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/magic.h>
18#include <linux/pipe_fs_i.h>
19#include <linux/uio.h>
20#include <linux/highmem.h>
21#include <linux/pagemap.h>
22#include <linux/audit.h>
23#include <linux/syscalls.h>
24#include <linux/fcntl.h>
25#include <linux/memcontrol.h>
26
27#include <linux/uaccess.h>
28#include <asm/ioctls.h>
29
30#include "internal.h"
31
32/*
33 * The max size that a non-root user is allowed to grow the pipe. Can
34 * be set by root in /proc/sys/fs/pipe-max-size
35 */
36unsigned int pipe_max_size = 1048576;
37
38/* Maximum allocatable pages per user. Hard limit is unset by default, soft
39 * matches default values.
40 */
41unsigned long pipe_user_pages_hard;
42unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
43
44/*
45 * We use a start+len construction, which provides full use of the
46 * allocated memory.
47 * -- Florian Coosmann (FGC)
48 *
49 * Reads with count = 0 should always return 0.
50 * -- Julian Bradfield 1999-06-07.
51 *
52 * FIFOs and Pipes now generate SIGIO for both readers and writers.
53 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
54 *
55 * pipe_read & write cleanup
56 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
57 */
58
59static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
60{
61 if (pipe->files)
62 mutex_lock_nested(&pipe->mutex, subclass);
63}
64
65void pipe_lock(struct pipe_inode_info *pipe)
66{
67 /*
68 * pipe_lock() nests non-pipe inode locks (for writing to a file)
69 */
70 pipe_lock_nested(pipe, I_MUTEX_PARENT);
71}
72EXPORT_SYMBOL(pipe_lock);
73
74void pipe_unlock(struct pipe_inode_info *pipe)
75{
76 if (pipe->files)
77 mutex_unlock(&pipe->mutex);
78}
79EXPORT_SYMBOL(pipe_unlock);
80
81static inline void __pipe_lock(struct pipe_inode_info *pipe)
82{
83 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
84}
85
86static inline void __pipe_unlock(struct pipe_inode_info *pipe)
87{
88 mutex_unlock(&pipe->mutex);
89}
90
91void pipe_double_lock(struct pipe_inode_info *pipe1,
92 struct pipe_inode_info *pipe2)
93{
94 BUG_ON(pipe1 == pipe2);
95
96 if (pipe1 < pipe2) {
97 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
98 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
99 } else {
100 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
101 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
102 }
103}
104
105/* Drop the inode semaphore and wait for a pipe event, atomically */
106void pipe_wait(struct pipe_inode_info *pipe)
107{
108 DEFINE_WAIT(wait);
109
110 /*
111 * Pipes are system-local resources, so sleeping on them
112 * is considered a noninteractive wait:
113 */
114 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
115 pipe_unlock(pipe);
116 schedule();
117 finish_wait(&pipe->wait, &wait);
118 pipe_lock(pipe);
119}
120
121static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
122 struct pipe_buffer *buf)
123{
124 struct page *page = buf->page;
125
126 /*
127 * If nobody else uses this page, and we don't already have a
128 * temporary page, let's keep track of it as a one-deep
129 * allocation cache. (Otherwise just release our reference to it)
130 */
131 if (page_count(page) == 1 && !pipe->tmp_page)
132 pipe->tmp_page = page;
133 else
134 put_page(page);
135}
136
137static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
138 struct pipe_buffer *buf)
139{
140 struct page *page = buf->page;
141
142 if (page_count(page) == 1) {
143 if (memcg_kmem_enabled())
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 */
192void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
193{
194 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
229static const struct pipe_buf_operations anon_pipe_buf_ops = {
230 .can_merge = 1,
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 packet_pipe_buf_ops = {
238 .can_merge = 0,
239 .confirm = generic_pipe_buf_confirm,
240 .release = anon_pipe_buf_release,
241 .steal = anon_pipe_buf_steal,
242 .get = generic_pipe_buf_get,
243};
244
245static ssize_t
246pipe_read(struct kiocb *iocb, struct iov_iter *to)
247{
248 size_t total_len = iov_iter_count(to);
249 struct file *filp = iocb->ki_filp;
250 struct pipe_inode_info *pipe = filp->private_data;
251 int do_wakeup;
252 ssize_t ret;
253
254 /* Null read succeeds. */
255 if (unlikely(total_len == 0))
256 return 0;
257
258 do_wakeup = 0;
259 ret = 0;
260 __pipe_lock(pipe);
261 for (;;) {
262 int bufs = pipe->nrbufs;
263 if (bufs) {
264 int curbuf = pipe->curbuf;
265 struct pipe_buffer *buf = pipe->bufs + curbuf;
266 size_t chars = buf->len;
267 size_t written;
268 int error;
269
270 if (chars > total_len)
271 chars = total_len;
272
273 error = pipe_buf_confirm(pipe, buf);
274 if (error) {
275 if (!ret)
276 ret = error;
277 break;
278 }
279
280 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
281 if (unlikely(written < chars)) {
282 if (!ret)
283 ret = -EFAULT;
284 break;
285 }
286 ret += chars;
287 buf->offset += chars;
288 buf->len -= chars;
289
290 /* Was it a packet buffer? Clean up and exit */
291 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
292 total_len = chars;
293 buf->len = 0;
294 }
295
296 if (!buf->len) {
297 pipe_buf_release(pipe, buf);
298 curbuf = (curbuf + 1) & (pipe->buffers - 1);
299 pipe->curbuf = curbuf;
300 pipe->nrbufs = --bufs;
301 do_wakeup = 1;
302 }
303 total_len -= chars;
304 if (!total_len)
305 break; /* common path: read succeeded */
306 }
307 if (bufs) /* More to do? */
308 continue;
309 if (!pipe->writers)
310 break;
311 if (!pipe->waiting_writers) {
312 /* syscall merging: Usually we must not sleep
313 * if O_NONBLOCK is set, or if we got some data.
314 * But if a writer sleeps in kernel space, then
315 * we can wait for that data without violating POSIX.
316 */
317 if (ret)
318 break;
319 if (filp->f_flags & O_NONBLOCK) {
320 ret = -EAGAIN;
321 break;
322 }
323 }
324 if (signal_pending(current)) {
325 if (!ret)
326 ret = -ERESTARTSYS;
327 break;
328 }
329 if (do_wakeup) {
330 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
331 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
332 }
333 pipe_wait(pipe);
334 }
335 __pipe_unlock(pipe);
336
337 /* Signal writers asynchronously that there is more room. */
338 if (do_wakeup) {
339 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
340 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
341 }
342 if (ret > 0)
343 file_accessed(filp);
344 return ret;
345}
346
347static inline int is_packetized(struct file *file)
348{
349 return (file->f_flags & O_DIRECT) != 0;
350}
351
352static ssize_t
353pipe_write(struct kiocb *iocb, struct iov_iter *from)
354{
355 struct file *filp = iocb->ki_filp;
356 struct pipe_inode_info *pipe = filp->private_data;
357 ssize_t ret = 0;
358 int do_wakeup = 0;
359 size_t total_len = iov_iter_count(from);
360 ssize_t chars;
361
362 /* Null write succeeds. */
363 if (unlikely(total_len == 0))
364 return 0;
365
366 __pipe_lock(pipe);
367
368 if (!pipe->readers) {
369 send_sig(SIGPIPE, current, 0);
370 ret = -EPIPE;
371 goto out;
372 }
373
374 /* We try to merge small writes */
375 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
376 if (pipe->nrbufs && chars != 0) {
377 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
378 (pipe->buffers - 1);
379 struct pipe_buffer *buf = pipe->bufs + lastbuf;
380 int offset = buf->offset + buf->len;
381
382 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
383 ret = pipe_buf_confirm(pipe, buf);
384 if (ret)
385 goto out;
386
387 ret = copy_page_from_iter(buf->page, offset, chars, from);
388 if (unlikely(ret < chars)) {
389 ret = -EFAULT;
390 goto out;
391 }
392 do_wakeup = 1;
393 buf->len += ret;
394 if (!iov_iter_count(from))
395 goto out;
396 }
397 }
398
399 for (;;) {
400 int bufs;
401
402 if (!pipe->readers) {
403 send_sig(SIGPIPE, current, 0);
404 if (!ret)
405 ret = -EPIPE;
406 break;
407 }
408 bufs = pipe->nrbufs;
409 if (bufs < pipe->buffers) {
410 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
411 struct pipe_buffer *buf = pipe->bufs + newbuf;
412 struct page *page = pipe->tmp_page;
413 int copied;
414
415 if (!page) {
416 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
417 if (unlikely(!page)) {
418 ret = ret ? : -ENOMEM;
419 break;
420 }
421 pipe->tmp_page = page;
422 }
423 /* Always wake up, even if the copy fails. Otherwise
424 * we lock up (O_NONBLOCK-)readers that sleep due to
425 * syscall merging.
426 * FIXME! Is this really true?
427 */
428 do_wakeup = 1;
429 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
430 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
431 if (!ret)
432 ret = -EFAULT;
433 break;
434 }
435 ret += copied;
436
437 /* Insert it into the buffer array */
438 buf->page = page;
439 buf->ops = &anon_pipe_buf_ops;
440 buf->offset = 0;
441 buf->len = copied;
442 buf->flags = 0;
443 if (is_packetized(filp)) {
444 buf->ops = &packet_pipe_buf_ops;
445 buf->flags = PIPE_BUF_FLAG_PACKET;
446 }
447 pipe->nrbufs = ++bufs;
448 pipe->tmp_page = NULL;
449
450 if (!iov_iter_count(from))
451 break;
452 }
453 if (bufs < pipe->buffers)
454 continue;
455 if (filp->f_flags & O_NONBLOCK) {
456 if (!ret)
457 ret = -EAGAIN;
458 break;
459 }
460 if (signal_pending(current)) {
461 if (!ret)
462 ret = -ERESTARTSYS;
463 break;
464 }
465 if (do_wakeup) {
466 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
467 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
468 do_wakeup = 0;
469 }
470 pipe->waiting_writers++;
471 pipe_wait(pipe);
472 pipe->waiting_writers--;
473 }
474out:
475 __pipe_unlock(pipe);
476 if (do_wakeup) {
477 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
478 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
479 }
480 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
481 int err = file_update_time(filp);
482 if (err)
483 ret = err;
484 sb_end_write(file_inode(filp)->i_sb);
485 }
486 return ret;
487}
488
489static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
490{
491 struct pipe_inode_info *pipe = filp->private_data;
492 int count, buf, nrbufs;
493
494 switch (cmd) {
495 case FIONREAD:
496 __pipe_lock(pipe);
497 count = 0;
498 buf = pipe->curbuf;
499 nrbufs = pipe->nrbufs;
500 while (--nrbufs >= 0) {
501 count += pipe->bufs[buf].len;
502 buf = (buf+1) & (pipe->buffers - 1);
503 }
504 __pipe_unlock(pipe);
505
506 return put_user(count, (int __user *)arg);
507 default:
508 return -ENOIOCTLCMD;
509 }
510}
511
512/* No kernel lock held - fine */
513static __poll_t
514pipe_poll(struct file *filp, poll_table *wait)
515{
516 __poll_t mask;
517 struct pipe_inode_info *pipe = filp->private_data;
518 int nrbufs;
519
520 poll_wait(filp, &pipe->wait, wait);
521
522 /* Reading only -- no need for acquiring the semaphore. */
523 nrbufs = pipe->nrbufs;
524 mask = 0;
525 if (filp->f_mode & FMODE_READ) {
526 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
527 if (!pipe->writers && filp->f_version != pipe->w_counter)
528 mask |= EPOLLHUP;
529 }
530
531 if (filp->f_mode & FMODE_WRITE) {
532 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
533 /*
534 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
535 * behave exactly like pipes for poll().
536 */
537 if (!pipe->readers)
538 mask |= EPOLLERR;
539 }
540
541 return mask;
542}
543
544static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
545{
546 int kill = 0;
547
548 spin_lock(&inode->i_lock);
549 if (!--pipe->files) {
550 inode->i_pipe = NULL;
551 kill = 1;
552 }
553 spin_unlock(&inode->i_lock);
554
555 if (kill)
556 free_pipe_info(pipe);
557}
558
559static int
560pipe_release(struct inode *inode, struct file *file)
561{
562 struct pipe_inode_info *pipe = file->private_data;
563
564 __pipe_lock(pipe);
565 if (file->f_mode & FMODE_READ)
566 pipe->readers--;
567 if (file->f_mode & FMODE_WRITE)
568 pipe->writers--;
569
570 if (pipe->readers || pipe->writers) {
571 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
572 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
573 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
574 }
575 __pipe_unlock(pipe);
576
577 put_pipe_info(inode, pipe);
578 return 0;
579}
580
581static int
582pipe_fasync(int fd, struct file *filp, int on)
583{
584 struct pipe_inode_info *pipe = filp->private_data;
585 int retval = 0;
586
587 __pipe_lock(pipe);
588 if (filp->f_mode & FMODE_READ)
589 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
590 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
591 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
592 if (retval < 0 && (filp->f_mode & FMODE_READ))
593 /* this can happen only if on == T */
594 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
595 }
596 __pipe_unlock(pipe);
597 return retval;
598}
599
600static unsigned long account_pipe_buffers(struct user_struct *user,
601 unsigned long old, unsigned long new)
602{
603 return atomic_long_add_return(new - old, &user->pipe_bufs);
604}
605
606static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
607{
608 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
609
610 return soft_limit && user_bufs > soft_limit;
611}
612
613static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
614{
615 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
616
617 return hard_limit && user_bufs > hard_limit;
618}
619
620static bool is_unprivileged_user(void)
621{
622 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
623}
624
625struct pipe_inode_info *alloc_pipe_info(void)
626{
627 struct pipe_inode_info *pipe;
628 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
629 struct user_struct *user = get_current_user();
630 unsigned long user_bufs;
631 unsigned int max_size = READ_ONCE(pipe_max_size);
632
633 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
634 if (pipe == NULL)
635 goto out_free_uid;
636
637 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
638 pipe_bufs = max_size >> PAGE_SHIFT;
639
640 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
641
642 if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
643 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
644 pipe_bufs = 1;
645 }
646
647 if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
648 goto out_revert_acct;
649
650 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
651 GFP_KERNEL_ACCOUNT);
652
653 if (pipe->bufs) {
654 init_waitqueue_head(&pipe->wait);
655 pipe->r_counter = pipe->w_counter = 1;
656 pipe->buffers = pipe_bufs;
657 pipe->user = user;
658 mutex_init(&pipe->mutex);
659 return pipe;
660 }
661
662out_revert_acct:
663 (void) account_pipe_buffers(user, pipe_bufs, 0);
664 kfree(pipe);
665out_free_uid:
666 free_uid(user);
667 return NULL;
668}
669
670void free_pipe_info(struct pipe_inode_info *pipe)
671{
672 int i;
673
674 (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
675 free_uid(pipe->user);
676 for (i = 0; i < pipe->buffers; i++) {
677 struct pipe_buffer *buf = pipe->bufs + i;
678 if (buf->ops)
679 pipe_buf_release(pipe, buf);
680 }
681 if (pipe->tmp_page)
682 __free_page(pipe->tmp_page);
683 kfree(pipe->bufs);
684 kfree(pipe);
685}
686
687static struct vfsmount *pipe_mnt __read_mostly;
688
689/*
690 * pipefs_dname() is called from d_path().
691 */
692static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
693{
694 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
695 d_inode(dentry)->i_ino);
696}
697
698static const struct dentry_operations pipefs_dentry_operations = {
699 .d_dname = pipefs_dname,
700};
701
702static struct inode * get_pipe_inode(void)
703{
704 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
705 struct pipe_inode_info *pipe;
706
707 if (!inode)
708 goto fail_inode;
709
710 inode->i_ino = get_next_ino();
711
712 pipe = alloc_pipe_info();
713 if (!pipe)
714 goto fail_iput;
715
716 inode->i_pipe = pipe;
717 pipe->files = 2;
718 pipe->readers = pipe->writers = 1;
719 inode->i_fop = &pipefifo_fops;
720
721 /*
722 * Mark the inode dirty from the very beginning,
723 * that way it will never be moved to the dirty
724 * list because "mark_inode_dirty()" will think
725 * that it already _is_ on the dirty list.
726 */
727 inode->i_state = I_DIRTY;
728 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
729 inode->i_uid = current_fsuid();
730 inode->i_gid = current_fsgid();
731 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
732
733 return inode;
734
735fail_iput:
736 iput(inode);
737
738fail_inode:
739 return NULL;
740}
741
742int create_pipe_files(struct file **res, int flags)
743{
744 int err;
745 struct inode *inode = get_pipe_inode();
746 struct file *f;
747 struct path path;
748
749 if (!inode)
750 return -ENFILE;
751
752 err = -ENOMEM;
753 path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &empty_name);
754 if (!path.dentry)
755 goto err_inode;
756 path.mnt = mntget(pipe_mnt);
757
758 d_instantiate(path.dentry, inode);
759
760 f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
761 if (IS_ERR(f)) {
762 err = PTR_ERR(f);
763 goto err_dentry;
764 }
765
766 f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
767 f->private_data = inode->i_pipe;
768
769 res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
770 if (IS_ERR(res[0])) {
771 err = PTR_ERR(res[0]);
772 goto err_file;
773 }
774
775 path_get(&path);
776 res[0]->private_data = inode->i_pipe;
777 res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
778 res[1] = f;
779 return 0;
780
781err_file:
782 put_filp(f);
783err_dentry:
784 free_pipe_info(inode->i_pipe);
785 path_put(&path);
786 return err;
787
788err_inode:
789 free_pipe_info(inode->i_pipe);
790 iput(inode);
791 return err;
792}
793
794static int __do_pipe_flags(int *fd, struct file **files, int flags)
795{
796 int error;
797 int fdw, fdr;
798
799 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
800 return -EINVAL;
801
802 error = create_pipe_files(files, flags);
803 if (error)
804 return error;
805
806 error = get_unused_fd_flags(flags);
807 if (error < 0)
808 goto err_read_pipe;
809 fdr = error;
810
811 error = get_unused_fd_flags(flags);
812 if (error < 0)
813 goto err_fdr;
814 fdw = error;
815
816 audit_fd_pair(fdr, fdw);
817 fd[0] = fdr;
818 fd[1] = fdw;
819 return 0;
820
821 err_fdr:
822 put_unused_fd(fdr);
823 err_read_pipe:
824 fput(files[0]);
825 fput(files[1]);
826 return error;
827}
828
829int do_pipe_flags(int *fd, int flags)
830{
831 struct file *files[2];
832 int error = __do_pipe_flags(fd, files, flags);
833 if (!error) {
834 fd_install(fd[0], files[0]);
835 fd_install(fd[1], files[1]);
836 }
837 return error;
838}
839
840/*
841 * sys_pipe() is the normal C calling standard for creating
842 * a pipe. It's not the way Unix traditionally does this, though.
843 */
844static int do_pipe2(int __user *fildes, int flags)
845{
846 struct file *files[2];
847 int fd[2];
848 int error;
849
850 error = __do_pipe_flags(fd, files, flags);
851 if (!error) {
852 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
853 fput(files[0]);
854 fput(files[1]);
855 put_unused_fd(fd[0]);
856 put_unused_fd(fd[1]);
857 error = -EFAULT;
858 } else {
859 fd_install(fd[0], files[0]);
860 fd_install(fd[1], files[1]);
861 }
862 }
863 return error;
864}
865
866SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
867{
868 return do_pipe2(fildes, flags);
869}
870
871SYSCALL_DEFINE1(pipe, int __user *, fildes)
872{
873 return do_pipe2(fildes, 0);
874}
875
876static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
877{
878 int cur = *cnt;
879
880 while (cur == *cnt) {
881 pipe_wait(pipe);
882 if (signal_pending(current))
883 break;
884 }
885 return cur == *cnt ? -ERESTARTSYS : 0;
886}
887
888static void wake_up_partner(struct pipe_inode_info *pipe)
889{
890 wake_up_interruptible(&pipe->wait);
891}
892
893static int fifo_open(struct inode *inode, struct file *filp)
894{
895 struct pipe_inode_info *pipe;
896 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
897 int ret;
898
899 filp->f_version = 0;
900
901 spin_lock(&inode->i_lock);
902 if (inode->i_pipe) {
903 pipe = inode->i_pipe;
904 pipe->files++;
905 spin_unlock(&inode->i_lock);
906 } else {
907 spin_unlock(&inode->i_lock);
908 pipe = alloc_pipe_info();
909 if (!pipe)
910 return -ENOMEM;
911 pipe->files = 1;
912 spin_lock(&inode->i_lock);
913 if (unlikely(inode->i_pipe)) {
914 inode->i_pipe->files++;
915 spin_unlock(&inode->i_lock);
916 free_pipe_info(pipe);
917 pipe = inode->i_pipe;
918 } else {
919 inode->i_pipe = pipe;
920 spin_unlock(&inode->i_lock);
921 }
922 }
923 filp->private_data = pipe;
924 /* OK, we have a pipe and it's pinned down */
925
926 __pipe_lock(pipe);
927
928 /* We can only do regular read/write on fifos */
929 filp->f_mode &= (FMODE_READ | FMODE_WRITE);
930
931 switch (filp->f_mode) {
932 case FMODE_READ:
933 /*
934 * O_RDONLY
935 * POSIX.1 says that O_NONBLOCK means return with the FIFO
936 * opened, even when there is no process writing the FIFO.
937 */
938 pipe->r_counter++;
939 if (pipe->readers++ == 0)
940 wake_up_partner(pipe);
941
942 if (!is_pipe && !pipe->writers) {
943 if ((filp->f_flags & O_NONBLOCK)) {
944 /* suppress EPOLLHUP until we have
945 * seen a writer */
946 filp->f_version = pipe->w_counter;
947 } else {
948 if (wait_for_partner(pipe, &pipe->w_counter))
949 goto err_rd;
950 }
951 }
952 break;
953
954 case FMODE_WRITE:
955 /*
956 * O_WRONLY
957 * POSIX.1 says that O_NONBLOCK means return -1 with
958 * errno=ENXIO when there is no process reading the FIFO.
959 */
960 ret = -ENXIO;
961 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
962 goto err;
963
964 pipe->w_counter++;
965 if (!pipe->writers++)
966 wake_up_partner(pipe);
967
968 if (!is_pipe && !pipe->readers) {
969 if (wait_for_partner(pipe, &pipe->r_counter))
970 goto err_wr;
971 }
972 break;
973
974 case FMODE_READ | FMODE_WRITE:
975 /*
976 * O_RDWR
977 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
978 * This implementation will NEVER block on a O_RDWR open, since
979 * the process can at least talk to itself.
980 */
981
982 pipe->readers++;
983 pipe->writers++;
984 pipe->r_counter++;
985 pipe->w_counter++;
986 if (pipe->readers == 1 || pipe->writers == 1)
987 wake_up_partner(pipe);
988 break;
989
990 default:
991 ret = -EINVAL;
992 goto err;
993 }
994
995 /* Ok! */
996 __pipe_unlock(pipe);
997 return 0;
998
999err_rd:
1000 if (!--pipe->readers)
1001 wake_up_interruptible(&pipe->wait);
1002 ret = -ERESTARTSYS;
1003 goto err;
1004
1005err_wr:
1006 if (!--pipe->writers)
1007 wake_up_interruptible(&pipe->wait);
1008 ret = -ERESTARTSYS;
1009 goto err;
1010
1011err:
1012 __pipe_unlock(pipe);
1013
1014 put_pipe_info(inode, pipe);
1015 return ret;
1016}
1017
1018const struct file_operations pipefifo_fops = {
1019 .open = fifo_open,
1020 .llseek = no_llseek,
1021 .read_iter = pipe_read,
1022 .write_iter = pipe_write,
1023 .poll = pipe_poll,
1024 .unlocked_ioctl = pipe_ioctl,
1025 .release = pipe_release,
1026 .fasync = pipe_fasync,
1027};
1028
1029/*
1030 * Currently we rely on the pipe array holding a power-of-2 number
1031 * of pages. Returns 0 on error.
1032 */
1033unsigned int round_pipe_size(unsigned long size)
1034{
1035 if (size > (1U << 31))
1036 return 0;
1037
1038 /* Minimum pipe size, as required by POSIX */
1039 if (size < PAGE_SIZE)
1040 return PAGE_SIZE;
1041
1042 return roundup_pow_of_two(size);
1043}
1044
1045/*
1046 * Allocate a new array of pipe buffers and copy the info over. Returns the
1047 * pipe size if successful, or return -ERROR on error.
1048 */
1049static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1050{
1051 struct pipe_buffer *bufs;
1052 unsigned int size, nr_pages;
1053 unsigned long user_bufs;
1054 long ret = 0;
1055
1056 size = round_pipe_size(arg);
1057 nr_pages = size >> PAGE_SHIFT;
1058
1059 if (!nr_pages)
1060 return -EINVAL;
1061
1062 /*
1063 * If trying to increase the pipe capacity, check that an
1064 * unprivileged user is not trying to exceed various limits
1065 * (soft limit check here, hard limit check just below).
1066 * Decreasing the pipe capacity is always permitted, even
1067 * if the user is currently over a limit.
1068 */
1069 if (nr_pages > pipe->buffers &&
1070 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1071 return -EPERM;
1072
1073 user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1074
1075 if (nr_pages > pipe->buffers &&
1076 (too_many_pipe_buffers_hard(user_bufs) ||
1077 too_many_pipe_buffers_soft(user_bufs)) &&
1078 is_unprivileged_user()) {
1079 ret = -EPERM;
1080 goto out_revert_acct;
1081 }
1082
1083 /*
1084 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1085 * expect a lot of shrink+grow operations, just free and allocate
1086 * again like we would do for growing. If the pipe currently
1087 * contains more buffers than arg, then return busy.
1088 */
1089 if (nr_pages < pipe->nrbufs) {
1090 ret = -EBUSY;
1091 goto out_revert_acct;
1092 }
1093
1094 bufs = kcalloc(nr_pages, sizeof(*bufs),
1095 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1096 if (unlikely(!bufs)) {
1097 ret = -ENOMEM;
1098 goto out_revert_acct;
1099 }
1100
1101 /*
1102 * The pipe array wraps around, so just start the new one at zero
1103 * and adjust the indexes.
1104 */
1105 if (pipe->nrbufs) {
1106 unsigned int tail;
1107 unsigned int head;
1108
1109 tail = pipe->curbuf + pipe->nrbufs;
1110 if (tail < pipe->buffers)
1111 tail = 0;
1112 else
1113 tail &= (pipe->buffers - 1);
1114
1115 head = pipe->nrbufs - tail;
1116 if (head)
1117 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1118 if (tail)
1119 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1120 }
1121
1122 pipe->curbuf = 0;
1123 kfree(pipe->bufs);
1124 pipe->bufs = bufs;
1125 pipe->buffers = nr_pages;
1126 return nr_pages * PAGE_SIZE;
1127
1128out_revert_acct:
1129 (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1130 return ret;
1131}
1132
1133/*
1134 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1135 * location, so checking ->i_pipe is not enough to verify that this is a
1136 * pipe.
1137 */
1138struct pipe_inode_info *get_pipe_info(struct file *file)
1139{
1140 return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1141}
1142
1143long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1144{
1145 struct pipe_inode_info *pipe;
1146 long ret;
1147
1148 pipe = get_pipe_info(file);
1149 if (!pipe)
1150 return -EBADF;
1151
1152 __pipe_lock(pipe);
1153
1154 switch (cmd) {
1155 case F_SETPIPE_SZ:
1156 ret = pipe_set_size(pipe, arg);
1157 break;
1158 case F_GETPIPE_SZ:
1159 ret = pipe->buffers * PAGE_SIZE;
1160 break;
1161 default:
1162 ret = -EINVAL;
1163 break;
1164 }
1165
1166 __pipe_unlock(pipe);
1167 return ret;
1168}
1169
1170static const struct super_operations pipefs_ops = {
1171 .destroy_inode = free_inode_nonrcu,
1172 .statfs = simple_statfs,
1173};
1174
1175/*
1176 * pipefs should _never_ be mounted by userland - too much of security hassle,
1177 * no real gain from having the whole whorehouse mounted. So we don't need
1178 * any operations on the root directory. However, we need a non-trivial
1179 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1180 */
1181static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1182 int flags, const char *dev_name, void *data)
1183{
1184 return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1185 &pipefs_dentry_operations, PIPEFS_MAGIC);
1186}
1187
1188static struct file_system_type pipe_fs_type = {
1189 .name = "pipefs",
1190 .mount = pipefs_mount,
1191 .kill_sb = kill_anon_super,
1192};
1193
1194static int __init init_pipe_fs(void)
1195{
1196 int err = register_filesystem(&pipe_fs_type);
1197
1198 if (!err) {
1199 pipe_mnt = kern_mount(&pipe_fs_type);
1200 if (IS_ERR(pipe_mnt)) {
1201 err = PTR_ERR(pipe_mnt);
1202 unregister_filesystem(&pipe_fs_type);
1203 }
1204 }
1205 return err;
1206}
1207
1208fs_initcall(init_pipe_fs);