xref: /linux/fs/pipe.c (revision 2624f124b3b5d550ab2fbef7ee3bc0e1fed09722)
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/mount.h>
15 #include <linux/pipe_fs_i.h>
16 #include <linux/uio.h>
17 #include <linux/highmem.h>
18 
19 #include <asm/uaccess.h>
20 #include <asm/ioctls.h>
21 
22 /*
23  * We use a start+len construction, which provides full use of the
24  * allocated memory.
25  * -- Florian Coosmann (FGC)
26  *
27  * Reads with count = 0 should always return 0.
28  * -- Julian Bradfield 1999-06-07.
29  *
30  * FIFOs and Pipes now generate SIGIO for both readers and writers.
31  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
32  *
33  * pipe_read & write cleanup
34  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
35  */
36 
37 /* Drop the inode semaphore and wait for a pipe event, atomically */
38 void pipe_wait(struct inode * inode)
39 {
40 	DEFINE_WAIT(wait);
41 
42 	prepare_to_wait(PIPE_WAIT(*inode), &wait, TASK_INTERRUPTIBLE);
43 	up(PIPE_SEM(*inode));
44 	schedule();
45 	finish_wait(PIPE_WAIT(*inode), &wait);
46 	down(PIPE_SEM(*inode));
47 }
48 
49 static inline int
50 pipe_iov_copy_from_user(void *to, struct iovec *iov, unsigned long len)
51 {
52 	unsigned long copy;
53 
54 	while (len > 0) {
55 		while (!iov->iov_len)
56 			iov++;
57 		copy = min_t(unsigned long, len, iov->iov_len);
58 
59 		if (copy_from_user(to, iov->iov_base, copy))
60 			return -EFAULT;
61 		to += copy;
62 		len -= copy;
63 		iov->iov_base += copy;
64 		iov->iov_len -= copy;
65 	}
66 	return 0;
67 }
68 
69 static inline int
70 pipe_iov_copy_to_user(struct iovec *iov, const void *from, unsigned long len)
71 {
72 	unsigned long copy;
73 
74 	while (len > 0) {
75 		while (!iov->iov_len)
76 			iov++;
77 		copy = min_t(unsigned long, len, iov->iov_len);
78 
79 		if (copy_to_user(iov->iov_base, from, copy))
80 			return -EFAULT;
81 		from += copy;
82 		len -= copy;
83 		iov->iov_base += copy;
84 		iov->iov_len -= copy;
85 	}
86 	return 0;
87 }
88 
89 static void anon_pipe_buf_release(struct pipe_inode_info *info, struct pipe_buffer *buf)
90 {
91 	struct page *page = buf->page;
92 
93 	if (info->tmp_page) {
94 		__free_page(page);
95 		return;
96 	}
97 	info->tmp_page = page;
98 }
99 
100 static void *anon_pipe_buf_map(struct file *file, struct pipe_inode_info *info, struct pipe_buffer *buf)
101 {
102 	return kmap(buf->page);
103 }
104 
105 static void anon_pipe_buf_unmap(struct pipe_inode_info *info, struct pipe_buffer *buf)
106 {
107 	kunmap(buf->page);
108 }
109 
110 static struct pipe_buf_operations anon_pipe_buf_ops = {
111 	.can_merge = 1,
112 	.map = anon_pipe_buf_map,
113 	.unmap = anon_pipe_buf_unmap,
114 	.release = anon_pipe_buf_release,
115 };
116 
117 static ssize_t
118 pipe_readv(struct file *filp, const struct iovec *_iov,
119 	   unsigned long nr_segs, loff_t *ppos)
120 {
121 	struct inode *inode = filp->f_dentry->d_inode;
122 	struct pipe_inode_info *info;
123 	int do_wakeup;
124 	ssize_t ret;
125 	struct iovec *iov = (struct iovec *)_iov;
126 	size_t total_len;
127 
128 	total_len = iov_length(iov, nr_segs);
129 	/* Null read succeeds. */
130 	if (unlikely(total_len == 0))
131 		return 0;
132 
133 	do_wakeup = 0;
134 	ret = 0;
135 	down(PIPE_SEM(*inode));
136 	info = inode->i_pipe;
137 	for (;;) {
138 		int bufs = info->nrbufs;
139 		if (bufs) {
140 			int curbuf = info->curbuf;
141 			struct pipe_buffer *buf = info->bufs + curbuf;
142 			struct pipe_buf_operations *ops = buf->ops;
143 			void *addr;
144 			size_t chars = buf->len;
145 			int error;
146 
147 			if (chars > total_len)
148 				chars = total_len;
149 
150 			addr = ops->map(filp, info, buf);
151 			error = pipe_iov_copy_to_user(iov, addr + buf->offset, chars);
152 			ops->unmap(info, buf);
153 			if (unlikely(error)) {
154 				if (!ret) ret = -EFAULT;
155 				break;
156 			}
157 			ret += chars;
158 			buf->offset += chars;
159 			buf->len -= chars;
160 			if (!buf->len) {
161 				buf->ops = NULL;
162 				ops->release(info, buf);
163 				curbuf = (curbuf + 1) & (PIPE_BUFFERS-1);
164 				info->curbuf = curbuf;
165 				info->nrbufs = --bufs;
166 				do_wakeup = 1;
167 			}
168 			total_len -= chars;
169 			if (!total_len)
170 				break;	/* common path: read succeeded */
171 		}
172 		if (bufs)	/* More to do? */
173 			continue;
174 		if (!PIPE_WRITERS(*inode))
175 			break;
176 		if (!PIPE_WAITING_WRITERS(*inode)) {
177 			/* syscall merging: Usually we must not sleep
178 			 * if O_NONBLOCK is set, or if we got some data.
179 			 * But if a writer sleeps in kernel space, then
180 			 * we can wait for that data without violating POSIX.
181 			 */
182 			if (ret)
183 				break;
184 			if (filp->f_flags & O_NONBLOCK) {
185 				ret = -EAGAIN;
186 				break;
187 			}
188 		}
189 		if (signal_pending(current)) {
190 			if (!ret) ret = -ERESTARTSYS;
191 			break;
192 		}
193 		if (do_wakeup) {
194 			wake_up_interruptible_sync(PIPE_WAIT(*inode));
195  			kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
196 		}
197 		pipe_wait(inode);
198 	}
199 	up(PIPE_SEM(*inode));
200 	/* Signal writers asynchronously that there is more room.  */
201 	if (do_wakeup) {
202 		wake_up_interruptible(PIPE_WAIT(*inode));
203 		kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
204 	}
205 	if (ret > 0)
206 		file_accessed(filp);
207 	return ret;
208 }
209 
210 static ssize_t
211 pipe_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
212 {
213 	struct iovec iov = { .iov_base = buf, .iov_len = count };
214 	return pipe_readv(filp, &iov, 1, ppos);
215 }
216 
217 static ssize_t
218 pipe_writev(struct file *filp, const struct iovec *_iov,
219 	    unsigned long nr_segs, loff_t *ppos)
220 {
221 	struct inode *inode = filp->f_dentry->d_inode;
222 	struct pipe_inode_info *info;
223 	ssize_t ret;
224 	int do_wakeup;
225 	struct iovec *iov = (struct iovec *)_iov;
226 	size_t total_len;
227 	ssize_t chars;
228 
229 	total_len = iov_length(iov, nr_segs);
230 	/* Null write succeeds. */
231 	if (unlikely(total_len == 0))
232 		return 0;
233 
234 	do_wakeup = 0;
235 	ret = 0;
236 	down(PIPE_SEM(*inode));
237 	info = inode->i_pipe;
238 
239 	if (!PIPE_READERS(*inode)) {
240 		send_sig(SIGPIPE, current, 0);
241 		ret = -EPIPE;
242 		goto out;
243 	}
244 
245 	/* We try to merge small writes */
246 	chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
247 	if (info->nrbufs && chars != 0) {
248 		int lastbuf = (info->curbuf + info->nrbufs - 1) & (PIPE_BUFFERS-1);
249 		struct pipe_buffer *buf = info->bufs + lastbuf;
250 		struct pipe_buf_operations *ops = buf->ops;
251 		int offset = buf->offset + buf->len;
252 		if (ops->can_merge && offset + chars <= PAGE_SIZE) {
253 			void *addr = ops->map(filp, info, buf);
254 			int error = pipe_iov_copy_from_user(offset + addr, iov, chars);
255 			ops->unmap(info, buf);
256 			ret = error;
257 			do_wakeup = 1;
258 			if (error)
259 				goto out;
260 			buf->len += chars;
261 			total_len -= chars;
262 			ret = chars;
263 			if (!total_len)
264 				goto out;
265 		}
266 	}
267 
268 	for (;;) {
269 		int bufs;
270 		if (!PIPE_READERS(*inode)) {
271 			send_sig(SIGPIPE, current, 0);
272 			if (!ret) ret = -EPIPE;
273 			break;
274 		}
275 		bufs = info->nrbufs;
276 		if (bufs < PIPE_BUFFERS) {
277 			int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS-1);
278 			struct pipe_buffer *buf = info->bufs + newbuf;
279 			struct page *page = info->tmp_page;
280 			int error;
281 
282 			if (!page) {
283 				page = alloc_page(GFP_HIGHUSER);
284 				if (unlikely(!page)) {
285 					ret = ret ? : -ENOMEM;
286 					break;
287 				}
288 				info->tmp_page = page;
289 			}
290 			/* Always wakeup, even if the copy fails. Otherwise
291 			 * we lock up (O_NONBLOCK-)readers that sleep due to
292 			 * syscall merging.
293 			 * FIXME! Is this really true?
294 			 */
295 			do_wakeup = 1;
296 			chars = PAGE_SIZE;
297 			if (chars > total_len)
298 				chars = total_len;
299 
300 			error = pipe_iov_copy_from_user(kmap(page), iov, chars);
301 			kunmap(page);
302 			if (unlikely(error)) {
303 				if (!ret) ret = -EFAULT;
304 				break;
305 			}
306 			ret += chars;
307 
308 			/* Insert it into the buffer array */
309 			buf->page = page;
310 			buf->ops = &anon_pipe_buf_ops;
311 			buf->offset = 0;
312 			buf->len = chars;
313 			info->nrbufs = ++bufs;
314 			info->tmp_page = NULL;
315 
316 			total_len -= chars;
317 			if (!total_len)
318 				break;
319 		}
320 		if (bufs < PIPE_BUFFERS)
321 			continue;
322 		if (filp->f_flags & O_NONBLOCK) {
323 			if (!ret) ret = -EAGAIN;
324 			break;
325 		}
326 		if (signal_pending(current)) {
327 			if (!ret) ret = -ERESTARTSYS;
328 			break;
329 		}
330 		if (do_wakeup) {
331 			wake_up_interruptible_sync(PIPE_WAIT(*inode));
332 			kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
333 			do_wakeup = 0;
334 		}
335 		PIPE_WAITING_WRITERS(*inode)++;
336 		pipe_wait(inode);
337 		PIPE_WAITING_WRITERS(*inode)--;
338 	}
339 out:
340 	up(PIPE_SEM(*inode));
341 	if (do_wakeup) {
342 		wake_up_interruptible(PIPE_WAIT(*inode));
343 		kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
344 	}
345 	if (ret > 0)
346 		inode_update_time(inode, 1);	/* mtime and ctime */
347 	return ret;
348 }
349 
350 static ssize_t
351 pipe_write(struct file *filp, const char __user *buf,
352 	   size_t count, loff_t *ppos)
353 {
354 	struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
355 	return pipe_writev(filp, &iov, 1, ppos);
356 }
357 
358 static ssize_t
359 bad_pipe_r(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
360 {
361 	return -EBADF;
362 }
363 
364 static ssize_t
365 bad_pipe_w(struct file *filp, const char __user *buf, size_t count, loff_t *ppos)
366 {
367 	return -EBADF;
368 }
369 
370 static int
371 pipe_ioctl(struct inode *pino, struct file *filp,
372 	   unsigned int cmd, unsigned long arg)
373 {
374 	struct inode *inode = filp->f_dentry->d_inode;
375 	struct pipe_inode_info *info;
376 	int count, buf, nrbufs;
377 
378 	switch (cmd) {
379 		case FIONREAD:
380 			down(PIPE_SEM(*inode));
381 			info =  inode->i_pipe;
382 			count = 0;
383 			buf = info->curbuf;
384 			nrbufs = info->nrbufs;
385 			while (--nrbufs >= 0) {
386 				count += info->bufs[buf].len;
387 				buf = (buf+1) & (PIPE_BUFFERS-1);
388 			}
389 			up(PIPE_SEM(*inode));
390 			return put_user(count, (int __user *)arg);
391 		default:
392 			return -EINVAL;
393 	}
394 }
395 
396 /* No kernel lock held - fine */
397 static unsigned int
398 pipe_poll(struct file *filp, poll_table *wait)
399 {
400 	unsigned int mask;
401 	struct inode *inode = filp->f_dentry->d_inode;
402 	struct pipe_inode_info *info = inode->i_pipe;
403 	int nrbufs;
404 
405 	poll_wait(filp, PIPE_WAIT(*inode), wait);
406 
407 	/* Reading only -- no need for acquiring the semaphore.  */
408 	nrbufs = info->nrbufs;
409 	mask = 0;
410 	if (filp->f_mode & FMODE_READ) {
411 		mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
412 		if (!PIPE_WRITERS(*inode) && filp->f_version != PIPE_WCOUNTER(*inode))
413 			mask |= POLLHUP;
414 	}
415 
416 	if (filp->f_mode & FMODE_WRITE) {
417 		mask |= (nrbufs < PIPE_BUFFERS) ? POLLOUT | POLLWRNORM : 0;
418 		/*
419 		 * Most Unices do not set POLLERR for FIFOs but on Linux they
420 		 * behave exactly like pipes for poll().
421 		 */
422 		if (!PIPE_READERS(*inode))
423 			mask |= POLLERR;
424 	}
425 
426 	return mask;
427 }
428 
429 static int
430 pipe_release(struct inode *inode, int decr, int decw)
431 {
432 	down(PIPE_SEM(*inode));
433 	PIPE_READERS(*inode) -= decr;
434 	PIPE_WRITERS(*inode) -= decw;
435 	if (!PIPE_READERS(*inode) && !PIPE_WRITERS(*inode)) {
436 		free_pipe_info(inode);
437 	} else {
438 		wake_up_interruptible(PIPE_WAIT(*inode));
439 		kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
440 		kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
441 	}
442 	up(PIPE_SEM(*inode));
443 
444 	return 0;
445 }
446 
447 static int
448 pipe_read_fasync(int fd, struct file *filp, int on)
449 {
450 	struct inode *inode = filp->f_dentry->d_inode;
451 	int retval;
452 
453 	down(PIPE_SEM(*inode));
454 	retval = fasync_helper(fd, filp, on, PIPE_FASYNC_READERS(*inode));
455 	up(PIPE_SEM(*inode));
456 
457 	if (retval < 0)
458 		return retval;
459 
460 	return 0;
461 }
462 
463 
464 static int
465 pipe_write_fasync(int fd, struct file *filp, int on)
466 {
467 	struct inode *inode = filp->f_dentry->d_inode;
468 	int retval;
469 
470 	down(PIPE_SEM(*inode));
471 	retval = fasync_helper(fd, filp, on, PIPE_FASYNC_WRITERS(*inode));
472 	up(PIPE_SEM(*inode));
473 
474 	if (retval < 0)
475 		return retval;
476 
477 	return 0;
478 }
479 
480 
481 static int
482 pipe_rdwr_fasync(int fd, struct file *filp, int on)
483 {
484 	struct inode *inode = filp->f_dentry->d_inode;
485 	int retval;
486 
487 	down(PIPE_SEM(*inode));
488 
489 	retval = fasync_helper(fd, filp, on, PIPE_FASYNC_READERS(*inode));
490 
491 	if (retval >= 0)
492 		retval = fasync_helper(fd, filp, on, PIPE_FASYNC_WRITERS(*inode));
493 
494 	up(PIPE_SEM(*inode));
495 
496 	if (retval < 0)
497 		return retval;
498 
499 	return 0;
500 }
501 
502 
503 static int
504 pipe_read_release(struct inode *inode, struct file *filp)
505 {
506 	pipe_read_fasync(-1, filp, 0);
507 	return pipe_release(inode, 1, 0);
508 }
509 
510 static int
511 pipe_write_release(struct inode *inode, struct file *filp)
512 {
513 	pipe_write_fasync(-1, filp, 0);
514 	return pipe_release(inode, 0, 1);
515 }
516 
517 static int
518 pipe_rdwr_release(struct inode *inode, struct file *filp)
519 {
520 	int decr, decw;
521 
522 	pipe_rdwr_fasync(-1, filp, 0);
523 	decr = (filp->f_mode & FMODE_READ) != 0;
524 	decw = (filp->f_mode & FMODE_WRITE) != 0;
525 	return pipe_release(inode, decr, decw);
526 }
527 
528 static int
529 pipe_read_open(struct inode *inode, struct file *filp)
530 {
531 	/* We could have perhaps used atomic_t, but this and friends
532 	   below are the only places.  So it doesn't seem worthwhile.  */
533 	down(PIPE_SEM(*inode));
534 	PIPE_READERS(*inode)++;
535 	up(PIPE_SEM(*inode));
536 
537 	return 0;
538 }
539 
540 static int
541 pipe_write_open(struct inode *inode, struct file *filp)
542 {
543 	down(PIPE_SEM(*inode));
544 	PIPE_WRITERS(*inode)++;
545 	up(PIPE_SEM(*inode));
546 
547 	return 0;
548 }
549 
550 static int
551 pipe_rdwr_open(struct inode *inode, struct file *filp)
552 {
553 	down(PIPE_SEM(*inode));
554 	if (filp->f_mode & FMODE_READ)
555 		PIPE_READERS(*inode)++;
556 	if (filp->f_mode & FMODE_WRITE)
557 		PIPE_WRITERS(*inode)++;
558 	up(PIPE_SEM(*inode));
559 
560 	return 0;
561 }
562 
563 /*
564  * The file_operations structs are not static because they
565  * are also used in linux/fs/fifo.c to do operations on FIFOs.
566  */
567 struct file_operations read_fifo_fops = {
568 	.llseek		= no_llseek,
569 	.read		= pipe_read,
570 	.readv		= pipe_readv,
571 	.write		= bad_pipe_w,
572 	.poll		= pipe_poll,
573 	.ioctl		= pipe_ioctl,
574 	.open		= pipe_read_open,
575 	.release	= pipe_read_release,
576 	.fasync		= pipe_read_fasync,
577 };
578 
579 struct file_operations write_fifo_fops = {
580 	.llseek		= no_llseek,
581 	.read		= bad_pipe_r,
582 	.write		= pipe_write,
583 	.writev		= pipe_writev,
584 	.poll		= pipe_poll,
585 	.ioctl		= pipe_ioctl,
586 	.open		= pipe_write_open,
587 	.release	= pipe_write_release,
588 	.fasync		= pipe_write_fasync,
589 };
590 
591 struct file_operations rdwr_fifo_fops = {
592 	.llseek		= no_llseek,
593 	.read		= pipe_read,
594 	.readv		= pipe_readv,
595 	.write		= pipe_write,
596 	.writev		= pipe_writev,
597 	.poll		= pipe_poll,
598 	.ioctl		= pipe_ioctl,
599 	.open		= pipe_rdwr_open,
600 	.release	= pipe_rdwr_release,
601 	.fasync		= pipe_rdwr_fasync,
602 };
603 
604 struct file_operations read_pipe_fops = {
605 	.llseek		= no_llseek,
606 	.read		= pipe_read,
607 	.readv		= pipe_readv,
608 	.write		= bad_pipe_w,
609 	.poll		= pipe_poll,
610 	.ioctl		= pipe_ioctl,
611 	.open		= pipe_read_open,
612 	.release	= pipe_read_release,
613 	.fasync		= pipe_read_fasync,
614 };
615 
616 struct file_operations write_pipe_fops = {
617 	.llseek		= no_llseek,
618 	.read		= bad_pipe_r,
619 	.write		= pipe_write,
620 	.writev		= pipe_writev,
621 	.poll		= pipe_poll,
622 	.ioctl		= pipe_ioctl,
623 	.open		= pipe_write_open,
624 	.release	= pipe_write_release,
625 	.fasync		= pipe_write_fasync,
626 };
627 
628 struct file_operations rdwr_pipe_fops = {
629 	.llseek		= no_llseek,
630 	.read		= pipe_read,
631 	.readv		= pipe_readv,
632 	.write		= pipe_write,
633 	.writev		= pipe_writev,
634 	.poll		= pipe_poll,
635 	.ioctl		= pipe_ioctl,
636 	.open		= pipe_rdwr_open,
637 	.release	= pipe_rdwr_release,
638 	.fasync		= pipe_rdwr_fasync,
639 };
640 
641 void free_pipe_info(struct inode *inode)
642 {
643 	int i;
644 	struct pipe_inode_info *info = inode->i_pipe;
645 
646 	inode->i_pipe = NULL;
647 	for (i = 0; i < PIPE_BUFFERS; i++) {
648 		struct pipe_buffer *buf = info->bufs + i;
649 		if (buf->ops)
650 			buf->ops->release(info, buf);
651 	}
652 	if (info->tmp_page)
653 		__free_page(info->tmp_page);
654 	kfree(info);
655 }
656 
657 struct inode* pipe_new(struct inode* inode)
658 {
659 	struct pipe_inode_info *info;
660 
661 	info = kmalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
662 	if (!info)
663 		goto fail_page;
664 	memset(info, 0, sizeof(*info));
665 	inode->i_pipe = info;
666 
667 	init_waitqueue_head(PIPE_WAIT(*inode));
668 	PIPE_RCOUNTER(*inode) = PIPE_WCOUNTER(*inode) = 1;
669 
670 	return inode;
671 fail_page:
672 	return NULL;
673 }
674 
675 static struct vfsmount *pipe_mnt;
676 static int pipefs_delete_dentry(struct dentry *dentry)
677 {
678 	return 1;
679 }
680 static struct dentry_operations pipefs_dentry_operations = {
681 	.d_delete	= pipefs_delete_dentry,
682 };
683 
684 static struct inode * get_pipe_inode(void)
685 {
686 	struct inode *inode = new_inode(pipe_mnt->mnt_sb);
687 
688 	if (!inode)
689 		goto fail_inode;
690 
691 	if(!pipe_new(inode))
692 		goto fail_iput;
693 	PIPE_READERS(*inode) = PIPE_WRITERS(*inode) = 1;
694 	inode->i_fop = &rdwr_pipe_fops;
695 
696 	/*
697 	 * Mark the inode dirty from the very beginning,
698 	 * that way it will never be moved to the dirty
699 	 * list because "mark_inode_dirty()" will think
700 	 * that it already _is_ on the dirty list.
701 	 */
702 	inode->i_state = I_DIRTY;
703 	inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
704 	inode->i_uid = current->fsuid;
705 	inode->i_gid = current->fsgid;
706 	inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
707 	inode->i_blksize = PAGE_SIZE;
708 	return inode;
709 
710 fail_iput:
711 	iput(inode);
712 fail_inode:
713 	return NULL;
714 }
715 
716 int do_pipe(int *fd)
717 {
718 	struct qstr this;
719 	char name[32];
720 	struct dentry *dentry;
721 	struct inode * inode;
722 	struct file *f1, *f2;
723 	int error;
724 	int i,j;
725 
726 	error = -ENFILE;
727 	f1 = get_empty_filp();
728 	if (!f1)
729 		goto no_files;
730 
731 	f2 = get_empty_filp();
732 	if (!f2)
733 		goto close_f1;
734 
735 	inode = get_pipe_inode();
736 	if (!inode)
737 		goto close_f12;
738 
739 	error = get_unused_fd();
740 	if (error < 0)
741 		goto close_f12_inode;
742 	i = error;
743 
744 	error = get_unused_fd();
745 	if (error < 0)
746 		goto close_f12_inode_i;
747 	j = error;
748 
749 	error = -ENOMEM;
750 	sprintf(name, "[%lu]", inode->i_ino);
751 	this.name = name;
752 	this.len = strlen(name);
753 	this.hash = inode->i_ino; /* will go */
754 	dentry = d_alloc(pipe_mnt->mnt_sb->s_root, &this);
755 	if (!dentry)
756 		goto close_f12_inode_i_j;
757 	dentry->d_op = &pipefs_dentry_operations;
758 	d_add(dentry, inode);
759 	f1->f_vfsmnt = f2->f_vfsmnt = mntget(mntget(pipe_mnt));
760 	f1->f_dentry = f2->f_dentry = dget(dentry);
761 	f1->f_mapping = f2->f_mapping = inode->i_mapping;
762 
763 	/* read file */
764 	f1->f_pos = f2->f_pos = 0;
765 	f1->f_flags = O_RDONLY;
766 	f1->f_op = &read_pipe_fops;
767 	f1->f_mode = FMODE_READ;
768 	f1->f_version = 0;
769 
770 	/* write file */
771 	f2->f_flags = O_WRONLY;
772 	f2->f_op = &write_pipe_fops;
773 	f2->f_mode = FMODE_WRITE;
774 	f2->f_version = 0;
775 
776 	fd_install(i, f1);
777 	fd_install(j, f2);
778 	fd[0] = i;
779 	fd[1] = j;
780 	return 0;
781 
782 close_f12_inode_i_j:
783 	put_unused_fd(j);
784 close_f12_inode_i:
785 	put_unused_fd(i);
786 close_f12_inode:
787 	free_pipe_info(inode);
788 	iput(inode);
789 close_f12:
790 	put_filp(f2);
791 close_f1:
792 	put_filp(f1);
793 no_files:
794 	return error;
795 }
796 
797 /*
798  * pipefs should _never_ be mounted by userland - too much of security hassle,
799  * no real gain from having the whole whorehouse mounted. So we don't need
800  * any operations on the root directory. However, we need a non-trivial
801  * d_name - pipe: will go nicely and kill the special-casing in procfs.
802  */
803 
804 static struct super_block *pipefs_get_sb(struct file_system_type *fs_type,
805 	int flags, const char *dev_name, void *data)
806 {
807 	return get_sb_pseudo(fs_type, "pipe:", NULL, PIPEFS_MAGIC);
808 }
809 
810 static struct file_system_type pipe_fs_type = {
811 	.name		= "pipefs",
812 	.get_sb		= pipefs_get_sb,
813 	.kill_sb	= kill_anon_super,
814 };
815 
816 static int __init init_pipe_fs(void)
817 {
818 	int err = register_filesystem(&pipe_fs_type);
819 	if (!err) {
820 		pipe_mnt = kern_mount(&pipe_fs_type);
821 		if (IS_ERR(pipe_mnt)) {
822 			err = PTR_ERR(pipe_mnt);
823 			unregister_filesystem(&pipe_fs_type);
824 		}
825 	}
826 	return err;
827 }
828 
829 static void __exit exit_pipe_fs(void)
830 {
831 	unregister_filesystem(&pipe_fs_type);
832 	mntput(pipe_mnt);
833 }
834 
835 fs_initcall(init_pipe_fs);
836 module_exit(exit_pipe_fs);
837