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
28#include <linux/uaccess.h>
29#include <asm/ioctls.h>
30
31#include "internal.h"
32
33/*
34 * The max size that a non-root user is allowed to grow the pipe. Can
35 * be set by root in /proc/sys/fs/pipe-max-size
36 */
37unsigned int pipe_max_size = 1048576;
38
39/* Maximum allocatable pages per user. Hard limit is unset by default, soft
40 * matches default values.
41 */
42unsigned long pipe_user_pages_hard;
43unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
44
45/*
46 * We use a start+len construction, which provides full use of the
47 * allocated memory.
48 * -- Florian Coosmann (FGC)
49 *
50 * Reads with count = 0 should always return 0.
51 * -- Julian Bradfield 1999-06-07.
52 *
53 * FIFOs and Pipes now generate SIGIO for both readers and writers.
54 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
55 *
56 * pipe_read & write cleanup
57 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
58 */
59
60static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
61{
62 if (pipe->files)
63 mutex_lock_nested(&pipe->mutex, subclass);
64}
65
66void pipe_lock(struct pipe_inode_info *pipe)
67{
68 /*
69 * pipe_lock() nests non-pipe inode locks (for writing to a file)
70 */
71 pipe_lock_nested(pipe, I_MUTEX_PARENT);
72}
73EXPORT_SYMBOL(pipe_lock);
74
75void pipe_unlock(struct pipe_inode_info *pipe)
76{
77 if (pipe->files)
78 mutex_unlock(&pipe->mutex);
79}
80EXPORT_SYMBOL(pipe_unlock);
81
82static inline void __pipe_lock(struct pipe_inode_info *pipe)
83{
84 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
85}
86
87static inline void __pipe_unlock(struct pipe_inode_info *pipe)
88{
89 mutex_unlock(&pipe->mutex);
90}
91
92void pipe_double_lock(struct pipe_inode_info *pipe1,
93 struct pipe_inode_info *pipe2)
94{
95 BUG_ON(pipe1 == pipe2);
96
97 if (pipe1 < pipe2) {
98 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
99 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
100 } else {
101 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
102 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
103 }
104}
105
106/* Drop the inode semaphore and wait for a pipe event, atomically */
107void pipe_wait(struct pipe_inode_info *pipe)
108{
109 DEFINE_WAIT(wait);
110
111 /*
112 * Pipes are system-local resources, so sleeping on them
113 * is considered a noninteractive wait:
114 */
115 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
116 pipe_unlock(pipe);
117 schedule();
118 finish_wait(&pipe->wait, &wait);
119 pipe_lock(pipe);
120}
121
122static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
123 struct pipe_buffer *buf)
124{
125 struct page *page = buf->page;
126
127 /*
128 * If nobody else uses this page, and we don't already have a
129 * temporary page, let's keep track of it as a one-deep
130 * allocation cache. (Otherwise just release our reference to it)
131 */
132 if (page_count(page) == 1 && !pipe->tmp_page)
133 pipe->tmp_page = page;
134 else
135 put_page(page);
136}
137
138static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
139 struct pipe_buffer *buf)
140{
141 struct page *page = buf->page;
142
143 if (page_count(page) == 1) {
144 memcg_kmem_uncharge(page, 0);
145 __SetPageLocked(page);
146 return 0;
147 }
148 return 1;
149}
150
151/**
152 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
153 * @pipe: the pipe that the buffer belongs to
154 * @buf: the buffer to attempt to steal
155 *
156 * Description:
157 * This function attempts to steal the &struct page attached to
158 * @buf. If successful, this function returns 0 and returns with
159 * the page locked. The caller may then reuse the page for whatever
160 * he wishes; the typical use is insertion into a different file
161 * page cache.
162 */
163int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
164 struct pipe_buffer *buf)
165{
166 struct page *page = buf->page;
167
168 /*
169 * A reference of one is golden, that means that the owner of this
170 * page is the only one holding a reference to it. lock the page
171 * and return OK.
172 */
173 if (page_count(page) == 1) {
174 lock_page(page);
175 return 0;
176 }
177
178 return 1;
179}
180EXPORT_SYMBOL(generic_pipe_buf_steal);
181
182/**
183 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
184 * @pipe: the pipe that the buffer belongs to
185 * @buf: the buffer to get a reference to
186 *
187 * Description:
188 * This function grabs an extra reference to @buf. It's used in
189 * in the tee() system call, when we duplicate the buffers in one
190 * pipe into another.
191 */
192bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
193{
194 return try_get_page(buf->page);
195}
196EXPORT_SYMBOL(generic_pipe_buf_get);
197
198/**
199 * generic_pipe_buf_confirm - verify contents of the pipe buffer
200 * @info: the pipe that the buffer belongs to
201 * @buf: the buffer to confirm
202 *
203 * Description:
204 * This function does nothing, because the generic pipe code uses
205 * pages that are always good when inserted into the pipe.
206 */
207int generic_pipe_buf_confirm(struct pipe_inode_info *info,
208 struct pipe_buffer *buf)
209{
210 return 0;
211}
212EXPORT_SYMBOL(generic_pipe_buf_confirm);
213
214/**
215 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
216 * @pipe: the pipe that the buffer belongs to
217 * @buf: the buffer to put a reference to
218 *
219 * Description:
220 * This function releases a reference to @buf.
221 */
222void generic_pipe_buf_release(struct pipe_inode_info *pipe,
223 struct pipe_buffer *buf)
224{
225 put_page(buf->page);
226}
227EXPORT_SYMBOL(generic_pipe_buf_release);
228
229/* New data written to a pipe may be appended to a buffer with this type. */
230static const struct pipe_buf_operations anon_pipe_buf_ops = {
231 .confirm = generic_pipe_buf_confirm,
232 .release = anon_pipe_buf_release,
233 .steal = anon_pipe_buf_steal,
234 .get = generic_pipe_buf_get,
235};
236
237static const struct pipe_buf_operations anon_pipe_buf_nomerge_ops = {
238 .confirm = generic_pipe_buf_confirm,
239 .release = anon_pipe_buf_release,
240 .steal = anon_pipe_buf_steal,
241 .get = generic_pipe_buf_get,
242};
243
244static const struct pipe_buf_operations packet_pipe_buf_ops = {
245 .confirm = generic_pipe_buf_confirm,
246 .release = anon_pipe_buf_release,
247 .steal = anon_pipe_buf_steal,
248 .get = generic_pipe_buf_get,
249};
250
251/**
252 * pipe_buf_mark_unmergeable - mark a &struct pipe_buffer as unmergeable
253 * @buf: the buffer to mark
254 *
255 * Description:
256 * This function ensures that no future writes will be merged into the
257 * given &struct pipe_buffer. This is necessary when multiple pipe buffers
258 * share the same backing page.
259 */
260void pipe_buf_mark_unmergeable(struct pipe_buffer *buf)
261{
262 if (buf->ops == &anon_pipe_buf_ops)
263 buf->ops = &anon_pipe_buf_nomerge_ops;
264}
265
266static bool pipe_buf_can_merge(struct pipe_buffer *buf)
267{
268 return buf->ops == &anon_pipe_buf_ops;
269}
270
271static ssize_t
272pipe_read(struct kiocb *iocb, struct iov_iter *to)
273{
274 size_t total_len = iov_iter_count(to);
275 struct file *filp = iocb->ki_filp;
276 struct pipe_inode_info *pipe = filp->private_data;
277 int do_wakeup;
278 ssize_t ret;
279
280 /* Null read succeeds. */
281 if (unlikely(total_len == 0))
282 return 0;
283
284 do_wakeup = 0;
285 ret = 0;
286 __pipe_lock(pipe);
287 for (;;) {
288 int bufs = pipe->nrbufs;
289 if (bufs) {
290 int curbuf = pipe->curbuf;
291 struct pipe_buffer *buf = pipe->bufs + curbuf;
292 size_t chars = buf->len;
293 size_t written;
294 int error;
295
296 if (chars > total_len)
297 chars = total_len;
298
299 error = pipe_buf_confirm(pipe, buf);
300 if (error) {
301 if (!ret)
302 ret = error;
303 break;
304 }
305
306 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
307 if (unlikely(written < chars)) {
308 if (!ret)
309 ret = -EFAULT;
310 break;
311 }
312 ret += chars;
313 buf->offset += chars;
314 buf->len -= chars;
315
316 /* Was it a packet buffer? Clean up and exit */
317 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
318 total_len = chars;
319 buf->len = 0;
320 }
321
322 if (!buf->len) {
323 pipe_buf_release(pipe, buf);
324 curbuf = (curbuf + 1) & (pipe->buffers - 1);
325 pipe->curbuf = curbuf;
326 pipe->nrbufs = --bufs;
327 do_wakeup = 1;
328 }
329 total_len -= chars;
330 if (!total_len)
331 break; /* common path: read succeeded */
332 }
333 if (bufs) /* More to do? */
334 continue;
335 if (!pipe->writers)
336 break;
337 if (!pipe->waiting_writers) {
338 /* syscall merging: Usually we must not sleep
339 * if O_NONBLOCK is set, or if we got some data.
340 * But if a writer sleeps in kernel space, then
341 * we can wait for that data without violating POSIX.
342 */
343 if (ret)
344 break;
345 if (filp->f_flags & O_NONBLOCK) {
346 ret = -EAGAIN;
347 break;
348 }
349 }
350 if (signal_pending(current)) {
351 if (!ret)
352 ret = -ERESTARTSYS;
353 break;
354 }
355 if (do_wakeup) {
356 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
357 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
358 }
359 pipe_wait(pipe);
360 }
361 __pipe_unlock(pipe);
362
363 /* Signal writers asynchronously that there is more room. */
364 if (do_wakeup) {
365 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
366 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
367 }
368 if (ret > 0)
369 file_accessed(filp);
370 return ret;
371}
372
373static inline int is_packetized(struct file *file)
374{
375 return (file->f_flags & O_DIRECT) != 0;
376}
377
378static ssize_t
379pipe_write(struct kiocb *iocb, struct iov_iter *from)
380{
381 struct file *filp = iocb->ki_filp;
382 struct pipe_inode_info *pipe = filp->private_data;
383 ssize_t ret = 0;
384 int do_wakeup = 0;
385 size_t total_len = iov_iter_count(from);
386 ssize_t chars;
387
388 /* Null write succeeds. */
389 if (unlikely(total_len == 0))
390 return 0;
391
392 __pipe_lock(pipe);
393
394 if (!pipe->readers) {
395 send_sig(SIGPIPE, current, 0);
396 ret = -EPIPE;
397 goto out;
398 }
399
400 /* We try to merge small writes */
401 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
402 if (pipe->nrbufs && chars != 0) {
403 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
404 (pipe->buffers - 1);
405 struct pipe_buffer *buf = pipe->bufs + lastbuf;
406 int offset = buf->offset + buf->len;
407
408 if (pipe_buf_can_merge(buf) && offset + chars <= PAGE_SIZE) {
409 ret = pipe_buf_confirm(pipe, buf);
410 if (ret)
411 goto out;
412
413 ret = copy_page_from_iter(buf->page, offset, chars, from);
414 if (unlikely(ret < chars)) {
415 ret = -EFAULT;
416 goto out;
417 }
418 do_wakeup = 1;
419 buf->len += ret;
420 if (!iov_iter_count(from))
421 goto out;
422 }
423 }
424
425 for (;;) {
426 int bufs;
427
428 if (!pipe->readers) {
429 send_sig(SIGPIPE, current, 0);
430 if (!ret)
431 ret = -EPIPE;
432 break;
433 }
434 bufs = pipe->nrbufs;
435 if (bufs < pipe->buffers) {
436 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
437 struct pipe_buffer *buf = pipe->bufs + newbuf;
438 struct page *page = pipe->tmp_page;
439 int copied;
440
441 if (!page) {
442 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
443 if (unlikely(!page)) {
444 ret = ret ? : -ENOMEM;
445 break;
446 }
447 pipe->tmp_page = page;
448 }
449 /* Always wake up, even if the copy fails. Otherwise
450 * we lock up (O_NONBLOCK-)readers that sleep due to
451 * syscall merging.
452 * FIXME! Is this really true?
453 */
454 do_wakeup = 1;
455 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
456 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
457 if (!ret)
458 ret = -EFAULT;
459 break;
460 }
461 ret += copied;
462
463 /* Insert it into the buffer array */
464 buf->page = page;
465 buf->ops = &anon_pipe_buf_ops;
466 buf->offset = 0;
467 buf->len = copied;
468 buf->flags = 0;
469 if (is_packetized(filp)) {
470 buf->ops = &packet_pipe_buf_ops;
471 buf->flags = PIPE_BUF_FLAG_PACKET;
472 }
473 pipe->nrbufs = ++bufs;
474 pipe->tmp_page = NULL;
475
476 if (!iov_iter_count(from))
477 break;
478 }
479 if (bufs < pipe->buffers)
480 continue;
481 if (filp->f_flags & O_NONBLOCK) {
482 if (!ret)
483 ret = -EAGAIN;
484 break;
485 }
486 if (signal_pending(current)) {
487 if (!ret)
488 ret = -ERESTARTSYS;
489 break;
490 }
491 if (do_wakeup) {
492 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
493 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
494 do_wakeup = 0;
495 }
496 pipe->waiting_writers++;
497 pipe_wait(pipe);
498 pipe->waiting_writers--;
499 }
500out:
501 __pipe_unlock(pipe);
502 if (do_wakeup) {
503 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
504 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
505 }
506 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
507 int err = file_update_time(filp);
508 if (err)
509 ret = err;
510 sb_end_write(file_inode(filp)->i_sb);
511 }
512 return ret;
513}
514
515static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
516{
517 struct pipe_inode_info *pipe = filp->private_data;
518 int count, buf, nrbufs;
519
520 switch (cmd) {
521 case FIONREAD:
522 __pipe_lock(pipe);
523 count = 0;
524 buf = pipe->curbuf;
525 nrbufs = pipe->nrbufs;
526 while (--nrbufs >= 0) {
527 count += pipe->bufs[buf].len;
528 buf = (buf+1) & (pipe->buffers - 1);
529 }
530 __pipe_unlock(pipe);
531
532 return put_user(count, (int __user *)arg);
533 default:
534 return -ENOIOCTLCMD;
535 }
536}
537
538/* No kernel lock held - fine */
539static __poll_t
540pipe_poll(struct file *filp, poll_table *wait)
541{
542 __poll_t mask;
543 struct pipe_inode_info *pipe = filp->private_data;
544 int nrbufs;
545
546 poll_wait(filp, &pipe->wait, wait);
547
548 /* Reading only -- no need for acquiring the semaphore. */
549 nrbufs = pipe->nrbufs;
550 mask = 0;
551 if (filp->f_mode & FMODE_READ) {
552 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
553 if (!pipe->writers && filp->f_version != pipe->w_counter)
554 mask |= EPOLLHUP;
555 }
556
557 if (filp->f_mode & FMODE_WRITE) {
558 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
559 /*
560 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
561 * behave exactly like pipes for poll().
562 */
563 if (!pipe->readers)
564 mask |= EPOLLERR;
565 }
566
567 return mask;
568}
569
570static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
571{
572 int kill = 0;
573
574 spin_lock(&inode->i_lock);
575 if (!--pipe->files) {
576 inode->i_pipe = NULL;
577 kill = 1;
578 }
579 spin_unlock(&inode->i_lock);
580
581 if (kill)
582 free_pipe_info(pipe);
583}
584
585static int
586pipe_release(struct inode *inode, struct file *file)
587{
588 struct pipe_inode_info *pipe = file->private_data;
589
590 __pipe_lock(pipe);
591 if (file->f_mode & FMODE_READ)
592 pipe->readers--;
593 if (file->f_mode & FMODE_WRITE)
594 pipe->writers--;
595
596 if (pipe->readers || pipe->writers) {
597 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
598 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
599 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
600 }
601 __pipe_unlock(pipe);
602
603 put_pipe_info(inode, pipe);
604 return 0;
605}
606
607static int
608pipe_fasync(int fd, struct file *filp, int on)
609{
610 struct pipe_inode_info *pipe = filp->private_data;
611 int retval = 0;
612
613 __pipe_lock(pipe);
614 if (filp->f_mode & FMODE_READ)
615 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
616 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
617 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
618 if (retval < 0 && (filp->f_mode & FMODE_READ))
619 /* this can happen only if on == T */
620 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
621 }
622 __pipe_unlock(pipe);
623 return retval;
624}
625
626static unsigned long account_pipe_buffers(struct user_struct *user,
627 unsigned long old, unsigned long new)
628{
629 return atomic_long_add_return(new - old, &user->pipe_bufs);
630}
631
632static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
633{
634 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
635
636 return soft_limit && user_bufs > soft_limit;
637}
638
639static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
640{
641 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
642
643 return hard_limit && user_bufs > hard_limit;
644}
645
646static bool is_unprivileged_user(void)
647{
648 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
649}
650
651struct pipe_inode_info *alloc_pipe_info(void)
652{
653 struct pipe_inode_info *pipe;
654 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
655 struct user_struct *user = get_current_user();
656 unsigned long user_bufs;
657 unsigned int max_size = READ_ONCE(pipe_max_size);
658
659 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
660 if (pipe == NULL)
661 goto out_free_uid;
662
663 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
664 pipe_bufs = max_size >> PAGE_SHIFT;
665
666 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
667
668 if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
669 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
670 pipe_bufs = 1;
671 }
672
673 if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
674 goto out_revert_acct;
675
676 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
677 GFP_KERNEL_ACCOUNT);
678
679 if (pipe->bufs) {
680 init_waitqueue_head(&pipe->wait);
681 pipe->r_counter = pipe->w_counter = 1;
682 pipe->buffers = pipe_bufs;
683 pipe->user = user;
684 mutex_init(&pipe->mutex);
685 return pipe;
686 }
687
688out_revert_acct:
689 (void) account_pipe_buffers(user, pipe_bufs, 0);
690 kfree(pipe);
691out_free_uid:
692 free_uid(user);
693 return NULL;
694}
695
696void free_pipe_info(struct pipe_inode_info *pipe)
697{
698 int i;
699
700 (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
701 free_uid(pipe->user);
702 for (i = 0; i < pipe->buffers; i++) {
703 struct pipe_buffer *buf = pipe->bufs + i;
704 if (buf->ops)
705 pipe_buf_release(pipe, buf);
706 }
707 if (pipe->tmp_page)
708 __free_page(pipe->tmp_page);
709 kfree(pipe->bufs);
710 kfree(pipe);
711}
712
713static struct vfsmount *pipe_mnt __read_mostly;
714
715/*
716 * pipefs_dname() is called from d_path().
717 */
718static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
719{
720 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
721 d_inode(dentry)->i_ino);
722}
723
724static const struct dentry_operations pipefs_dentry_operations = {
725 .d_dname = pipefs_dname,
726};
727
728static struct inode * get_pipe_inode(void)
729{
730 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
731 struct pipe_inode_info *pipe;
732
733 if (!inode)
734 goto fail_inode;
735
736 inode->i_ino = get_next_ino();
737
738 pipe = alloc_pipe_info();
739 if (!pipe)
740 goto fail_iput;
741
742 inode->i_pipe = pipe;
743 pipe->files = 2;
744 pipe->readers = pipe->writers = 1;
745 inode->i_fop = &pipefifo_fops;
746
747 /*
748 * Mark the inode dirty from the very beginning,
749 * that way it will never be moved to the dirty
750 * list because "mark_inode_dirty()" will think
751 * that it already _is_ on the dirty list.
752 */
753 inode->i_state = I_DIRTY;
754 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
755 inode->i_uid = current_fsuid();
756 inode->i_gid = current_fsgid();
757 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
758
759 return inode;
760
761fail_iput:
762 iput(inode);
763
764fail_inode:
765 return NULL;
766}
767
768int create_pipe_files(struct file **res, int flags)
769{
770 struct inode *inode = get_pipe_inode();
771 struct file *f;
772
773 if (!inode)
774 return -ENFILE;
775
776 f = alloc_file_pseudo(inode, pipe_mnt, "",
777 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
778 &pipefifo_fops);
779 if (IS_ERR(f)) {
780 free_pipe_info(inode->i_pipe);
781 iput(inode);
782 return PTR_ERR(f);
783 }
784
785 f->private_data = inode->i_pipe;
786
787 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
788 &pipefifo_fops);
789 if (IS_ERR(res[0])) {
790 put_pipe_info(inode, inode->i_pipe);
791 fput(f);
792 return PTR_ERR(res[0]);
793 }
794 res[0]->private_data = inode->i_pipe;
795 res[1] = f;
796 return 0;
797}
798
799static int __do_pipe_flags(int *fd, struct file **files, int flags)
800{
801 int error;
802 int fdw, fdr;
803
804 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
805 return -EINVAL;
806
807 error = create_pipe_files(files, flags);
808 if (error)
809 return error;
810
811 error = get_unused_fd_flags(flags);
812 if (error < 0)
813 goto err_read_pipe;
814 fdr = error;
815
816 error = get_unused_fd_flags(flags);
817 if (error < 0)
818 goto err_fdr;
819 fdw = error;
820
821 audit_fd_pair(fdr, fdw);
822 fd[0] = fdr;
823 fd[1] = fdw;
824 return 0;
825
826 err_fdr:
827 put_unused_fd(fdr);
828 err_read_pipe:
829 fput(files[0]);
830 fput(files[1]);
831 return error;
832}
833
834int do_pipe_flags(int *fd, int flags)
835{
836 struct file *files[2];
837 int error = __do_pipe_flags(fd, files, flags);
838 if (!error) {
839 fd_install(fd[0], files[0]);
840 fd_install(fd[1], files[1]);
841 }
842 return error;
843}
844
845/*
846 * sys_pipe() is the normal C calling standard for creating
847 * a pipe. It's not the way Unix traditionally does this, though.
848 */
849static int do_pipe2(int __user *fildes, int flags)
850{
851 struct file *files[2];
852 int fd[2];
853 int error;
854
855 error = __do_pipe_flags(fd, files, flags);
856 if (!error) {
857 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
858 fput(files[0]);
859 fput(files[1]);
860 put_unused_fd(fd[0]);
861 put_unused_fd(fd[1]);
862 error = -EFAULT;
863 } else {
864 fd_install(fd[0], files[0]);
865 fd_install(fd[1], files[1]);
866 }
867 }
868 return error;
869}
870
871SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
872{
873 return do_pipe2(fildes, flags);
874}
875
876SYSCALL_DEFINE1(pipe, int __user *, fildes)
877{
878 return do_pipe2(fildes, 0);
879}
880
881static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
882{
883 int cur = *cnt;
884
885 while (cur == *cnt) {
886 pipe_wait(pipe);
887 if (signal_pending(current))
888 break;
889 }
890 return cur == *cnt ? -ERESTARTSYS : 0;
891}
892
893static void wake_up_partner(struct pipe_inode_info *pipe)
894{
895 wake_up_interruptible(&pipe->wait);
896}
897
898static int fifo_open(struct inode *inode, struct file *filp)
899{
900 struct pipe_inode_info *pipe;
901 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
902 int ret;
903
904 filp->f_version = 0;
905
906 spin_lock(&inode->i_lock);
907 if (inode->i_pipe) {
908 pipe = inode->i_pipe;
909 pipe->files++;
910 spin_unlock(&inode->i_lock);
911 } else {
912 spin_unlock(&inode->i_lock);
913 pipe = alloc_pipe_info();
914 if (!pipe)
915 return -ENOMEM;
916 pipe->files = 1;
917 spin_lock(&inode->i_lock);
918 if (unlikely(inode->i_pipe)) {
919 inode->i_pipe->files++;
920 spin_unlock(&inode->i_lock);
921 free_pipe_info(pipe);
922 pipe = inode->i_pipe;
923 } else {
924 inode->i_pipe = pipe;
925 spin_unlock(&inode->i_lock);
926 }
927 }
928 filp->private_data = pipe;
929 /* OK, we have a pipe and it's pinned down */
930
931 __pipe_lock(pipe);
932
933 /* We can only do regular read/write on fifos */
934 filp->f_mode &= (FMODE_READ | FMODE_WRITE);
935
936 switch (filp->f_mode) {
937 case FMODE_READ:
938 /*
939 * O_RDONLY
940 * POSIX.1 says that O_NONBLOCK means return with the FIFO
941 * opened, even when there is no process writing the FIFO.
942 */
943 pipe->r_counter++;
944 if (pipe->readers++ == 0)
945 wake_up_partner(pipe);
946
947 if (!is_pipe && !pipe->writers) {
948 if ((filp->f_flags & O_NONBLOCK)) {
949 /* suppress EPOLLHUP until we have
950 * seen a writer */
951 filp->f_version = pipe->w_counter;
952 } else {
953 if (wait_for_partner(pipe, &pipe->w_counter))
954 goto err_rd;
955 }
956 }
957 break;
958
959 case FMODE_WRITE:
960 /*
961 * O_WRONLY
962 * POSIX.1 says that O_NONBLOCK means return -1 with
963 * errno=ENXIO when there is no process reading the FIFO.
964 */
965 ret = -ENXIO;
966 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
967 goto err;
968
969 pipe->w_counter++;
970 if (!pipe->writers++)
971 wake_up_partner(pipe);
972
973 if (!is_pipe && !pipe->readers) {
974 if (wait_for_partner(pipe, &pipe->r_counter))
975 goto err_wr;
976 }
977 break;
978
979 case FMODE_READ | FMODE_WRITE:
980 /*
981 * O_RDWR
982 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
983 * This implementation will NEVER block on a O_RDWR open, since
984 * the process can at least talk to itself.
985 */
986
987 pipe->readers++;
988 pipe->writers++;
989 pipe->r_counter++;
990 pipe->w_counter++;
991 if (pipe->readers == 1 || pipe->writers == 1)
992 wake_up_partner(pipe);
993 break;
994
995 default:
996 ret = -EINVAL;
997 goto err;
998 }
999
1000 /* Ok! */
1001 __pipe_unlock(pipe);
1002 return 0;
1003
1004err_rd:
1005 if (!--pipe->readers)
1006 wake_up_interruptible(&pipe->wait);
1007 ret = -ERESTARTSYS;
1008 goto err;
1009
1010err_wr:
1011 if (!--pipe->writers)
1012 wake_up_interruptible(&pipe->wait);
1013 ret = -ERESTARTSYS;
1014 goto err;
1015
1016err:
1017 __pipe_unlock(pipe);
1018
1019 put_pipe_info(inode, pipe);
1020 return ret;
1021}
1022
1023const struct file_operations pipefifo_fops = {
1024 .open = fifo_open,
1025 .llseek = no_llseek,
1026 .read_iter = pipe_read,
1027 .write_iter = pipe_write,
1028 .poll = pipe_poll,
1029 .unlocked_ioctl = pipe_ioctl,
1030 .release = pipe_release,
1031 .fasync = pipe_fasync,
1032};
1033
1034/*
1035 * Currently we rely on the pipe array holding a power-of-2 number
1036 * of pages. Returns 0 on error.
1037 */
1038unsigned int round_pipe_size(unsigned long size)
1039{
1040 if (size > (1U << 31))
1041 return 0;
1042
1043 /* Minimum pipe size, as required by POSIX */
1044 if (size < PAGE_SIZE)
1045 return PAGE_SIZE;
1046
1047 return roundup_pow_of_two(size);
1048}
1049
1050/*
1051 * Allocate a new array of pipe buffers and copy the info over. Returns the
1052 * pipe size if successful, or return -ERROR on error.
1053 */
1054static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1055{
1056 struct pipe_buffer *bufs;
1057 unsigned int size, nr_pages;
1058 unsigned long user_bufs;
1059 long ret = 0;
1060
1061 size = round_pipe_size(arg);
1062 nr_pages = size >> PAGE_SHIFT;
1063
1064 if (!nr_pages)
1065 return -EINVAL;
1066
1067 /*
1068 * If trying to increase the pipe capacity, check that an
1069 * unprivileged user is not trying to exceed various limits
1070 * (soft limit check here, hard limit check just below).
1071 * Decreasing the pipe capacity is always permitted, even
1072 * if the user is currently over a limit.
1073 */
1074 if (nr_pages > pipe->buffers &&
1075 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1076 return -EPERM;
1077
1078 user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1079
1080 if (nr_pages > pipe->buffers &&
1081 (too_many_pipe_buffers_hard(user_bufs) ||
1082 too_many_pipe_buffers_soft(user_bufs)) &&
1083 is_unprivileged_user()) {
1084 ret = -EPERM;
1085 goto out_revert_acct;
1086 }
1087
1088 /*
1089 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1090 * expect a lot of shrink+grow operations, just free and allocate
1091 * again like we would do for growing. If the pipe currently
1092 * contains more buffers than arg, then return busy.
1093 */
1094 if (nr_pages < pipe->nrbufs) {
1095 ret = -EBUSY;
1096 goto out_revert_acct;
1097 }
1098
1099 bufs = kcalloc(nr_pages, sizeof(*bufs),
1100 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1101 if (unlikely(!bufs)) {
1102 ret = -ENOMEM;
1103 goto out_revert_acct;
1104 }
1105
1106 /*
1107 * The pipe array wraps around, so just start the new one at zero
1108 * and adjust the indexes.
1109 */
1110 if (pipe->nrbufs) {
1111 unsigned int tail;
1112 unsigned int head;
1113
1114 tail = pipe->curbuf + pipe->nrbufs;
1115 if (tail < pipe->buffers)
1116 tail = 0;
1117 else
1118 tail &= (pipe->buffers - 1);
1119
1120 head = pipe->nrbufs - tail;
1121 if (head)
1122 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1123 if (tail)
1124 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1125 }
1126
1127 pipe->curbuf = 0;
1128 kfree(pipe->bufs);
1129 pipe->bufs = bufs;
1130 pipe->buffers = nr_pages;
1131 return nr_pages * PAGE_SIZE;
1132
1133out_revert_acct:
1134 (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1135 return ret;
1136}
1137
1138/*
1139 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1140 * location, so checking ->i_pipe is not enough to verify that this is a
1141 * pipe.
1142 */
1143struct pipe_inode_info *get_pipe_info(struct file *file)
1144{
1145 return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1146}
1147
1148long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1149{
1150 struct pipe_inode_info *pipe;
1151 long ret;
1152
1153 pipe = get_pipe_info(file);
1154 if (!pipe)
1155 return -EBADF;
1156
1157 __pipe_lock(pipe);
1158
1159 switch (cmd) {
1160 case F_SETPIPE_SZ:
1161 ret = pipe_set_size(pipe, arg);
1162 break;
1163 case F_GETPIPE_SZ:
1164 ret = pipe->buffers * PAGE_SIZE;
1165 break;
1166 default:
1167 ret = -EINVAL;
1168 break;
1169 }
1170
1171 __pipe_unlock(pipe);
1172 return ret;
1173}
1174
1175static const struct super_operations pipefs_ops = {
1176 .destroy_inode = free_inode_nonrcu,
1177 .statfs = simple_statfs,
1178};
1179
1180/*
1181 * pipefs should _never_ be mounted by userland - too much of security hassle,
1182 * no real gain from having the whole whorehouse mounted. So we don't need
1183 * any operations on the root directory. However, we need a non-trivial
1184 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1185 */
1186
1187static int pipefs_init_fs_context(struct fs_context *fc)
1188{
1189 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1190 if (!ctx)
1191 return -ENOMEM;
1192 ctx->ops = &pipefs_ops;
1193 ctx->dops = &pipefs_dentry_operations;
1194 return 0;
1195}
1196
1197static struct file_system_type pipe_fs_type = {
1198 .name = "pipefs",
1199 .init_fs_context = pipefs_init_fs_context,
1200 .kill_sb = kill_anon_super,
1201};
1202
1203static int __init init_pipe_fs(void)
1204{
1205 int err = register_filesystem(&pipe_fs_type);
1206
1207 if (!err) {
1208 pipe_mnt = kern_mount(&pipe_fs_type);
1209 if (IS_ERR(pipe_mnt)) {
1210 err = PTR_ERR(pipe_mnt);
1211 unregister_filesystem(&pipe_fs_type);
1212 }
1213 }
1214 return err;
1215}
1216
1217fs_initcall(init_pipe_fs);