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
29#include <linux/uaccess.h>
30#include <asm/ioctls.h>
31
32#include "internal.h"
33
34/*
35 * The max size that a non-root user is allowed to grow the pipe. Can
36 * be set by root in /proc/sys/fs/pipe-max-size
37 */
38unsigned int pipe_max_size = 1048576;
39
40/* Maximum allocatable pages per user. Hard limit is unset by default, soft
41 * matches default values.
42 */
43unsigned long pipe_user_pages_hard;
44unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
45
46/*
47 * We use head and tail indices that aren't masked off, except at the point of
48 * dereference, but rather they're allowed to wrap naturally. This means there
49 * isn't a dead spot in the buffer, but the ring has to be a power of two and
50 * <= 2^31.
51 * -- David Howells 2019-09-23.
52 *
53 * Reads with count = 0 should always return 0.
54 * -- Julian Bradfield 1999-06-07.
55 *
56 * FIFOs and Pipes now generate SIGIO for both readers and writers.
57 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
58 *
59 * pipe_read & write cleanup
60 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
61 */
62
63static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
64{
65 if (pipe->files)
66 mutex_lock_nested(&pipe->mutex, subclass);
67}
68
69void pipe_lock(struct pipe_inode_info *pipe)
70{
71 /*
72 * pipe_lock() nests non-pipe inode locks (for writing to a file)
73 */
74 pipe_lock_nested(pipe, I_MUTEX_PARENT);
75}
76EXPORT_SYMBOL(pipe_lock);
77
78void pipe_unlock(struct pipe_inode_info *pipe)
79{
80 if (pipe->files)
81 mutex_unlock(&pipe->mutex);
82}
83EXPORT_SYMBOL(pipe_unlock);
84
85static inline void __pipe_lock(struct pipe_inode_info *pipe)
86{
87 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
88}
89
90static inline void __pipe_unlock(struct pipe_inode_info *pipe)
91{
92 mutex_unlock(&pipe->mutex);
93}
94
95void pipe_double_lock(struct pipe_inode_info *pipe1,
96 struct pipe_inode_info *pipe2)
97{
98 BUG_ON(pipe1 == pipe2);
99
100 if (pipe1 < pipe2) {
101 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
102 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
103 } else {
104 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
105 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
106 }
107}
108
109static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
110 struct pipe_buffer *buf)
111{
112 struct page *page = buf->page;
113
114 /*
115 * If nobody else uses this page, and we don't already have a
116 * temporary page, let's keep track of it as a one-deep
117 * allocation cache. (Otherwise just release our reference to it)
118 */
119 if (page_count(page) == 1 && !pipe->tmp_page)
120 pipe->tmp_page = page;
121 else
122 put_page(page);
123}
124
125static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
126 struct pipe_buffer *buf)
127{
128 struct page *page = buf->page;
129
130 if (page_count(page) != 1)
131 return false;
132 memcg_kmem_uncharge_page(page, 0);
133 __SetPageLocked(page);
134 return true;
135}
136
137/**
138 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
139 * @pipe: the pipe that the buffer belongs to
140 * @buf: the buffer to attempt to steal
141 *
142 * Description:
143 * This function attempts to steal the &struct page attached to
144 * @buf. If successful, this function returns 0 and returns with
145 * the page locked. The caller may then reuse the page for whatever
146 * he wishes; the typical use is insertion into a different file
147 * page cache.
148 */
149bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
150 struct pipe_buffer *buf)
151{
152 struct page *page = buf->page;
153
154 /*
155 * A reference of one is golden, that means that the owner of this
156 * page is the only one holding a reference to it. lock the page
157 * and return OK.
158 */
159 if (page_count(page) == 1) {
160 lock_page(page);
161 return true;
162 }
163 return false;
164}
165EXPORT_SYMBOL(generic_pipe_buf_try_steal);
166
167/**
168 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
169 * @pipe: the pipe that the buffer belongs to
170 * @buf: the buffer to get a reference to
171 *
172 * Description:
173 * This function grabs an extra reference to @buf. It's used in
174 * in the tee() system call, when we duplicate the buffers in one
175 * pipe into another.
176 */
177bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
178{
179 return try_get_page(buf->page);
180}
181EXPORT_SYMBOL(generic_pipe_buf_get);
182
183/**
184 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
185 * @pipe: the pipe that the buffer belongs to
186 * @buf: the buffer to put a reference to
187 *
188 * Description:
189 * This function releases a reference to @buf.
190 */
191void generic_pipe_buf_release(struct pipe_inode_info *pipe,
192 struct pipe_buffer *buf)
193{
194 put_page(buf->page);
195}
196EXPORT_SYMBOL(generic_pipe_buf_release);
197
198static const struct pipe_buf_operations anon_pipe_buf_ops = {
199 .release = anon_pipe_buf_release,
200 .try_steal = anon_pipe_buf_try_steal,
201 .get = generic_pipe_buf_get,
202};
203
204/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
205static inline bool pipe_readable(const struct pipe_inode_info *pipe)
206{
207 unsigned int head = READ_ONCE(pipe->head);
208 unsigned int tail = READ_ONCE(pipe->tail);
209 unsigned int writers = READ_ONCE(pipe->writers);
210
211 return !pipe_empty(head, tail) || !writers;
212}
213
214static ssize_t
215pipe_read(struct kiocb *iocb, struct iov_iter *to)
216{
217 size_t total_len = iov_iter_count(to);
218 struct file *filp = iocb->ki_filp;
219 struct pipe_inode_info *pipe = filp->private_data;
220 bool was_full, wake_next_reader = false;
221 ssize_t ret;
222
223 /* Null read succeeds. */
224 if (unlikely(total_len == 0))
225 return 0;
226
227 ret = 0;
228 __pipe_lock(pipe);
229
230 /*
231 * We only wake up writers if the pipe was full when we started
232 * reading in order to avoid unnecessary wakeups.
233 *
234 * But when we do wake up writers, we do so using a sync wakeup
235 * (WF_SYNC), because we want them to get going and generate more
236 * data for us.
237 */
238 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
239 for (;;) {
240 unsigned int head = pipe->head;
241 unsigned int tail = pipe->tail;
242 unsigned int mask = pipe->ring_size - 1;
243
244#ifdef CONFIG_WATCH_QUEUE
245 if (pipe->note_loss) {
246 struct watch_notification n;
247
248 if (total_len < 8) {
249 if (ret == 0)
250 ret = -ENOBUFS;
251 break;
252 }
253
254 n.type = WATCH_TYPE_META;
255 n.subtype = WATCH_META_LOSS_NOTIFICATION;
256 n.info = watch_sizeof(n);
257 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
258 if (ret == 0)
259 ret = -EFAULT;
260 break;
261 }
262 ret += sizeof(n);
263 total_len -= sizeof(n);
264 pipe->note_loss = false;
265 }
266#endif
267
268 if (!pipe_empty(head, tail)) {
269 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
270 size_t chars = buf->len;
271 size_t written;
272 int error;
273
274 if (chars > total_len) {
275 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
276 if (ret == 0)
277 ret = -ENOBUFS;
278 break;
279 }
280 chars = total_len;
281 }
282
283 error = pipe_buf_confirm(pipe, buf);
284 if (error) {
285 if (!ret)
286 ret = error;
287 break;
288 }
289
290 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
291 if (unlikely(written < chars)) {
292 if (!ret)
293 ret = -EFAULT;
294 break;
295 }
296 ret += chars;
297 buf->offset += chars;
298 buf->len -= chars;
299
300 /* Was it a packet buffer? Clean up and exit */
301 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
302 total_len = chars;
303 buf->len = 0;
304 }
305
306 if (!buf->len) {
307 pipe_buf_release(pipe, buf);
308 spin_lock_irq(&pipe->rd_wait.lock);
309#ifdef CONFIG_WATCH_QUEUE
310 if (buf->flags & PIPE_BUF_FLAG_LOSS)
311 pipe->note_loss = true;
312#endif
313 tail++;
314 pipe->tail = tail;
315 spin_unlock_irq(&pipe->rd_wait.lock);
316 }
317 total_len -= chars;
318 if (!total_len)
319 break; /* common path: read succeeded */
320 if (!pipe_empty(head, tail)) /* More to do? */
321 continue;
322 }
323
324 if (!pipe->writers)
325 break;
326 if (ret)
327 break;
328 if (filp->f_flags & O_NONBLOCK) {
329 ret = -EAGAIN;
330 break;
331 }
332 __pipe_unlock(pipe);
333
334 /*
335 * We only get here if we didn't actually read anything.
336 *
337 * However, we could have seen (and removed) a zero-sized
338 * pipe buffer, and might have made space in the buffers
339 * that way.
340 *
341 * You can't make zero-sized pipe buffers by doing an empty
342 * write (not even in packet mode), but they can happen if
343 * the writer gets an EFAULT when trying to fill a buffer
344 * that already got allocated and inserted in the buffer
345 * array.
346 *
347 * So we still need to wake up any pending writers in the
348 * _very_ unlikely case that the pipe was full, but we got
349 * no data.
350 */
351 if (unlikely(was_full)) {
352 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
353 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
354 }
355
356 /*
357 * But because we didn't read anything, at this point we can
358 * just return directly with -ERESTARTSYS if we're interrupted,
359 * since we've done any required wakeups and there's no need
360 * to mark anything accessed. And we've dropped the lock.
361 */
362 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
363 return -ERESTARTSYS;
364
365 __pipe_lock(pipe);
366 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
367 wake_next_reader = true;
368 }
369 if (pipe_empty(pipe->head, pipe->tail))
370 wake_next_reader = false;
371 __pipe_unlock(pipe);
372
373 if (was_full) {
374 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
375 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
376 }
377 if (wake_next_reader)
378 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
379 if (ret > 0)
380 file_accessed(filp);
381 return ret;
382}
383
384static inline int is_packetized(struct file *file)
385{
386 return (file->f_flags & O_DIRECT) != 0;
387}
388
389/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
390static inline bool pipe_writable(const struct pipe_inode_info *pipe)
391{
392 unsigned int head = READ_ONCE(pipe->head);
393 unsigned int tail = READ_ONCE(pipe->tail);
394 unsigned int max_usage = READ_ONCE(pipe->max_usage);
395
396 return !pipe_full(head, tail, max_usage) ||
397 !READ_ONCE(pipe->readers);
398}
399
400static ssize_t
401pipe_write(struct kiocb *iocb, struct iov_iter *from)
402{
403 struct file *filp = iocb->ki_filp;
404 struct pipe_inode_info *pipe = filp->private_data;
405 unsigned int head;
406 ssize_t ret = 0;
407 size_t total_len = iov_iter_count(from);
408 ssize_t chars;
409 bool was_empty = false;
410 bool wake_next_writer = false;
411
412 /* Null write succeeds. */
413 if (unlikely(total_len == 0))
414 return 0;
415
416 __pipe_lock(pipe);
417
418 if (!pipe->readers) {
419 send_sig(SIGPIPE, current, 0);
420 ret = -EPIPE;
421 goto out;
422 }
423
424#ifdef CONFIG_WATCH_QUEUE
425 if (pipe->watch_queue) {
426 ret = -EXDEV;
427 goto out;
428 }
429#endif
430
431 /*
432 * Only wake up if the pipe started out empty, since
433 * otherwise there should be no readers waiting.
434 *
435 * If it wasn't empty we try to merge new data into
436 * the last buffer.
437 *
438 * That naturally merges small writes, but it also
439 * page-aligs the rest of the writes for large writes
440 * spanning multiple pages.
441 */
442 head = pipe->head;
443 was_empty = pipe_empty(head, pipe->tail);
444 chars = total_len & (PAGE_SIZE-1);
445 if (chars && !was_empty) {
446 unsigned int mask = pipe->ring_size - 1;
447 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
448 int offset = buf->offset + buf->len;
449
450 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
451 offset + chars <= PAGE_SIZE) {
452 ret = pipe_buf_confirm(pipe, buf);
453 if (ret)
454 goto out;
455
456 ret = copy_page_from_iter(buf->page, offset, chars, from);
457 if (unlikely(ret < chars)) {
458 ret = -EFAULT;
459 goto out;
460 }
461
462 buf->len += ret;
463 if (!iov_iter_count(from))
464 goto out;
465 }
466 }
467
468 for (;;) {
469 if (!pipe->readers) {
470 send_sig(SIGPIPE, current, 0);
471 if (!ret)
472 ret = -EPIPE;
473 break;
474 }
475
476 head = pipe->head;
477 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
478 unsigned int mask = pipe->ring_size - 1;
479 struct pipe_buffer *buf = &pipe->bufs[head & mask];
480 struct page *page = pipe->tmp_page;
481 int copied;
482
483 if (!page) {
484 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
485 if (unlikely(!page)) {
486 ret = ret ? : -ENOMEM;
487 break;
488 }
489 pipe->tmp_page = page;
490 }
491
492 /* Allocate a slot in the ring in advance and attach an
493 * empty buffer. If we fault or otherwise fail to use
494 * it, either the reader will consume it or it'll still
495 * be there for the next write.
496 */
497 spin_lock_irq(&pipe->rd_wait.lock);
498
499 head = pipe->head;
500 if (pipe_full(head, pipe->tail, pipe->max_usage)) {
501 spin_unlock_irq(&pipe->rd_wait.lock);
502 continue;
503 }
504
505 pipe->head = head + 1;
506 spin_unlock_irq(&pipe->rd_wait.lock);
507
508 /* Insert it into the buffer array */
509 buf = &pipe->bufs[head & mask];
510 buf->page = page;
511 buf->ops = &anon_pipe_buf_ops;
512 buf->offset = 0;
513 buf->len = 0;
514 if (is_packetized(filp))
515 buf->flags = PIPE_BUF_FLAG_PACKET;
516 else
517 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
518 pipe->tmp_page = NULL;
519
520 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
521 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
522 if (!ret)
523 ret = -EFAULT;
524 break;
525 }
526 ret += copied;
527 buf->offset = 0;
528 buf->len = copied;
529
530 if (!iov_iter_count(from))
531 break;
532 }
533
534 if (!pipe_full(head, pipe->tail, pipe->max_usage))
535 continue;
536
537 /* Wait for buffer space to become available. */
538 if (filp->f_flags & O_NONBLOCK) {
539 if (!ret)
540 ret = -EAGAIN;
541 break;
542 }
543 if (signal_pending(current)) {
544 if (!ret)
545 ret = -ERESTARTSYS;
546 break;
547 }
548
549 /*
550 * We're going to release the pipe lock and wait for more
551 * space. We wake up any readers if necessary, and then
552 * after waiting we need to re-check whether the pipe
553 * become empty while we dropped the lock.
554 */
555 __pipe_unlock(pipe);
556 if (was_empty) {
557 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
558 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
559 }
560 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
561 __pipe_lock(pipe);
562 was_empty = pipe_empty(pipe->head, pipe->tail);
563 wake_next_writer = true;
564 }
565out:
566 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
567 wake_next_writer = false;
568 __pipe_unlock(pipe);
569
570 /*
571 * If we do do a wakeup event, we do a 'sync' wakeup, because we
572 * want the reader to start processing things asap, rather than
573 * leave the data pending.
574 *
575 * This is particularly important for small writes, because of
576 * how (for example) the GNU make jobserver uses small writes to
577 * wake up pending jobs
578 */
579 if (was_empty) {
580 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
581 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
582 }
583 if (wake_next_writer)
584 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
585 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
586 int err = file_update_time(filp);
587 if (err)
588 ret = err;
589 sb_end_write(file_inode(filp)->i_sb);
590 }
591 return ret;
592}
593
594static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
595{
596 struct pipe_inode_info *pipe = filp->private_data;
597 int count, head, tail, mask;
598
599 switch (cmd) {
600 case FIONREAD:
601 __pipe_lock(pipe);
602 count = 0;
603 head = pipe->head;
604 tail = pipe->tail;
605 mask = pipe->ring_size - 1;
606
607 while (tail != head) {
608 count += pipe->bufs[tail & mask].len;
609 tail++;
610 }
611 __pipe_unlock(pipe);
612
613 return put_user(count, (int __user *)arg);
614
615#ifdef CONFIG_WATCH_QUEUE
616 case IOC_WATCH_QUEUE_SET_SIZE: {
617 int ret;
618 __pipe_lock(pipe);
619 ret = watch_queue_set_size(pipe, arg);
620 __pipe_unlock(pipe);
621 return ret;
622 }
623
624 case IOC_WATCH_QUEUE_SET_FILTER:
625 return watch_queue_set_filter(
626 pipe, (struct watch_notification_filter __user *)arg);
627#endif
628
629 default:
630 return -ENOIOCTLCMD;
631 }
632}
633
634/* No kernel lock held - fine */
635static __poll_t
636pipe_poll(struct file *filp, poll_table *wait)
637{
638 __poll_t mask;
639 struct pipe_inode_info *pipe = filp->private_data;
640 unsigned int head, tail;
641
642 /*
643 * Reading pipe state only -- no need for acquiring the semaphore.
644 *
645 * But because this is racy, the code has to add the
646 * entry to the poll table _first_ ..
647 */
648 if (filp->f_mode & FMODE_READ)
649 poll_wait(filp, &pipe->rd_wait, wait);
650 if (filp->f_mode & FMODE_WRITE)
651 poll_wait(filp, &pipe->wr_wait, wait);
652
653 /*
654 * .. and only then can you do the racy tests. That way,
655 * if something changes and you got it wrong, the poll
656 * table entry will wake you up and fix it.
657 */
658 head = READ_ONCE(pipe->head);
659 tail = READ_ONCE(pipe->tail);
660
661 mask = 0;
662 if (filp->f_mode & FMODE_READ) {
663 if (!pipe_empty(head, tail))
664 mask |= EPOLLIN | EPOLLRDNORM;
665 if (!pipe->writers && filp->f_version != pipe->w_counter)
666 mask |= EPOLLHUP;
667 }
668
669 if (filp->f_mode & FMODE_WRITE) {
670 if (!pipe_full(head, tail, pipe->max_usage))
671 mask |= EPOLLOUT | EPOLLWRNORM;
672 /*
673 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
674 * behave exactly like pipes for poll().
675 */
676 if (!pipe->readers)
677 mask |= EPOLLERR;
678 }
679
680 return mask;
681}
682
683static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
684{
685 int kill = 0;
686
687 spin_lock(&inode->i_lock);
688 if (!--pipe->files) {
689 inode->i_pipe = NULL;
690 kill = 1;
691 }
692 spin_unlock(&inode->i_lock);
693
694 if (kill)
695 free_pipe_info(pipe);
696}
697
698static int
699pipe_release(struct inode *inode, struct file *file)
700{
701 struct pipe_inode_info *pipe = file->private_data;
702
703 __pipe_lock(pipe);
704 if (file->f_mode & FMODE_READ)
705 pipe->readers--;
706 if (file->f_mode & FMODE_WRITE)
707 pipe->writers--;
708
709 /* Was that the last reader or writer, but not the other side? */
710 if (!pipe->readers != !pipe->writers) {
711 wake_up_interruptible_all(&pipe->rd_wait);
712 wake_up_interruptible_all(&pipe->wr_wait);
713 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
714 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
715 }
716 __pipe_unlock(pipe);
717
718 put_pipe_info(inode, pipe);
719 return 0;
720}
721
722static int
723pipe_fasync(int fd, struct file *filp, int on)
724{
725 struct pipe_inode_info *pipe = filp->private_data;
726 int retval = 0;
727
728 __pipe_lock(pipe);
729 if (filp->f_mode & FMODE_READ)
730 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
731 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
732 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
733 if (retval < 0 && (filp->f_mode & FMODE_READ))
734 /* this can happen only if on == T */
735 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
736 }
737 __pipe_unlock(pipe);
738 return retval;
739}
740
741unsigned long account_pipe_buffers(struct user_struct *user,
742 unsigned long old, unsigned long new)
743{
744 return atomic_long_add_return(new - old, &user->pipe_bufs);
745}
746
747bool too_many_pipe_buffers_soft(unsigned long user_bufs)
748{
749 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
750
751 return soft_limit && user_bufs > soft_limit;
752}
753
754bool too_many_pipe_buffers_hard(unsigned long user_bufs)
755{
756 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
757
758 return hard_limit && user_bufs > hard_limit;
759}
760
761bool pipe_is_unprivileged_user(void)
762{
763 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
764}
765
766struct pipe_inode_info *alloc_pipe_info(void)
767{
768 struct pipe_inode_info *pipe;
769 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
770 struct user_struct *user = get_current_user();
771 unsigned long user_bufs;
772 unsigned int max_size = READ_ONCE(pipe_max_size);
773
774 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
775 if (pipe == NULL)
776 goto out_free_uid;
777
778 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
779 pipe_bufs = max_size >> PAGE_SHIFT;
780
781 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
782
783 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
784 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
785 pipe_bufs = 1;
786 }
787
788 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
789 goto out_revert_acct;
790
791 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
792 GFP_KERNEL_ACCOUNT);
793
794 if (pipe->bufs) {
795 init_waitqueue_head(&pipe->rd_wait);
796 init_waitqueue_head(&pipe->wr_wait);
797 pipe->r_counter = pipe->w_counter = 1;
798 pipe->max_usage = pipe_bufs;
799 pipe->ring_size = pipe_bufs;
800 pipe->nr_accounted = pipe_bufs;
801 pipe->user = user;
802 mutex_init(&pipe->mutex);
803 return pipe;
804 }
805
806out_revert_acct:
807 (void) account_pipe_buffers(user, pipe_bufs, 0);
808 kfree(pipe);
809out_free_uid:
810 free_uid(user);
811 return NULL;
812}
813
814void free_pipe_info(struct pipe_inode_info *pipe)
815{
816 int i;
817
818#ifdef CONFIG_WATCH_QUEUE
819 if (pipe->watch_queue) {
820 watch_queue_clear(pipe->watch_queue);
821 put_watch_queue(pipe->watch_queue);
822 }
823#endif
824
825 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
826 free_uid(pipe->user);
827 for (i = 0; i < pipe->ring_size; i++) {
828 struct pipe_buffer *buf = pipe->bufs + i;
829 if (buf->ops)
830 pipe_buf_release(pipe, buf);
831 }
832 if (pipe->tmp_page)
833 __free_page(pipe->tmp_page);
834 kfree(pipe->bufs);
835 kfree(pipe);
836}
837
838static struct vfsmount *pipe_mnt __read_mostly;
839
840/*
841 * pipefs_dname() is called from d_path().
842 */
843static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
844{
845 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
846 d_inode(dentry)->i_ino);
847}
848
849static const struct dentry_operations pipefs_dentry_operations = {
850 .d_dname = pipefs_dname,
851};
852
853static struct inode * get_pipe_inode(void)
854{
855 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
856 struct pipe_inode_info *pipe;
857
858 if (!inode)
859 goto fail_inode;
860
861 inode->i_ino = get_next_ino();
862
863 pipe = alloc_pipe_info();
864 if (!pipe)
865 goto fail_iput;
866
867 inode->i_pipe = pipe;
868 pipe->files = 2;
869 pipe->readers = pipe->writers = 1;
870 inode->i_fop = &pipefifo_fops;
871
872 /*
873 * Mark the inode dirty from the very beginning,
874 * that way it will never be moved to the dirty
875 * list because "mark_inode_dirty()" will think
876 * that it already _is_ on the dirty list.
877 */
878 inode->i_state = I_DIRTY;
879 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
880 inode->i_uid = current_fsuid();
881 inode->i_gid = current_fsgid();
882 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
883
884 return inode;
885
886fail_iput:
887 iput(inode);
888
889fail_inode:
890 return NULL;
891}
892
893int create_pipe_files(struct file **res, int flags)
894{
895 struct inode *inode = get_pipe_inode();
896 struct file *f;
897 int error;
898
899 if (!inode)
900 return -ENFILE;
901
902 if (flags & O_NOTIFICATION_PIPE) {
903 error = watch_queue_init(inode->i_pipe);
904 if (error) {
905 free_pipe_info(inode->i_pipe);
906 iput(inode);
907 return error;
908 }
909 }
910
911 f = alloc_file_pseudo(inode, pipe_mnt, "",
912 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
913 &pipefifo_fops);
914 if (IS_ERR(f)) {
915 free_pipe_info(inode->i_pipe);
916 iput(inode);
917 return PTR_ERR(f);
918 }
919
920 f->private_data = inode->i_pipe;
921
922 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
923 &pipefifo_fops);
924 if (IS_ERR(res[0])) {
925 put_pipe_info(inode, inode->i_pipe);
926 fput(f);
927 return PTR_ERR(res[0]);
928 }
929 res[0]->private_data = inode->i_pipe;
930 res[1] = f;
931 stream_open(inode, res[0]);
932 stream_open(inode, res[1]);
933 return 0;
934}
935
936static int __do_pipe_flags(int *fd, struct file **files, int flags)
937{
938 int error;
939 int fdw, fdr;
940
941 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
942 return -EINVAL;
943
944 error = create_pipe_files(files, flags);
945 if (error)
946 return error;
947
948 error = get_unused_fd_flags(flags);
949 if (error < 0)
950 goto err_read_pipe;
951 fdr = error;
952
953 error = get_unused_fd_flags(flags);
954 if (error < 0)
955 goto err_fdr;
956 fdw = error;
957
958 audit_fd_pair(fdr, fdw);
959 fd[0] = fdr;
960 fd[1] = fdw;
961 return 0;
962
963 err_fdr:
964 put_unused_fd(fdr);
965 err_read_pipe:
966 fput(files[0]);
967 fput(files[1]);
968 return error;
969}
970
971int do_pipe_flags(int *fd, int flags)
972{
973 struct file *files[2];
974 int error = __do_pipe_flags(fd, files, flags);
975 if (!error) {
976 fd_install(fd[0], files[0]);
977 fd_install(fd[1], files[1]);
978 }
979 return error;
980}
981
982/*
983 * sys_pipe() is the normal C calling standard for creating
984 * a pipe. It's not the way Unix traditionally does this, though.
985 */
986static int do_pipe2(int __user *fildes, int flags)
987{
988 struct file *files[2];
989 int fd[2];
990 int error;
991
992 error = __do_pipe_flags(fd, files, flags);
993 if (!error) {
994 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
995 fput(files[0]);
996 fput(files[1]);
997 put_unused_fd(fd[0]);
998 put_unused_fd(fd[1]);
999 error = -EFAULT;
1000 } else {
1001 fd_install(fd[0], files[0]);
1002 fd_install(fd[1], files[1]);
1003 }
1004 }
1005 return error;
1006}
1007
1008SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1009{
1010 return do_pipe2(fildes, flags);
1011}
1012
1013SYSCALL_DEFINE1(pipe, int __user *, fildes)
1014{
1015 return do_pipe2(fildes, 0);
1016}
1017
1018/*
1019 * This is the stupid "wait for pipe to be readable or writable"
1020 * model.
1021 *
1022 * See pipe_read/write() for the proper kind of exclusive wait,
1023 * but that requires that we wake up any other readers/writers
1024 * if we then do not end up reading everything (ie the whole
1025 * "wake_next_reader/writer" logic in pipe_read/write()).
1026 */
1027void pipe_wait_readable(struct pipe_inode_info *pipe)
1028{
1029 pipe_unlock(pipe);
1030 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1031 pipe_lock(pipe);
1032}
1033
1034void pipe_wait_writable(struct pipe_inode_info *pipe)
1035{
1036 pipe_unlock(pipe);
1037 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1038 pipe_lock(pipe);
1039}
1040
1041/*
1042 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1043 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1044 * race with the count check and waitqueue prep.
1045 *
1046 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1047 * then check the condition you're waiting for, and only then sleep. But
1048 * because of the pipe lock, we can check the condition before being on
1049 * the wait queue.
1050 *
1051 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1052 */
1053static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1054{
1055 DEFINE_WAIT(rdwait);
1056 int cur = *cnt;
1057
1058 while (cur == *cnt) {
1059 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1060 pipe_unlock(pipe);
1061 schedule();
1062 finish_wait(&pipe->rd_wait, &rdwait);
1063 pipe_lock(pipe);
1064 if (signal_pending(current))
1065 break;
1066 }
1067 return cur == *cnt ? -ERESTARTSYS : 0;
1068}
1069
1070static void wake_up_partner(struct pipe_inode_info *pipe)
1071{
1072 wake_up_interruptible_all(&pipe->rd_wait);
1073}
1074
1075static int fifo_open(struct inode *inode, struct file *filp)
1076{
1077 struct pipe_inode_info *pipe;
1078 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1079 int ret;
1080
1081 filp->f_version = 0;
1082
1083 spin_lock(&inode->i_lock);
1084 if (inode->i_pipe) {
1085 pipe = inode->i_pipe;
1086 pipe->files++;
1087 spin_unlock(&inode->i_lock);
1088 } else {
1089 spin_unlock(&inode->i_lock);
1090 pipe = alloc_pipe_info();
1091 if (!pipe)
1092 return -ENOMEM;
1093 pipe->files = 1;
1094 spin_lock(&inode->i_lock);
1095 if (unlikely(inode->i_pipe)) {
1096 inode->i_pipe->files++;
1097 spin_unlock(&inode->i_lock);
1098 free_pipe_info(pipe);
1099 pipe = inode->i_pipe;
1100 } else {
1101 inode->i_pipe = pipe;
1102 spin_unlock(&inode->i_lock);
1103 }
1104 }
1105 filp->private_data = pipe;
1106 /* OK, we have a pipe and it's pinned down */
1107
1108 __pipe_lock(pipe);
1109
1110 /* We can only do regular read/write on fifos */
1111 stream_open(inode, filp);
1112
1113 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1114 case FMODE_READ:
1115 /*
1116 * O_RDONLY
1117 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1118 * opened, even when there is no process writing the FIFO.
1119 */
1120 pipe->r_counter++;
1121 if (pipe->readers++ == 0)
1122 wake_up_partner(pipe);
1123
1124 if (!is_pipe && !pipe->writers) {
1125 if ((filp->f_flags & O_NONBLOCK)) {
1126 /* suppress EPOLLHUP until we have
1127 * seen a writer */
1128 filp->f_version = pipe->w_counter;
1129 } else {
1130 if (wait_for_partner(pipe, &pipe->w_counter))
1131 goto err_rd;
1132 }
1133 }
1134 break;
1135
1136 case FMODE_WRITE:
1137 /*
1138 * O_WRONLY
1139 * POSIX.1 says that O_NONBLOCK means return -1 with
1140 * errno=ENXIO when there is no process reading the FIFO.
1141 */
1142 ret = -ENXIO;
1143 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1144 goto err;
1145
1146 pipe->w_counter++;
1147 if (!pipe->writers++)
1148 wake_up_partner(pipe);
1149
1150 if (!is_pipe && !pipe->readers) {
1151 if (wait_for_partner(pipe, &pipe->r_counter))
1152 goto err_wr;
1153 }
1154 break;
1155
1156 case FMODE_READ | FMODE_WRITE:
1157 /*
1158 * O_RDWR
1159 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1160 * This implementation will NEVER block on a O_RDWR open, since
1161 * the process can at least talk to itself.
1162 */
1163
1164 pipe->readers++;
1165 pipe->writers++;
1166 pipe->r_counter++;
1167 pipe->w_counter++;
1168 if (pipe->readers == 1 || pipe->writers == 1)
1169 wake_up_partner(pipe);
1170 break;
1171
1172 default:
1173 ret = -EINVAL;
1174 goto err;
1175 }
1176
1177 /* Ok! */
1178 __pipe_unlock(pipe);
1179 return 0;
1180
1181err_rd:
1182 if (!--pipe->readers)
1183 wake_up_interruptible(&pipe->wr_wait);
1184 ret = -ERESTARTSYS;
1185 goto err;
1186
1187err_wr:
1188 if (!--pipe->writers)
1189 wake_up_interruptible_all(&pipe->rd_wait);
1190 ret = -ERESTARTSYS;
1191 goto err;
1192
1193err:
1194 __pipe_unlock(pipe);
1195
1196 put_pipe_info(inode, pipe);
1197 return ret;
1198}
1199
1200const struct file_operations pipefifo_fops = {
1201 .open = fifo_open,
1202 .llseek = no_llseek,
1203 .read_iter = pipe_read,
1204 .write_iter = pipe_write,
1205 .poll = pipe_poll,
1206 .unlocked_ioctl = pipe_ioctl,
1207 .release = pipe_release,
1208 .fasync = pipe_fasync,
1209};
1210
1211/*
1212 * Currently we rely on the pipe array holding a power-of-2 number
1213 * of pages. Returns 0 on error.
1214 */
1215unsigned int round_pipe_size(unsigned long size)
1216{
1217 if (size > (1U << 31))
1218 return 0;
1219
1220 /* Minimum pipe size, as required by POSIX */
1221 if (size < PAGE_SIZE)
1222 return PAGE_SIZE;
1223
1224 return roundup_pow_of_two(size);
1225}
1226
1227/*
1228 * Resize the pipe ring to a number of slots.
1229 */
1230int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1231{
1232 struct pipe_buffer *bufs;
1233 unsigned int head, tail, mask, n;
1234
1235 /*
1236 * We can shrink the pipe, if arg is greater than the ring occupancy.
1237 * Since we don't expect a lot of shrink+grow operations, just free and
1238 * allocate again like we would do for growing. If the pipe currently
1239 * contains more buffers than arg, then return busy.
1240 */
1241 mask = pipe->ring_size - 1;
1242 head = pipe->head;
1243 tail = pipe->tail;
1244 n = pipe_occupancy(pipe->head, pipe->tail);
1245 if (nr_slots < n)
1246 return -EBUSY;
1247
1248 bufs = kcalloc(nr_slots, sizeof(*bufs),
1249 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1250 if (unlikely(!bufs))
1251 return -ENOMEM;
1252
1253 /*
1254 * The pipe array wraps around, so just start the new one at zero
1255 * and adjust the indices.
1256 */
1257 if (n > 0) {
1258 unsigned int h = head & mask;
1259 unsigned int t = tail & mask;
1260 if (h > t) {
1261 memcpy(bufs, pipe->bufs + t,
1262 n * sizeof(struct pipe_buffer));
1263 } else {
1264 unsigned int tsize = pipe->ring_size - t;
1265 if (h > 0)
1266 memcpy(bufs + tsize, pipe->bufs,
1267 h * sizeof(struct pipe_buffer));
1268 memcpy(bufs, pipe->bufs + t,
1269 tsize * sizeof(struct pipe_buffer));
1270 }
1271 }
1272
1273 head = n;
1274 tail = 0;
1275
1276 kfree(pipe->bufs);
1277 pipe->bufs = bufs;
1278 pipe->ring_size = nr_slots;
1279 if (pipe->max_usage > nr_slots)
1280 pipe->max_usage = nr_slots;
1281 pipe->tail = tail;
1282 pipe->head = head;
1283
1284 /* This might have made more room for writers */
1285 wake_up_interruptible(&pipe->wr_wait);
1286 return 0;
1287}
1288
1289/*
1290 * Allocate a new array of pipe buffers and copy the info over. Returns the
1291 * pipe size if successful, or return -ERROR on error.
1292 */
1293static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1294{
1295 unsigned long user_bufs;
1296 unsigned int nr_slots, size;
1297 long ret = 0;
1298
1299#ifdef CONFIG_WATCH_QUEUE
1300 if (pipe->watch_queue)
1301 return -EBUSY;
1302#endif
1303
1304 size = round_pipe_size(arg);
1305 nr_slots = size >> PAGE_SHIFT;
1306
1307 if (!nr_slots)
1308 return -EINVAL;
1309
1310 /*
1311 * If trying to increase the pipe capacity, check that an
1312 * unprivileged user is not trying to exceed various limits
1313 * (soft limit check here, hard limit check just below).
1314 * Decreasing the pipe capacity is always permitted, even
1315 * if the user is currently over a limit.
1316 */
1317 if (nr_slots > pipe->max_usage &&
1318 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1319 return -EPERM;
1320
1321 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1322
1323 if (nr_slots > pipe->max_usage &&
1324 (too_many_pipe_buffers_hard(user_bufs) ||
1325 too_many_pipe_buffers_soft(user_bufs)) &&
1326 pipe_is_unprivileged_user()) {
1327 ret = -EPERM;
1328 goto out_revert_acct;
1329 }
1330
1331 ret = pipe_resize_ring(pipe, nr_slots);
1332 if (ret < 0)
1333 goto out_revert_acct;
1334
1335 pipe->max_usage = nr_slots;
1336 pipe->nr_accounted = nr_slots;
1337 return pipe->max_usage * PAGE_SIZE;
1338
1339out_revert_acct:
1340 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1341 return ret;
1342}
1343
1344/*
1345 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1346 * location, so checking ->i_pipe is not enough to verify that this is a
1347 * pipe.
1348 */
1349struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1350{
1351 struct pipe_inode_info *pipe = file->private_data;
1352
1353 if (file->f_op != &pipefifo_fops || !pipe)
1354 return NULL;
1355#ifdef CONFIG_WATCH_QUEUE
1356 if (for_splice && pipe->watch_queue)
1357 return NULL;
1358#endif
1359 return pipe;
1360}
1361
1362long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1363{
1364 struct pipe_inode_info *pipe;
1365 long ret;
1366
1367 pipe = get_pipe_info(file, false);
1368 if (!pipe)
1369 return -EBADF;
1370
1371 __pipe_lock(pipe);
1372
1373 switch (cmd) {
1374 case F_SETPIPE_SZ:
1375 ret = pipe_set_size(pipe, arg);
1376 break;
1377 case F_GETPIPE_SZ:
1378 ret = pipe->max_usage * PAGE_SIZE;
1379 break;
1380 default:
1381 ret = -EINVAL;
1382 break;
1383 }
1384
1385 __pipe_unlock(pipe);
1386 return ret;
1387}
1388
1389static const struct super_operations pipefs_ops = {
1390 .destroy_inode = free_inode_nonrcu,
1391 .statfs = simple_statfs,
1392};
1393
1394/*
1395 * pipefs should _never_ be mounted by userland - too much of security hassle,
1396 * no real gain from having the whole whorehouse mounted. So we don't need
1397 * any operations on the root directory. However, we need a non-trivial
1398 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1399 */
1400
1401static int pipefs_init_fs_context(struct fs_context *fc)
1402{
1403 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1404 if (!ctx)
1405 return -ENOMEM;
1406 ctx->ops = &pipefs_ops;
1407 ctx->dops = &pipefs_dentry_operations;
1408 return 0;
1409}
1410
1411static struct file_system_type pipe_fs_type = {
1412 .name = "pipefs",
1413 .init_fs_context = pipefs_init_fs_context,
1414 .kill_sb = kill_anon_super,
1415};
1416
1417static int __init init_pipe_fs(void)
1418{
1419 int err = register_filesystem(&pipe_fs_type);
1420
1421 if (!err) {
1422 pipe_mnt = kern_mount(&pipe_fs_type);
1423 if (IS_ERR(pipe_mnt)) {
1424 err = PTR_ERR(pipe_mnt);
1425 unregister_filesystem(&pipe_fs_type);
1426 }
1427 }
1428 return err;
1429}
1430
1431fs_initcall(init_pipe_fs);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/pipe.c
4 *
5 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
6 */
7
8#include <linux/mm.h>
9#include <linux/file.h>
10#include <linux/poll.h>
11#include <linux/slab.h>
12#include <linux/module.h>
13#include <linux/init.h>
14#include <linux/fs.h>
15#include <linux/log2.h>
16#include <linux/mount.h>
17#include <linux/pseudo_fs.h>
18#include <linux/magic.h>
19#include <linux/pipe_fs_i.h>
20#include <linux/uio.h>
21#include <linux/highmem.h>
22#include <linux/pagemap.h>
23#include <linux/audit.h>
24#include <linux/syscalls.h>
25#include <linux/fcntl.h>
26#include <linux/memcontrol.h>
27#include <linux/watch_queue.h>
28#include <linux/sysctl.h>
29
30#include <linux/uaccess.h>
31#include <asm/ioctls.h>
32
33#include "internal.h"
34
35/*
36 * New pipe buffers will be restricted to this size while the user is exceeding
37 * their pipe buffer quota. The general pipe use case needs at least two
38 * buffers: one for data yet to be read, and one for new data. If this is less
39 * than two, then a write to a non-empty pipe may block even if the pipe is not
40 * full. This can occur with GNU make jobserver or similar uses of pipes as
41 * semaphores: multiple processes may be waiting to write tokens back to the
42 * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
43 *
44 * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
45 * own risk, namely: pipe writes to non-full pipes may block until the pipe is
46 * emptied.
47 */
48#define PIPE_MIN_DEF_BUFFERS 2
49
50/*
51 * The max size that a non-root user is allowed to grow the pipe. Can
52 * be set by root in /proc/sys/fs/pipe-max-size
53 */
54static unsigned int pipe_max_size = 1048576;
55
56/* Maximum allocatable pages per user. Hard limit is unset by default, soft
57 * matches default values.
58 */
59static unsigned long pipe_user_pages_hard;
60static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
61
62/*
63 * We use head and tail indices that aren't masked off, except at the point of
64 * dereference, but rather they're allowed to wrap naturally. This means there
65 * isn't a dead spot in the buffer, but the ring has to be a power of two and
66 * <= 2^31.
67 * -- David Howells 2019-09-23.
68 *
69 * Reads with count = 0 should always return 0.
70 * -- Julian Bradfield 1999-06-07.
71 *
72 * FIFOs and Pipes now generate SIGIO for both readers and writers.
73 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
74 *
75 * pipe_read & write cleanup
76 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
77 */
78
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);