create_pipe_files(): switch the first allocation to alloc_file_pseudo()
[muen/linux.git] / fs / pipe.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  linux/fs/pipe.c
4  *
5  *  Copyright (C) 1991, 1992, 1999  Linus Torvalds
6  */
7
8 #include <linux/mm.h>
9 #include <linux/file.h>
10 #include <linux/poll.h>
11 #include <linux/slab.h>
12 #include <linux/module.h>
13 #include <linux/init.h>
14 #include <linux/fs.h>
15 #include <linux/log2.h>
16 #include <linux/mount.h>
17 #include <linux/magic.h>
18 #include <linux/pipe_fs_i.h>
19 #include <linux/uio.h>
20 #include <linux/highmem.h>
21 #include <linux/pagemap.h>
22 #include <linux/audit.h>
23 #include <linux/syscalls.h>
24 #include <linux/fcntl.h>
25 #include <linux/memcontrol.h>
26
27 #include <linux/uaccess.h>
28 #include <asm/ioctls.h>
29
30 #include "internal.h"
31
32 /*
33  * The max size that a non-root user is allowed to grow the pipe. Can
34  * be set by root in /proc/sys/fs/pipe-max-size
35  */
36 unsigned int pipe_max_size = 1048576;
37
38 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
39  * matches default values.
40  */
41 unsigned long pipe_user_pages_hard;
42 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
43
44 /*
45  * We use a start+len construction, which provides full use of the 
46  * allocated memory.
47  * -- Florian Coosmann (FGC)
48  * 
49  * Reads with count = 0 should always return 0.
50  * -- Julian Bradfield 1999-06-07.
51  *
52  * FIFOs and Pipes now generate SIGIO for both readers and writers.
53  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
54  *
55  * pipe_read & write cleanup
56  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
57  */
58
59 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
60 {
61         if (pipe->files)
62                 mutex_lock_nested(&pipe->mutex, subclass);
63 }
64
65 void pipe_lock(struct pipe_inode_info *pipe)
66 {
67         /*
68          * pipe_lock() nests non-pipe inode locks (for writing to a file)
69          */
70         pipe_lock_nested(pipe, I_MUTEX_PARENT);
71 }
72 EXPORT_SYMBOL(pipe_lock);
73
74 void pipe_unlock(struct pipe_inode_info *pipe)
75 {
76         if (pipe->files)
77                 mutex_unlock(&pipe->mutex);
78 }
79 EXPORT_SYMBOL(pipe_unlock);
80
81 static inline void __pipe_lock(struct pipe_inode_info *pipe)
82 {
83         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
84 }
85
86 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
87 {
88         mutex_unlock(&pipe->mutex);
89 }
90
91 void pipe_double_lock(struct pipe_inode_info *pipe1,
92                       struct pipe_inode_info *pipe2)
93 {
94         BUG_ON(pipe1 == pipe2);
95
96         if (pipe1 < pipe2) {
97                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
98                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
99         } else {
100                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
101                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
102         }
103 }
104
105 /* Drop the inode semaphore and wait for a pipe event, atomically */
106 void pipe_wait(struct pipe_inode_info *pipe)
107 {
108         DEFINE_WAIT(wait);
109
110         /*
111          * Pipes are system-local resources, so sleeping on them
112          * is considered a noninteractive wait:
113          */
114         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
115         pipe_unlock(pipe);
116         schedule();
117         finish_wait(&pipe->wait, &wait);
118         pipe_lock(pipe);
119 }
120
121 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
122                                   struct pipe_buffer *buf)
123 {
124         struct page *page = buf->page;
125
126         /*
127          * If nobody else uses this page, and we don't already have a
128          * temporary page, let's keep track of it as a one-deep
129          * allocation cache. (Otherwise just release our reference to it)
130          */
131         if (page_count(page) == 1 && !pipe->tmp_page)
132                 pipe->tmp_page = page;
133         else
134                 put_page(page);
135 }
136
137 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
138                                struct pipe_buffer *buf)
139 {
140         struct page *page = buf->page;
141
142         if (page_count(page) == 1) {
143                 if (memcg_kmem_enabled())
144                         memcg_kmem_uncharge(page, 0);
145                 __SetPageLocked(page);
146                 return 0;
147         }
148         return 1;
149 }
150
151 /**
152  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
153  * @pipe:       the pipe that the buffer belongs to
154  * @buf:        the buffer to attempt to steal
155  *
156  * Description:
157  *      This function attempts to steal the &struct page attached to
158  *      @buf. If successful, this function returns 0 and returns with
159  *      the page locked. The caller may then reuse the page for whatever
160  *      he wishes; the typical use is insertion into a different file
161  *      page cache.
162  */
163 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
164                            struct pipe_buffer *buf)
165 {
166         struct page *page = buf->page;
167
168         /*
169          * A reference of one is golden, that means that the owner of this
170          * page is the only one holding a reference to it. lock the page
171          * and return OK.
172          */
173         if (page_count(page) == 1) {
174                 lock_page(page);
175                 return 0;
176         }
177
178         return 1;
179 }
180 EXPORT_SYMBOL(generic_pipe_buf_steal);
181
182 /**
183  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
184  * @pipe:       the pipe that the buffer belongs to
185  * @buf:        the buffer to get a reference to
186  *
187  * Description:
188  *      This function grabs an extra reference to @buf. It's used in
189  *      in the tee() system call, when we duplicate the buffers in one
190  *      pipe into another.
191  */
192 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
193 {
194         get_page(buf->page);
195 }
196 EXPORT_SYMBOL(generic_pipe_buf_get);
197
198 /**
199  * generic_pipe_buf_confirm - verify contents of the pipe buffer
200  * @info:       the pipe that the buffer belongs to
201  * @buf:        the buffer to confirm
202  *
203  * Description:
204  *      This function does nothing, because the generic pipe code uses
205  *      pages that are always good when inserted into the pipe.
206  */
207 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
208                              struct pipe_buffer *buf)
209 {
210         return 0;
211 }
212 EXPORT_SYMBOL(generic_pipe_buf_confirm);
213
214 /**
215  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
216  * @pipe:       the pipe that the buffer belongs to
217  * @buf:        the buffer to put a reference to
218  *
219  * Description:
220  *      This function releases a reference to @buf.
221  */
222 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
223                               struct pipe_buffer *buf)
224 {
225         put_page(buf->page);
226 }
227 EXPORT_SYMBOL(generic_pipe_buf_release);
228
229 static const struct pipe_buf_operations anon_pipe_buf_ops = {
230         .can_merge = 1,
231         .confirm = generic_pipe_buf_confirm,
232         .release = anon_pipe_buf_release,
233         .steal = anon_pipe_buf_steal,
234         .get = generic_pipe_buf_get,
235 };
236
237 static const struct pipe_buf_operations packet_pipe_buf_ops = {
238         .can_merge = 0,
239         .confirm = generic_pipe_buf_confirm,
240         .release = anon_pipe_buf_release,
241         .steal = anon_pipe_buf_steal,
242         .get = generic_pipe_buf_get,
243 };
244
245 static ssize_t
246 pipe_read(struct kiocb *iocb, struct iov_iter *to)
247 {
248         size_t total_len = iov_iter_count(to);
249         struct file *filp = iocb->ki_filp;
250         struct pipe_inode_info *pipe = filp->private_data;
251         int do_wakeup;
252         ssize_t ret;
253
254         /* Null read succeeds. */
255         if (unlikely(total_len == 0))
256                 return 0;
257
258         do_wakeup = 0;
259         ret = 0;
260         __pipe_lock(pipe);
261         for (;;) {
262                 int bufs = pipe->nrbufs;
263                 if (bufs) {
264                         int curbuf = pipe->curbuf;
265                         struct pipe_buffer *buf = pipe->bufs + curbuf;
266                         size_t chars = buf->len;
267                         size_t written;
268                         int error;
269
270                         if (chars > total_len)
271                                 chars = total_len;
272
273                         error = pipe_buf_confirm(pipe, buf);
274                         if (error) {
275                                 if (!ret)
276                                         ret = error;
277                                 break;
278                         }
279
280                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
281                         if (unlikely(written < chars)) {
282                                 if (!ret)
283                                         ret = -EFAULT;
284                                 break;
285                         }
286                         ret += chars;
287                         buf->offset += chars;
288                         buf->len -= chars;
289
290                         /* Was it a packet buffer? Clean up and exit */
291                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
292                                 total_len = chars;
293                                 buf->len = 0;
294                         }
295
296                         if (!buf->len) {
297                                 pipe_buf_release(pipe, buf);
298                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
299                                 pipe->curbuf = curbuf;
300                                 pipe->nrbufs = --bufs;
301                                 do_wakeup = 1;
302                         }
303                         total_len -= chars;
304                         if (!total_len)
305                                 break;  /* common path: read succeeded */
306                 }
307                 if (bufs)       /* More to do? */
308                         continue;
309                 if (!pipe->writers)
310                         break;
311                 if (!pipe->waiting_writers) {
312                         /* syscall merging: Usually we must not sleep
313                          * if O_NONBLOCK is set, or if we got some data.
314                          * But if a writer sleeps in kernel space, then
315                          * we can wait for that data without violating POSIX.
316                          */
317                         if (ret)
318                                 break;
319                         if (filp->f_flags & O_NONBLOCK) {
320                                 ret = -EAGAIN;
321                                 break;
322                         }
323                 }
324                 if (signal_pending(current)) {
325                         if (!ret)
326                                 ret = -ERESTARTSYS;
327                         break;
328                 }
329                 if (do_wakeup) {
330                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
331                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
332                 }
333                 pipe_wait(pipe);
334         }
335         __pipe_unlock(pipe);
336
337         /* Signal writers asynchronously that there is more room. */
338         if (do_wakeup) {
339                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLOUT | EPOLLWRNORM);
340                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
341         }
342         if (ret > 0)
343                 file_accessed(filp);
344         return ret;
345 }
346
347 static inline int is_packetized(struct file *file)
348 {
349         return (file->f_flags & O_DIRECT) != 0;
350 }
351
352 static ssize_t
353 pipe_write(struct kiocb *iocb, struct iov_iter *from)
354 {
355         struct file *filp = iocb->ki_filp;
356         struct pipe_inode_info *pipe = filp->private_data;
357         ssize_t ret = 0;
358         int do_wakeup = 0;
359         size_t total_len = iov_iter_count(from);
360         ssize_t chars;
361
362         /* Null write succeeds. */
363         if (unlikely(total_len == 0))
364                 return 0;
365
366         __pipe_lock(pipe);
367
368         if (!pipe->readers) {
369                 send_sig(SIGPIPE, current, 0);
370                 ret = -EPIPE;
371                 goto out;
372         }
373
374         /* We try to merge small writes */
375         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
376         if (pipe->nrbufs && chars != 0) {
377                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
378                                                         (pipe->buffers - 1);
379                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
380                 int offset = buf->offset + buf->len;
381
382                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
383                         ret = pipe_buf_confirm(pipe, buf);
384                         if (ret)
385                                 goto out;
386
387                         ret = copy_page_from_iter(buf->page, offset, chars, from);
388                         if (unlikely(ret < chars)) {
389                                 ret = -EFAULT;
390                                 goto out;
391                         }
392                         do_wakeup = 1;
393                         buf->len += ret;
394                         if (!iov_iter_count(from))
395                                 goto out;
396                 }
397         }
398
399         for (;;) {
400                 int bufs;
401
402                 if (!pipe->readers) {
403                         send_sig(SIGPIPE, current, 0);
404                         if (!ret)
405                                 ret = -EPIPE;
406                         break;
407                 }
408                 bufs = pipe->nrbufs;
409                 if (bufs < pipe->buffers) {
410                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
411                         struct pipe_buffer *buf = pipe->bufs + newbuf;
412                         struct page *page = pipe->tmp_page;
413                         int copied;
414
415                         if (!page) {
416                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
417                                 if (unlikely(!page)) {
418                                         ret = ret ? : -ENOMEM;
419                                         break;
420                                 }
421                                 pipe->tmp_page = page;
422                         }
423                         /* Always wake up, even if the copy fails. Otherwise
424                          * we lock up (O_NONBLOCK-)readers that sleep due to
425                          * syscall merging.
426                          * FIXME! Is this really true?
427                          */
428                         do_wakeup = 1;
429                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
430                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
431                                 if (!ret)
432                                         ret = -EFAULT;
433                                 break;
434                         }
435                         ret += copied;
436
437                         /* Insert it into the buffer array */
438                         buf->page = page;
439                         buf->ops = &anon_pipe_buf_ops;
440                         buf->offset = 0;
441                         buf->len = copied;
442                         buf->flags = 0;
443                         if (is_packetized(filp)) {
444                                 buf->ops = &packet_pipe_buf_ops;
445                                 buf->flags = PIPE_BUF_FLAG_PACKET;
446                         }
447                         pipe->nrbufs = ++bufs;
448                         pipe->tmp_page = NULL;
449
450                         if (!iov_iter_count(from))
451                                 break;
452                 }
453                 if (bufs < pipe->buffers)
454                         continue;
455                 if (filp->f_flags & O_NONBLOCK) {
456                         if (!ret)
457                                 ret = -EAGAIN;
458                         break;
459                 }
460                 if (signal_pending(current)) {
461                         if (!ret)
462                                 ret = -ERESTARTSYS;
463                         break;
464                 }
465                 if (do_wakeup) {
466                         wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
467                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
468                         do_wakeup = 0;
469                 }
470                 pipe->waiting_writers++;
471                 pipe_wait(pipe);
472                 pipe->waiting_writers--;
473         }
474 out:
475         __pipe_unlock(pipe);
476         if (do_wakeup) {
477                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLRDNORM);
478                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
479         }
480         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
481                 int err = file_update_time(filp);
482                 if (err)
483                         ret = err;
484                 sb_end_write(file_inode(filp)->i_sb);
485         }
486         return ret;
487 }
488
489 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
490 {
491         struct pipe_inode_info *pipe = filp->private_data;
492         int count, buf, nrbufs;
493
494         switch (cmd) {
495                 case FIONREAD:
496                         __pipe_lock(pipe);
497                         count = 0;
498                         buf = pipe->curbuf;
499                         nrbufs = pipe->nrbufs;
500                         while (--nrbufs >= 0) {
501                                 count += pipe->bufs[buf].len;
502                                 buf = (buf+1) & (pipe->buffers - 1);
503                         }
504                         __pipe_unlock(pipe);
505
506                         return put_user(count, (int __user *)arg);
507                 default:
508                         return -ENOIOCTLCMD;
509         }
510 }
511
512 static struct wait_queue_head *
513 pipe_get_poll_head(struct file *filp, __poll_t events)
514 {
515         struct pipe_inode_info *pipe = filp->private_data;
516
517         return &pipe->wait;
518 }
519
520 /* No kernel lock held - fine */
521 static __poll_t pipe_poll_mask(struct file *filp, __poll_t events)
522 {
523         struct pipe_inode_info *pipe = filp->private_data;
524         int nrbufs = pipe->nrbufs;
525         __poll_t mask = 0;
526
527         /* Reading only -- no need for acquiring the semaphore.  */
528         if (filp->f_mode & FMODE_READ) {
529                 mask = (nrbufs > 0) ? EPOLLIN | EPOLLRDNORM : 0;
530                 if (!pipe->writers && filp->f_version != pipe->w_counter)
531                         mask |= EPOLLHUP;
532         }
533
534         if (filp->f_mode & FMODE_WRITE) {
535                 mask |= (nrbufs < pipe->buffers) ? EPOLLOUT | EPOLLWRNORM : 0;
536                 /*
537                  * Most Unices do not set EPOLLERR for FIFOs but on Linux they
538                  * behave exactly like pipes for poll().
539                  */
540                 if (!pipe->readers)
541                         mask |= EPOLLERR;
542         }
543
544         return mask;
545 }
546
547 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
548 {
549         int kill = 0;
550
551         spin_lock(&inode->i_lock);
552         if (!--pipe->files) {
553                 inode->i_pipe = NULL;
554                 kill = 1;
555         }
556         spin_unlock(&inode->i_lock);
557
558         if (kill)
559                 free_pipe_info(pipe);
560 }
561
562 static int
563 pipe_release(struct inode *inode, struct file *file)
564 {
565         struct pipe_inode_info *pipe = file->private_data;
566
567         __pipe_lock(pipe);
568         if (file->f_mode & FMODE_READ)
569                 pipe->readers--;
570         if (file->f_mode & FMODE_WRITE)
571                 pipe->writers--;
572
573         if (pipe->readers || pipe->writers) {
574                 wake_up_interruptible_sync_poll(&pipe->wait, EPOLLIN | EPOLLOUT | EPOLLRDNORM | EPOLLWRNORM | EPOLLERR | EPOLLHUP);
575                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
576                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
577         }
578         __pipe_unlock(pipe);
579
580         put_pipe_info(inode, pipe);
581         return 0;
582 }
583
584 static int
585 pipe_fasync(int fd, struct file *filp, int on)
586 {
587         struct pipe_inode_info *pipe = filp->private_data;
588         int retval = 0;
589
590         __pipe_lock(pipe);
591         if (filp->f_mode & FMODE_READ)
592                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
593         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
594                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
595                 if (retval < 0 && (filp->f_mode & FMODE_READ))
596                         /* this can happen only if on == T */
597                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
598         }
599         __pipe_unlock(pipe);
600         return retval;
601 }
602
603 static unsigned long account_pipe_buffers(struct user_struct *user,
604                                  unsigned long old, unsigned long new)
605 {
606         return atomic_long_add_return(new - old, &user->pipe_bufs);
607 }
608
609 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
610 {
611         unsigned long soft_limit = READ_ONCE(pipe_user_pages_soft);
612
613         return soft_limit && user_bufs > soft_limit;
614 }
615
616 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
617 {
618         unsigned long hard_limit = READ_ONCE(pipe_user_pages_hard);
619
620         return hard_limit && user_bufs > hard_limit;
621 }
622
623 static bool is_unprivileged_user(void)
624 {
625         return !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN);
626 }
627
628 struct pipe_inode_info *alloc_pipe_info(void)
629 {
630         struct pipe_inode_info *pipe;
631         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
632         struct user_struct *user = get_current_user();
633         unsigned long user_bufs;
634         unsigned int max_size = READ_ONCE(pipe_max_size);
635
636         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
637         if (pipe == NULL)
638                 goto out_free_uid;
639
640         if (pipe_bufs * PAGE_SIZE > max_size && !capable(CAP_SYS_RESOURCE))
641                 pipe_bufs = max_size >> PAGE_SHIFT;
642
643         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
644
645         if (too_many_pipe_buffers_soft(user_bufs) && is_unprivileged_user()) {
646                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
647                 pipe_bufs = 1;
648         }
649
650         if (too_many_pipe_buffers_hard(user_bufs) && is_unprivileged_user())
651                 goto out_revert_acct;
652
653         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
654                              GFP_KERNEL_ACCOUNT);
655
656         if (pipe->bufs) {
657                 init_waitqueue_head(&pipe->wait);
658                 pipe->r_counter = pipe->w_counter = 1;
659                 pipe->buffers = pipe_bufs;
660                 pipe->user = user;
661                 mutex_init(&pipe->mutex);
662                 return pipe;
663         }
664
665 out_revert_acct:
666         (void) account_pipe_buffers(user, pipe_bufs, 0);
667         kfree(pipe);
668 out_free_uid:
669         free_uid(user);
670         return NULL;
671 }
672
673 void free_pipe_info(struct pipe_inode_info *pipe)
674 {
675         int i;
676
677         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
678         free_uid(pipe->user);
679         for (i = 0; i < pipe->buffers; i++) {
680                 struct pipe_buffer *buf = pipe->bufs + i;
681                 if (buf->ops)
682                         pipe_buf_release(pipe, buf);
683         }
684         if (pipe->tmp_page)
685                 __free_page(pipe->tmp_page);
686         kfree(pipe->bufs);
687         kfree(pipe);
688 }
689
690 static struct vfsmount *pipe_mnt __read_mostly;
691
692 /*
693  * pipefs_dname() is called from d_path().
694  */
695 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
696 {
697         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
698                                 d_inode(dentry)->i_ino);
699 }
700
701 static const struct dentry_operations pipefs_dentry_operations = {
702         .d_dname        = pipefs_dname,
703 };
704
705 static struct inode * get_pipe_inode(void)
706 {
707         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
708         struct pipe_inode_info *pipe;
709
710         if (!inode)
711                 goto fail_inode;
712
713         inode->i_ino = get_next_ino();
714
715         pipe = alloc_pipe_info();
716         if (!pipe)
717                 goto fail_iput;
718
719         inode->i_pipe = pipe;
720         pipe->files = 2;
721         pipe->readers = pipe->writers = 1;
722         inode->i_fop = &pipefifo_fops;
723
724         /*
725          * Mark the inode dirty from the very beginning,
726          * that way it will never be moved to the dirty
727          * list because "mark_inode_dirty()" will think
728          * that it already _is_ on the dirty list.
729          */
730         inode->i_state = I_DIRTY;
731         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
732         inode->i_uid = current_fsuid();
733         inode->i_gid = current_fsgid();
734         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
735
736         return inode;
737
738 fail_iput:
739         iput(inode);
740
741 fail_inode:
742         return NULL;
743 }
744
745 int create_pipe_files(struct file **res, int flags)
746 {
747         struct inode *inode = get_pipe_inode();
748         struct file *f;
749
750         if (!inode)
751                 return -ENFILE;
752
753         f = alloc_file_pseudo(inode, pipe_mnt, "",
754                                 O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT)),
755                                 &pipefifo_fops);
756         if (IS_ERR(f)) {
757                 free_pipe_info(inode->i_pipe);
758                 iput(inode);
759                 return PTR_ERR(f);
760         }
761
762         f->private_data = inode->i_pipe;
763
764         res[0] = alloc_file(&f->f_path, O_RDONLY | (flags & O_NONBLOCK),
765                         &pipefifo_fops);
766         if (IS_ERR(res[0])) {
767                 put_pipe_info(inode, inode->i_pipe);
768                 fput(f);
769                 return PTR_ERR(res[0]);
770         }
771
772         path_get(&f->f_path);
773         res[0]->private_data = inode->i_pipe;
774         res[1] = f;
775         return 0;
776 }
777
778 static int __do_pipe_flags(int *fd, struct file **files, int flags)
779 {
780         int error;
781         int fdw, fdr;
782
783         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
784                 return -EINVAL;
785
786         error = create_pipe_files(files, flags);
787         if (error)
788                 return error;
789
790         error = get_unused_fd_flags(flags);
791         if (error < 0)
792                 goto err_read_pipe;
793         fdr = error;
794
795         error = get_unused_fd_flags(flags);
796         if (error < 0)
797                 goto err_fdr;
798         fdw = error;
799
800         audit_fd_pair(fdr, fdw);
801         fd[0] = fdr;
802         fd[1] = fdw;
803         return 0;
804
805  err_fdr:
806         put_unused_fd(fdr);
807  err_read_pipe:
808         fput(files[0]);
809         fput(files[1]);
810         return error;
811 }
812
813 int do_pipe_flags(int *fd, int flags)
814 {
815         struct file *files[2];
816         int error = __do_pipe_flags(fd, files, flags);
817         if (!error) {
818                 fd_install(fd[0], files[0]);
819                 fd_install(fd[1], files[1]);
820         }
821         return error;
822 }
823
824 /*
825  * sys_pipe() is the normal C calling standard for creating
826  * a pipe. It's not the way Unix traditionally does this, though.
827  */
828 static int do_pipe2(int __user *fildes, int flags)
829 {
830         struct file *files[2];
831         int fd[2];
832         int error;
833
834         error = __do_pipe_flags(fd, files, flags);
835         if (!error) {
836                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
837                         fput(files[0]);
838                         fput(files[1]);
839                         put_unused_fd(fd[0]);
840                         put_unused_fd(fd[1]);
841                         error = -EFAULT;
842                 } else {
843                         fd_install(fd[0], files[0]);
844                         fd_install(fd[1], files[1]);
845                 }
846         }
847         return error;
848 }
849
850 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
851 {
852         return do_pipe2(fildes, flags);
853 }
854
855 SYSCALL_DEFINE1(pipe, int __user *, fildes)
856 {
857         return do_pipe2(fildes, 0);
858 }
859
860 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
861 {
862         int cur = *cnt; 
863
864         while (cur == *cnt) {
865                 pipe_wait(pipe);
866                 if (signal_pending(current))
867                         break;
868         }
869         return cur == *cnt ? -ERESTARTSYS : 0;
870 }
871
872 static void wake_up_partner(struct pipe_inode_info *pipe)
873 {
874         wake_up_interruptible(&pipe->wait);
875 }
876
877 static int fifo_open(struct inode *inode, struct file *filp)
878 {
879         struct pipe_inode_info *pipe;
880         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
881         int ret;
882
883         filp->f_version = 0;
884
885         spin_lock(&inode->i_lock);
886         if (inode->i_pipe) {
887                 pipe = inode->i_pipe;
888                 pipe->files++;
889                 spin_unlock(&inode->i_lock);
890         } else {
891                 spin_unlock(&inode->i_lock);
892                 pipe = alloc_pipe_info();
893                 if (!pipe)
894                         return -ENOMEM;
895                 pipe->files = 1;
896                 spin_lock(&inode->i_lock);
897                 if (unlikely(inode->i_pipe)) {
898                         inode->i_pipe->files++;
899                         spin_unlock(&inode->i_lock);
900                         free_pipe_info(pipe);
901                         pipe = inode->i_pipe;
902                 } else {
903                         inode->i_pipe = pipe;
904                         spin_unlock(&inode->i_lock);
905                 }
906         }
907         filp->private_data = pipe;
908         /* OK, we have a pipe and it's pinned down */
909
910         __pipe_lock(pipe);
911
912         /* We can only do regular read/write on fifos */
913         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
914
915         switch (filp->f_mode) {
916         case FMODE_READ:
917         /*
918          *  O_RDONLY
919          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
920          *  opened, even when there is no process writing the FIFO.
921          */
922                 pipe->r_counter++;
923                 if (pipe->readers++ == 0)
924                         wake_up_partner(pipe);
925
926                 if (!is_pipe && !pipe->writers) {
927                         if ((filp->f_flags & O_NONBLOCK)) {
928                                 /* suppress EPOLLHUP until we have
929                                  * seen a writer */
930                                 filp->f_version = pipe->w_counter;
931                         } else {
932                                 if (wait_for_partner(pipe, &pipe->w_counter))
933                                         goto err_rd;
934                         }
935                 }
936                 break;
937         
938         case FMODE_WRITE:
939         /*
940          *  O_WRONLY
941          *  POSIX.1 says that O_NONBLOCK means return -1 with
942          *  errno=ENXIO when there is no process reading the FIFO.
943          */
944                 ret = -ENXIO;
945                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
946                         goto err;
947
948                 pipe->w_counter++;
949                 if (!pipe->writers++)
950                         wake_up_partner(pipe);
951
952                 if (!is_pipe && !pipe->readers) {
953                         if (wait_for_partner(pipe, &pipe->r_counter))
954                                 goto err_wr;
955                 }
956                 break;
957         
958         case FMODE_READ | FMODE_WRITE:
959         /*
960          *  O_RDWR
961          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
962          *  This implementation will NEVER block on a O_RDWR open, since
963          *  the process can at least talk to itself.
964          */
965
966                 pipe->readers++;
967                 pipe->writers++;
968                 pipe->r_counter++;
969                 pipe->w_counter++;
970                 if (pipe->readers == 1 || pipe->writers == 1)
971                         wake_up_partner(pipe);
972                 break;
973
974         default:
975                 ret = -EINVAL;
976                 goto err;
977         }
978
979         /* Ok! */
980         __pipe_unlock(pipe);
981         return 0;
982
983 err_rd:
984         if (!--pipe->readers)
985                 wake_up_interruptible(&pipe->wait);
986         ret = -ERESTARTSYS;
987         goto err;
988
989 err_wr:
990         if (!--pipe->writers)
991                 wake_up_interruptible(&pipe->wait);
992         ret = -ERESTARTSYS;
993         goto err;
994
995 err:
996         __pipe_unlock(pipe);
997
998         put_pipe_info(inode, pipe);
999         return ret;
1000 }
1001
1002 const struct file_operations pipefifo_fops = {
1003         .open           = fifo_open,
1004         .llseek         = no_llseek,
1005         .read_iter      = pipe_read,
1006         .write_iter     = pipe_write,
1007         .get_poll_head  = pipe_get_poll_head,
1008         .poll_mask      = pipe_poll_mask,
1009         .unlocked_ioctl = pipe_ioctl,
1010         .release        = pipe_release,
1011         .fasync         = pipe_fasync,
1012 };
1013
1014 /*
1015  * Currently we rely on the pipe array holding a power-of-2 number
1016  * of pages. Returns 0 on error.
1017  */
1018 unsigned int round_pipe_size(unsigned long size)
1019 {
1020         if (size > (1U << 31))
1021                 return 0;
1022
1023         /* Minimum pipe size, as required by POSIX */
1024         if (size < PAGE_SIZE)
1025                 return PAGE_SIZE;
1026
1027         return roundup_pow_of_two(size);
1028 }
1029
1030 /*
1031  * Allocate a new array of pipe buffers and copy the info over. Returns the
1032  * pipe size if successful, or return -ERROR on error.
1033  */
1034 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1035 {
1036         struct pipe_buffer *bufs;
1037         unsigned int size, nr_pages;
1038         unsigned long user_bufs;
1039         long ret = 0;
1040
1041         size = round_pipe_size(arg);
1042         nr_pages = size >> PAGE_SHIFT;
1043
1044         if (!nr_pages)
1045                 return -EINVAL;
1046
1047         /*
1048          * If trying to increase the pipe capacity, check that an
1049          * unprivileged user is not trying to exceed various limits
1050          * (soft limit check here, hard limit check just below).
1051          * Decreasing the pipe capacity is always permitted, even
1052          * if the user is currently over a limit.
1053          */
1054         if (nr_pages > pipe->buffers &&
1055                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1056                 return -EPERM;
1057
1058         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1059
1060         if (nr_pages > pipe->buffers &&
1061                         (too_many_pipe_buffers_hard(user_bufs) ||
1062                          too_many_pipe_buffers_soft(user_bufs)) &&
1063                         is_unprivileged_user()) {
1064                 ret = -EPERM;
1065                 goto out_revert_acct;
1066         }
1067
1068         /*
1069          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1070          * expect a lot of shrink+grow operations, just free and allocate
1071          * again like we would do for growing. If the pipe currently
1072          * contains more buffers than arg, then return busy.
1073          */
1074         if (nr_pages < pipe->nrbufs) {
1075                 ret = -EBUSY;
1076                 goto out_revert_acct;
1077         }
1078
1079         bufs = kcalloc(nr_pages, sizeof(*bufs),
1080                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1081         if (unlikely(!bufs)) {
1082                 ret = -ENOMEM;
1083                 goto out_revert_acct;
1084         }
1085
1086         /*
1087          * The pipe array wraps around, so just start the new one at zero
1088          * and adjust the indexes.
1089          */
1090         if (pipe->nrbufs) {
1091                 unsigned int tail;
1092                 unsigned int head;
1093
1094                 tail = pipe->curbuf + pipe->nrbufs;
1095                 if (tail < pipe->buffers)
1096                         tail = 0;
1097                 else
1098                         tail &= (pipe->buffers - 1);
1099
1100                 head = pipe->nrbufs - tail;
1101                 if (head)
1102                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1103                 if (tail)
1104                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1105         }
1106
1107         pipe->curbuf = 0;
1108         kfree(pipe->bufs);
1109         pipe->bufs = bufs;
1110         pipe->buffers = nr_pages;
1111         return nr_pages * PAGE_SIZE;
1112
1113 out_revert_acct:
1114         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1115         return ret;
1116 }
1117
1118 /*
1119  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1120  * location, so checking ->i_pipe is not enough to verify that this is a
1121  * pipe.
1122  */
1123 struct pipe_inode_info *get_pipe_info(struct file *file)
1124 {
1125         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1126 }
1127
1128 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1129 {
1130         struct pipe_inode_info *pipe;
1131         long ret;
1132
1133         pipe = get_pipe_info(file);
1134         if (!pipe)
1135                 return -EBADF;
1136
1137         __pipe_lock(pipe);
1138
1139         switch (cmd) {
1140         case F_SETPIPE_SZ:
1141                 ret = pipe_set_size(pipe, arg);
1142                 break;
1143         case F_GETPIPE_SZ:
1144                 ret = pipe->buffers * PAGE_SIZE;
1145                 break;
1146         default:
1147                 ret = -EINVAL;
1148                 break;
1149         }
1150
1151         __pipe_unlock(pipe);
1152         return ret;
1153 }
1154
1155 static const struct super_operations pipefs_ops = {
1156         .destroy_inode = free_inode_nonrcu,
1157         .statfs = simple_statfs,
1158 };
1159
1160 /*
1161  * pipefs should _never_ be mounted by userland - too much of security hassle,
1162  * no real gain from having the whole whorehouse mounted. So we don't need
1163  * any operations on the root directory. However, we need a non-trivial
1164  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1165  */
1166 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1167                          int flags, const char *dev_name, void *data)
1168 {
1169         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1170                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1171 }
1172
1173 static struct file_system_type pipe_fs_type = {
1174         .name           = "pipefs",
1175         .mount          = pipefs_mount,
1176         .kill_sb        = kill_anon_super,
1177 };
1178
1179 static int __init init_pipe_fs(void)
1180 {
1181         int err = register_filesystem(&pipe_fs_type);
1182
1183         if (!err) {
1184                 pipe_mnt = kern_mount(&pipe_fs_type);
1185                 if (IS_ERR(pipe_mnt)) {
1186                         err = PTR_ERR(pipe_mnt);
1187                         unregister_filesystem(&pipe_fs_type);
1188                 }
1189         }
1190         return err;
1191 }
1192
1193 fs_initcall(init_pipe_fs);