Loading...
1/*
2 * linux/fs/pipe.c
3 *
4 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
5 */
6
7#include <linux/mm.h>
8#include <linux/file.h>
9#include <linux/poll.h>
10#include <linux/slab.h>
11#include <linux/module.h>
12#include <linux/init.h>
13#include <linux/fs.h>
14#include <linux/log2.h>
15#include <linux/mount.h>
16#include <linux/magic.h>
17#include <linux/pipe_fs_i.h>
18#include <linux/uio.h>
19#include <linux/highmem.h>
20#include <linux/pagemap.h>
21#include <linux/audit.h>
22#include <linux/syscalls.h>
23#include <linux/fcntl.h>
24
25#include <asm/uaccess.h>
26#include <asm/ioctls.h>
27
28#include "internal.h"
29
30/*
31 * The max size that a non-root user is allowed to grow the pipe. Can
32 * be set by root in /proc/sys/fs/pipe-max-size
33 */
34unsigned int pipe_max_size = 1048576;
35
36/*
37 * Minimum pipe size, as required by POSIX
38 */
39unsigned int pipe_min_size = PAGE_SIZE;
40
41/* Maximum allocatable pages per user. Hard limit is unset by default, soft
42 * matches default values.
43 */
44unsigned long pipe_user_pages_hard;
45unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
46
47/*
48 * We use a start+len construction, which provides full use of the
49 * allocated memory.
50 * -- Florian Coosmann (FGC)
51 *
52 * Reads with count = 0 should always return 0.
53 * -- Julian Bradfield 1999-06-07.
54 *
55 * FIFOs and Pipes now generate SIGIO for both readers and writers.
56 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
57 *
58 * pipe_read & write cleanup
59 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
60 */
61
62static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
63{
64 if (pipe->files)
65 mutex_lock_nested(&pipe->mutex, subclass);
66}
67
68void pipe_lock(struct pipe_inode_info *pipe)
69{
70 /*
71 * pipe_lock() nests non-pipe inode locks (for writing to a file)
72 */
73 pipe_lock_nested(pipe, I_MUTEX_PARENT);
74}
75EXPORT_SYMBOL(pipe_lock);
76
77void pipe_unlock(struct pipe_inode_info *pipe)
78{
79 if (pipe->files)
80 mutex_unlock(&pipe->mutex);
81}
82EXPORT_SYMBOL(pipe_unlock);
83
84static inline void __pipe_lock(struct pipe_inode_info *pipe)
85{
86 mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
87}
88
89static inline void __pipe_unlock(struct pipe_inode_info *pipe)
90{
91 mutex_unlock(&pipe->mutex);
92}
93
94void pipe_double_lock(struct pipe_inode_info *pipe1,
95 struct pipe_inode_info *pipe2)
96{
97 BUG_ON(pipe1 == pipe2);
98
99 if (pipe1 < pipe2) {
100 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
101 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
102 } else {
103 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
104 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
105 }
106}
107
108/* Drop the inode semaphore and wait for a pipe event, atomically */
109void pipe_wait(struct pipe_inode_info *pipe)
110{
111 DEFINE_WAIT(wait);
112
113 /*
114 * Pipes are system-local resources, so sleeping on them
115 * is considered a noninteractive wait:
116 */
117 prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
118 pipe_unlock(pipe);
119 schedule();
120 finish_wait(&pipe->wait, &wait);
121 pipe_lock(pipe);
122}
123
124static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
125 struct pipe_buffer *buf)
126{
127 struct page *page = buf->page;
128
129 /*
130 * If nobody else uses this page, and we don't already have a
131 * temporary page, let's keep track of it as a one-deep
132 * allocation cache. (Otherwise just release our reference to it)
133 */
134 if (page_count(page) == 1 && !pipe->tmp_page)
135 pipe->tmp_page = page;
136 else
137 put_page(page);
138}
139
140/**
141 * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
142 * @pipe: the pipe that the buffer belongs to
143 * @buf: the buffer to attempt to steal
144 *
145 * Description:
146 * This function attempts to steal the &struct page attached to
147 * @buf. If successful, this function returns 0 and returns with
148 * the page locked. The caller may then reuse the page for whatever
149 * he wishes; the typical use is insertion into a different file
150 * page cache.
151 */
152int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
153 struct pipe_buffer *buf)
154{
155 struct page *page = buf->page;
156
157 /*
158 * A reference of one is golden, that means that the owner of this
159 * page is the only one holding a reference to it. lock the page
160 * and return OK.
161 */
162 if (page_count(page) == 1) {
163 lock_page(page);
164 return 0;
165 }
166
167 return 1;
168}
169EXPORT_SYMBOL(generic_pipe_buf_steal);
170
171/**
172 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
173 * @pipe: the pipe that the buffer belongs to
174 * @buf: the buffer to get a reference to
175 *
176 * Description:
177 * This function grabs an extra reference to @buf. It's used in
178 * in the tee() system call, when we duplicate the buffers in one
179 * pipe into another.
180 */
181void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
182{
183 get_page(buf->page);
184}
185EXPORT_SYMBOL(generic_pipe_buf_get);
186
187/**
188 * generic_pipe_buf_confirm - verify contents of the pipe buffer
189 * @info: the pipe that the buffer belongs to
190 * @buf: the buffer to confirm
191 *
192 * Description:
193 * This function does nothing, because the generic pipe code uses
194 * pages that are always good when inserted into the pipe.
195 */
196int generic_pipe_buf_confirm(struct pipe_inode_info *info,
197 struct pipe_buffer *buf)
198{
199 return 0;
200}
201EXPORT_SYMBOL(generic_pipe_buf_confirm);
202
203/**
204 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
205 * @pipe: the pipe that the buffer belongs to
206 * @buf: the buffer to put a reference to
207 *
208 * Description:
209 * This function releases a reference to @buf.
210 */
211void generic_pipe_buf_release(struct pipe_inode_info *pipe,
212 struct pipe_buffer *buf)
213{
214 put_page(buf->page);
215}
216EXPORT_SYMBOL(generic_pipe_buf_release);
217
218static const struct pipe_buf_operations anon_pipe_buf_ops = {
219 .can_merge = 1,
220 .confirm = generic_pipe_buf_confirm,
221 .release = anon_pipe_buf_release,
222 .steal = generic_pipe_buf_steal,
223 .get = generic_pipe_buf_get,
224};
225
226static const struct pipe_buf_operations packet_pipe_buf_ops = {
227 .can_merge = 0,
228 .confirm = generic_pipe_buf_confirm,
229 .release = anon_pipe_buf_release,
230 .steal = generic_pipe_buf_steal,
231 .get = generic_pipe_buf_get,
232};
233
234static ssize_t
235pipe_read(struct kiocb *iocb, struct iov_iter *to)
236{
237 size_t total_len = iov_iter_count(to);
238 struct file *filp = iocb->ki_filp;
239 struct pipe_inode_info *pipe = filp->private_data;
240 int do_wakeup;
241 ssize_t ret;
242
243 /* Null read succeeds. */
244 if (unlikely(total_len == 0))
245 return 0;
246
247 do_wakeup = 0;
248 ret = 0;
249 __pipe_lock(pipe);
250 for (;;) {
251 int bufs = pipe->nrbufs;
252 if (bufs) {
253 int curbuf = pipe->curbuf;
254 struct pipe_buffer *buf = pipe->bufs + curbuf;
255 const struct pipe_buf_operations *ops = buf->ops;
256 size_t chars = buf->len;
257 size_t written;
258 int error;
259
260 if (chars > total_len)
261 chars = total_len;
262
263 error = ops->confirm(pipe, buf);
264 if (error) {
265 if (!ret)
266 ret = error;
267 break;
268 }
269
270 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
271 if (unlikely(written < chars)) {
272 if (!ret)
273 ret = -EFAULT;
274 break;
275 }
276 ret += chars;
277 buf->offset += chars;
278 buf->len -= chars;
279
280 /* Was it a packet buffer? Clean up and exit */
281 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
282 total_len = chars;
283 buf->len = 0;
284 }
285
286 if (!buf->len) {
287 buf->ops = NULL;
288 ops->release(pipe, buf);
289 curbuf = (curbuf + 1) & (pipe->buffers - 1);
290 pipe->curbuf = curbuf;
291 pipe->nrbufs = --bufs;
292 do_wakeup = 1;
293 }
294 total_len -= chars;
295 if (!total_len)
296 break; /* common path: read succeeded */
297 }
298 if (bufs) /* More to do? */
299 continue;
300 if (!pipe->writers)
301 break;
302 if (!pipe->waiting_writers) {
303 /* syscall merging: Usually we must not sleep
304 * if O_NONBLOCK is set, or if we got some data.
305 * But if a writer sleeps in kernel space, then
306 * we can wait for that data without violating POSIX.
307 */
308 if (ret)
309 break;
310 if (filp->f_flags & O_NONBLOCK) {
311 ret = -EAGAIN;
312 break;
313 }
314 }
315 if (signal_pending(current)) {
316 if (!ret)
317 ret = -ERESTARTSYS;
318 break;
319 }
320 if (do_wakeup) {
321 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
322 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
323 }
324 pipe_wait(pipe);
325 }
326 __pipe_unlock(pipe);
327
328 /* Signal writers asynchronously that there is more room. */
329 if (do_wakeup) {
330 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
331 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
332 }
333 if (ret > 0)
334 file_accessed(filp);
335 return ret;
336}
337
338static inline int is_packetized(struct file *file)
339{
340 return (file->f_flags & O_DIRECT) != 0;
341}
342
343static ssize_t
344pipe_write(struct kiocb *iocb, struct iov_iter *from)
345{
346 struct file *filp = iocb->ki_filp;
347 struct pipe_inode_info *pipe = filp->private_data;
348 ssize_t ret = 0;
349 int do_wakeup = 0;
350 size_t total_len = iov_iter_count(from);
351 ssize_t chars;
352
353 /* Null write succeeds. */
354 if (unlikely(total_len == 0))
355 return 0;
356
357 __pipe_lock(pipe);
358
359 if (!pipe->readers) {
360 send_sig(SIGPIPE, current, 0);
361 ret = -EPIPE;
362 goto out;
363 }
364
365 /* We try to merge small writes */
366 chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
367 if (pipe->nrbufs && chars != 0) {
368 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
369 (pipe->buffers - 1);
370 struct pipe_buffer *buf = pipe->bufs + lastbuf;
371 const struct pipe_buf_operations *ops = buf->ops;
372 int offset = buf->offset + buf->len;
373
374 if (ops->can_merge && offset + chars <= PAGE_SIZE) {
375 ret = ops->confirm(pipe, buf);
376 if (ret)
377 goto out;
378
379 ret = copy_page_from_iter(buf->page, offset, chars, from);
380 if (unlikely(ret < chars)) {
381 ret = -EFAULT;
382 goto out;
383 }
384 do_wakeup = 1;
385 buf->len += ret;
386 if (!iov_iter_count(from))
387 goto out;
388 }
389 }
390
391 for (;;) {
392 int bufs;
393
394 if (!pipe->readers) {
395 send_sig(SIGPIPE, current, 0);
396 if (!ret)
397 ret = -EPIPE;
398 break;
399 }
400 bufs = pipe->nrbufs;
401 if (bufs < pipe->buffers) {
402 int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
403 struct pipe_buffer *buf = pipe->bufs + newbuf;
404 struct page *page = pipe->tmp_page;
405 int copied;
406
407 if (!page) {
408 page = alloc_page(GFP_HIGHUSER);
409 if (unlikely(!page)) {
410 ret = ret ? : -ENOMEM;
411 break;
412 }
413 pipe->tmp_page = page;
414 }
415 /* Always wake up, even if the copy fails. Otherwise
416 * we lock up (O_NONBLOCK-)readers that sleep due to
417 * syscall merging.
418 * FIXME! Is this really true?
419 */
420 do_wakeup = 1;
421 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
422 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
423 if (!ret)
424 ret = -EFAULT;
425 break;
426 }
427 ret += copied;
428
429 /* Insert it into the buffer array */
430 buf->page = page;
431 buf->ops = &anon_pipe_buf_ops;
432 buf->offset = 0;
433 buf->len = copied;
434 buf->flags = 0;
435 if (is_packetized(filp)) {
436 buf->ops = &packet_pipe_buf_ops;
437 buf->flags = PIPE_BUF_FLAG_PACKET;
438 }
439 pipe->nrbufs = ++bufs;
440 pipe->tmp_page = NULL;
441
442 if (!iov_iter_count(from))
443 break;
444 }
445 if (bufs < pipe->buffers)
446 continue;
447 if (filp->f_flags & O_NONBLOCK) {
448 if (!ret)
449 ret = -EAGAIN;
450 break;
451 }
452 if (signal_pending(current)) {
453 if (!ret)
454 ret = -ERESTARTSYS;
455 break;
456 }
457 if (do_wakeup) {
458 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
459 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
460 do_wakeup = 0;
461 }
462 pipe->waiting_writers++;
463 pipe_wait(pipe);
464 pipe->waiting_writers--;
465 }
466out:
467 __pipe_unlock(pipe);
468 if (do_wakeup) {
469 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
470 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
471 }
472 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
473 int err = file_update_time(filp);
474 if (err)
475 ret = err;
476 sb_end_write(file_inode(filp)->i_sb);
477 }
478 return ret;
479}
480
481static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
482{
483 struct pipe_inode_info *pipe = filp->private_data;
484 int count, buf, nrbufs;
485
486 switch (cmd) {
487 case FIONREAD:
488 __pipe_lock(pipe);
489 count = 0;
490 buf = pipe->curbuf;
491 nrbufs = pipe->nrbufs;
492 while (--nrbufs >= 0) {
493 count += pipe->bufs[buf].len;
494 buf = (buf+1) & (pipe->buffers - 1);
495 }
496 __pipe_unlock(pipe);
497
498 return put_user(count, (int __user *)arg);
499 default:
500 return -ENOIOCTLCMD;
501 }
502}
503
504/* No kernel lock held - fine */
505static unsigned int
506pipe_poll(struct file *filp, poll_table *wait)
507{
508 unsigned int mask;
509 struct pipe_inode_info *pipe = filp->private_data;
510 int nrbufs;
511
512 poll_wait(filp, &pipe->wait, wait);
513
514 /* Reading only -- no need for acquiring the semaphore. */
515 nrbufs = pipe->nrbufs;
516 mask = 0;
517 if (filp->f_mode & FMODE_READ) {
518 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
519 if (!pipe->writers && filp->f_version != pipe->w_counter)
520 mask |= POLLHUP;
521 }
522
523 if (filp->f_mode & FMODE_WRITE) {
524 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
525 /*
526 * Most Unices do not set POLLERR for FIFOs but on Linux they
527 * behave exactly like pipes for poll().
528 */
529 if (!pipe->readers)
530 mask |= POLLERR;
531 }
532
533 return mask;
534}
535
536static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
537{
538 int kill = 0;
539
540 spin_lock(&inode->i_lock);
541 if (!--pipe->files) {
542 inode->i_pipe = NULL;
543 kill = 1;
544 }
545 spin_unlock(&inode->i_lock);
546
547 if (kill)
548 free_pipe_info(pipe);
549}
550
551static int
552pipe_release(struct inode *inode, struct file *file)
553{
554 struct pipe_inode_info *pipe = file->private_data;
555
556 __pipe_lock(pipe);
557 if (file->f_mode & FMODE_READ)
558 pipe->readers--;
559 if (file->f_mode & FMODE_WRITE)
560 pipe->writers--;
561
562 if (pipe->readers || pipe->writers) {
563 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
564 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
565 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
566 }
567 __pipe_unlock(pipe);
568
569 put_pipe_info(inode, pipe);
570 return 0;
571}
572
573static int
574pipe_fasync(int fd, struct file *filp, int on)
575{
576 struct pipe_inode_info *pipe = filp->private_data;
577 int retval = 0;
578
579 __pipe_lock(pipe);
580 if (filp->f_mode & FMODE_READ)
581 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
582 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
583 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
584 if (retval < 0 && (filp->f_mode & FMODE_READ))
585 /* this can happen only if on == T */
586 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
587 }
588 __pipe_unlock(pipe);
589 return retval;
590}
591
592static void account_pipe_buffers(struct pipe_inode_info *pipe,
593 unsigned long old, unsigned long new)
594{
595 atomic_long_add(new - old, &pipe->user->pipe_bufs);
596}
597
598static bool too_many_pipe_buffers_soft(struct user_struct *user)
599{
600 return pipe_user_pages_soft &&
601 atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_soft;
602}
603
604static bool too_many_pipe_buffers_hard(struct user_struct *user)
605{
606 return pipe_user_pages_hard &&
607 atomic_long_read(&user->pipe_bufs) >= pipe_user_pages_hard;
608}
609
610struct pipe_inode_info *alloc_pipe_info(void)
611{
612 struct pipe_inode_info *pipe;
613
614 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
615 if (pipe) {
616 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
617 struct user_struct *user = get_current_user();
618
619 if (!too_many_pipe_buffers_hard(user)) {
620 if (too_many_pipe_buffers_soft(user))
621 pipe_bufs = 1;
622 pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * pipe_bufs, GFP_KERNEL);
623 }
624
625 if (pipe->bufs) {
626 init_waitqueue_head(&pipe->wait);
627 pipe->r_counter = pipe->w_counter = 1;
628 pipe->buffers = pipe_bufs;
629 pipe->user = user;
630 account_pipe_buffers(pipe, 0, pipe_bufs);
631 mutex_init(&pipe->mutex);
632 return pipe;
633 }
634 free_uid(user);
635 kfree(pipe);
636 }
637
638 return NULL;
639}
640
641void free_pipe_info(struct pipe_inode_info *pipe)
642{
643 int i;
644
645 account_pipe_buffers(pipe, pipe->buffers, 0);
646 free_uid(pipe->user);
647 for (i = 0; i < pipe->buffers; i++) {
648 struct pipe_buffer *buf = pipe->bufs + i;
649 if (buf->ops)
650 buf->ops->release(pipe, buf);
651 }
652 if (pipe->tmp_page)
653 __free_page(pipe->tmp_page);
654 kfree(pipe->bufs);
655 kfree(pipe);
656}
657
658static struct vfsmount *pipe_mnt __read_mostly;
659
660/*
661 * pipefs_dname() is called from d_path().
662 */
663static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
664{
665 return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
666 d_inode(dentry)->i_ino);
667}
668
669static const struct dentry_operations pipefs_dentry_operations = {
670 .d_dname = pipefs_dname,
671};
672
673static struct inode * get_pipe_inode(void)
674{
675 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
676 struct pipe_inode_info *pipe;
677
678 if (!inode)
679 goto fail_inode;
680
681 inode->i_ino = get_next_ino();
682
683 pipe = alloc_pipe_info();
684 if (!pipe)
685 goto fail_iput;
686
687 inode->i_pipe = pipe;
688 pipe->files = 2;
689 pipe->readers = pipe->writers = 1;
690 inode->i_fop = &pipefifo_fops;
691
692 /*
693 * Mark the inode dirty from the very beginning,
694 * that way it will never be moved to the dirty
695 * list because "mark_inode_dirty()" will think
696 * that it already _is_ on the dirty list.
697 */
698 inode->i_state = I_DIRTY;
699 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
700 inode->i_uid = current_fsuid();
701 inode->i_gid = current_fsgid();
702 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
703
704 return inode;
705
706fail_iput:
707 iput(inode);
708
709fail_inode:
710 return NULL;
711}
712
713int create_pipe_files(struct file **res, int flags)
714{
715 int err;
716 struct inode *inode = get_pipe_inode();
717 struct file *f;
718 struct path path;
719 static struct qstr name = { .name = "" };
720
721 if (!inode)
722 return -ENFILE;
723
724 err = -ENOMEM;
725 path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
726 if (!path.dentry)
727 goto err_inode;
728 path.mnt = mntget(pipe_mnt);
729
730 d_instantiate(path.dentry, inode);
731
732 f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
733 if (IS_ERR(f)) {
734 err = PTR_ERR(f);
735 goto err_dentry;
736 }
737
738 f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
739 f->private_data = inode->i_pipe;
740
741 res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
742 if (IS_ERR(res[0])) {
743 err = PTR_ERR(res[0]);
744 goto err_file;
745 }
746
747 path_get(&path);
748 res[0]->private_data = inode->i_pipe;
749 res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
750 res[1] = f;
751 return 0;
752
753err_file:
754 put_filp(f);
755err_dentry:
756 free_pipe_info(inode->i_pipe);
757 path_put(&path);
758 return err;
759
760err_inode:
761 free_pipe_info(inode->i_pipe);
762 iput(inode);
763 return err;
764}
765
766static int __do_pipe_flags(int *fd, struct file **files, int flags)
767{
768 int error;
769 int fdw, fdr;
770
771 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
772 return -EINVAL;
773
774 error = create_pipe_files(files, flags);
775 if (error)
776 return error;
777
778 error = get_unused_fd_flags(flags);
779 if (error < 0)
780 goto err_read_pipe;
781 fdr = error;
782
783 error = get_unused_fd_flags(flags);
784 if (error < 0)
785 goto err_fdr;
786 fdw = error;
787
788 audit_fd_pair(fdr, fdw);
789 fd[0] = fdr;
790 fd[1] = fdw;
791 return 0;
792
793 err_fdr:
794 put_unused_fd(fdr);
795 err_read_pipe:
796 fput(files[0]);
797 fput(files[1]);
798 return error;
799}
800
801int do_pipe_flags(int *fd, int flags)
802{
803 struct file *files[2];
804 int error = __do_pipe_flags(fd, files, flags);
805 if (!error) {
806 fd_install(fd[0], files[0]);
807 fd_install(fd[1], files[1]);
808 }
809 return error;
810}
811
812/*
813 * sys_pipe() is the normal C calling standard for creating
814 * a pipe. It's not the way Unix traditionally does this, though.
815 */
816SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
817{
818 struct file *files[2];
819 int fd[2];
820 int error;
821
822 error = __do_pipe_flags(fd, files, flags);
823 if (!error) {
824 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
825 fput(files[0]);
826 fput(files[1]);
827 put_unused_fd(fd[0]);
828 put_unused_fd(fd[1]);
829 error = -EFAULT;
830 } else {
831 fd_install(fd[0], files[0]);
832 fd_install(fd[1], files[1]);
833 }
834 }
835 return error;
836}
837
838SYSCALL_DEFINE1(pipe, int __user *, fildes)
839{
840 return sys_pipe2(fildes, 0);
841}
842
843static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
844{
845 int cur = *cnt;
846
847 while (cur == *cnt) {
848 pipe_wait(pipe);
849 if (signal_pending(current))
850 break;
851 }
852 return cur == *cnt ? -ERESTARTSYS : 0;
853}
854
855static void wake_up_partner(struct pipe_inode_info *pipe)
856{
857 wake_up_interruptible(&pipe->wait);
858}
859
860static int fifo_open(struct inode *inode, struct file *filp)
861{
862 struct pipe_inode_info *pipe;
863 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
864 int ret;
865
866 filp->f_version = 0;
867
868 spin_lock(&inode->i_lock);
869 if (inode->i_pipe) {
870 pipe = inode->i_pipe;
871 pipe->files++;
872 spin_unlock(&inode->i_lock);
873 } else {
874 spin_unlock(&inode->i_lock);
875 pipe = alloc_pipe_info();
876 if (!pipe)
877 return -ENOMEM;
878 pipe->files = 1;
879 spin_lock(&inode->i_lock);
880 if (unlikely(inode->i_pipe)) {
881 inode->i_pipe->files++;
882 spin_unlock(&inode->i_lock);
883 free_pipe_info(pipe);
884 pipe = inode->i_pipe;
885 } else {
886 inode->i_pipe = pipe;
887 spin_unlock(&inode->i_lock);
888 }
889 }
890 filp->private_data = pipe;
891 /* OK, we have a pipe and it's pinned down */
892
893 __pipe_lock(pipe);
894
895 /* We can only do regular read/write on fifos */
896 filp->f_mode &= (FMODE_READ | FMODE_WRITE);
897
898 switch (filp->f_mode) {
899 case FMODE_READ:
900 /*
901 * O_RDONLY
902 * POSIX.1 says that O_NONBLOCK means return with the FIFO
903 * opened, even when there is no process writing the FIFO.
904 */
905 pipe->r_counter++;
906 if (pipe->readers++ == 0)
907 wake_up_partner(pipe);
908
909 if (!is_pipe && !pipe->writers) {
910 if ((filp->f_flags & O_NONBLOCK)) {
911 /* suppress POLLHUP until we have
912 * seen a writer */
913 filp->f_version = pipe->w_counter;
914 } else {
915 if (wait_for_partner(pipe, &pipe->w_counter))
916 goto err_rd;
917 }
918 }
919 break;
920
921 case FMODE_WRITE:
922 /*
923 * O_WRONLY
924 * POSIX.1 says that O_NONBLOCK means return -1 with
925 * errno=ENXIO when there is no process reading the FIFO.
926 */
927 ret = -ENXIO;
928 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
929 goto err;
930
931 pipe->w_counter++;
932 if (!pipe->writers++)
933 wake_up_partner(pipe);
934
935 if (!is_pipe && !pipe->readers) {
936 if (wait_for_partner(pipe, &pipe->r_counter))
937 goto err_wr;
938 }
939 break;
940
941 case FMODE_READ | FMODE_WRITE:
942 /*
943 * O_RDWR
944 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
945 * This implementation will NEVER block on a O_RDWR open, since
946 * the process can at least talk to itself.
947 */
948
949 pipe->readers++;
950 pipe->writers++;
951 pipe->r_counter++;
952 pipe->w_counter++;
953 if (pipe->readers == 1 || pipe->writers == 1)
954 wake_up_partner(pipe);
955 break;
956
957 default:
958 ret = -EINVAL;
959 goto err;
960 }
961
962 /* Ok! */
963 __pipe_unlock(pipe);
964 return 0;
965
966err_rd:
967 if (!--pipe->readers)
968 wake_up_interruptible(&pipe->wait);
969 ret = -ERESTARTSYS;
970 goto err;
971
972err_wr:
973 if (!--pipe->writers)
974 wake_up_interruptible(&pipe->wait);
975 ret = -ERESTARTSYS;
976 goto err;
977
978err:
979 __pipe_unlock(pipe);
980
981 put_pipe_info(inode, pipe);
982 return ret;
983}
984
985const struct file_operations pipefifo_fops = {
986 .open = fifo_open,
987 .llseek = no_llseek,
988 .read_iter = pipe_read,
989 .write_iter = pipe_write,
990 .poll = pipe_poll,
991 .unlocked_ioctl = pipe_ioctl,
992 .release = pipe_release,
993 .fasync = pipe_fasync,
994};
995
996/*
997 * Allocate a new array of pipe buffers and copy the info over. Returns the
998 * pipe size if successful, or return -ERROR on error.
999 */
1000static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1001{
1002 struct pipe_buffer *bufs;
1003
1004 /*
1005 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1006 * expect a lot of shrink+grow operations, just free and allocate
1007 * again like we would do for growing. If the pipe currently
1008 * contains more buffers than arg, then return busy.
1009 */
1010 if (nr_pages < pipe->nrbufs)
1011 return -EBUSY;
1012
1013 bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1014 if (unlikely(!bufs))
1015 return -ENOMEM;
1016
1017 /*
1018 * The pipe array wraps around, so just start the new one at zero
1019 * and adjust the indexes.
1020 */
1021 if (pipe->nrbufs) {
1022 unsigned int tail;
1023 unsigned int head;
1024
1025 tail = pipe->curbuf + pipe->nrbufs;
1026 if (tail < pipe->buffers)
1027 tail = 0;
1028 else
1029 tail &= (pipe->buffers - 1);
1030
1031 head = pipe->nrbufs - tail;
1032 if (head)
1033 memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1034 if (tail)
1035 memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1036 }
1037
1038 account_pipe_buffers(pipe, pipe->buffers, nr_pages);
1039 pipe->curbuf = 0;
1040 kfree(pipe->bufs);
1041 pipe->bufs = bufs;
1042 pipe->buffers = nr_pages;
1043 return nr_pages * PAGE_SIZE;
1044}
1045
1046/*
1047 * Currently we rely on the pipe array holding a power-of-2 number
1048 * of pages.
1049 */
1050static inline unsigned int round_pipe_size(unsigned int size)
1051{
1052 unsigned long nr_pages;
1053
1054 nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1055 return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1056}
1057
1058/*
1059 * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1060 * will return an error.
1061 */
1062int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1063 size_t *lenp, loff_t *ppos)
1064{
1065 int ret;
1066
1067 ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1068 if (ret < 0 || !write)
1069 return ret;
1070
1071 pipe_max_size = round_pipe_size(pipe_max_size);
1072 return ret;
1073}
1074
1075/*
1076 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1077 * location, so checking ->i_pipe is not enough to verify that this is a
1078 * pipe.
1079 */
1080struct pipe_inode_info *get_pipe_info(struct file *file)
1081{
1082 return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1083}
1084
1085long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1086{
1087 struct pipe_inode_info *pipe;
1088 long ret;
1089
1090 pipe = get_pipe_info(file);
1091 if (!pipe)
1092 return -EBADF;
1093
1094 __pipe_lock(pipe);
1095
1096 switch (cmd) {
1097 case F_SETPIPE_SZ: {
1098 unsigned int size, nr_pages;
1099
1100 size = round_pipe_size(arg);
1101 nr_pages = size >> PAGE_SHIFT;
1102
1103 ret = -EINVAL;
1104 if (!nr_pages)
1105 goto out;
1106
1107 if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1108 ret = -EPERM;
1109 goto out;
1110 } else if ((too_many_pipe_buffers_hard(pipe->user) ||
1111 too_many_pipe_buffers_soft(pipe->user)) &&
1112 !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1113 ret = -EPERM;
1114 goto out;
1115 }
1116 ret = pipe_set_size(pipe, nr_pages);
1117 break;
1118 }
1119 case F_GETPIPE_SZ:
1120 ret = pipe->buffers * PAGE_SIZE;
1121 break;
1122 default:
1123 ret = -EINVAL;
1124 break;
1125 }
1126
1127out:
1128 __pipe_unlock(pipe);
1129 return ret;
1130}
1131
1132static const struct super_operations pipefs_ops = {
1133 .destroy_inode = free_inode_nonrcu,
1134 .statfs = simple_statfs,
1135};
1136
1137/*
1138 * pipefs should _never_ be mounted by userland - too much of security hassle,
1139 * no real gain from having the whole whorehouse mounted. So we don't need
1140 * any operations on the root directory. However, we need a non-trivial
1141 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1142 */
1143static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1144 int flags, const char *dev_name, void *data)
1145{
1146 return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1147 &pipefs_dentry_operations, PIPEFS_MAGIC);
1148}
1149
1150static struct file_system_type pipe_fs_type = {
1151 .name = "pipefs",
1152 .mount = pipefs_mount,
1153 .kill_sb = kill_anon_super,
1154};
1155
1156static int __init init_pipe_fs(void)
1157{
1158 int err = register_filesystem(&pipe_fs_type);
1159
1160 if (!err) {
1161 pipe_mnt = kern_mount(&pipe_fs_type);
1162 if (IS_ERR(pipe_mnt)) {
1163 err = PTR_ERR(pipe_mnt);
1164 unregister_filesystem(&pipe_fs_type);
1165 }
1166 }
1167 return err;
1168}
1169
1170fs_initcall(init_pipe_fs);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/pipe.c
4 *
5 * Copyright (C) 1991, 1992, 1999 Linus Torvalds
6 */
7
8#include <linux/mm.h>
9#include <linux/file.h>
10#include <linux/poll.h>
11#include <linux/slab.h>
12#include <linux/module.h>
13#include <linux/init.h>
14#include <linux/fs.h>
15#include <linux/log2.h>
16#include <linux/mount.h>
17#include <linux/pseudo_fs.h>
18#include <linux/magic.h>
19#include <linux/pipe_fs_i.h>
20#include <linux/uio.h>
21#include <linux/highmem.h>
22#include <linux/pagemap.h>
23#include <linux/audit.h>
24#include <linux/syscalls.h>
25#include <linux/fcntl.h>
26#include <linux/memcontrol.h>
27#include <linux/watch_queue.h>
28#include <linux/sysctl.h>
29
30#include <linux/uaccess.h>
31#include <asm/ioctls.h>
32
33#include "internal.h"
34
35/*
36 * New pipe buffers will be restricted to this size while the user is exceeding
37 * their pipe buffer quota. The general pipe use case needs at least two
38 * buffers: one for data yet to be read, and one for new data. If this is less
39 * than two, then a write to a non-empty pipe may block even if the pipe is not
40 * full. This can occur with GNU make jobserver or similar uses of pipes as
41 * semaphores: multiple processes may be waiting to write tokens back to the
42 * pipe before reading tokens: https://lore.kernel.org/lkml/1628086770.5rn8p04n6j.none@localhost/.
43 *
44 * Users can reduce their pipe buffers with F_SETPIPE_SZ below this at their
45 * own risk, namely: pipe writes to non-full pipes may block until the pipe is
46 * emptied.
47 */
48#define PIPE_MIN_DEF_BUFFERS 2
49
50/*
51 * The max size that a non-root user is allowed to grow the pipe. Can
52 * be set by root in /proc/sys/fs/pipe-max-size
53 */
54static unsigned int pipe_max_size = 1048576;
55
56/* Maximum allocatable pages per user. Hard limit is unset by default, soft
57 * matches default values.
58 */
59static unsigned long pipe_user_pages_hard;
60static unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
61
62/*
63 * We use head and tail indices that aren't masked off, except at the point of
64 * dereference, but rather they're allowed to wrap naturally. This means there
65 * isn't a dead spot in the buffer, but the ring has to be a power of two and
66 * <= 2^31.
67 * -- David Howells 2019-09-23.
68 *
69 * Reads with count = 0 should always return 0.
70 * -- Julian Bradfield 1999-06-07.
71 *
72 * FIFOs and Pipes now generate SIGIO for both readers and writers.
73 * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
74 *
75 * pipe_read & write cleanup
76 * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
77 */
78
79#define cmp_int(l, r) ((l > r) - (l < r))
80
81#ifdef CONFIG_PROVE_LOCKING
82static int pipe_lock_cmp_fn(const struct lockdep_map *a,
83 const struct lockdep_map *b)
84{
85 return cmp_int((unsigned long) a, (unsigned long) b);
86}
87#endif
88
89void pipe_lock(struct pipe_inode_info *pipe)
90{
91 if (pipe->files)
92 mutex_lock(&pipe->mutex);
93}
94EXPORT_SYMBOL(pipe_lock);
95
96void pipe_unlock(struct pipe_inode_info *pipe)
97{
98 if (pipe->files)
99 mutex_unlock(&pipe->mutex);
100}
101EXPORT_SYMBOL(pipe_unlock);
102
103void pipe_double_lock(struct pipe_inode_info *pipe1,
104 struct pipe_inode_info *pipe2)
105{
106 BUG_ON(pipe1 == pipe2);
107
108 if (pipe1 > pipe2)
109 swap(pipe1, pipe2);
110
111 pipe_lock(pipe1);
112 pipe_lock(pipe2);
113}
114
115static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
116 struct pipe_buffer *buf)
117{
118 struct page *page = buf->page;
119
120 /*
121 * If nobody else uses this page, and we don't already have a
122 * temporary page, let's keep track of it as a one-deep
123 * allocation cache. (Otherwise just release our reference to it)
124 */
125 if (page_count(page) == 1 && !pipe->tmp_page)
126 pipe->tmp_page = page;
127 else
128 put_page(page);
129}
130
131static bool anon_pipe_buf_try_steal(struct pipe_inode_info *pipe,
132 struct pipe_buffer *buf)
133{
134 struct page *page = buf->page;
135
136 if (page_count(page) != 1)
137 return false;
138 memcg_kmem_uncharge_page(page, 0);
139 __SetPageLocked(page);
140 return true;
141}
142
143/**
144 * generic_pipe_buf_try_steal - attempt to take ownership of a &pipe_buffer
145 * @pipe: the pipe that the buffer belongs to
146 * @buf: the buffer to attempt to steal
147 *
148 * Description:
149 * This function attempts to steal the &struct page attached to
150 * @buf. If successful, this function returns 0 and returns with
151 * the page locked. The caller may then reuse the page for whatever
152 * he wishes; the typical use is insertion into a different file
153 * page cache.
154 */
155bool generic_pipe_buf_try_steal(struct pipe_inode_info *pipe,
156 struct pipe_buffer *buf)
157{
158 struct page *page = buf->page;
159
160 /*
161 * A reference of one is golden, that means that the owner of this
162 * page is the only one holding a reference to it. lock the page
163 * and return OK.
164 */
165 if (page_count(page) == 1) {
166 lock_page(page);
167 return true;
168 }
169 return false;
170}
171EXPORT_SYMBOL(generic_pipe_buf_try_steal);
172
173/**
174 * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
175 * @pipe: the pipe that the buffer belongs to
176 * @buf: the buffer to get a reference to
177 *
178 * Description:
179 * This function grabs an extra reference to @buf. It's used in
180 * the tee() system call, when we duplicate the buffers in one
181 * pipe into another.
182 */
183bool generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
184{
185 return try_get_page(buf->page);
186}
187EXPORT_SYMBOL(generic_pipe_buf_get);
188
189/**
190 * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
191 * @pipe: the pipe that the buffer belongs to
192 * @buf: the buffer to put a reference to
193 *
194 * Description:
195 * This function releases a reference to @buf.
196 */
197void generic_pipe_buf_release(struct pipe_inode_info *pipe,
198 struct pipe_buffer *buf)
199{
200 put_page(buf->page);
201}
202EXPORT_SYMBOL(generic_pipe_buf_release);
203
204static const struct pipe_buf_operations anon_pipe_buf_ops = {
205 .release = anon_pipe_buf_release,
206 .try_steal = anon_pipe_buf_try_steal,
207 .get = generic_pipe_buf_get,
208};
209
210/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
211static inline bool pipe_readable(const struct pipe_inode_info *pipe)
212{
213 unsigned int head = READ_ONCE(pipe->head);
214 unsigned int tail = READ_ONCE(pipe->tail);
215 unsigned int writers = READ_ONCE(pipe->writers);
216
217 return !pipe_empty(head, tail) || !writers;
218}
219
220static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
221 struct pipe_buffer *buf,
222 unsigned int tail)
223{
224 pipe_buf_release(pipe, buf);
225
226 /*
227 * If the pipe has a watch_queue, we need additional protection
228 * by the spinlock because notifications get posted with only
229 * this spinlock, no mutex
230 */
231 if (pipe_has_watch_queue(pipe)) {
232 spin_lock_irq(&pipe->rd_wait.lock);
233#ifdef CONFIG_WATCH_QUEUE
234 if (buf->flags & PIPE_BUF_FLAG_LOSS)
235 pipe->note_loss = true;
236#endif
237 pipe->tail = ++tail;
238 spin_unlock_irq(&pipe->rd_wait.lock);
239 return tail;
240 }
241
242 /*
243 * Without a watch_queue, we can simply increment the tail
244 * without the spinlock - the mutex is enough.
245 */
246 pipe->tail = ++tail;
247 return tail;
248}
249
250static ssize_t
251pipe_read(struct kiocb *iocb, struct iov_iter *to)
252{
253 size_t total_len = iov_iter_count(to);
254 struct file *filp = iocb->ki_filp;
255 struct pipe_inode_info *pipe = filp->private_data;
256 bool was_full, wake_next_reader = false;
257 ssize_t ret;
258
259 /* Null read succeeds. */
260 if (unlikely(total_len == 0))
261 return 0;
262
263 ret = 0;
264 mutex_lock(&pipe->mutex);
265
266 /*
267 * We only wake up writers if the pipe was full when we started
268 * reading in order to avoid unnecessary wakeups.
269 *
270 * But when we do wake up writers, we do so using a sync wakeup
271 * (WF_SYNC), because we want them to get going and generate more
272 * data for us.
273 */
274 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
275 for (;;) {
276 /* Read ->head with a barrier vs post_one_notification() */
277 unsigned int head = smp_load_acquire(&pipe->head);
278 unsigned int tail = pipe->tail;
279 unsigned int mask = pipe->ring_size - 1;
280
281#ifdef CONFIG_WATCH_QUEUE
282 if (pipe->note_loss) {
283 struct watch_notification n;
284
285 if (total_len < 8) {
286 if (ret == 0)
287 ret = -ENOBUFS;
288 break;
289 }
290
291 n.type = WATCH_TYPE_META;
292 n.subtype = WATCH_META_LOSS_NOTIFICATION;
293 n.info = watch_sizeof(n);
294 if (copy_to_iter(&n, sizeof(n), to) != sizeof(n)) {
295 if (ret == 0)
296 ret = -EFAULT;
297 break;
298 }
299 ret += sizeof(n);
300 total_len -= sizeof(n);
301 pipe->note_loss = false;
302 }
303#endif
304
305 if (!pipe_empty(head, tail)) {
306 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
307 size_t chars = buf->len;
308 size_t written;
309 int error;
310
311 if (chars > total_len) {
312 if (buf->flags & PIPE_BUF_FLAG_WHOLE) {
313 if (ret == 0)
314 ret = -ENOBUFS;
315 break;
316 }
317 chars = total_len;
318 }
319
320 error = pipe_buf_confirm(pipe, buf);
321 if (error) {
322 if (!ret)
323 ret = error;
324 break;
325 }
326
327 written = copy_page_to_iter(buf->page, buf->offset, chars, to);
328 if (unlikely(written < chars)) {
329 if (!ret)
330 ret = -EFAULT;
331 break;
332 }
333 ret += chars;
334 buf->offset += chars;
335 buf->len -= chars;
336
337 /* Was it a packet buffer? Clean up and exit */
338 if (buf->flags & PIPE_BUF_FLAG_PACKET) {
339 total_len = chars;
340 buf->len = 0;
341 }
342
343 if (!buf->len)
344 tail = pipe_update_tail(pipe, buf, tail);
345 total_len -= chars;
346 if (!total_len)
347 break; /* common path: read succeeded */
348 if (!pipe_empty(head, tail)) /* More to do? */
349 continue;
350 }
351
352 if (!pipe->writers)
353 break;
354 if (ret)
355 break;
356 if ((filp->f_flags & O_NONBLOCK) ||
357 (iocb->ki_flags & IOCB_NOWAIT)) {
358 ret = -EAGAIN;
359 break;
360 }
361 mutex_unlock(&pipe->mutex);
362
363 /*
364 * We only get here if we didn't actually read anything.
365 *
366 * However, we could have seen (and removed) a zero-sized
367 * pipe buffer, and might have made space in the buffers
368 * that way.
369 *
370 * You can't make zero-sized pipe buffers by doing an empty
371 * write (not even in packet mode), but they can happen if
372 * the writer gets an EFAULT when trying to fill a buffer
373 * that already got allocated and inserted in the buffer
374 * array.
375 *
376 * So we still need to wake up any pending writers in the
377 * _very_ unlikely case that the pipe was full, but we got
378 * no data.
379 */
380 if (unlikely(was_full))
381 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
382 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
383
384 /*
385 * But because we didn't read anything, at this point we can
386 * just return directly with -ERESTARTSYS if we're interrupted,
387 * since we've done any required wakeups and there's no need
388 * to mark anything accessed. And we've dropped the lock.
389 */
390 if (wait_event_interruptible_exclusive(pipe->rd_wait, pipe_readable(pipe)) < 0)
391 return -ERESTARTSYS;
392
393 mutex_lock(&pipe->mutex);
394 was_full = pipe_full(pipe->head, pipe->tail, pipe->max_usage);
395 wake_next_reader = true;
396 }
397 if (pipe_empty(pipe->head, pipe->tail))
398 wake_next_reader = false;
399 mutex_unlock(&pipe->mutex);
400
401 if (was_full)
402 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
403 if (wake_next_reader)
404 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
405 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
406 if (ret > 0)
407 file_accessed(filp);
408 return ret;
409}
410
411static inline int is_packetized(struct file *file)
412{
413 return (file->f_flags & O_DIRECT) != 0;
414}
415
416/* Done while waiting without holding the pipe lock - thus the READ_ONCE() */
417static inline bool pipe_writable(const struct pipe_inode_info *pipe)
418{
419 unsigned int head = READ_ONCE(pipe->head);
420 unsigned int tail = READ_ONCE(pipe->tail);
421 unsigned int max_usage = READ_ONCE(pipe->max_usage);
422
423 return !pipe_full(head, tail, max_usage) ||
424 !READ_ONCE(pipe->readers);
425}
426
427static ssize_t
428pipe_write(struct kiocb *iocb, struct iov_iter *from)
429{
430 struct file *filp = iocb->ki_filp;
431 struct pipe_inode_info *pipe = filp->private_data;
432 unsigned int head;
433 ssize_t ret = 0;
434 size_t total_len = iov_iter_count(from);
435 ssize_t chars;
436 bool was_empty = false;
437 bool wake_next_writer = false;
438
439 /*
440 * Reject writing to watch queue pipes before the point where we lock
441 * the pipe.
442 * Otherwise, lockdep would be unhappy if the caller already has another
443 * pipe locked.
444 * If we had to support locking a normal pipe and a notification pipe at
445 * the same time, we could set up lockdep annotations for that, but
446 * since we don't actually need that, it's simpler to just bail here.
447 */
448 if (pipe_has_watch_queue(pipe))
449 return -EXDEV;
450
451 /* Null write succeeds. */
452 if (unlikely(total_len == 0))
453 return 0;
454
455 mutex_lock(&pipe->mutex);
456
457 if (!pipe->readers) {
458 send_sig(SIGPIPE, current, 0);
459 ret = -EPIPE;
460 goto out;
461 }
462
463 /*
464 * If it wasn't empty we try to merge new data into
465 * the last buffer.
466 *
467 * That naturally merges small writes, but it also
468 * page-aligns the rest of the writes for large writes
469 * spanning multiple pages.
470 */
471 head = pipe->head;
472 was_empty = pipe_empty(head, pipe->tail);
473 chars = total_len & (PAGE_SIZE-1);
474 if (chars && !was_empty) {
475 unsigned int mask = pipe->ring_size - 1;
476 struct pipe_buffer *buf = &pipe->bufs[(head - 1) & mask];
477 int offset = buf->offset + buf->len;
478
479 if ((buf->flags & PIPE_BUF_FLAG_CAN_MERGE) &&
480 offset + chars <= PAGE_SIZE) {
481 ret = pipe_buf_confirm(pipe, buf);
482 if (ret)
483 goto out;
484
485 ret = copy_page_from_iter(buf->page, offset, chars, from);
486 if (unlikely(ret < chars)) {
487 ret = -EFAULT;
488 goto out;
489 }
490
491 buf->len += ret;
492 if (!iov_iter_count(from))
493 goto out;
494 }
495 }
496
497 for (;;) {
498 if (!pipe->readers) {
499 send_sig(SIGPIPE, current, 0);
500 if (!ret)
501 ret = -EPIPE;
502 break;
503 }
504
505 head = pipe->head;
506 if (!pipe_full(head, pipe->tail, pipe->max_usage)) {
507 unsigned int mask = pipe->ring_size - 1;
508 struct pipe_buffer *buf;
509 struct page *page = pipe->tmp_page;
510 int copied;
511
512 if (!page) {
513 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
514 if (unlikely(!page)) {
515 ret = ret ? : -ENOMEM;
516 break;
517 }
518 pipe->tmp_page = page;
519 }
520
521 /* Allocate a slot in the ring in advance and attach an
522 * empty buffer. If we fault or otherwise fail to use
523 * it, either the reader will consume it or it'll still
524 * be there for the next write.
525 */
526 pipe->head = head + 1;
527
528 /* Insert it into the buffer array */
529 buf = &pipe->bufs[head & mask];
530 buf->page = page;
531 buf->ops = &anon_pipe_buf_ops;
532 buf->offset = 0;
533 buf->len = 0;
534 if (is_packetized(filp))
535 buf->flags = PIPE_BUF_FLAG_PACKET;
536 else
537 buf->flags = PIPE_BUF_FLAG_CAN_MERGE;
538 pipe->tmp_page = NULL;
539
540 copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
541 if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
542 if (!ret)
543 ret = -EFAULT;
544 break;
545 }
546 ret += copied;
547 buf->len = copied;
548
549 if (!iov_iter_count(from))
550 break;
551 }
552
553 if (!pipe_full(head, pipe->tail, pipe->max_usage))
554 continue;
555
556 /* Wait for buffer space to become available. */
557 if ((filp->f_flags & O_NONBLOCK) ||
558 (iocb->ki_flags & IOCB_NOWAIT)) {
559 if (!ret)
560 ret = -EAGAIN;
561 break;
562 }
563 if (signal_pending(current)) {
564 if (!ret)
565 ret = -ERESTARTSYS;
566 break;
567 }
568
569 /*
570 * We're going to release the pipe lock and wait for more
571 * space. We wake up any readers if necessary, and then
572 * after waiting we need to re-check whether the pipe
573 * become empty while we dropped the lock.
574 */
575 mutex_unlock(&pipe->mutex);
576 if (was_empty)
577 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
578 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
579 wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));
580 mutex_lock(&pipe->mutex);
581 was_empty = pipe_empty(pipe->head, pipe->tail);
582 wake_next_writer = true;
583 }
584out:
585 if (pipe_full(pipe->head, pipe->tail, pipe->max_usage))
586 wake_next_writer = false;
587 mutex_unlock(&pipe->mutex);
588
589 /*
590 * If we do do a wakeup event, we do a 'sync' wakeup, because we
591 * want the reader to start processing things asap, rather than
592 * leave the data pending.
593 *
594 * This is particularly important for small writes, because of
595 * how (for example) the GNU make jobserver uses small writes to
596 * wake up pending jobs
597 *
598 * Epoll nonsensically wants a wakeup whether the pipe
599 * was already empty or not.
600 */
601 if (was_empty || pipe->poll_usage)
602 wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
603 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
604 if (wake_next_writer)
605 wake_up_interruptible_sync_poll(&pipe->wr_wait, EPOLLOUT | EPOLLWRNORM);
606 if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
607 int err = file_update_time(filp);
608 if (err)
609 ret = err;
610 sb_end_write(file_inode(filp)->i_sb);
611 }
612 return ret;
613}
614
615static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
616{
617 struct pipe_inode_info *pipe = filp->private_data;
618 unsigned int count, head, tail, mask;
619
620 switch (cmd) {
621 case FIONREAD:
622 mutex_lock(&pipe->mutex);
623 count = 0;
624 head = pipe->head;
625 tail = pipe->tail;
626 mask = pipe->ring_size - 1;
627
628 while (tail != head) {
629 count += pipe->bufs[tail & mask].len;
630 tail++;
631 }
632 mutex_unlock(&pipe->mutex);
633
634 return put_user(count, (int __user *)arg);
635
636#ifdef CONFIG_WATCH_QUEUE
637 case IOC_WATCH_QUEUE_SET_SIZE: {
638 int ret;
639 mutex_lock(&pipe->mutex);
640 ret = watch_queue_set_size(pipe, arg);
641 mutex_unlock(&pipe->mutex);
642 return ret;
643 }
644
645 case IOC_WATCH_QUEUE_SET_FILTER:
646 return watch_queue_set_filter(
647 pipe, (struct watch_notification_filter __user *)arg);
648#endif
649
650 default:
651 return -ENOIOCTLCMD;
652 }
653}
654
655/* No kernel lock held - fine */
656static __poll_t
657pipe_poll(struct file *filp, poll_table *wait)
658{
659 __poll_t mask;
660 struct pipe_inode_info *pipe = filp->private_data;
661 unsigned int head, tail;
662
663 /* Epoll has some historical nasty semantics, this enables them */
664 WRITE_ONCE(pipe->poll_usage, true);
665
666 /*
667 * Reading pipe state only -- no need for acquiring the semaphore.
668 *
669 * But because this is racy, the code has to add the
670 * entry to the poll table _first_ ..
671 */
672 if (filp->f_mode & FMODE_READ)
673 poll_wait(filp, &pipe->rd_wait, wait);
674 if (filp->f_mode & FMODE_WRITE)
675 poll_wait(filp, &pipe->wr_wait, wait);
676
677 /*
678 * .. and only then can you do the racy tests. That way,
679 * if something changes and you got it wrong, the poll
680 * table entry will wake you up and fix it.
681 */
682 head = READ_ONCE(pipe->head);
683 tail = READ_ONCE(pipe->tail);
684
685 mask = 0;
686 if (filp->f_mode & FMODE_READ) {
687 if (!pipe_empty(head, tail))
688 mask |= EPOLLIN | EPOLLRDNORM;
689 if (!pipe->writers && filp->f_pipe != pipe->w_counter)
690 mask |= EPOLLHUP;
691 }
692
693 if (filp->f_mode & FMODE_WRITE) {
694 if (!pipe_full(head, tail, pipe->max_usage))
695 mask |= EPOLLOUT | EPOLLWRNORM;
696 /*
697 * Most Unices do not set EPOLLERR for FIFOs but on Linux they
698 * behave exactly like pipes for poll().
699 */
700 if (!pipe->readers)
701 mask |= EPOLLERR;
702 }
703
704 return mask;
705}
706
707static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
708{
709 int kill = 0;
710
711 spin_lock(&inode->i_lock);
712 if (!--pipe->files) {
713 inode->i_pipe = NULL;
714 kill = 1;
715 }
716 spin_unlock(&inode->i_lock);
717
718 if (kill)
719 free_pipe_info(pipe);
720}
721
722static int
723pipe_release(struct inode *inode, struct file *file)
724{
725 struct pipe_inode_info *pipe = file->private_data;
726
727 mutex_lock(&pipe->mutex);
728 if (file->f_mode & FMODE_READ)
729 pipe->readers--;
730 if (file->f_mode & FMODE_WRITE)
731 pipe->writers--;
732
733 /* Was that the last reader or writer, but not the other side? */
734 if (!pipe->readers != !pipe->writers) {
735 wake_up_interruptible_all(&pipe->rd_wait);
736 wake_up_interruptible_all(&pipe->wr_wait);
737 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
738 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
739 }
740 mutex_unlock(&pipe->mutex);
741
742 put_pipe_info(inode, pipe);
743 return 0;
744}
745
746static int
747pipe_fasync(int fd, struct file *filp, int on)
748{
749 struct pipe_inode_info *pipe = filp->private_data;
750 int retval = 0;
751
752 mutex_lock(&pipe->mutex);
753 if (filp->f_mode & FMODE_READ)
754 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
755 if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
756 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
757 if (retval < 0 && (filp->f_mode & FMODE_READ))
758 /* this can happen only if on == T */
759 fasync_helper(-1, filp, 0, &pipe->fasync_readers);
760 }
761 mutex_unlock(&pipe->mutex);
762 return retval;
763}
764
765unsigned long account_pipe_buffers(struct user_struct *user,
766 unsigned long old, unsigned long new)
767{
768 return atomic_long_add_return(new - old, &user->pipe_bufs);
769}
770
771bool too_many_pipe_buffers_soft(unsigned long user_bufs)
772{
773 unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
774
775 return soft_limit && user_bufs > soft_limit;
776}
777
778bool too_many_pipe_buffers_hard(unsigned long user_bufs)
779{
780 unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
781
782 return hard_limit && user_bufs > hard_limit;
783}
784
785bool pipe_is_unprivileged_user(void)
786{
787 return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
788}
789
790struct pipe_inode_info *alloc_pipe_info(void)
791{
792 struct pipe_inode_info *pipe;
793 unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
794 struct user_struct *user = get_current_user();
795 unsigned long user_bufs;
796 unsigned int max_size = READ_ONCE(pipe_max_size);
797
798 pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
799 if (pipe == NULL)
800 goto out_free_uid;
801
802 if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
803 pipe_bufs = max_size >> PAGE_SHIFT;
804
805 user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
806
807 if (too_many_pipe_buffers_soft(user_bufs) && pipe_is_unprivileged_user()) {
808 user_bufs = account_pipe_buffers(user, pipe_bufs, PIPE_MIN_DEF_BUFFERS);
809 pipe_bufs = PIPE_MIN_DEF_BUFFERS;
810 }
811
812 if (too_many_pipe_buffers_hard(user_bufs) && pipe_is_unprivileged_user())
813 goto out_revert_acct;
814
815 pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
816 GFP_KERNEL_ACCOUNT);
817
818 if (pipe->bufs) {
819 init_waitqueue_head(&pipe->rd_wait);
820 init_waitqueue_head(&pipe->wr_wait);
821 pipe->r_counter = pipe->w_counter = 1;
822 pipe->max_usage = pipe_bufs;
823 pipe->ring_size = pipe_bufs;
824 pipe->nr_accounted = pipe_bufs;
825 pipe->user = user;
826 mutex_init(&pipe->mutex);
827 lock_set_cmp_fn(&pipe->mutex, pipe_lock_cmp_fn, NULL);
828 return pipe;
829 }
830
831out_revert_acct:
832 (void) account_pipe_buffers(user, pipe_bufs, 0);
833 kfree(pipe);
834out_free_uid:
835 free_uid(user);
836 return NULL;
837}
838
839void free_pipe_info(struct pipe_inode_info *pipe)
840{
841 unsigned int i;
842
843#ifdef CONFIG_WATCH_QUEUE
844 if (pipe->watch_queue)
845 watch_queue_clear(pipe->watch_queue);
846#endif
847
848 (void) account_pipe_buffers(pipe->user, pipe->nr_accounted, 0);
849 free_uid(pipe->user);
850 for (i = 0; i < pipe->ring_size; i++) {
851 struct pipe_buffer *buf = pipe->bufs + i;
852 if (buf->ops)
853 pipe_buf_release(pipe, buf);
854 }
855#ifdef CONFIG_WATCH_QUEUE
856 if (pipe->watch_queue)
857 put_watch_queue(pipe->watch_queue);
858#endif
859 if (pipe->tmp_page)
860 __free_page(pipe->tmp_page);
861 kfree(pipe->bufs);
862 kfree(pipe);
863}
864
865static struct vfsmount *pipe_mnt __ro_after_init;
866
867/*
868 * pipefs_dname() is called from d_path().
869 */
870static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
871{
872 return dynamic_dname(buffer, buflen, "pipe:[%lu]",
873 d_inode(dentry)->i_ino);
874}
875
876static const struct dentry_operations pipefs_dentry_operations = {
877 .d_dname = pipefs_dname,
878};
879
880static struct inode * get_pipe_inode(void)
881{
882 struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
883 struct pipe_inode_info *pipe;
884
885 if (!inode)
886 goto fail_inode;
887
888 inode->i_ino = get_next_ino();
889
890 pipe = alloc_pipe_info();
891 if (!pipe)
892 goto fail_iput;
893
894 inode->i_pipe = pipe;
895 pipe->files = 2;
896 pipe->readers = pipe->writers = 1;
897 inode->i_fop = &pipefifo_fops;
898
899 /*
900 * Mark the inode dirty from the very beginning,
901 * that way it will never be moved to the dirty
902 * list because "mark_inode_dirty()" will think
903 * that it already _is_ on the dirty list.
904 */
905 inode->i_state = I_DIRTY;
906 inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
907 inode->i_uid = current_fsuid();
908 inode->i_gid = current_fsgid();
909 simple_inode_init_ts(inode);
910
911 return inode;
912
913fail_iput:
914 iput(inode);
915
916fail_inode:
917 return NULL;
918}
919
920int create_pipe_files(struct file **res, int flags)
921{
922 struct inode *inode = get_pipe_inode();
923 struct file *f;
924 int error;
925
926 if (!inode)
927 return -ENFILE;
928
929 if (flags & O_NOTIFICATION_PIPE) {
930 error = watch_queue_init(inode->i_pipe);
931 if (error) {
932 free_pipe_info(inode->i_pipe);
933 iput(inode);
934 return error;
935 }
936 }
937
938 f = alloc_file_pseudo(inode, pipe_mnt, "",
939 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
940 &pipefifo_fops);
941 if (IS_ERR(f)) {
942 free_pipe_info(inode->i_pipe);
943 iput(inode);
944 return PTR_ERR(f);
945 }
946
947 f->private_data = inode->i_pipe;
948 f->f_pipe = 0;
949
950 res[0] = alloc_file_clone(f, O_RDONLY | (flags & O_NONBLOCK),
951 &pipefifo_fops);
952 if (IS_ERR(res[0])) {
953 put_pipe_info(inode, inode->i_pipe);
954 fput(f);
955 return PTR_ERR(res[0]);
956 }
957 res[0]->private_data = inode->i_pipe;
958 res[0]->f_pipe = 0;
959 res[1] = f;
960 stream_open(inode, res[0]);
961 stream_open(inode, res[1]);
962 return 0;
963}
964
965static int __do_pipe_flags(int *fd, struct file **files, int flags)
966{
967 int error;
968 int fdw, fdr;
969
970 if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT | O_NOTIFICATION_PIPE))
971 return -EINVAL;
972
973 error = create_pipe_files(files, flags);
974 if (error)
975 return error;
976
977 error = get_unused_fd_flags(flags);
978 if (error < 0)
979 goto err_read_pipe;
980 fdr = error;
981
982 error = get_unused_fd_flags(flags);
983 if (error < 0)
984 goto err_fdr;
985 fdw = error;
986
987 audit_fd_pair(fdr, fdw);
988 fd[0] = fdr;
989 fd[1] = fdw;
990 /* pipe groks IOCB_NOWAIT */
991 files[0]->f_mode |= FMODE_NOWAIT;
992 files[1]->f_mode |= FMODE_NOWAIT;
993 return 0;
994
995 err_fdr:
996 put_unused_fd(fdr);
997 err_read_pipe:
998 fput(files[0]);
999 fput(files[1]);
1000 return error;
1001}
1002
1003int do_pipe_flags(int *fd, int flags)
1004{
1005 struct file *files[2];
1006 int error = __do_pipe_flags(fd, files, flags);
1007 if (!error) {
1008 fd_install(fd[0], files[0]);
1009 fd_install(fd[1], files[1]);
1010 }
1011 return error;
1012}
1013
1014/*
1015 * sys_pipe() is the normal C calling standard for creating
1016 * a pipe. It's not the way Unix traditionally does this, though.
1017 */
1018static int do_pipe2(int __user *fildes, int flags)
1019{
1020 struct file *files[2];
1021 int fd[2];
1022 int error;
1023
1024 error = __do_pipe_flags(fd, files, flags);
1025 if (!error) {
1026 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
1027 fput(files[0]);
1028 fput(files[1]);
1029 put_unused_fd(fd[0]);
1030 put_unused_fd(fd[1]);
1031 error = -EFAULT;
1032 } else {
1033 fd_install(fd[0], files[0]);
1034 fd_install(fd[1], files[1]);
1035 }
1036 }
1037 return error;
1038}
1039
1040SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
1041{
1042 return do_pipe2(fildes, flags);
1043}
1044
1045SYSCALL_DEFINE1(pipe, int __user *, fildes)
1046{
1047 return do_pipe2(fildes, 0);
1048}
1049
1050/*
1051 * This is the stupid "wait for pipe to be readable or writable"
1052 * model.
1053 *
1054 * See pipe_read/write() for the proper kind of exclusive wait,
1055 * but that requires that we wake up any other readers/writers
1056 * if we then do not end up reading everything (ie the whole
1057 * "wake_next_reader/writer" logic in pipe_read/write()).
1058 */
1059void pipe_wait_readable(struct pipe_inode_info *pipe)
1060{
1061 pipe_unlock(pipe);
1062 wait_event_interruptible(pipe->rd_wait, pipe_readable(pipe));
1063 pipe_lock(pipe);
1064}
1065
1066void pipe_wait_writable(struct pipe_inode_info *pipe)
1067{
1068 pipe_unlock(pipe);
1069 wait_event_interruptible(pipe->wr_wait, pipe_writable(pipe));
1070 pipe_lock(pipe);
1071}
1072
1073/*
1074 * This depends on both the wait (here) and the wakeup (wake_up_partner)
1075 * holding the pipe lock, so "*cnt" is stable and we know a wakeup cannot
1076 * race with the count check and waitqueue prep.
1077 *
1078 * Normally in order to avoid races, you'd do the prepare_to_wait() first,
1079 * then check the condition you're waiting for, and only then sleep. But
1080 * because of the pipe lock, we can check the condition before being on
1081 * the wait queue.
1082 *
1083 * We use the 'rd_wait' waitqueue for pipe partner waiting.
1084 */
1085static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1086{
1087 DEFINE_WAIT(rdwait);
1088 int cur = *cnt;
1089
1090 while (cur == *cnt) {
1091 prepare_to_wait(&pipe->rd_wait, &rdwait, TASK_INTERRUPTIBLE);
1092 pipe_unlock(pipe);
1093 schedule();
1094 finish_wait(&pipe->rd_wait, &rdwait);
1095 pipe_lock(pipe);
1096 if (signal_pending(current))
1097 break;
1098 }
1099 return cur == *cnt ? -ERESTARTSYS : 0;
1100}
1101
1102static void wake_up_partner(struct pipe_inode_info *pipe)
1103{
1104 wake_up_interruptible_all(&pipe->rd_wait);
1105}
1106
1107static int fifo_open(struct inode *inode, struct file *filp)
1108{
1109 struct pipe_inode_info *pipe;
1110 bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1111 int ret;
1112
1113 filp->f_pipe = 0;
1114
1115 spin_lock(&inode->i_lock);
1116 if (inode->i_pipe) {
1117 pipe = inode->i_pipe;
1118 pipe->files++;
1119 spin_unlock(&inode->i_lock);
1120 } else {
1121 spin_unlock(&inode->i_lock);
1122 pipe = alloc_pipe_info();
1123 if (!pipe)
1124 return -ENOMEM;
1125 pipe->files = 1;
1126 spin_lock(&inode->i_lock);
1127 if (unlikely(inode->i_pipe)) {
1128 inode->i_pipe->files++;
1129 spin_unlock(&inode->i_lock);
1130 free_pipe_info(pipe);
1131 pipe = inode->i_pipe;
1132 } else {
1133 inode->i_pipe = pipe;
1134 spin_unlock(&inode->i_lock);
1135 }
1136 }
1137 filp->private_data = pipe;
1138 /* OK, we have a pipe and it's pinned down */
1139
1140 mutex_lock(&pipe->mutex);
1141
1142 /* We can only do regular read/write on fifos */
1143 stream_open(inode, filp);
1144
1145 switch (filp->f_mode & (FMODE_READ | FMODE_WRITE)) {
1146 case FMODE_READ:
1147 /*
1148 * O_RDONLY
1149 * POSIX.1 says that O_NONBLOCK means return with the FIFO
1150 * opened, even when there is no process writing the FIFO.
1151 */
1152 pipe->r_counter++;
1153 if (pipe->readers++ == 0)
1154 wake_up_partner(pipe);
1155
1156 if (!is_pipe && !pipe->writers) {
1157 if ((filp->f_flags & O_NONBLOCK)) {
1158 /* suppress EPOLLHUP until we have
1159 * seen a writer */
1160 filp->f_pipe = pipe->w_counter;
1161 } else {
1162 if (wait_for_partner(pipe, &pipe->w_counter))
1163 goto err_rd;
1164 }
1165 }
1166 break;
1167
1168 case FMODE_WRITE:
1169 /*
1170 * O_WRONLY
1171 * POSIX.1 says that O_NONBLOCK means return -1 with
1172 * errno=ENXIO when there is no process reading the FIFO.
1173 */
1174 ret = -ENXIO;
1175 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1176 goto err;
1177
1178 pipe->w_counter++;
1179 if (!pipe->writers++)
1180 wake_up_partner(pipe);
1181
1182 if (!is_pipe && !pipe->readers) {
1183 if (wait_for_partner(pipe, &pipe->r_counter))
1184 goto err_wr;
1185 }
1186 break;
1187
1188 case FMODE_READ | FMODE_WRITE:
1189 /*
1190 * O_RDWR
1191 * POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1192 * This implementation will NEVER block on a O_RDWR open, since
1193 * the process can at least talk to itself.
1194 */
1195
1196 pipe->readers++;
1197 pipe->writers++;
1198 pipe->r_counter++;
1199 pipe->w_counter++;
1200 if (pipe->readers == 1 || pipe->writers == 1)
1201 wake_up_partner(pipe);
1202 break;
1203
1204 default:
1205 ret = -EINVAL;
1206 goto err;
1207 }
1208
1209 /* Ok! */
1210 mutex_unlock(&pipe->mutex);
1211 return 0;
1212
1213err_rd:
1214 if (!--pipe->readers)
1215 wake_up_interruptible(&pipe->wr_wait);
1216 ret = -ERESTARTSYS;
1217 goto err;
1218
1219err_wr:
1220 if (!--pipe->writers)
1221 wake_up_interruptible_all(&pipe->rd_wait);
1222 ret = -ERESTARTSYS;
1223 goto err;
1224
1225err:
1226 mutex_unlock(&pipe->mutex);
1227
1228 put_pipe_info(inode, pipe);
1229 return ret;
1230}
1231
1232const struct file_operations pipefifo_fops = {
1233 .open = fifo_open,
1234 .read_iter = pipe_read,
1235 .write_iter = pipe_write,
1236 .poll = pipe_poll,
1237 .unlocked_ioctl = pipe_ioctl,
1238 .release = pipe_release,
1239 .fasync = pipe_fasync,
1240 .splice_write = iter_file_splice_write,
1241};
1242
1243/*
1244 * Currently we rely on the pipe array holding a power-of-2 number
1245 * of pages. Returns 0 on error.
1246 */
1247unsigned int round_pipe_size(unsigned int size)
1248{
1249 if (size > (1U << 31))
1250 return 0;
1251
1252 /* Minimum pipe size, as required by POSIX */
1253 if (size < PAGE_SIZE)
1254 return PAGE_SIZE;
1255
1256 return roundup_pow_of_two(size);
1257}
1258
1259/*
1260 * Resize the pipe ring to a number of slots.
1261 *
1262 * Note the pipe can be reduced in capacity, but only if the current
1263 * occupancy doesn't exceed nr_slots; if it does, EBUSY will be
1264 * returned instead.
1265 */
1266int pipe_resize_ring(struct pipe_inode_info *pipe, unsigned int nr_slots)
1267{
1268 struct pipe_buffer *bufs;
1269 unsigned int head, tail, mask, n;
1270
1271 bufs = kcalloc(nr_slots, sizeof(*bufs),
1272 GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1273 if (unlikely(!bufs))
1274 return -ENOMEM;
1275
1276 spin_lock_irq(&pipe->rd_wait.lock);
1277 mask = pipe->ring_size - 1;
1278 head = pipe->head;
1279 tail = pipe->tail;
1280
1281 n = pipe_occupancy(head, tail);
1282 if (nr_slots < n) {
1283 spin_unlock_irq(&pipe->rd_wait.lock);
1284 kfree(bufs);
1285 return -EBUSY;
1286 }
1287
1288 /*
1289 * The pipe array wraps around, so just start the new one at zero
1290 * and adjust the indices.
1291 */
1292 if (n > 0) {
1293 unsigned int h = head & mask;
1294 unsigned int t = tail & mask;
1295 if (h > t) {
1296 memcpy(bufs, pipe->bufs + t,
1297 n * sizeof(struct pipe_buffer));
1298 } else {
1299 unsigned int tsize = pipe->ring_size - t;
1300 if (h > 0)
1301 memcpy(bufs + tsize, pipe->bufs,
1302 h * sizeof(struct pipe_buffer));
1303 memcpy(bufs, pipe->bufs + t,
1304 tsize * sizeof(struct pipe_buffer));
1305 }
1306 }
1307
1308 head = n;
1309 tail = 0;
1310
1311 kfree(pipe->bufs);
1312 pipe->bufs = bufs;
1313 pipe->ring_size = nr_slots;
1314 if (pipe->max_usage > nr_slots)
1315 pipe->max_usage = nr_slots;
1316 pipe->tail = tail;
1317 pipe->head = head;
1318
1319 if (!pipe_has_watch_queue(pipe)) {
1320 pipe->max_usage = nr_slots;
1321 pipe->nr_accounted = nr_slots;
1322 }
1323
1324 spin_unlock_irq(&pipe->rd_wait.lock);
1325
1326 /* This might have made more room for writers */
1327 wake_up_interruptible(&pipe->wr_wait);
1328 return 0;
1329}
1330
1331/*
1332 * Allocate a new array of pipe buffers and copy the info over. Returns the
1333 * pipe size if successful, or return -ERROR on error.
1334 */
1335static long pipe_set_size(struct pipe_inode_info *pipe, unsigned int arg)
1336{
1337 unsigned long user_bufs;
1338 unsigned int nr_slots, size;
1339 long ret = 0;
1340
1341 if (pipe_has_watch_queue(pipe))
1342 return -EBUSY;
1343
1344 size = round_pipe_size(arg);
1345 nr_slots = size >> PAGE_SHIFT;
1346
1347 if (!nr_slots)
1348 return -EINVAL;
1349
1350 /*
1351 * If trying to increase the pipe capacity, check that an
1352 * unprivileged user is not trying to exceed various limits
1353 * (soft limit check here, hard limit check just below).
1354 * Decreasing the pipe capacity is always permitted, even
1355 * if the user is currently over a limit.
1356 */
1357 if (nr_slots > pipe->max_usage &&
1358 size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1359 return -EPERM;
1360
1361 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_slots);
1362
1363 if (nr_slots > pipe->max_usage &&
1364 (too_many_pipe_buffers_hard(user_bufs) ||
1365 too_many_pipe_buffers_soft(user_bufs)) &&
1366 pipe_is_unprivileged_user()) {
1367 ret = -EPERM;
1368 goto out_revert_acct;
1369 }
1370
1371 ret = pipe_resize_ring(pipe, nr_slots);
1372 if (ret < 0)
1373 goto out_revert_acct;
1374
1375 return pipe->max_usage * PAGE_SIZE;
1376
1377out_revert_acct:
1378 (void) account_pipe_buffers(pipe->user, nr_slots, pipe->nr_accounted);
1379 return ret;
1380}
1381
1382/*
1383 * Note that i_pipe and i_cdev share the same location, so checking ->i_pipe is
1384 * not enough to verify that this is a pipe.
1385 */
1386struct pipe_inode_info *get_pipe_info(struct file *file, bool for_splice)
1387{
1388 struct pipe_inode_info *pipe = file->private_data;
1389
1390 if (file->f_op != &pipefifo_fops || !pipe)
1391 return NULL;
1392 if (for_splice && pipe_has_watch_queue(pipe))
1393 return NULL;
1394 return pipe;
1395}
1396
1397long pipe_fcntl(struct file *file, unsigned int cmd, unsigned int arg)
1398{
1399 struct pipe_inode_info *pipe;
1400 long ret;
1401
1402 pipe = get_pipe_info(file, false);
1403 if (!pipe)
1404 return -EBADF;
1405
1406 mutex_lock(&pipe->mutex);
1407
1408 switch (cmd) {
1409 case F_SETPIPE_SZ:
1410 ret = pipe_set_size(pipe, arg);
1411 break;
1412 case F_GETPIPE_SZ:
1413 ret = pipe->max_usage * PAGE_SIZE;
1414 break;
1415 default:
1416 ret = -EINVAL;
1417 break;
1418 }
1419
1420 mutex_unlock(&pipe->mutex);
1421 return ret;
1422}
1423
1424static const struct super_operations pipefs_ops = {
1425 .destroy_inode = free_inode_nonrcu,
1426 .statfs = simple_statfs,
1427};
1428
1429/*
1430 * pipefs should _never_ be mounted by userland - too much of security hassle,
1431 * no real gain from having the whole file system mounted. So we don't need
1432 * any operations on the root directory. However, we need a non-trivial
1433 * d_name - pipe: will go nicely and kill the special-casing in procfs.
1434 */
1435
1436static int pipefs_init_fs_context(struct fs_context *fc)
1437{
1438 struct pseudo_fs_context *ctx = init_pseudo(fc, PIPEFS_MAGIC);
1439 if (!ctx)
1440 return -ENOMEM;
1441 ctx->ops = &pipefs_ops;
1442 ctx->dops = &pipefs_dentry_operations;
1443 return 0;
1444}
1445
1446static struct file_system_type pipe_fs_type = {
1447 .name = "pipefs",
1448 .init_fs_context = pipefs_init_fs_context,
1449 .kill_sb = kill_anon_super,
1450};
1451
1452#ifdef CONFIG_SYSCTL
1453static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1454 unsigned int *valp,
1455 int write, void *data)
1456{
1457 if (write) {
1458 unsigned int val;
1459
1460 val = round_pipe_size(*lvalp);
1461 if (val == 0)
1462 return -EINVAL;
1463
1464 *valp = val;
1465 } else {
1466 unsigned int val = *valp;
1467 *lvalp = (unsigned long) val;
1468 }
1469
1470 return 0;
1471}
1472
1473static int proc_dopipe_max_size(const struct ctl_table *table, int write,
1474 void *buffer, size_t *lenp, loff_t *ppos)
1475{
1476 return do_proc_douintvec(table, write, buffer, lenp, ppos,
1477 do_proc_dopipe_max_size_conv, NULL);
1478}
1479
1480static struct ctl_table fs_pipe_sysctls[] = {
1481 {
1482 .procname = "pipe-max-size",
1483 .data = &pipe_max_size,
1484 .maxlen = sizeof(pipe_max_size),
1485 .mode = 0644,
1486 .proc_handler = proc_dopipe_max_size,
1487 },
1488 {
1489 .procname = "pipe-user-pages-hard",
1490 .data = &pipe_user_pages_hard,
1491 .maxlen = sizeof(pipe_user_pages_hard),
1492 .mode = 0644,
1493 .proc_handler = proc_doulongvec_minmax,
1494 },
1495 {
1496 .procname = "pipe-user-pages-soft",
1497 .data = &pipe_user_pages_soft,
1498 .maxlen = sizeof(pipe_user_pages_soft),
1499 .mode = 0644,
1500 .proc_handler = proc_doulongvec_minmax,
1501 },
1502};
1503#endif
1504
1505static int __init init_pipe_fs(void)
1506{
1507 int err = register_filesystem(&pipe_fs_type);
1508
1509 if (!err) {
1510 pipe_mnt = kern_mount(&pipe_fs_type);
1511 if (IS_ERR(pipe_mnt)) {
1512 err = PTR_ERR(pipe_mnt);
1513 unregister_filesystem(&pipe_fs_type);
1514 }
1515 }
1516#ifdef CONFIG_SYSCTL
1517 register_sysctl_init("fs", fs_pipe_sysctls);
1518#endif
1519 return err;
1520}
1521
1522fs_initcall(init_pipe_fs);