xref: /openbmc/linux/fs/pipe.c (revision 5f32c314)
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 #include <linux/aio.h>
25 
26 #include <asm/uaccess.h>
27 #include <asm/ioctls.h>
28 
29 #include "internal.h"
30 
31 /*
32  * The max size that a non-root user is allowed to grow the pipe. Can
33  * be set by root in /proc/sys/fs/pipe-max-size
34  */
35 unsigned int pipe_max_size = 1048576;
36 
37 /*
38  * Minimum pipe size, as required by POSIX
39  */
40 unsigned int pipe_min_size = PAGE_SIZE;
41 
42 /*
43  * We use a start+len construction, which provides full use of the
44  * allocated memory.
45  * -- Florian Coosmann (FGC)
46  *
47  * Reads with count = 0 should always return 0.
48  * -- Julian Bradfield 1999-06-07.
49  *
50  * FIFOs and Pipes now generate SIGIO for both readers and writers.
51  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
52  *
53  * pipe_read & write cleanup
54  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
55  */
56 
57 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
58 {
59 	if (pipe->files)
60 		mutex_lock_nested(&pipe->mutex, subclass);
61 }
62 
63 void pipe_lock(struct pipe_inode_info *pipe)
64 {
65 	/*
66 	 * pipe_lock() nests non-pipe inode locks (for writing to a file)
67 	 */
68 	pipe_lock_nested(pipe, I_MUTEX_PARENT);
69 }
70 EXPORT_SYMBOL(pipe_lock);
71 
72 void pipe_unlock(struct pipe_inode_info *pipe)
73 {
74 	if (pipe->files)
75 		mutex_unlock(&pipe->mutex);
76 }
77 EXPORT_SYMBOL(pipe_unlock);
78 
79 static inline void __pipe_lock(struct pipe_inode_info *pipe)
80 {
81 	mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
82 }
83 
84 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
85 {
86 	mutex_unlock(&pipe->mutex);
87 }
88 
89 void pipe_double_lock(struct pipe_inode_info *pipe1,
90 		      struct pipe_inode_info *pipe2)
91 {
92 	BUG_ON(pipe1 == pipe2);
93 
94 	if (pipe1 < pipe2) {
95 		pipe_lock_nested(pipe1, I_MUTEX_PARENT);
96 		pipe_lock_nested(pipe2, I_MUTEX_CHILD);
97 	} else {
98 		pipe_lock_nested(pipe2, I_MUTEX_PARENT);
99 		pipe_lock_nested(pipe1, I_MUTEX_CHILD);
100 	}
101 }
102 
103 /* Drop the inode semaphore and wait for a pipe event, atomically */
104 void pipe_wait(struct pipe_inode_info *pipe)
105 {
106 	DEFINE_WAIT(wait);
107 
108 	/*
109 	 * Pipes are system-local resources, so sleeping on them
110 	 * is considered a noninteractive wait:
111 	 */
112 	prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
113 	pipe_unlock(pipe);
114 	schedule();
115 	finish_wait(&pipe->wait, &wait);
116 	pipe_lock(pipe);
117 }
118 
119 static int
120 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len,
121 			int atomic)
122 {
123 	unsigned long copy;
124 
125 	while (len > 0) {
126 		while (!iov->iov_len)
127 			iov++;
128 		copy = min_t(unsigned long, len, iov->iov_len);
129 
130 		if (atomic) {
131 			if (__copy_from_user_inatomic(to, iov->iov_base, copy))
132 				return -EFAULT;
133 		} else {
134 			if (copy_from_user(to, iov->iov_base, copy))
135 				return -EFAULT;
136 		}
137 		to += copy;
138 		len -= copy;
139 		iov->iov_base += copy;
140 		iov->iov_len -= copy;
141 	}
142 	return 0;
143 }
144 
145 static int
146 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len,
147 		      int atomic)
148 {
149 	unsigned long copy;
150 
151 	while (len > 0) {
152 		while (!iov->iov_len)
153 			iov++;
154 		copy = min_t(unsigned long, len, iov->iov_len);
155 
156 		if (atomic) {
157 			if (__copy_to_user_inatomic(iov->iov_base, from, copy))
158 				return -EFAULT;
159 		} else {
160 			if (copy_to_user(iov->iov_base, from, copy))
161 				return -EFAULT;
162 		}
163 		from += copy;
164 		len -= copy;
165 		iov->iov_base += copy;
166 		iov->iov_len -= copy;
167 	}
168 	return 0;
169 }
170 
171 /*
172  * Attempt to pre-fault in the user memory, so we can use atomic copies.
173  * Returns the number of bytes not faulted in.
174  */
175 static int iov_fault_in_pages_write(struct iovec *iov, unsigned long len)
176 {
177 	while (!iov->iov_len)
178 		iov++;
179 
180 	while (len > 0) {
181 		unsigned long this_len;
182 
183 		this_len = min_t(unsigned long, len, iov->iov_len);
184 		if (fault_in_pages_writeable(iov->iov_base, this_len))
185 			break;
186 
187 		len -= this_len;
188 		iov++;
189 	}
190 
191 	return len;
192 }
193 
194 /*
195  * Pre-fault in the user memory, so we can use atomic copies.
196  */
197 static void iov_fault_in_pages_read(struct iovec *iov, unsigned long len)
198 {
199 	while (!iov->iov_len)
200 		iov++;
201 
202 	while (len > 0) {
203 		unsigned long this_len;
204 
205 		this_len = min_t(unsigned long, len, iov->iov_len);
206 		fault_in_pages_readable(iov->iov_base, this_len);
207 		len -= this_len;
208 		iov++;
209 	}
210 }
211 
212 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
213 				  struct pipe_buffer *buf)
214 {
215 	struct page *page = buf->page;
216 
217 	/*
218 	 * If nobody else uses this page, and we don't already have a
219 	 * temporary page, let's keep track of it as a one-deep
220 	 * allocation cache. (Otherwise just release our reference to it)
221 	 */
222 	if (page_count(page) == 1 && !pipe->tmp_page)
223 		pipe->tmp_page = page;
224 	else
225 		page_cache_release(page);
226 }
227 
228 /**
229  * generic_pipe_buf_map - virtually map a pipe buffer
230  * @pipe:	the pipe that the buffer belongs to
231  * @buf:	the buffer that should be mapped
232  * @atomic:	whether to use an atomic map
233  *
234  * Description:
235  *	This function returns a kernel virtual address mapping for the
236  *	pipe_buffer passed in @buf. If @atomic is set, an atomic map is provided
237  *	and the caller has to be careful not to fault before calling
238  *	the unmap function.
239  *
240  *	Note that this function calls kmap_atomic() if @atomic != 0.
241  */
242 void *generic_pipe_buf_map(struct pipe_inode_info *pipe,
243 			   struct pipe_buffer *buf, int atomic)
244 {
245 	if (atomic) {
246 		buf->flags |= PIPE_BUF_FLAG_ATOMIC;
247 		return kmap_atomic(buf->page);
248 	}
249 
250 	return kmap(buf->page);
251 }
252 EXPORT_SYMBOL(generic_pipe_buf_map);
253 
254 /**
255  * generic_pipe_buf_unmap - unmap a previously mapped pipe buffer
256  * @pipe:	the pipe that the buffer belongs to
257  * @buf:	the buffer that should be unmapped
258  * @map_data:	the data that the mapping function returned
259  *
260  * Description:
261  *	This function undoes the mapping that ->map() provided.
262  */
263 void generic_pipe_buf_unmap(struct pipe_inode_info *pipe,
264 			    struct pipe_buffer *buf, void *map_data)
265 {
266 	if (buf->flags & PIPE_BUF_FLAG_ATOMIC) {
267 		buf->flags &= ~PIPE_BUF_FLAG_ATOMIC;
268 		kunmap_atomic(map_data);
269 	} else
270 		kunmap(buf->page);
271 }
272 EXPORT_SYMBOL(generic_pipe_buf_unmap);
273 
274 /**
275  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
276  * @pipe:	the pipe that the buffer belongs to
277  * @buf:	the buffer to attempt to steal
278  *
279  * Description:
280  *	This function attempts to steal the &struct page attached to
281  *	@buf. If successful, this function returns 0 and returns with
282  *	the page locked. The caller may then reuse the page for whatever
283  *	he wishes; the typical use is insertion into a different file
284  *	page cache.
285  */
286 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
287 			   struct pipe_buffer *buf)
288 {
289 	struct page *page = buf->page;
290 
291 	/*
292 	 * A reference of one is golden, that means that the owner of this
293 	 * page is the only one holding a reference to it. lock the page
294 	 * and return OK.
295 	 */
296 	if (page_count(page) == 1) {
297 		lock_page(page);
298 		return 0;
299 	}
300 
301 	return 1;
302 }
303 EXPORT_SYMBOL(generic_pipe_buf_steal);
304 
305 /**
306  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
307  * @pipe:	the pipe that the buffer belongs to
308  * @buf:	the buffer to get a reference to
309  *
310  * Description:
311  *	This function grabs an extra reference to @buf. It's used in
312  *	in the tee() system call, when we duplicate the buffers in one
313  *	pipe into another.
314  */
315 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
316 {
317 	page_cache_get(buf->page);
318 }
319 EXPORT_SYMBOL(generic_pipe_buf_get);
320 
321 /**
322  * generic_pipe_buf_confirm - verify contents of the pipe buffer
323  * @info:	the pipe that the buffer belongs to
324  * @buf:	the buffer to confirm
325  *
326  * Description:
327  *	This function does nothing, because the generic pipe code uses
328  *	pages that are always good when inserted into the pipe.
329  */
330 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
331 			     struct pipe_buffer *buf)
332 {
333 	return 0;
334 }
335 EXPORT_SYMBOL(generic_pipe_buf_confirm);
336 
337 /**
338  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
339  * @pipe:	the pipe that the buffer belongs to
340  * @buf:	the buffer to put a reference to
341  *
342  * Description:
343  *	This function releases a reference to @buf.
344  */
345 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
346 			      struct pipe_buffer *buf)
347 {
348 	page_cache_release(buf->page);
349 }
350 EXPORT_SYMBOL(generic_pipe_buf_release);
351 
352 static const struct pipe_buf_operations anon_pipe_buf_ops = {
353 	.can_merge = 1,
354 	.map = generic_pipe_buf_map,
355 	.unmap = generic_pipe_buf_unmap,
356 	.confirm = generic_pipe_buf_confirm,
357 	.release = anon_pipe_buf_release,
358 	.steal = generic_pipe_buf_steal,
359 	.get = generic_pipe_buf_get,
360 };
361 
362 static const struct pipe_buf_operations packet_pipe_buf_ops = {
363 	.can_merge = 0,
364 	.map = generic_pipe_buf_map,
365 	.unmap = generic_pipe_buf_unmap,
366 	.confirm = generic_pipe_buf_confirm,
367 	.release = anon_pipe_buf_release,
368 	.steal = generic_pipe_buf_steal,
369 	.get = generic_pipe_buf_get,
370 };
371 
372 static ssize_t
373 pipe_read(struct kiocb *iocb, const struct iovec *_iov,
374 	   unsigned long nr_segs, loff_t pos)
375 {
376 	struct file *filp = iocb->ki_filp;
377 	struct pipe_inode_info *pipe = filp->private_data;
378 	int do_wakeup;
379 	ssize_t ret;
380 	struct iovec *iov = (struct iovec *)_iov;
381 	size_t total_len;
382 
383 	total_len = iov_length(iov, nr_segs);
384 	/* Null read succeeds. */
385 	if (unlikely(total_len == 0))
386 		return 0;
387 
388 	do_wakeup = 0;
389 	ret = 0;
390 	__pipe_lock(pipe);
391 	for (;;) {
392 		int bufs = pipe->nrbufs;
393 		if (bufs) {
394 			int curbuf = pipe->curbuf;
395 			struct pipe_buffer *buf = pipe->bufs + curbuf;
396 			const struct pipe_buf_operations *ops = buf->ops;
397 			void *addr;
398 			size_t chars = buf->len;
399 			int error, atomic;
400 
401 			if (chars > total_len)
402 				chars = total_len;
403 
404 			error = ops->confirm(pipe, buf);
405 			if (error) {
406 				if (!ret)
407 					ret = error;
408 				break;
409 			}
410 
411 			atomic = !iov_fault_in_pages_write(iov, chars);
412 redo:
413 			addr = ops->map(pipe, buf, atomic);
414 			error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars, atomic);
415 			ops->unmap(pipe, buf, addr);
416 			if (unlikely(error)) {
417 				/*
418 				 * Just retry with the slow path if we failed.
419 				 */
420 				if (atomic) {
421 					atomic = 0;
422 					goto redo;
423 				}
424 				if (!ret)
425 					ret = error;
426 				break;
427 			}
428 			ret += chars;
429 			buf->offset += chars;
430 			buf->len -= chars;
431 
432 			/* Was it a packet buffer? Clean up and exit */
433 			if (buf->flags & PIPE_BUF_FLAG_PACKET) {
434 				total_len = chars;
435 				buf->len = 0;
436 			}
437 
438 			if (!buf->len) {
439 				buf->ops = NULL;
440 				ops->release(pipe, buf);
441 				curbuf = (curbuf + 1) & (pipe->buffers - 1);
442 				pipe->curbuf = curbuf;
443 				pipe->nrbufs = --bufs;
444 				do_wakeup = 1;
445 			}
446 			total_len -= chars;
447 			if (!total_len)
448 				break;	/* common path: read succeeded */
449 		}
450 		if (bufs)	/* More to do? */
451 			continue;
452 		if (!pipe->writers)
453 			break;
454 		if (!pipe->waiting_writers) {
455 			/* syscall merging: Usually we must not sleep
456 			 * if O_NONBLOCK is set, or if we got some data.
457 			 * But if a writer sleeps in kernel space, then
458 			 * we can wait for that data without violating POSIX.
459 			 */
460 			if (ret)
461 				break;
462 			if (filp->f_flags & O_NONBLOCK) {
463 				ret = -EAGAIN;
464 				break;
465 			}
466 		}
467 		if (signal_pending(current)) {
468 			if (!ret)
469 				ret = -ERESTARTSYS;
470 			break;
471 		}
472 		if (do_wakeup) {
473 			wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
474  			kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
475 		}
476 		pipe_wait(pipe);
477 	}
478 	__pipe_unlock(pipe);
479 
480 	/* Signal writers asynchronously that there is more room. */
481 	if (do_wakeup) {
482 		wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
483 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
484 	}
485 	if (ret > 0)
486 		file_accessed(filp);
487 	return ret;
488 }
489 
490 static inline int is_packetized(struct file *file)
491 {
492 	return (file->f_flags & O_DIRECT) != 0;
493 }
494 
495 static ssize_t
496 pipe_write(struct kiocb *iocb, const struct iovec *_iov,
497 	    unsigned long nr_segs, loff_t ppos)
498 {
499 	struct file *filp = iocb->ki_filp;
500 	struct pipe_inode_info *pipe = filp->private_data;
501 	ssize_t ret;
502 	int do_wakeup;
503 	struct iovec *iov = (struct iovec *)_iov;
504 	size_t total_len;
505 	ssize_t chars;
506 
507 	total_len = iov_length(iov, nr_segs);
508 	/* Null write succeeds. */
509 	if (unlikely(total_len == 0))
510 		return 0;
511 
512 	do_wakeup = 0;
513 	ret = 0;
514 	__pipe_lock(pipe);
515 
516 	if (!pipe->readers) {
517 		send_sig(SIGPIPE, current, 0);
518 		ret = -EPIPE;
519 		goto out;
520 	}
521 
522 	/* We try to merge small writes */
523 	chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
524 	if (pipe->nrbufs && chars != 0) {
525 		int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
526 							(pipe->buffers - 1);
527 		struct pipe_buffer *buf = pipe->bufs + lastbuf;
528 		const struct pipe_buf_operations *ops = buf->ops;
529 		int offset = buf->offset + buf->len;
530 
531 		if (ops->can_merge && offset + chars <= PAGE_SIZE) {
532 			int error, atomic = 1;
533 			void *addr;
534 
535 			error = ops->confirm(pipe, buf);
536 			if (error)
537 				goto out;
538 
539 			iov_fault_in_pages_read(iov, chars);
540 redo1:
541 			addr = ops->map(pipe, buf, atomic);
542 			error = pipe_iov_copy_from_user(offset + addr, iov,
543 							chars, atomic);
544 			ops->unmap(pipe, buf, addr);
545 			ret = error;
546 			do_wakeup = 1;
547 			if (error) {
548 				if (atomic) {
549 					atomic = 0;
550 					goto redo1;
551 				}
552 				goto out;
553 			}
554 			buf->len += chars;
555 			total_len -= chars;
556 			ret = chars;
557 			if (!total_len)
558 				goto out;
559 		}
560 	}
561 
562 	for (;;) {
563 		int bufs;
564 
565 		if (!pipe->readers) {
566 			send_sig(SIGPIPE, current, 0);
567 			if (!ret)
568 				ret = -EPIPE;
569 			break;
570 		}
571 		bufs = pipe->nrbufs;
572 		if (bufs < pipe->buffers) {
573 			int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
574 			struct pipe_buffer *buf = pipe->bufs + newbuf;
575 			struct page *page = pipe->tmp_page;
576 			char *src;
577 			int error, atomic = 1;
578 
579 			if (!page) {
580 				page = alloc_page(GFP_HIGHUSER);
581 				if (unlikely(!page)) {
582 					ret = ret ? : -ENOMEM;
583 					break;
584 				}
585 				pipe->tmp_page = page;
586 			}
587 			/* Always wake up, even if the copy fails. Otherwise
588 			 * we lock up (O_NONBLOCK-)readers that sleep due to
589 			 * syscall merging.
590 			 * FIXME! Is this really true?
591 			 */
592 			do_wakeup = 1;
593 			chars = PAGE_SIZE;
594 			if (chars > total_len)
595 				chars = total_len;
596 
597 			iov_fault_in_pages_read(iov, chars);
598 redo2:
599 			if (atomic)
600 				src = kmap_atomic(page);
601 			else
602 				src = kmap(page);
603 
604 			error = pipe_iov_copy_from_user(src, iov, chars,
605 							atomic);
606 			if (atomic)
607 				kunmap_atomic(src);
608 			else
609 				kunmap(page);
610 
611 			if (unlikely(error)) {
612 				if (atomic) {
613 					atomic = 0;
614 					goto redo2;
615 				}
616 				if (!ret)
617 					ret = error;
618 				break;
619 			}
620 			ret += chars;
621 
622 			/* Insert it into the buffer array */
623 			buf->page = page;
624 			buf->ops = &anon_pipe_buf_ops;
625 			buf->offset = 0;
626 			buf->len = chars;
627 			buf->flags = 0;
628 			if (is_packetized(filp)) {
629 				buf->ops = &packet_pipe_buf_ops;
630 				buf->flags = PIPE_BUF_FLAG_PACKET;
631 			}
632 			pipe->nrbufs = ++bufs;
633 			pipe->tmp_page = NULL;
634 
635 			total_len -= chars;
636 			if (!total_len)
637 				break;
638 		}
639 		if (bufs < pipe->buffers)
640 			continue;
641 		if (filp->f_flags & O_NONBLOCK) {
642 			if (!ret)
643 				ret = -EAGAIN;
644 			break;
645 		}
646 		if (signal_pending(current)) {
647 			if (!ret)
648 				ret = -ERESTARTSYS;
649 			break;
650 		}
651 		if (do_wakeup) {
652 			wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
653 			kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
654 			do_wakeup = 0;
655 		}
656 		pipe->waiting_writers++;
657 		pipe_wait(pipe);
658 		pipe->waiting_writers--;
659 	}
660 out:
661 	__pipe_unlock(pipe);
662 	if (do_wakeup) {
663 		wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
664 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
665 	}
666 	if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
667 		int err = file_update_time(filp);
668 		if (err)
669 			ret = err;
670 		sb_end_write(file_inode(filp)->i_sb);
671 	}
672 	return ret;
673 }
674 
675 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
676 {
677 	struct pipe_inode_info *pipe = filp->private_data;
678 	int count, buf, nrbufs;
679 
680 	switch (cmd) {
681 		case FIONREAD:
682 			__pipe_lock(pipe);
683 			count = 0;
684 			buf = pipe->curbuf;
685 			nrbufs = pipe->nrbufs;
686 			while (--nrbufs >= 0) {
687 				count += pipe->bufs[buf].len;
688 				buf = (buf+1) & (pipe->buffers - 1);
689 			}
690 			__pipe_unlock(pipe);
691 
692 			return put_user(count, (int __user *)arg);
693 		default:
694 			return -ENOIOCTLCMD;
695 	}
696 }
697 
698 /* No kernel lock held - fine */
699 static unsigned int
700 pipe_poll(struct file *filp, poll_table *wait)
701 {
702 	unsigned int mask;
703 	struct pipe_inode_info *pipe = filp->private_data;
704 	int nrbufs;
705 
706 	poll_wait(filp, &pipe->wait, wait);
707 
708 	/* Reading only -- no need for acquiring the semaphore.  */
709 	nrbufs = pipe->nrbufs;
710 	mask = 0;
711 	if (filp->f_mode & FMODE_READ) {
712 		mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
713 		if (!pipe->writers && filp->f_version != pipe->w_counter)
714 			mask |= POLLHUP;
715 	}
716 
717 	if (filp->f_mode & FMODE_WRITE) {
718 		mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
719 		/*
720 		 * Most Unices do not set POLLERR for FIFOs but on Linux they
721 		 * behave exactly like pipes for poll().
722 		 */
723 		if (!pipe->readers)
724 			mask |= POLLERR;
725 	}
726 
727 	return mask;
728 }
729 
730 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
731 {
732 	int kill = 0;
733 
734 	spin_lock(&inode->i_lock);
735 	if (!--pipe->files) {
736 		inode->i_pipe = NULL;
737 		kill = 1;
738 	}
739 	spin_unlock(&inode->i_lock);
740 
741 	if (kill)
742 		free_pipe_info(pipe);
743 }
744 
745 static int
746 pipe_release(struct inode *inode, struct file *file)
747 {
748 	struct pipe_inode_info *pipe = file->private_data;
749 
750 	__pipe_lock(pipe);
751 	if (file->f_mode & FMODE_READ)
752 		pipe->readers--;
753 	if (file->f_mode & FMODE_WRITE)
754 		pipe->writers--;
755 
756 	if (pipe->readers || pipe->writers) {
757 		wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
758 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
759 		kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
760 	}
761 	__pipe_unlock(pipe);
762 
763 	put_pipe_info(inode, pipe);
764 	return 0;
765 }
766 
767 static int
768 pipe_fasync(int fd, struct file *filp, int on)
769 {
770 	struct pipe_inode_info *pipe = filp->private_data;
771 	int retval = 0;
772 
773 	__pipe_lock(pipe);
774 	if (filp->f_mode & FMODE_READ)
775 		retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
776 	if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
777 		retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
778 		if (retval < 0 && (filp->f_mode & FMODE_READ))
779 			/* this can happen only if on == T */
780 			fasync_helper(-1, filp, 0, &pipe->fasync_readers);
781 	}
782 	__pipe_unlock(pipe);
783 	return retval;
784 }
785 
786 struct pipe_inode_info *alloc_pipe_info(void)
787 {
788 	struct pipe_inode_info *pipe;
789 
790 	pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
791 	if (pipe) {
792 		pipe->bufs = kzalloc(sizeof(struct pipe_buffer) * PIPE_DEF_BUFFERS, GFP_KERNEL);
793 		if (pipe->bufs) {
794 			init_waitqueue_head(&pipe->wait);
795 			pipe->r_counter = pipe->w_counter = 1;
796 			pipe->buffers = PIPE_DEF_BUFFERS;
797 			mutex_init(&pipe->mutex);
798 			return pipe;
799 		}
800 		kfree(pipe);
801 	}
802 
803 	return NULL;
804 }
805 
806 void free_pipe_info(struct pipe_inode_info *pipe)
807 {
808 	int i;
809 
810 	for (i = 0; i < pipe->buffers; i++) {
811 		struct pipe_buffer *buf = pipe->bufs + i;
812 		if (buf->ops)
813 			buf->ops->release(pipe, buf);
814 	}
815 	if (pipe->tmp_page)
816 		__free_page(pipe->tmp_page);
817 	kfree(pipe->bufs);
818 	kfree(pipe);
819 }
820 
821 static struct vfsmount *pipe_mnt __read_mostly;
822 
823 /*
824  * pipefs_dname() is called from d_path().
825  */
826 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
827 {
828 	return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
829 				dentry->d_inode->i_ino);
830 }
831 
832 static const struct dentry_operations pipefs_dentry_operations = {
833 	.d_dname	= pipefs_dname,
834 };
835 
836 static struct inode * get_pipe_inode(void)
837 {
838 	struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
839 	struct pipe_inode_info *pipe;
840 
841 	if (!inode)
842 		goto fail_inode;
843 
844 	inode->i_ino = get_next_ino();
845 
846 	pipe = alloc_pipe_info();
847 	if (!pipe)
848 		goto fail_iput;
849 
850 	inode->i_pipe = pipe;
851 	pipe->files = 2;
852 	pipe->readers = pipe->writers = 1;
853 	inode->i_fop = &pipefifo_fops;
854 
855 	/*
856 	 * Mark the inode dirty from the very beginning,
857 	 * that way it will never be moved to the dirty
858 	 * list because "mark_inode_dirty()" will think
859 	 * that it already _is_ on the dirty list.
860 	 */
861 	inode->i_state = I_DIRTY;
862 	inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
863 	inode->i_uid = current_fsuid();
864 	inode->i_gid = current_fsgid();
865 	inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
866 
867 	return inode;
868 
869 fail_iput:
870 	iput(inode);
871 
872 fail_inode:
873 	return NULL;
874 }
875 
876 int create_pipe_files(struct file **res, int flags)
877 {
878 	int err;
879 	struct inode *inode = get_pipe_inode();
880 	struct file *f;
881 	struct path path;
882 	static struct qstr name = { .name = "" };
883 
884 	if (!inode)
885 		return -ENFILE;
886 
887 	err = -ENOMEM;
888 	path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
889 	if (!path.dentry)
890 		goto err_inode;
891 	path.mnt = mntget(pipe_mnt);
892 
893 	d_instantiate(path.dentry, inode);
894 
895 	err = -ENFILE;
896 	f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
897 	if (IS_ERR(f))
898 		goto err_dentry;
899 
900 	f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
901 	f->private_data = inode->i_pipe;
902 
903 	res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
904 	if (IS_ERR(res[0]))
905 		goto err_file;
906 
907 	path_get(&path);
908 	res[0]->private_data = inode->i_pipe;
909 	res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
910 	res[1] = f;
911 	return 0;
912 
913 err_file:
914 	put_filp(f);
915 err_dentry:
916 	free_pipe_info(inode->i_pipe);
917 	path_put(&path);
918 	return err;
919 
920 err_inode:
921 	free_pipe_info(inode->i_pipe);
922 	iput(inode);
923 	return err;
924 }
925 
926 static int __do_pipe_flags(int *fd, struct file **files, int flags)
927 {
928 	int error;
929 	int fdw, fdr;
930 
931 	if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
932 		return -EINVAL;
933 
934 	error = create_pipe_files(files, flags);
935 	if (error)
936 		return error;
937 
938 	error = get_unused_fd_flags(flags);
939 	if (error < 0)
940 		goto err_read_pipe;
941 	fdr = error;
942 
943 	error = get_unused_fd_flags(flags);
944 	if (error < 0)
945 		goto err_fdr;
946 	fdw = error;
947 
948 	audit_fd_pair(fdr, fdw);
949 	fd[0] = fdr;
950 	fd[1] = fdw;
951 	return 0;
952 
953  err_fdr:
954 	put_unused_fd(fdr);
955  err_read_pipe:
956 	fput(files[0]);
957 	fput(files[1]);
958 	return error;
959 }
960 
961 int do_pipe_flags(int *fd, int flags)
962 {
963 	struct file *files[2];
964 	int error = __do_pipe_flags(fd, files, flags);
965 	if (!error) {
966 		fd_install(fd[0], files[0]);
967 		fd_install(fd[1], files[1]);
968 	}
969 	return error;
970 }
971 
972 /*
973  * sys_pipe() is the normal C calling standard for creating
974  * a pipe. It's not the way Unix traditionally does this, though.
975  */
976 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
977 {
978 	struct file *files[2];
979 	int fd[2];
980 	int error;
981 
982 	error = __do_pipe_flags(fd, files, flags);
983 	if (!error) {
984 		if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
985 			fput(files[0]);
986 			fput(files[1]);
987 			put_unused_fd(fd[0]);
988 			put_unused_fd(fd[1]);
989 			error = -EFAULT;
990 		} else {
991 			fd_install(fd[0], files[0]);
992 			fd_install(fd[1], files[1]);
993 		}
994 	}
995 	return error;
996 }
997 
998 SYSCALL_DEFINE1(pipe, int __user *, fildes)
999 {
1000 	return sys_pipe2(fildes, 0);
1001 }
1002 
1003 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
1004 {
1005 	int cur = *cnt;
1006 
1007 	while (cur == *cnt) {
1008 		pipe_wait(pipe);
1009 		if (signal_pending(current))
1010 			break;
1011 	}
1012 	return cur == *cnt ? -ERESTARTSYS : 0;
1013 }
1014 
1015 static void wake_up_partner(struct pipe_inode_info *pipe)
1016 {
1017 	wake_up_interruptible(&pipe->wait);
1018 }
1019 
1020 static int fifo_open(struct inode *inode, struct file *filp)
1021 {
1022 	struct pipe_inode_info *pipe;
1023 	bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
1024 	int ret;
1025 
1026 	filp->f_version = 0;
1027 
1028 	spin_lock(&inode->i_lock);
1029 	if (inode->i_pipe) {
1030 		pipe = inode->i_pipe;
1031 		pipe->files++;
1032 		spin_unlock(&inode->i_lock);
1033 	} else {
1034 		spin_unlock(&inode->i_lock);
1035 		pipe = alloc_pipe_info();
1036 		if (!pipe)
1037 			return -ENOMEM;
1038 		pipe->files = 1;
1039 		spin_lock(&inode->i_lock);
1040 		if (unlikely(inode->i_pipe)) {
1041 			inode->i_pipe->files++;
1042 			spin_unlock(&inode->i_lock);
1043 			free_pipe_info(pipe);
1044 			pipe = inode->i_pipe;
1045 		} else {
1046 			inode->i_pipe = pipe;
1047 			spin_unlock(&inode->i_lock);
1048 		}
1049 	}
1050 	filp->private_data = pipe;
1051 	/* OK, we have a pipe and it's pinned down */
1052 
1053 	__pipe_lock(pipe);
1054 
1055 	/* We can only do regular read/write on fifos */
1056 	filp->f_mode &= (FMODE_READ | FMODE_WRITE);
1057 
1058 	switch (filp->f_mode) {
1059 	case FMODE_READ:
1060 	/*
1061 	 *  O_RDONLY
1062 	 *  POSIX.1 says that O_NONBLOCK means return with the FIFO
1063 	 *  opened, even when there is no process writing the FIFO.
1064 	 */
1065 		pipe->r_counter++;
1066 		if (pipe->readers++ == 0)
1067 			wake_up_partner(pipe);
1068 
1069 		if (!is_pipe && !pipe->writers) {
1070 			if ((filp->f_flags & O_NONBLOCK)) {
1071 				/* suppress POLLHUP until we have
1072 				 * seen a writer */
1073 				filp->f_version = pipe->w_counter;
1074 			} else {
1075 				if (wait_for_partner(pipe, &pipe->w_counter))
1076 					goto err_rd;
1077 			}
1078 		}
1079 		break;
1080 
1081 	case FMODE_WRITE:
1082 	/*
1083 	 *  O_WRONLY
1084 	 *  POSIX.1 says that O_NONBLOCK means return -1 with
1085 	 *  errno=ENXIO when there is no process reading the FIFO.
1086 	 */
1087 		ret = -ENXIO;
1088 		if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
1089 			goto err;
1090 
1091 		pipe->w_counter++;
1092 		if (!pipe->writers++)
1093 			wake_up_partner(pipe);
1094 
1095 		if (!is_pipe && !pipe->readers) {
1096 			if (wait_for_partner(pipe, &pipe->r_counter))
1097 				goto err_wr;
1098 		}
1099 		break;
1100 
1101 	case FMODE_READ | FMODE_WRITE:
1102 	/*
1103 	 *  O_RDWR
1104 	 *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
1105 	 *  This implementation will NEVER block on a O_RDWR open, since
1106 	 *  the process can at least talk to itself.
1107 	 */
1108 
1109 		pipe->readers++;
1110 		pipe->writers++;
1111 		pipe->r_counter++;
1112 		pipe->w_counter++;
1113 		if (pipe->readers == 1 || pipe->writers == 1)
1114 			wake_up_partner(pipe);
1115 		break;
1116 
1117 	default:
1118 		ret = -EINVAL;
1119 		goto err;
1120 	}
1121 
1122 	/* Ok! */
1123 	__pipe_unlock(pipe);
1124 	return 0;
1125 
1126 err_rd:
1127 	if (!--pipe->readers)
1128 		wake_up_interruptible(&pipe->wait);
1129 	ret = -ERESTARTSYS;
1130 	goto err;
1131 
1132 err_wr:
1133 	if (!--pipe->writers)
1134 		wake_up_interruptible(&pipe->wait);
1135 	ret = -ERESTARTSYS;
1136 	goto err;
1137 
1138 err:
1139 	__pipe_unlock(pipe);
1140 
1141 	put_pipe_info(inode, pipe);
1142 	return ret;
1143 }
1144 
1145 const struct file_operations pipefifo_fops = {
1146 	.open		= fifo_open,
1147 	.llseek		= no_llseek,
1148 	.read		= do_sync_read,
1149 	.aio_read	= pipe_read,
1150 	.write		= do_sync_write,
1151 	.aio_write	= pipe_write,
1152 	.poll		= pipe_poll,
1153 	.unlocked_ioctl	= pipe_ioctl,
1154 	.release	= pipe_release,
1155 	.fasync		= pipe_fasync,
1156 };
1157 
1158 /*
1159  * Allocate a new array of pipe buffers and copy the info over. Returns the
1160  * pipe size if successful, or return -ERROR on error.
1161  */
1162 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long nr_pages)
1163 {
1164 	struct pipe_buffer *bufs;
1165 
1166 	/*
1167 	 * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1168 	 * expect a lot of shrink+grow operations, just free and allocate
1169 	 * again like we would do for growing. If the pipe currently
1170 	 * contains more buffers than arg, then return busy.
1171 	 */
1172 	if (nr_pages < pipe->nrbufs)
1173 		return -EBUSY;
1174 
1175 	bufs = kcalloc(nr_pages, sizeof(*bufs), GFP_KERNEL | __GFP_NOWARN);
1176 	if (unlikely(!bufs))
1177 		return -ENOMEM;
1178 
1179 	/*
1180 	 * The pipe array wraps around, so just start the new one at zero
1181 	 * and adjust the indexes.
1182 	 */
1183 	if (pipe->nrbufs) {
1184 		unsigned int tail;
1185 		unsigned int head;
1186 
1187 		tail = pipe->curbuf + pipe->nrbufs;
1188 		if (tail < pipe->buffers)
1189 			tail = 0;
1190 		else
1191 			tail &= (pipe->buffers - 1);
1192 
1193 		head = pipe->nrbufs - tail;
1194 		if (head)
1195 			memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1196 		if (tail)
1197 			memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1198 	}
1199 
1200 	pipe->curbuf = 0;
1201 	kfree(pipe->bufs);
1202 	pipe->bufs = bufs;
1203 	pipe->buffers = nr_pages;
1204 	return nr_pages * PAGE_SIZE;
1205 }
1206 
1207 /*
1208  * Currently we rely on the pipe array holding a power-of-2 number
1209  * of pages.
1210  */
1211 static inline unsigned int round_pipe_size(unsigned int size)
1212 {
1213 	unsigned long nr_pages;
1214 
1215 	nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1216 	return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1217 }
1218 
1219 /*
1220  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1221  * will return an error.
1222  */
1223 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1224 		 size_t *lenp, loff_t *ppos)
1225 {
1226 	int ret;
1227 
1228 	ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1229 	if (ret < 0 || !write)
1230 		return ret;
1231 
1232 	pipe_max_size = round_pipe_size(pipe_max_size);
1233 	return ret;
1234 }
1235 
1236 /*
1237  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1238  * location, so checking ->i_pipe is not enough to verify that this is a
1239  * pipe.
1240  */
1241 struct pipe_inode_info *get_pipe_info(struct file *file)
1242 {
1243 	return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1244 }
1245 
1246 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1247 {
1248 	struct pipe_inode_info *pipe;
1249 	long ret;
1250 
1251 	pipe = get_pipe_info(file);
1252 	if (!pipe)
1253 		return -EBADF;
1254 
1255 	__pipe_lock(pipe);
1256 
1257 	switch (cmd) {
1258 	case F_SETPIPE_SZ: {
1259 		unsigned int size, nr_pages;
1260 
1261 		size = round_pipe_size(arg);
1262 		nr_pages = size >> PAGE_SHIFT;
1263 
1264 		ret = -EINVAL;
1265 		if (!nr_pages)
1266 			goto out;
1267 
1268 		if (!capable(CAP_SYS_RESOURCE) && size > pipe_max_size) {
1269 			ret = -EPERM;
1270 			goto out;
1271 		}
1272 		ret = pipe_set_size(pipe, nr_pages);
1273 		break;
1274 		}
1275 	case F_GETPIPE_SZ:
1276 		ret = pipe->buffers * PAGE_SIZE;
1277 		break;
1278 	default:
1279 		ret = -EINVAL;
1280 		break;
1281 	}
1282 
1283 out:
1284 	__pipe_unlock(pipe);
1285 	return ret;
1286 }
1287 
1288 static const struct super_operations pipefs_ops = {
1289 	.destroy_inode = free_inode_nonrcu,
1290 	.statfs = simple_statfs,
1291 };
1292 
1293 /*
1294  * pipefs should _never_ be mounted by userland - too much of security hassle,
1295  * no real gain from having the whole whorehouse mounted. So we don't need
1296  * any operations on the root directory. However, we need a non-trivial
1297  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1298  */
1299 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1300 			 int flags, const char *dev_name, void *data)
1301 {
1302 	return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1303 			&pipefs_dentry_operations, PIPEFS_MAGIC);
1304 }
1305 
1306 static struct file_system_type pipe_fs_type = {
1307 	.name		= "pipefs",
1308 	.mount		= pipefs_mount,
1309 	.kill_sb	= kill_anon_super,
1310 };
1311 
1312 static int __init init_pipe_fs(void)
1313 {
1314 	int err = register_filesystem(&pipe_fs_type);
1315 
1316 	if (!err) {
1317 		pipe_mnt = kern_mount(&pipe_fs_type);
1318 		if (IS_ERR(pipe_mnt)) {
1319 			err = PTR_ERR(pipe_mnt);
1320 			unregister_filesystem(&pipe_fs_type);
1321 		}
1322 	}
1323 	return err;
1324 }
1325 
1326 fs_initcall(init_pipe_fs);
1327