *page;
589 int copied;
590
591 page = anon_pipe_get_page(pipe);
592 if (unlikely(!page)) {
593 if (!ret)
594 ret = -ENOMEM;
595 break;
596 }
597
598 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
599 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
600 anon_pipe_put_page(pipe, page);
601 if (!ret)
602 ret = -EFAULT;
603 break;
604 }
605
606 pipe->head = head + 1;
607 /* Insert it into the buffer array */
608 buf = pipe_buf(pipe, head);
609 buf->page = page;
610 buf->ops = &anon_pipe_buf_ops;
611 buf->offset = 0;
612 if (is_packetized(filp))
613 buf->flags = PIPE_BUF_FLAG_PACKET;
614 else
615 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
616
617 buf->len = copied;
618 ret += copied;
619
620 if (!iov_iter_count(from))
621 break;
622
623 continue;
624 }
625
626 /* Wait for buffer space to become available. */
627 if ((filp->f_flags & O_NONBLOCK) ||
628 (iocb->ki_flags & IOCB_NOWAIT)) {
629 if (!ret)
630 ret = -EAGAIN;
631 break;
632 }
633 if (signal_pending(current)) {
634 if (!ret)
635 ret = -ERESTARTSYS;
636 break;
637 }
638
639 /*
640 * We're going to release the pipe lock and wait for more
641 * space. We wake up any readers if necessary, and then
642 * after waiting we need to re-check whether the pipe
643 * become empty while we dropped the lock.
644 */
645 mutex_unlock(&pipe->mutex);
646 if (was_empty)
647 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
648 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
649 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
650 mutex_lock(&pipe->mutex);
651 was_empty = pipe_is_empty(pipe);
652 wake_next_writer = true;
653 }
654out:
655 if (pipe_is_full(pipe))
656 wake_next_writer = false;
657 anon_pipe_trim_and_unlock(pipe);
658
659 /*
660 * If we do do a wakeup event, we do a 'sync' wakeup, because we
661 * want the reader to start processing things asap, rather than
662 * leave the data pending.
663 *
664 * This is particularly important for small writes, because of
665 * how (for example) the GNU make jobserver uses small writes to
666 * wake up pending jobs
667 *
668 * ->pseudo_edgetrigger enables per-write wakeups, see pipe_poll()
669 */
670 if (was_empty || READ_ONCE(pipe->pseudo_edgetrigger))
671 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
672 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
673 if (wake_next_writer)
674 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
675 return ret;
676}
677
678static ssize_t
679fifo_pipe_write(struct kiocb *iocb, struct iov_iter *from)
680{
681 int ret = anon_pipe_write(iocb, from);
682 if (ret > 0) {
683 struct file *filp = iocb->ki_filp;
684 if (sb_start_write_trylock(file_inode(filp)->i_sb)) {
685 int err = file_update_time(filp);
686 if (err)
687 ret = err;
688 sb_end_write(file_inode(filp)->i_sb);
689 }
690 }
691 return ret;
692}
693
694static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
695{
696 struct pipe_inode_info *pipe = filp->private_data;
697 unsigned int count, head, tail;
698
699 switch (cmd) {
700 case FIONREAD:
701 mutex_lock(&pipe->mutex);
702 count = 0;
703 head = pipe->head;
704 tail = pipe->tail;
705
706 while (!pipe_empty(head, tail)) {
707 count += pipe_buf(pipe, tail)->len;
708 tail++;
709 }
710 mutex_unlock(&pipe->mutex);
711
712 return put_user(count, (int __user *)arg);
713
714#ifdef CONFIG_WATCH_QUEUE
715 case IOC_WATCH_QUEUE_SET_SIZE: {
716 int ret;
717 mutex_lock(&pipe->mutex);
718 ret = watch_queue_set_size(pipe, arg);
719 mutex_unlock(&pipe->mutex);
720 return ret;
721 }
722
723 case IOC_WATCH_QUEUE_SET_FILTER:
724 return watch_queue_set_filter(
725 pipe, (struct watch_notification_filter __user *)arg);
726#endif
727
728 default:
729 return -ENOIOCTLCMD;
730 }
731}
732
733static __poll_t
734pipe_poll(struct file *filp, poll_table *wait)
735{
736 __poll_t mask;
737 struct pipe_inode_info *pipe = filp->private_data;
738 union pipe_index idx;
739
740 /*
741 * Legacy epoll(EPOLLET) users depend on historical per-write wakeups,
742 * see 3a34b13a88ca ("pipe: make pipe writes always wake up readers")
743 * and the ->pseudo_edgetrigger check in anon_pipe_write().
744 * Currently io_uring sets EPOLLET for multishot polls, so it gets the
745 * same behaviour.
746 */
747 if ((filp->f_mode & FMODE_READ) &&
748 wait && (wait->_key & EPOLLET) &&
749 unlikely(!READ_ONCE(pipe->pseudo_edgetrigger)))
750 WRITE_ONCE(pipe->pseudo_edgetrigger, true);
751
752 /*
753 * Reading pipe state only -- no need for acquiring the semaphore.
754 *
755 * But because this is racy, the code has to add the
756 * entry to the poll table _first_ ..
757 */
758 if (filp->f_mode & FMODE_READ)
759 poll_wait(filp, &pipe->rd_wait, wait);
760 if (filp->f_mode & FMODE_WRITE)
761 poll_wait(filp, &pipe->wr_wait, wait);
762
763 /*
764 * .. and only then can you do the racy tests. That way,
765 * if something changes and you got it wrong, the poll
766 * table entry will wake you up and fix it.
767 */
768 idx.head_tail = READ_ONCE(pipe->head_tail);
769
770 mask = 0;
771 if (filp->f_mode & FMODE_READ) {
772 if (!pipe_empty(idx.head, idx.tail))
773 mask |= EPOLLIN | EPOLLRDNORM;
774 if (!pipe->writers && filp->f_pipe != pipe->w_counter)
775 mask |= EPOLLHUP;
776 }
777
778 if (filp->f_mode & FMODE_WRITE) {
779 if (!pipe_full(idx.head, idx.tail, pipe->max_usage))
780 mask |= EPOLLOUT | EPOLLWRNORM;
781 /*
782 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
783 * behave exactly like pipes for poll().
784 */
785 if (!pipe->readers)
786 mask |= EPOLLERR;
787 }
788
789 return mask;
790}
791
792static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
793{
794 int kill = 0;
795
796 spin_lock(&inode->i_lock);
797 if (!--pipe->files) {
798 inode->i_pipe = NULL;
799 kill = 1;
800 }
801 spin_unlock(&inode->i_lock);
802
803 if (kill)
804 free_pipe_info(pipe);
805}
806
807static int
808pipe_release(struct inode *inode, struct file *file)
809{
810 struct pipe_inode_info *pipe = file->private_data;
811
812 mutex_lock(&pipe->mutex);
813 if (file->f_mode & FMODE_READ)
814 pipe->readers--;
815 if (file->f_mode & FMODE_WRITE)
816 pipe->writers--;
817
818 /* Was that the last reader or writer, but not the other side? */
819 if (!pipe->readers != !pipe->writers) {
820 wake_up_interruptible_all(&pipe->rd_wait);
821 wake_up_interruptible_all(&pipe->wr_wait);
822 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
823 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
824 }
825 mutex_unlock(&pipe->mutex);
826
827 put_pipe_info(inode, pipe);
828 return 0;
829}
830
831static int
832pipe_fasync(int fd, struct file *filp, int on)
833{
834 struct pipe_inode_info *pipe = filp->private_data;
835 int retval = 0;
836
837 mutex_lock(&pipe->mutex);
838 if (filp->f_mode & FMODE_READ)
839 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
840 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
841 int read_ret = retval;
842
843 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
844 /* Undo only a reader entry added by this call. */
845 if (retval < 0 && read_ret > 0)
846 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
847 }
848 mutex_unlock(&pipe->mutex);
849 return retval;
850}
851
852unsigned long account_pipe_buffers(struct user_struct *user,
853 unsigned long old, unsigned long new)
854{
855 return atomic_long_add_return(new - old, &user->pipe_bufs);
856}
857
858bool too_many_pipe_buffers_soft(unsigned long user_bufs)
859{
860 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
861
862 return soft_limit && user_bufs > soft_limit;
863}
864
865bool too_many_pipe_buffers_hard(unsigned long user_bufs)
866{
867 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
868
869 return hard_limit && user_bufs > hard_limit;
870}
871
872bool pipe_is_unprivileged_user(void)
873{
874 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
875}
876
877struct pipe_inode_info *alloc_pipe_info(void)
878{
879 struct pipe_inode_info *pipe;
880 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
881 struct user_struct *user = get_current_user();
882 unsigned long user_bufs;
883 unsigned int max_size = READ_ONCE(pipe_max_size);
884
885 pipe = kzalloc_obj(struct pipe_inode_info, GFP_KERNEL_ACCOUNT);
886 if (pipe == NULL)
887 goto out_free_uid;
888
889 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
890 pipe_bufs = max_size >> PAGE_SHIFT;
891
892 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
893
894 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
895 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
896 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
897 }
898
899 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
900 goto out_revert_acct;
901
902 pipe->bufs = kzalloc_objs(struct pipe_buffer, pipe_bufs,
903 GFP_KERNEL_ACCOUNT);
904
905 if (pipe->bufs) {
906 init_waitqueue_head(&pipe->rd_wait);
907 init_waitqueue_head(&pipe->wr_wait);
908 pipe->r_counter = pipe->w_counter = 1;
909 pipe->max_usage = pipe_bufs;
910 pipe->ring_size = pipe_bufs;
911 pipe->nr_accounted = pipe_bufs;
912 pipe->user = user;
913 mutex_init(&pipe->mutex);
914 lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);
915 return pipe;
916 }
917
918out_revert_acct:
919 (void) account_pipe_buffers(user, pipe_bufs, 0);
920 kfree(pipe);
921out_free_uid:
922 free_uid(user);
923 return NULL;
924}
925
926void free_pipe_info(struct pipe_inode_info *pipe)
927{
928 unsigned int i;
929
930#ifdef CONFIG_WATCH_QUEUE
931 if (pipe->watch_queue)
932 watch_queue_clear(pipe->watch_queue);
933#endif
934
935 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
936 free_uid(pipe->user);
937 for (i = 0; i < pipe->ring_size; i++) {
938 struct pipe_buffer *buf = pipe->bufs + i;
939 if (buf->ops)
940 pipe_buf_release(pipe, buf);
941 }
942#ifdef CONFIG_WATCH_QUEUE
943 if (pipe->watch_queue)
944 put_watch_queue(pipe->watch_queue);
945#endif
946 for (i = 0; i < pipe->prealloc.count; i++)
947 __free_page(pipe->prealloc.pages[i]);
948 kfree(pipe->bufs);
949 kfree(pipe);
950}
951
952static struct vfsmount *pipe_mnt __ro_after_init;
953
954/*
955 * pipefs_dname() is called from d_path().
956 */
957static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
958{
959 return dynamic_dname(buffer, buflen, "pipe:[%llu]",
960 d_inode(dentry)->i_ino);
961}
962
963static const struct dentry_operations pipefs_dentry_operations = {
964 .d_dname = pipefs_dname,
965};
966
967static const struct file_operations pipeanon_fops;
968
969static struct inode * get_pipe_inode(void)
970{
971 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
972 struct pipe_inode_info *pipe;
973
974 if (!inode)
975 goto fail_inode;
976
977 inode->i_ino = get_next_ino();
978
979 pipe = alloc_pipe_info();
980 if (!pipe)
981 goto fail_iput;
982
983 inode->i_pipe = pipe;
984 pipe->files = 2;
985 pipe->readers = pipe->writers = 1;
986 inode->i_fop = &pipeanon_fops;
987
988 /*
989 * Mark the inode dirty from the very beginning,
990 * that way it will never be moved to the dirty
991 * list because "mark_inode_dirty()" will think
992 * that it already _is_ on the dirty list.
993 */
994 inode_state_assign_raw(inode, I_DIRTY);
995 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
996 inode->i_uid = current_fsuid();
997 inode->i_gid = current_fsgid();
998 simple_inode_init_ts(inode);
999
1000 return inode;
1001
1002fail_iput:
1003 iput(inode);
1004
1005fail_inode:
1006 return NULL;
1007}
1008
1009int create_pipe_files(struct file **res, int flags)
1010{
1011 struct inode *inode = get_pipe_inode();
1012 struct file *f;
1013 int error;
1014
1015 if (!inode)
1016 return -ENFILE;
1017
1018 if (flags & O_NOTIFICATION_PIPE) {
1019 error = watch_queue_init(inode->i_pipe);
1020 if (error) {
1021 free_pipe_info(inode->i_pipe);
1022 iput(inode);
1023 return error;
1024 }
1025 }
1026
1027 f = alloc_file_pseudo(inode, pipe_mnt, "",
1028 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
1029 &pipeanon_fops);
1030 if (IS_ERR(f)) {
1031 free_pipe_info(inode->i_pipe);
1032 iput(inode);
1033 return PTR_ERR(f);
1034 }
1035
1036 f->private_data = inode->i_pipe;
1037 f->f_pipe = 0;
1038
1039 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
1040 &pipeanon_fops);
1041 if (IS_ERR(res[0])) {
1042 put_pipe_info(inode, inode->i_pipe);
1043 fput(f);
1044 return PTR_ERR(res[0]);
1045 }
1046 res[0]->private_data = inode->i_pipe;
1047 res[0]->f_pipe = 0;
1048 res[1] = f;
1049 stream_open(inode, res[0]);
1050 stream_open(inode, res[1]);
1051
1052 /* pipe groks IOCB_NOWAIT */
1053 res[0]->f_mode |= FMODE_NOWAIT;
1054 res[1]->f_mode |= FMODE_NOWAIT;
1055
1056 /*
1057 * Disable permission and pre-content events, but enable legacy
1058 * inotify events for legacy users.
1059 */
1060 file_set_fsnotify_mode(res[0], FMODE_NONOTIFY_PERM);
1061 file_set_fsnotify_mode(res[1], FMODE_NONOTIFY_PERM);
1062 return 0;
1063}
1064
1065static int __do_pipe_flags(int *fd, struct file **files, int flags)
1066{
1067 int error;
1068 int fdw, fdr;
1069
1070 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
1071 return -EINVAL;
1072
1073 error = create_pipe_files(files, flags);
1074 if (error)
1075 return error;
1076
1077 error = get_unused_fd_flags(flags);
1078 if (error < 0)
1079 goto err_read_pipe;
1080 fdr = error;
1081
1082 error = get_unused_fd_flags(flags);
1083 if (error < 0)
1084 goto err_fdr;
1085 fdw = error;
1086
1087 audit_fd_pair(fdr, fdw);
1088 fd[0] = fdr;
1089 fd[1] = fdw;
1090 return 0;
1091
1092 err_fdr:
1093 put_unused_fd(fdr);
1094 err_read_pipe:
1095 fput(files[0]);
1096 fput(files[1]);
1097 return error;
1098}
1099
1100int do_pipe_flags(int *fd, int flags)
1101{
1102 struct file *files[2];
1103 int error = __do_pipe_flags(fd, files, flags);
1104 if (!error) {
1105 fd_install(fd[0], files[0]);
1106 fd_install(fd[1], files[1]);
1107 }
1108 return error;
1109}
1110
1111/*
1112 * sys_pipe() is the normal C calling standard for creating
1113 * a pipe. It's not the way Unix traditionally does this, though.
1114 */
1115static int do_pipe2(int __user *fildes, int flags)
1116{
1117 struct file *files[2];
1118 int fd[2];
1119 int error;
1120
1121 error = __do_pipe_flags(fd, files, flags);
1122 if (!error) {
1123 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1124 fput(files[0]);
1125 fput(files[1]);
1126 put_unused_fd(fd[0]);
1127 put_unused_fd(fd[1]);
1128 error = -EFAULT;
1129 } else {
1130 fd_install(fd[0], files[0]);
1131 fd_install(fd[1], files[1]);
1132 }
1133 }
1134 return error;
1135}
1136
1137SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1138{
1139 return do_pipe2(fildes, flags);
1140}
1141
1142SYSCALL_DEFINE1(pipe, int __user *, fildes)
1143{
1144 return do_pipe2(fildes, 0);
1145}
1146
1147/*
1148 * This is the stupid "wait for pipe to be readable or writable"
1149 * model.
1150 *
1151 * See pipe_read/write() for the proper kind of exclusive wait,
1152 * but that requires that we wake up any other readers/writers
1153 * if we then do not end up reading everything (ie the whole
1154 * "wake_next_reader/writer" logic in pipe_read/write()).
1155 */
1156void pipe_wait_readable(struct pipe_inode_info *pipe)
1157{
1158 pipe_unlock(pipe);
1159 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1160 pipe_lock(pipe);
1161}
1162
1163void pipe_wait_writable(struct pipe_inode_info *pipe)
1164{
1165 pipe_unlock(pipe);
1166 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1167 pipe_lock(pipe);
1168}
1169
1170/*
1171 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1172 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1173 * race with the count check and waitqueue prep.
1174 *
1175 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1176 * then check the condition you're waiting for, and only then sleep. But
1177 * because of the pipe lock, we can check the condition before being on
1178 * the wait queue.
1179 *
1180 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1181 */
1182static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1183{
1184 DEFINE_WAIT(rdwait);
1185 int cur = *cnt;
1186
1187 while (cur == *cnt) {
1188 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1189 pipe_unlock(pipe);
1190 schedule();
1191 finish_wait(&pipe->rd_wait, &rdwait);
1192 pipe_lock(pipe);
1193 if (signal_pending(current))
1194 break;
1195 }
1196 return cur == *cnt ? -ERESTARTSYS : 0;
1197}
1198
1199static void wake_up_partner(struct pipe_inode_info *pipe)
1200{
1201 wake_up_interruptible_all(&pipe->rd_wait);
1202}
1203
1204static int fifo_open(struct inode *inode, struct file *filp)
1205{
1206 bool is_pipe = inode->i_fop == &pipeanon_fops;
1207 struct pipe_inode_info *pipe;
1208 int ret;
1209
1210 filp->f_pipe = 0;
1211
1212 spin_lock(&inode->i_lock);
1213 if (inode->i_pipe) {
1214 pipe = inode->i_pipe;
1215 pipe->files++;
1216 spin_unlock(&inode->i_lock);
1217 } else {
1218 spin_unlock(&inode->i_lock);
1219 pipe = alloc_pipe_info();
1220 if (!pipe)
1221 return -ENOMEM;
1222 pipe->files = 1;
1223 spin_lock(&inode->i_lock);
1224 if (unlikely(inode->i_pipe)) {
1225 inode->i_pipe->files++;
1226 spin_unlock(&inode->i_lock);
1227 free_pipe_info(pipe);
1228 pipe = inode->i_pipe;
1229 } else {
1230 inode->i_pipe = pipe;
1231 spin_unlock(&inode->i_lock);
1232 }
1233 }
1234 filp->private_data = pipe;
1235 /* OK, we have a pipe and it's pinned down */
1236
1237 mutex_lock(&pipe->mutex);
1238
1239 /* We can only do regular read/write on fifos */
1240 stream_open(inode, filp);
1241
1242 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1243 case FMODE_READ:
1244 /*
1245 * O_RDONLY
1246 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1247 * opened, even when there is no process writing the FIFO.
1248 */
1249 pipe->r_counter++;
1250 if (pipe->readers++ == 0)
1251 wake_up_partner(pipe);
1252
1253 if (!is_pipe && !pipe->writers) {
1254 if ((filp->f_flags & O_NONBLOCK)) {
1255 /* suppress EPOLLHUP until we have
1256 * seen a writer */
1257 filp->f_pipe = pipe->w_counter;
1258 } else {
1259 if (wait_for_partner(pipe, &pipe->w_counter))
1260 goto err_rd;
1261 }
1262 }
1263 break;
1264
1265 case FMODE_WRITE:
1266 /*
1267 * O_WRONLY
1268 * POSIX.1 says that O_NONBLOCK means return -1 with
1269 * errno=ENXIO when there is no process reading the FIFO.
1270 */
1271 ret = -ENXIO;
1272 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1273 goto err;
1274
1275 pipe->w_counter++;
1276 if (!pipe->writers++)
1277 wake_up_partner(pipe);
1278
1279 if (!is_pipe && !pipe->readers) {
1280 if (wait_for_partner(pipe, &pipe->r_counter))
1281 goto err_wr;
1282 }
1283 break;
1284
1285 case FMODE_READ | FMODE_WRITE:
1286 /*
1287 * O_RDWR
1288 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1289 * This implementation will NEVER block on a O_RDWR open, since
1290 * the process can at least talk to itself.
1291 */
1292
1293 pipe->readers++;
1294 pipe->writers++;
1295 pipe->r_counter++;
1296 pipe->w_counter++;
1297 if (pipe->readers == 1 || pipe->writers == 1)
1298 wake_up_partner(pipe);
1299 break;
1300
1301 default:
1302 ret = -EINVAL;
1303 goto err;
1304 }
1305
1306 /* Ok! */
1307 mutex_unlock(&pipe->mutex);
1308 return 0;
1309
1310err_rd:
1311 if (!--pipe->readers)
1312 wake_up_interruptible(&pipe->wr_wait);
1313 ret = -ERESTARTSYS;
1314 goto err;
1315
1316err_wr:
1317 if (!--pipe->writers)
1318 wake_up_interruptible_all(&pipe->rd_wait);
1319 ret = -ERESTARTSYS;
1320 goto err;
1321
1322err:
1323 mutex_unlock(&pipe->mutex);
1324
1325 put_pipe_info(inode, pipe);
1326 return ret;
1327}
1328
1329const struct file_operations pipefifo_fops = {
1330 .open = fifo_open,
1331 .read_iter = fifo_pipe_read,
1332 .write_iter = fifo_pipe_write,
1333 .poll = pipe_poll,
1334 .unlocked_ioctl = pipe_ioctl,
1335 .release = pipe_release,
1336 .fasync = pipe_fasync,
1337 .splice_write = iter_file_splice_write,
1338};
1339
1340static const struct file_operations pipeanon_fops = {
1341 .open = fifo_open,
1342 .read_iter = anon_pipe_read,
1343 .write_iter = anon_pipe_write,
1344 .poll = pipe_poll,
1345 .unlocked_ioctl = pipe_ioctl,
1346 .release = pipe_release,
1347 .fasync = pipe_fasync,
1348 .splice_write = iter_file_splice_write,
1349};
1350
1351/*
1352 * Currently we rely on the pipe array holding a power-of-2 number
1353 * of pages. Returns 0 on error.
1354 */
1355unsigned int round_pipe_size(unsigned int size)
1356{
1357 if (size > (1U << 31))
1358 return 0;
1359
1360 /* Minimum pipe size, as required by POSIX */
1361 if (size < PAGE_SIZE)
1362 return PAGE_SIZE;
1363
1364 return roundup_pow_of_two(size);
1365}
1366
1367/*
1368 * Resize the pipe ring to a number of slots.
1369 *
1370 * Note the pipe can be reduced in capacity, but only if the current
1371 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1372 * returned instead.
1373 */
1374int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1375{
1376 struct pipe_buffer *bufs;
1377 unsigned int head, tail, mask, n;
1378
1379 /* nr_slots larger than limits of pipe->{head,tail} */
1380 if (unlikely(nr_slots > (pipe_index_t)-1u))
1381 return -EINVAL;
1382
1383 bufs = kzalloc_objs(*bufs, nr_slots, GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1384 if (unlikely(!bufs))
1385 return -ENOMEM;
1386
1387 spin_lock_irq(&pipe->rd_wait.lock);
1388 mask = pipe->ring_size - 1;
1389 head = pipe->head;
1390 tail = pipe->tail;
1391
1392 n = pipe_occupancy(head, tail);
1393 if (nr_slots < n) {
1394 spin_unlock_irq(&pipe->rd_wait.lock);
1395 kfree(bufs);
1396 return -EBUSY;
1397 }
1398
1399 /*
1400 * The pipe array wraps around, so just start the new one at zero
1401 * and adjust the indices.
1402 */
1403 if (n > 0) {
1404 unsigned int h = head & mask;
1405 unsigned int t = tail & mask;
1406 if (h > t) {
1407 memcpy(bufs, pipe->bufs + t,
1408 n * sizeof(struct pipe_buffer));
1409 } else {
1410 unsigned int tsize = pipe->ring_size - t;
1411 if (h > 0)
1412 memcpy(bufs + tsize, pipe->bufs,
1413 h * sizeof(struct pipe_buffer));
1414 memcpy(bufs, pipe->bufs + t,
1415 tsize * sizeof(struct pipe_buffer));
1416 }
1417 }
1418
1419 head = n;
1420 tail = 0;
1421
1422 kfree(pipe->bufs);
1423 pipe->bufs = bufs;
1424 pipe->ring_size = nr_slots;
1425 if (pipe->max_usage > nr_slots)
1426 pipe->max_usage = nr_slots;
1427 pipe->tail = tail;
1428 pipe->head = head;
1429
1430 if (!pipe_has_watch_queue(pipe)) {
1431 pipe->max_usage = nr_slots;
1432 pipe->nr_accounted = nr_slots;
1433 }
1434
1435 spin_unlock_irq(&pipe->rd_wait.lock);
1436
1437 /* This might have made more room for writers */
1438 wake_up_interruptible_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
1439 return 0;
1440}
1441
1442/*
1443 * Allocate a new array of pipe buffers and copy the info over. Returns the
1444 * pipe size if successful, or return -ERROR on error.
1445 */
1446static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1447{
1448 unsigned long user_bufs;
1449 unsigned int nr_slots, size;
1450 long ret = 0;
1451
1452 if (pipe_has_watch_queue(pipe))
1453 return -EBUSY;
1454
1455 size = round_pipe_size(arg);
1456 nr_slots = size >> PAGE_SHIFT;
1457
1458 if (!nr_slots)
1459 return -EINVAL;
1460
1461 /*
1462 * If trying to increase the pipe capacity, check that an
1463 * unprivileged user is not trying to exceed various limits
1464 * (soft limit check here, hard limit check just below).
1465 * Decreasing the pipe capacity is always permitted, even
1466 * if the user is currently over a limit.
1467 */
1468 if (nr_slots > pipe->max_usage &&
1469 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1470 return -EPERM;
1471
1472 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1473
1474 if (nr_slots > pipe->max_usage &&
1475 (too_many_pipe_buffers_hard(user_bufs) ||
1476 too_many_pipe_buffers_soft(user_bufs)) &&
1477 pipe_is_unprivileged_user()) {
1478 ret = -EPERM;
1479 goto out_revert_acct;
1480 }
1481
1482 ret = pipe_resize_ring(pipe, nr_slots);
1483 if (ret < 0)
1484 goto out_revert_acct;
1485
1486 return pipe->max_usage * PAGE_SIZE;
1487
1488out_revert_acct:
1489 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1490 return ret;
1491}
1492
1493/*
1494 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1495 * not enough to verify that this is a pipe.
1496 */
1497struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1498{
1499 struct pipe_inode_info *pipe = file->private_data;
1500
1501 if (!pipe)
1502 return NULL;
1503 if (file->f_op != &pipefifo_fops && file->f_op != &pipeanon_fops)
1504 return NULL;
1505 if (for_splice && pipe_has_watch_queue(pipe))
1506 return NULL;
1507 return pipe;
1508}
1509
1510long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1511{
1512 struct pipe_inode_info *pipe;
1513 long ret;
1514
1515 pipe = get_pipe_info(file, false);
1516 if (!pipe)
1517 return -EBADF;
1518
1519 mutex_lock(&pipe->mutex);
1520
1521 switch (cmd) {
1522 case F_SETPIPE_SZ:
1523 ret = pipe_set_size(pipe, arg);
1524 break;
1525 case F_GETPIPE_SZ:
1526 ret = pipe->max_usage * PAGE_SIZE;
1527 break;
1528 default:
1529 ret = -EINVAL;
1530 break;
1531 }
1532
1533 mutex_unlock(&pipe->mutex);
1534 return ret;
1535}
1536
1537static const struct super_operations pipefs_ops = {
1538 .destroy_inode = free_inode_nonrcu,
1539 .statfs = simple_statfs,
1540};
1541
1542/*
1543 * pipefs should _never_ be mounted by userland - too much of security hassle,
1544 * no real gain from having the whole file system mounted. So we don't need
1545 * any operations on the root directory. However, we need a non-trivial
1546 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1547 */
1548
1549static int pipefs_init_fs_context(struct fs_context *fc)
1550{
1551 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1552 if (!ctx)
1553 return -ENOMEM;
1554 ctx->ops = &pipefs_ops;
1555 ctx->dops = &pipefs_dentry_operations;
1556 return 0;
1557}
1558
1559static struct file_system_type pipe_fs_type = {
1560 .name = "pipefs",
1561 .init_fs_context = pipefs_init_fs_context,
1562 .kill_sb = kill_anon_super,
1563};
1564
1565#ifdef CONFIG_SYSCTL
1566
1567static ulong round_pipe_size_ul(ulong size)
1568{
1569 return round_pipe_size(size);
1570}
1571
1572static int u2k_pipe_maxsz(const ulong *u_ptr, uint *k_ptr)
1573{
1574 return proc_uint_u2k_conv_uop(u_ptr, k_ptr, round_pipe_size_ul);
1575}
1576
1577static int do_proc_uint_conv_pipe_maxsz(bool *negp, ulong *u_ptr, uint *k_ptr,
1578 int dir, const struct ctl_table *table)
1579{
1580 return proc_uint_conv(u_ptr, k_ptr, dir, table, true,
1581 u2k_pipe_maxsz,
1582 proc_uint_k2u_conv);
1583}
1584
1585static int proc_dopipe_max_size(const struct ctl_table *table, int write,
1586 void *buffer, size_t *lenp, loff_t *ppos)
1587{
1588 return proc_douintvec_conv(table, write, buffer, lenp, ppos,
1589 do_proc_uint_conv_pipe_maxsz);
1590}
1591
1592static const struct ctl_table fs_pipe_sysctls[] = {
1593 {
1594 .procname = "pipe-max-size",
1595 .data = &pipe_max_size,
1596 .maxlen = sizeof(pipe_max_size),
1597 .mode = 0644,
1598 .proc_handler = proc_dopipe_max_size,
1599 .extra1 = SYSCTL_ONE,
1600 },
1601 {
1602 .procname = "pipe-user-pages-hard",
1603 .data = &pipe_user_pages_hard,
1604 .maxlen = sizeof(pipe_user_pages_hard),
1605 .mode = 0644,
1606 .proc_handler = proc_doulongvec_minmax,
1607 },
1608 {
1609 .procname = "pipe-user-pages-soft",
1610 .data = &pipe_user_pages_soft,
1611 .maxlen = sizeof(pipe_user_pages_soft),
1612 .mode = 0644,
1613 .proc_handler = proc_doulongvec_minmax,
1614 },
1615};
1616#endif
1617
1618static int __init init_pipe_fs(void)
1619{
1620 int err = register_filesystem(&pipe_fs_type);
1621
1622 if (!err) {
1623 pipe_mnt = kern_mount(&pipe_fs_type);
1624 if (IS_ERR(pipe_mnt)) {
1625 err = PTR_ERR(pipe_mnt);
1626 unregister_filesystem(&pipe_fs_type);
1627 }
1628 }
1629#ifdef CONFIG_SYSCTL
1630 register_sysctl_init("fs", fs_pipe_sysctls);
1631#endif
1632 return err;
1633}
1634
1635fs_initcall(init_pipe_fs);