]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - fs/pipe.c
Merge branch 'stable/for-linus-4.10' of git://git.kernel.org/pub/scm/linux/kernel...
[karo-tx-linux.git] / fs / pipe.c
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/memcontrol.h>
25
26 #include <linux/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 /* Maximum allocatable pages per user. Hard limit is unset by default, soft
43  * matches default values.
44  */
45 unsigned long pipe_user_pages_hard;
46 unsigned long pipe_user_pages_soft = PIPE_DEF_BUFFERS * INR_OPEN_CUR;
47
48 /*
49  * We use a start+len construction, which provides full use of the 
50  * allocated memory.
51  * -- Florian Coosmann (FGC)
52  * 
53  * Reads with count = 0 should always return 0.
54  * -- Julian Bradfield 1999-06-07.
55  *
56  * FIFOs and Pipes now generate SIGIO for both readers and writers.
57  * -- Jeremy Elson <jelson@circlemud.org> 2001-08-16
58  *
59  * pipe_read & write cleanup
60  * -- Manfred Spraul <manfred@colorfullife.com> 2002-05-09
61  */
62
63 static void pipe_lock_nested(struct pipe_inode_info *pipe, int subclass)
64 {
65         if (pipe->files)
66                 mutex_lock_nested(&pipe->mutex, subclass);
67 }
68
69 void pipe_lock(struct pipe_inode_info *pipe)
70 {
71         /*
72          * pipe_lock() nests non-pipe inode locks (for writing to a file)
73          */
74         pipe_lock_nested(pipe, I_MUTEX_PARENT);
75 }
76 EXPORT_SYMBOL(pipe_lock);
77
78 void pipe_unlock(struct pipe_inode_info *pipe)
79 {
80         if (pipe->files)
81                 mutex_unlock(&pipe->mutex);
82 }
83 EXPORT_SYMBOL(pipe_unlock);
84
85 static inline void __pipe_lock(struct pipe_inode_info *pipe)
86 {
87         mutex_lock_nested(&pipe->mutex, I_MUTEX_PARENT);
88 }
89
90 static inline void __pipe_unlock(struct pipe_inode_info *pipe)
91 {
92         mutex_unlock(&pipe->mutex);
93 }
94
95 void pipe_double_lock(struct pipe_inode_info *pipe1,
96                       struct pipe_inode_info *pipe2)
97 {
98         BUG_ON(pipe1 == pipe2);
99
100         if (pipe1 < pipe2) {
101                 pipe_lock_nested(pipe1, I_MUTEX_PARENT);
102                 pipe_lock_nested(pipe2, I_MUTEX_CHILD);
103         } else {
104                 pipe_lock_nested(pipe2, I_MUTEX_PARENT);
105                 pipe_lock_nested(pipe1, I_MUTEX_CHILD);
106         }
107 }
108
109 /* Drop the inode semaphore and wait for a pipe event, atomically */
110 void pipe_wait(struct pipe_inode_info *pipe)
111 {
112         DEFINE_WAIT(wait);
113
114         /*
115          * Pipes are system-local resources, so sleeping on them
116          * is considered a noninteractive wait:
117          */
118         prepare_to_wait(&pipe->wait, &wait, TASK_INTERRUPTIBLE);
119         pipe_unlock(pipe);
120         schedule();
121         finish_wait(&pipe->wait, &wait);
122         pipe_lock(pipe);
123 }
124
125 static void anon_pipe_buf_release(struct pipe_inode_info *pipe,
126                                   struct pipe_buffer *buf)
127 {
128         struct page *page = buf->page;
129
130         /*
131          * If nobody else uses this page, and we don't already have a
132          * temporary page, let's keep track of it as a one-deep
133          * allocation cache. (Otherwise just release our reference to it)
134          */
135         if (page_count(page) == 1 && !pipe->tmp_page)
136                 pipe->tmp_page = page;
137         else
138                 put_page(page);
139 }
140
141 static int anon_pipe_buf_steal(struct pipe_inode_info *pipe,
142                                struct pipe_buffer *buf)
143 {
144         struct page *page = buf->page;
145
146         if (page_count(page) == 1) {
147                 if (memcg_kmem_enabled())
148                         memcg_kmem_uncharge(page, 0);
149                 __SetPageLocked(page);
150                 return 0;
151         }
152         return 1;
153 }
154
155 /**
156  * generic_pipe_buf_steal - attempt to take ownership of a &pipe_buffer
157  * @pipe:       the pipe that the buffer belongs to
158  * @buf:        the buffer to attempt to steal
159  *
160  * Description:
161  *      This function attempts to steal the &struct page attached to
162  *      @buf. If successful, this function returns 0 and returns with
163  *      the page locked. The caller may then reuse the page for whatever
164  *      he wishes; the typical use is insertion into a different file
165  *      page cache.
166  */
167 int generic_pipe_buf_steal(struct pipe_inode_info *pipe,
168                            struct pipe_buffer *buf)
169 {
170         struct page *page = buf->page;
171
172         /*
173          * A reference of one is golden, that means that the owner of this
174          * page is the only one holding a reference to it. lock the page
175          * and return OK.
176          */
177         if (page_count(page) == 1) {
178                 lock_page(page);
179                 return 0;
180         }
181
182         return 1;
183 }
184 EXPORT_SYMBOL(generic_pipe_buf_steal);
185
186 /**
187  * generic_pipe_buf_get - get a reference to a &struct pipe_buffer
188  * @pipe:       the pipe that the buffer belongs to
189  * @buf:        the buffer to get a reference to
190  *
191  * Description:
192  *      This function grabs an extra reference to @buf. It's used in
193  *      in the tee() system call, when we duplicate the buffers in one
194  *      pipe into another.
195  */
196 void generic_pipe_buf_get(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
197 {
198         get_page(buf->page);
199 }
200 EXPORT_SYMBOL(generic_pipe_buf_get);
201
202 /**
203  * generic_pipe_buf_confirm - verify contents of the pipe buffer
204  * @info:       the pipe that the buffer belongs to
205  * @buf:        the buffer to confirm
206  *
207  * Description:
208  *      This function does nothing, because the generic pipe code uses
209  *      pages that are always good when inserted into the pipe.
210  */
211 int generic_pipe_buf_confirm(struct pipe_inode_info *info,
212                              struct pipe_buffer *buf)
213 {
214         return 0;
215 }
216 EXPORT_SYMBOL(generic_pipe_buf_confirm);
217
218 /**
219  * generic_pipe_buf_release - put a reference to a &struct pipe_buffer
220  * @pipe:       the pipe that the buffer belongs to
221  * @buf:        the buffer to put a reference to
222  *
223  * Description:
224  *      This function releases a reference to @buf.
225  */
226 void generic_pipe_buf_release(struct pipe_inode_info *pipe,
227                               struct pipe_buffer *buf)
228 {
229         put_page(buf->page);
230 }
231 EXPORT_SYMBOL(generic_pipe_buf_release);
232
233 static const struct pipe_buf_operations anon_pipe_buf_ops = {
234         .can_merge = 1,
235         .confirm = generic_pipe_buf_confirm,
236         .release = anon_pipe_buf_release,
237         .steal = anon_pipe_buf_steal,
238         .get = generic_pipe_buf_get,
239 };
240
241 static const struct pipe_buf_operations packet_pipe_buf_ops = {
242         .can_merge = 0,
243         .confirm = generic_pipe_buf_confirm,
244         .release = anon_pipe_buf_release,
245         .steal = anon_pipe_buf_steal,
246         .get = generic_pipe_buf_get,
247 };
248
249 static ssize_t
250 pipe_read(struct kiocb *iocb, struct iov_iter *to)
251 {
252         size_t total_len = iov_iter_count(to);
253         struct file *filp = iocb->ki_filp;
254         struct pipe_inode_info *pipe = filp->private_data;
255         int do_wakeup;
256         ssize_t ret;
257
258         /* Null read succeeds. */
259         if (unlikely(total_len == 0))
260                 return 0;
261
262         do_wakeup = 0;
263         ret = 0;
264         __pipe_lock(pipe);
265         for (;;) {
266                 int bufs = pipe->nrbufs;
267                 if (bufs) {
268                         int curbuf = pipe->curbuf;
269                         struct pipe_buffer *buf = pipe->bufs + curbuf;
270                         size_t chars = buf->len;
271                         size_t written;
272                         int error;
273
274                         if (chars > total_len)
275                                 chars = total_len;
276
277                         error = pipe_buf_confirm(pipe, buf);
278                         if (error) {
279                                 if (!ret)
280                                         ret = error;
281                                 break;
282                         }
283
284                         written = copy_page_to_iter(buf->page, buf->offset, chars, to);
285                         if (unlikely(written < chars)) {
286                                 if (!ret)
287                                         ret = -EFAULT;
288                                 break;
289                         }
290                         ret += chars;
291                         buf->offset += chars;
292                         buf->len -= chars;
293
294                         /* Was it a packet buffer? Clean up and exit */
295                         if (buf->flags & PIPE_BUF_FLAG_PACKET) {
296                                 total_len = chars;
297                                 buf->len = 0;
298                         }
299
300                         if (!buf->len) {
301                                 pipe_buf_release(pipe, buf);
302                                 curbuf = (curbuf + 1) & (pipe->buffers - 1);
303                                 pipe->curbuf = curbuf;
304                                 pipe->nrbufs = --bufs;
305                                 do_wakeup = 1;
306                         }
307                         total_len -= chars;
308                         if (!total_len)
309                                 break;  /* common path: read succeeded */
310                 }
311                 if (bufs)       /* More to do? */
312                         continue;
313                 if (!pipe->writers)
314                         break;
315                 if (!pipe->waiting_writers) {
316                         /* syscall merging: Usually we must not sleep
317                          * if O_NONBLOCK is set, or if we got some data.
318                          * But if a writer sleeps in kernel space, then
319                          * we can wait for that data without violating POSIX.
320                          */
321                         if (ret)
322                                 break;
323                         if (filp->f_flags & O_NONBLOCK) {
324                                 ret = -EAGAIN;
325                                 break;
326                         }
327                 }
328                 if (signal_pending(current)) {
329                         if (!ret)
330                                 ret = -ERESTARTSYS;
331                         break;
332                 }
333                 if (do_wakeup) {
334                         wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
335                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
336                 }
337                 pipe_wait(pipe);
338         }
339         __pipe_unlock(pipe);
340
341         /* Signal writers asynchronously that there is more room. */
342         if (do_wakeup) {
343                 wake_up_interruptible_sync_poll(&pipe->wait, POLLOUT | POLLWRNORM);
344                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
345         }
346         if (ret > 0)
347                 file_accessed(filp);
348         return ret;
349 }
350
351 static inline int is_packetized(struct file *file)
352 {
353         return (file->f_flags & O_DIRECT) != 0;
354 }
355
356 static ssize_t
357 pipe_write(struct kiocb *iocb, struct iov_iter *from)
358 {
359         struct file *filp = iocb->ki_filp;
360         struct pipe_inode_info *pipe = filp->private_data;
361         ssize_t ret = 0;
362         int do_wakeup = 0;
363         size_t total_len = iov_iter_count(from);
364         ssize_t chars;
365
366         /* Null write succeeds. */
367         if (unlikely(total_len == 0))
368                 return 0;
369
370         __pipe_lock(pipe);
371
372         if (!pipe->readers) {
373                 send_sig(SIGPIPE, current, 0);
374                 ret = -EPIPE;
375                 goto out;
376         }
377
378         /* We try to merge small writes */
379         chars = total_len & (PAGE_SIZE-1); /* size of the last buffer */
380         if (pipe->nrbufs && chars != 0) {
381                 int lastbuf = (pipe->curbuf + pipe->nrbufs - 1) &
382                                                         (pipe->buffers - 1);
383                 struct pipe_buffer *buf = pipe->bufs + lastbuf;
384                 int offset = buf->offset + buf->len;
385
386                 if (buf->ops->can_merge && offset + chars <= PAGE_SIZE) {
387                         ret = pipe_buf_confirm(pipe, buf);
388                         if (ret)
389                                 goto out;
390
391                         ret = copy_page_from_iter(buf->page, offset, chars, from);
392                         if (unlikely(ret < chars)) {
393                                 ret = -EFAULT;
394                                 goto out;
395                         }
396                         do_wakeup = 1;
397                         buf->len += ret;
398                         if (!iov_iter_count(from))
399                                 goto out;
400                 }
401         }
402
403         for (;;) {
404                 int bufs;
405
406                 if (!pipe->readers) {
407                         send_sig(SIGPIPE, current, 0);
408                         if (!ret)
409                                 ret = -EPIPE;
410                         break;
411                 }
412                 bufs = pipe->nrbufs;
413                 if (bufs < pipe->buffers) {
414                         int newbuf = (pipe->curbuf + bufs) & (pipe->buffers-1);
415                         struct pipe_buffer *buf = pipe->bufs + newbuf;
416                         struct page *page = pipe->tmp_page;
417                         int copied;
418
419                         if (!page) {
420                                 page = alloc_page(GFP_HIGHUSER | __GFP_ACCOUNT);
421                                 if (unlikely(!page)) {
422                                         ret = ret ? : -ENOMEM;
423                                         break;
424                                 }
425                                 pipe->tmp_page = page;
426                         }
427                         /* Always wake up, even if the copy fails. Otherwise
428                          * we lock up (O_NONBLOCK-)readers that sleep due to
429                          * syscall merging.
430                          * FIXME! Is this really true?
431                          */
432                         do_wakeup = 1;
433                         copied = copy_page_from_iter(page, 0, PAGE_SIZE, from);
434                         if (unlikely(copied < PAGE_SIZE && iov_iter_count(from))) {
435                                 if (!ret)
436                                         ret = -EFAULT;
437                                 break;
438                         }
439                         ret += copied;
440
441                         /* Insert it into the buffer array */
442                         buf->page = page;
443                         buf->ops = &anon_pipe_buf_ops;
444                         buf->offset = 0;
445                         buf->len = copied;
446                         buf->flags = 0;
447                         if (is_packetized(filp)) {
448                                 buf->ops = &packet_pipe_buf_ops;
449                                 buf->flags = PIPE_BUF_FLAG_PACKET;
450                         }
451                         pipe->nrbufs = ++bufs;
452                         pipe->tmp_page = NULL;
453
454                         if (!iov_iter_count(from))
455                                 break;
456                 }
457                 if (bufs < pipe->buffers)
458                         continue;
459                 if (filp->f_flags & O_NONBLOCK) {
460                         if (!ret)
461                                 ret = -EAGAIN;
462                         break;
463                 }
464                 if (signal_pending(current)) {
465                         if (!ret)
466                                 ret = -ERESTARTSYS;
467                         break;
468                 }
469                 if (do_wakeup) {
470                         wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
471                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
472                         do_wakeup = 0;
473                 }
474                 pipe->waiting_writers++;
475                 pipe_wait(pipe);
476                 pipe->waiting_writers--;
477         }
478 out:
479         __pipe_unlock(pipe);
480         if (do_wakeup) {
481                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLRDNORM);
482                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
483         }
484         if (ret > 0 && sb_start_write_trylock(file_inode(filp)->i_sb)) {
485                 int err = file_update_time(filp);
486                 if (err)
487                         ret = err;
488                 sb_end_write(file_inode(filp)->i_sb);
489         }
490         return ret;
491 }
492
493 static long pipe_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
494 {
495         struct pipe_inode_info *pipe = filp->private_data;
496         int count, buf, nrbufs;
497
498         switch (cmd) {
499                 case FIONREAD:
500                         __pipe_lock(pipe);
501                         count = 0;
502                         buf = pipe->curbuf;
503                         nrbufs = pipe->nrbufs;
504                         while (--nrbufs >= 0) {
505                                 count += pipe->bufs[buf].len;
506                                 buf = (buf+1) & (pipe->buffers - 1);
507                         }
508                         __pipe_unlock(pipe);
509
510                         return put_user(count, (int __user *)arg);
511                 default:
512                         return -ENOIOCTLCMD;
513         }
514 }
515
516 /* No kernel lock held - fine */
517 static unsigned int
518 pipe_poll(struct file *filp, poll_table *wait)
519 {
520         unsigned int mask;
521         struct pipe_inode_info *pipe = filp->private_data;
522         int nrbufs;
523
524         poll_wait(filp, &pipe->wait, wait);
525
526         /* Reading only -- no need for acquiring the semaphore.  */
527         nrbufs = pipe->nrbufs;
528         mask = 0;
529         if (filp->f_mode & FMODE_READ) {
530                 mask = (nrbufs > 0) ? POLLIN | POLLRDNORM : 0;
531                 if (!pipe->writers && filp->f_version != pipe->w_counter)
532                         mask |= POLLHUP;
533         }
534
535         if (filp->f_mode & FMODE_WRITE) {
536                 mask |= (nrbufs < pipe->buffers) ? POLLOUT | POLLWRNORM : 0;
537                 /*
538                  * Most Unices do not set POLLERR for FIFOs but on Linux they
539                  * behave exactly like pipes for poll().
540                  */
541                 if (!pipe->readers)
542                         mask |= POLLERR;
543         }
544
545         return mask;
546 }
547
548 static void put_pipe_info(struct inode *inode, struct pipe_inode_info *pipe)
549 {
550         int kill = 0;
551
552         spin_lock(&inode->i_lock);
553         if (!--pipe->files) {
554                 inode->i_pipe = NULL;
555                 kill = 1;
556         }
557         spin_unlock(&inode->i_lock);
558
559         if (kill)
560                 free_pipe_info(pipe);
561 }
562
563 static int
564 pipe_release(struct inode *inode, struct file *file)
565 {
566         struct pipe_inode_info *pipe = file->private_data;
567
568         __pipe_lock(pipe);
569         if (file->f_mode & FMODE_READ)
570                 pipe->readers--;
571         if (file->f_mode & FMODE_WRITE)
572                 pipe->writers--;
573
574         if (pipe->readers || pipe->writers) {
575                 wake_up_interruptible_sync_poll(&pipe->wait, POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM | POLLERR | POLLHUP);
576                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
577                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
578         }
579         __pipe_unlock(pipe);
580
581         put_pipe_info(inode, pipe);
582         return 0;
583 }
584
585 static int
586 pipe_fasync(int fd, struct file *filp, int on)
587 {
588         struct pipe_inode_info *pipe = filp->private_data;
589         int retval = 0;
590
591         __pipe_lock(pipe);
592         if (filp->f_mode & FMODE_READ)
593                 retval = fasync_helper(fd, filp, on, &pipe->fasync_readers);
594         if ((filp->f_mode & FMODE_WRITE) && retval >= 0) {
595                 retval = fasync_helper(fd, filp, on, &pipe->fasync_writers);
596                 if (retval < 0 && (filp->f_mode & FMODE_READ))
597                         /* this can happen only if on == T */
598                         fasync_helper(-1, filp, 0, &pipe->fasync_readers);
599         }
600         __pipe_unlock(pipe);
601         return retval;
602 }
603
604 static unsigned long account_pipe_buffers(struct user_struct *user,
605                                  unsigned long old, unsigned long new)
606 {
607         return atomic_long_add_return(new - old, &user->pipe_bufs);
608 }
609
610 static bool too_many_pipe_buffers_soft(unsigned long user_bufs)
611 {
612         return pipe_user_pages_soft && user_bufs >= pipe_user_pages_soft;
613 }
614
615 static bool too_many_pipe_buffers_hard(unsigned long user_bufs)
616 {
617         return pipe_user_pages_hard && user_bufs >= pipe_user_pages_hard;
618 }
619
620 struct pipe_inode_info *alloc_pipe_info(void)
621 {
622         struct pipe_inode_info *pipe;
623         unsigned long pipe_bufs = PIPE_DEF_BUFFERS;
624         struct user_struct *user = get_current_user();
625         unsigned long user_bufs;
626
627         pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL_ACCOUNT);
628         if (pipe == NULL)
629                 goto out_free_uid;
630
631         if (pipe_bufs * PAGE_SIZE > pipe_max_size && !capable(CAP_SYS_RESOURCE))
632                 pipe_bufs = pipe_max_size >> PAGE_SHIFT;
633
634         user_bufs = account_pipe_buffers(user, 0, pipe_bufs);
635
636         if (too_many_pipe_buffers_soft(user_bufs)) {
637                 user_bufs = account_pipe_buffers(user, pipe_bufs, 1);
638                 pipe_bufs = 1;
639         }
640
641         if (too_many_pipe_buffers_hard(user_bufs))
642                 goto out_revert_acct;
643
644         pipe->bufs = kcalloc(pipe_bufs, sizeof(struct pipe_buffer),
645                              GFP_KERNEL_ACCOUNT);
646
647         if (pipe->bufs) {
648                 init_waitqueue_head(&pipe->wait);
649                 pipe->r_counter = pipe->w_counter = 1;
650                 pipe->buffers = pipe_bufs;
651                 pipe->user = user;
652                 mutex_init(&pipe->mutex);
653                 return pipe;
654         }
655
656 out_revert_acct:
657         (void) account_pipe_buffers(user, pipe_bufs, 0);
658         kfree(pipe);
659 out_free_uid:
660         free_uid(user);
661         return NULL;
662 }
663
664 void free_pipe_info(struct pipe_inode_info *pipe)
665 {
666         int i;
667
668         (void) account_pipe_buffers(pipe->user, pipe->buffers, 0);
669         free_uid(pipe->user);
670         for (i = 0; i < pipe->buffers; i++) {
671                 struct pipe_buffer *buf = pipe->bufs + i;
672                 if (buf->ops)
673                         pipe_buf_release(pipe, buf);
674         }
675         if (pipe->tmp_page)
676                 __free_page(pipe->tmp_page);
677         kfree(pipe->bufs);
678         kfree(pipe);
679 }
680
681 static struct vfsmount *pipe_mnt __read_mostly;
682
683 /*
684  * pipefs_dname() is called from d_path().
685  */
686 static char *pipefs_dname(struct dentry *dentry, char *buffer, int buflen)
687 {
688         return dynamic_dname(dentry, buffer, buflen, "pipe:[%lu]",
689                                 d_inode(dentry)->i_ino);
690 }
691
692 static const struct dentry_operations pipefs_dentry_operations = {
693         .d_dname        = pipefs_dname,
694 };
695
696 static struct inode * get_pipe_inode(void)
697 {
698         struct inode *inode = new_inode_pseudo(pipe_mnt->mnt_sb);
699         struct pipe_inode_info *pipe;
700
701         if (!inode)
702                 goto fail_inode;
703
704         inode->i_ino = get_next_ino();
705
706         pipe = alloc_pipe_info();
707         if (!pipe)
708                 goto fail_iput;
709
710         inode->i_pipe = pipe;
711         pipe->files = 2;
712         pipe->readers = pipe->writers = 1;
713         inode->i_fop = &pipefifo_fops;
714
715         /*
716          * Mark the inode dirty from the very beginning,
717          * that way it will never be moved to the dirty
718          * list because "mark_inode_dirty()" will think
719          * that it already _is_ on the dirty list.
720          */
721         inode->i_state = I_DIRTY;
722         inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
723         inode->i_uid = current_fsuid();
724         inode->i_gid = current_fsgid();
725         inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
726
727         return inode;
728
729 fail_iput:
730         iput(inode);
731
732 fail_inode:
733         return NULL;
734 }
735
736 int create_pipe_files(struct file **res, int flags)
737 {
738         int err;
739         struct inode *inode = get_pipe_inode();
740         struct file *f;
741         struct path path;
742         static struct qstr name = { .name = "" };
743
744         if (!inode)
745                 return -ENFILE;
746
747         err = -ENOMEM;
748         path.dentry = d_alloc_pseudo(pipe_mnt->mnt_sb, &name);
749         if (!path.dentry)
750                 goto err_inode;
751         path.mnt = mntget(pipe_mnt);
752
753         d_instantiate(path.dentry, inode);
754
755         f = alloc_file(&path, FMODE_WRITE, &pipefifo_fops);
756         if (IS_ERR(f)) {
757                 err = PTR_ERR(f);
758                 goto err_dentry;
759         }
760
761         f->f_flags = O_WRONLY | (flags & (O_NONBLOCK | O_DIRECT));
762         f->private_data = inode->i_pipe;
763
764         res[0] = alloc_file(&path, FMODE_READ, &pipefifo_fops);
765         if (IS_ERR(res[0])) {
766                 err = PTR_ERR(res[0]);
767                 goto err_file;
768         }
769
770         path_get(&path);
771         res[0]->private_data = inode->i_pipe;
772         res[0]->f_flags = O_RDONLY | (flags & O_NONBLOCK);
773         res[1] = f;
774         return 0;
775
776 err_file:
777         put_filp(f);
778 err_dentry:
779         free_pipe_info(inode->i_pipe);
780         path_put(&path);
781         return err;
782
783 err_inode:
784         free_pipe_info(inode->i_pipe);
785         iput(inode);
786         return err;
787 }
788
789 static int __do_pipe_flags(int *fd, struct file **files, int flags)
790 {
791         int error;
792         int fdw, fdr;
793
794         if (flags & ~(O_CLOEXEC | O_NONBLOCK | O_DIRECT))
795                 return -EINVAL;
796
797         error = create_pipe_files(files, flags);
798         if (error)
799                 return error;
800
801         error = get_unused_fd_flags(flags);
802         if (error < 0)
803                 goto err_read_pipe;
804         fdr = error;
805
806         error = get_unused_fd_flags(flags);
807         if (error < 0)
808                 goto err_fdr;
809         fdw = error;
810
811         audit_fd_pair(fdr, fdw);
812         fd[0] = fdr;
813         fd[1] = fdw;
814         return 0;
815
816  err_fdr:
817         put_unused_fd(fdr);
818  err_read_pipe:
819         fput(files[0]);
820         fput(files[1]);
821         return error;
822 }
823
824 int do_pipe_flags(int *fd, int flags)
825 {
826         struct file *files[2];
827         int error = __do_pipe_flags(fd, files, flags);
828         if (!error) {
829                 fd_install(fd[0], files[0]);
830                 fd_install(fd[1], files[1]);
831         }
832         return error;
833 }
834
835 /*
836  * sys_pipe() is the normal C calling standard for creating
837  * a pipe. It's not the way Unix traditionally does this, though.
838  */
839 SYSCALL_DEFINE2(pipe2, int __user *, fildes, int, flags)
840 {
841         struct file *files[2];
842         int fd[2];
843         int error;
844
845         error = __do_pipe_flags(fd, files, flags);
846         if (!error) {
847                 if (unlikely(copy_to_user(fildes, fd, sizeof(fd)))) {
848                         fput(files[0]);
849                         fput(files[1]);
850                         put_unused_fd(fd[0]);
851                         put_unused_fd(fd[1]);
852                         error = -EFAULT;
853                 } else {
854                         fd_install(fd[0], files[0]);
855                         fd_install(fd[1], files[1]);
856                 }
857         }
858         return error;
859 }
860
861 SYSCALL_DEFINE1(pipe, int __user *, fildes)
862 {
863         return sys_pipe2(fildes, 0);
864 }
865
866 static int wait_for_partner(struct pipe_inode_info *pipe, unsigned int *cnt)
867 {
868         int cur = *cnt; 
869
870         while (cur == *cnt) {
871                 pipe_wait(pipe);
872                 if (signal_pending(current))
873                         break;
874         }
875         return cur == *cnt ? -ERESTARTSYS : 0;
876 }
877
878 static void wake_up_partner(struct pipe_inode_info *pipe)
879 {
880         wake_up_interruptible(&pipe->wait);
881 }
882
883 static int fifo_open(struct inode *inode, struct file *filp)
884 {
885         struct pipe_inode_info *pipe;
886         bool is_pipe = inode->i_sb->s_magic == PIPEFS_MAGIC;
887         int ret;
888
889         filp->f_version = 0;
890
891         spin_lock(&inode->i_lock);
892         if (inode->i_pipe) {
893                 pipe = inode->i_pipe;
894                 pipe->files++;
895                 spin_unlock(&inode->i_lock);
896         } else {
897                 spin_unlock(&inode->i_lock);
898                 pipe = alloc_pipe_info();
899                 if (!pipe)
900                         return -ENOMEM;
901                 pipe->files = 1;
902                 spin_lock(&inode->i_lock);
903                 if (unlikely(inode->i_pipe)) {
904                         inode->i_pipe->files++;
905                         spin_unlock(&inode->i_lock);
906                         free_pipe_info(pipe);
907                         pipe = inode->i_pipe;
908                 } else {
909                         inode->i_pipe = pipe;
910                         spin_unlock(&inode->i_lock);
911                 }
912         }
913         filp->private_data = pipe;
914         /* OK, we have a pipe and it's pinned down */
915
916         __pipe_lock(pipe);
917
918         /* We can only do regular read/write on fifos */
919         filp->f_mode &= (FMODE_READ | FMODE_WRITE);
920
921         switch (filp->f_mode) {
922         case FMODE_READ:
923         /*
924          *  O_RDONLY
925          *  POSIX.1 says that O_NONBLOCK means return with the FIFO
926          *  opened, even when there is no process writing the FIFO.
927          */
928                 pipe->r_counter++;
929                 if (pipe->readers++ == 0)
930                         wake_up_partner(pipe);
931
932                 if (!is_pipe && !pipe->writers) {
933                         if ((filp->f_flags & O_NONBLOCK)) {
934                                 /* suppress POLLHUP until we have
935                                  * seen a writer */
936                                 filp->f_version = pipe->w_counter;
937                         } else {
938                                 if (wait_for_partner(pipe, &pipe->w_counter))
939                                         goto err_rd;
940                         }
941                 }
942                 break;
943         
944         case FMODE_WRITE:
945         /*
946          *  O_WRONLY
947          *  POSIX.1 says that O_NONBLOCK means return -1 with
948          *  errno=ENXIO when there is no process reading the FIFO.
949          */
950                 ret = -ENXIO;
951                 if (!is_pipe && (filp->f_flags & O_NONBLOCK) && !pipe->readers)
952                         goto err;
953
954                 pipe->w_counter++;
955                 if (!pipe->writers++)
956                         wake_up_partner(pipe);
957
958                 if (!is_pipe && !pipe->readers) {
959                         if (wait_for_partner(pipe, &pipe->r_counter))
960                                 goto err_wr;
961                 }
962                 break;
963         
964         case FMODE_READ | FMODE_WRITE:
965         /*
966          *  O_RDWR
967          *  POSIX.1 leaves this case "undefined" when O_NONBLOCK is set.
968          *  This implementation will NEVER block on a O_RDWR open, since
969          *  the process can at least talk to itself.
970          */
971
972                 pipe->readers++;
973                 pipe->writers++;
974                 pipe->r_counter++;
975                 pipe->w_counter++;
976                 if (pipe->readers == 1 || pipe->writers == 1)
977                         wake_up_partner(pipe);
978                 break;
979
980         default:
981                 ret = -EINVAL;
982                 goto err;
983         }
984
985         /* Ok! */
986         __pipe_unlock(pipe);
987         return 0;
988
989 err_rd:
990         if (!--pipe->readers)
991                 wake_up_interruptible(&pipe->wait);
992         ret = -ERESTARTSYS;
993         goto err;
994
995 err_wr:
996         if (!--pipe->writers)
997                 wake_up_interruptible(&pipe->wait);
998         ret = -ERESTARTSYS;
999         goto err;
1000
1001 err:
1002         __pipe_unlock(pipe);
1003
1004         put_pipe_info(inode, pipe);
1005         return ret;
1006 }
1007
1008 const struct file_operations pipefifo_fops = {
1009         .open           = fifo_open,
1010         .llseek         = no_llseek,
1011         .read_iter      = pipe_read,
1012         .write_iter     = pipe_write,
1013         .poll           = pipe_poll,
1014         .unlocked_ioctl = pipe_ioctl,
1015         .release        = pipe_release,
1016         .fasync         = pipe_fasync,
1017 };
1018
1019 /*
1020  * Currently we rely on the pipe array holding a power-of-2 number
1021  * of pages.
1022  */
1023 static inline unsigned int round_pipe_size(unsigned int size)
1024 {
1025         unsigned long nr_pages;
1026
1027         nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1028         return roundup_pow_of_two(nr_pages) << PAGE_SHIFT;
1029 }
1030
1031 /*
1032  * Allocate a new array of pipe buffers and copy the info over. Returns the
1033  * pipe size if successful, or return -ERROR on error.
1034  */
1035 static long pipe_set_size(struct pipe_inode_info *pipe, unsigned long arg)
1036 {
1037         struct pipe_buffer *bufs;
1038         unsigned int size, nr_pages;
1039         unsigned long user_bufs;
1040         long ret = 0;
1041
1042         size = round_pipe_size(arg);
1043         nr_pages = size >> PAGE_SHIFT;
1044
1045         if (!nr_pages)
1046                 return -EINVAL;
1047
1048         /*
1049          * If trying to increase the pipe capacity, check that an
1050          * unprivileged user is not trying to exceed various limits
1051          * (soft limit check here, hard limit check just below).
1052          * Decreasing the pipe capacity is always permitted, even
1053          * if the user is currently over a limit.
1054          */
1055         if (nr_pages > pipe->buffers &&
1056                         size > pipe_max_size && !capable(CAP_SYS_RESOURCE))
1057                 return -EPERM;
1058
1059         user_bufs = account_pipe_buffers(pipe->user, pipe->buffers, nr_pages);
1060
1061         if (nr_pages > pipe->buffers &&
1062                         (too_many_pipe_buffers_hard(user_bufs) ||
1063                          too_many_pipe_buffers_soft(user_bufs)) &&
1064                         !capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
1065                 ret = -EPERM;
1066                 goto out_revert_acct;
1067         }
1068
1069         /*
1070          * We can shrink the pipe, if arg >= pipe->nrbufs. Since we don't
1071          * expect a lot of shrink+grow operations, just free and allocate
1072          * again like we would do for growing. If the pipe currently
1073          * contains more buffers than arg, then return busy.
1074          */
1075         if (nr_pages < pipe->nrbufs) {
1076                 ret = -EBUSY;
1077                 goto out_revert_acct;
1078         }
1079
1080         bufs = kcalloc(nr_pages, sizeof(*bufs),
1081                        GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
1082         if (unlikely(!bufs)) {
1083                 ret = -ENOMEM;
1084                 goto out_revert_acct;
1085         }
1086
1087         /*
1088          * The pipe array wraps around, so just start the new one at zero
1089          * and adjust the indexes.
1090          */
1091         if (pipe->nrbufs) {
1092                 unsigned int tail;
1093                 unsigned int head;
1094
1095                 tail = pipe->curbuf + pipe->nrbufs;
1096                 if (tail < pipe->buffers)
1097                         tail = 0;
1098                 else
1099                         tail &= (pipe->buffers - 1);
1100
1101                 head = pipe->nrbufs - tail;
1102                 if (head)
1103                         memcpy(bufs, pipe->bufs + pipe->curbuf, head * sizeof(struct pipe_buffer));
1104                 if (tail)
1105                         memcpy(bufs + head, pipe->bufs, tail * sizeof(struct pipe_buffer));
1106         }
1107
1108         pipe->curbuf = 0;
1109         kfree(pipe->bufs);
1110         pipe->bufs = bufs;
1111         pipe->buffers = nr_pages;
1112         return nr_pages * PAGE_SIZE;
1113
1114 out_revert_acct:
1115         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->buffers);
1116         return ret;
1117 }
1118
1119 /*
1120  * This should work even if CONFIG_PROC_FS isn't set, as proc_dointvec_minmax
1121  * will return an error.
1122  */
1123 int pipe_proc_fn(struct ctl_table *table, int write, void __user *buf,
1124                  size_t *lenp, loff_t *ppos)
1125 {
1126         int ret;
1127
1128         ret = proc_dointvec_minmax(table, write, buf, lenp, ppos);
1129         if (ret < 0 || !write)
1130                 return ret;
1131
1132         pipe_max_size = round_pipe_size(pipe_max_size);
1133         return ret;
1134 }
1135
1136 /*
1137  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1138  * location, so checking ->i_pipe is not enough to verify that this is a
1139  * pipe.
1140  */
1141 struct pipe_inode_info *get_pipe_info(struct file *file)
1142 {
1143         return file->f_op == &pipefifo_fops ? file->private_data : NULL;
1144 }
1145
1146 long pipe_fcntl(struct file *file, unsigned int cmd, unsigned long arg)
1147 {
1148         struct pipe_inode_info *pipe;
1149         long ret;
1150
1151         pipe = get_pipe_info(file);
1152         if (!pipe)
1153                 return -EBADF;
1154
1155         __pipe_lock(pipe);
1156
1157         switch (cmd) {
1158         case F_SETPIPE_SZ:
1159                 ret = pipe_set_size(pipe, arg);
1160                 break;
1161         case F_GETPIPE_SZ:
1162                 ret = pipe->buffers * PAGE_SIZE;
1163                 break;
1164         default:
1165                 ret = -EINVAL;
1166                 break;
1167         }
1168
1169         __pipe_unlock(pipe);
1170         return ret;
1171 }
1172
1173 static const struct super_operations pipefs_ops = {
1174         .destroy_inode = free_inode_nonrcu,
1175         .statfs = simple_statfs,
1176 };
1177
1178 /*
1179  * pipefs should _never_ be mounted by userland - too much of security hassle,
1180  * no real gain from having the whole whorehouse mounted. So we don't need
1181  * any operations on the root directory. However, we need a non-trivial
1182  * d_name - pipe: will go nicely and kill the special-casing in procfs.
1183  */
1184 static struct dentry *pipefs_mount(struct file_system_type *fs_type,
1185                          int flags, const char *dev_name, void *data)
1186 {
1187         return mount_pseudo(fs_type, "pipe:", &pipefs_ops,
1188                         &pipefs_dentry_operations, PIPEFS_MAGIC);
1189 }
1190
1191 static struct file_system_type pipe_fs_type = {
1192         .name           = "pipefs",
1193         .mount          = pipefs_mount,
1194         .kill_sb        = kill_anon_super,
1195 };
1196
1197 static int __init init_pipe_fs(void)
1198 {
1199         int err = register_filesystem(&pipe_fs_type);
1200
1201         if (!err) {
1202                 pipe_mnt = kern_mount(&pipe_fs_type);
1203                 if (IS_ERR(pipe_mnt)) {
1204                         err = PTR_ERR(pipe_mnt);
1205                         unregister_filesystem(&pipe_fs_type);
1206                 }
1207         }
1208         return err;
1209 }
1210
1211 fs_initcall(init_pipe_fs);