splice: fix sendfile() issue with relay
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32
33 /*
34  * Attempt to steal a page from a pipe buffer. This should perhaps go into
35  * a vm helper function, it's already simplified quite a bit by the
36  * addition of remove_mapping(). If success is returned, the caller may
37  * attempt to reuse this page for another destination.
38  */
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
40                                      struct pipe_buffer *buf)
41 {
42         struct page *page = buf->page;
43         struct address_space *mapping;
44
45         lock_page(page);
46
47         mapping = page_mapping(page);
48         if (mapping) {
49                 WARN_ON(!PageUptodate(page));
50
51                 /*
52                  * At least for ext2 with nobh option, we need to wait on
53                  * writeback completing on this page, since we'll remove it
54                  * from the pagecache.  Otherwise truncate wont wait on the
55                  * page, allowing the disk blocks to be reused by someone else
56                  * before we actually wrote our data to them. fs corruption
57                  * ensues.
58                  */
59                 wait_on_page_writeback(page);
60
61                 if (PagePrivate(page))
62                         try_to_release_page(page, GFP_KERNEL);
63
64                 /*
65                  * If we succeeded in removing the mapping, set LRU flag
66                  * and return good.
67                  */
68                 if (remove_mapping(mapping, page)) {
69                         buf->flags |= PIPE_BUF_FLAG_LRU;
70                         return 0;
71                 }
72         }
73
74         /*
75          * Raced with truncate or failed to remove page from current
76          * address space, unlock and return failure.
77          */
78         unlock_page(page);
79         return 1;
80 }
81
82 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
83                                         struct pipe_buffer *buf)
84 {
85         page_cache_release(buf->page);
86         buf->flags &= ~PIPE_BUF_FLAG_LRU;
87 }
88
89 /*
90  * Check whether the contents of buf is OK to access. Since the content
91  * is a page cache page, IO may be in flight.
92  */
93 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
94                                        struct pipe_buffer *buf)
95 {
96         struct page *page = buf->page;
97         int err;
98
99         if (!PageUptodate(page)) {
100                 lock_page(page);
101
102                 /*
103                  * Page got truncated/unhashed. This will cause a 0-byte
104                  * splice, if this is the first page.
105                  */
106                 if (!page->mapping) {
107                         err = -ENODATA;
108                         goto error;
109                 }
110
111                 /*
112                  * Uh oh, read-error from disk.
113                  */
114                 if (!PageUptodate(page)) {
115                         err = -EIO;
116                         goto error;
117                 }
118
119                 /*
120                  * Page is ok afterall, we are done.
121                  */
122                 unlock_page(page);
123         }
124
125         return 0;
126 error:
127         unlock_page(page);
128         return err;
129 }
130
131 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
132         .can_merge = 0,
133         .map = generic_pipe_buf_map,
134         .unmap = generic_pipe_buf_unmap,
135         .confirm = page_cache_pipe_buf_confirm,
136         .release = page_cache_pipe_buf_release,
137         .steal = page_cache_pipe_buf_steal,
138         .get = generic_pipe_buf_get,
139 };
140
141 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
142                                     struct pipe_buffer *buf)
143 {
144         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
145                 return 1;
146
147         buf->flags |= PIPE_BUF_FLAG_LRU;
148         return generic_pipe_buf_steal(pipe, buf);
149 }
150
151 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
152         .can_merge = 0,
153         .map = generic_pipe_buf_map,
154         .unmap = generic_pipe_buf_unmap,
155         .confirm = generic_pipe_buf_confirm,
156         .release = page_cache_pipe_buf_release,
157         .steal = user_page_pipe_buf_steal,
158         .get = generic_pipe_buf_get,
159 };
160
161 /**
162  * splice_to_pipe - fill passed data into a pipe
163  * @pipe:       pipe to fill
164  * @spd:        data to fill
165  *
166  * Description:
167  *    @spd contains a map of pages and len/offset tuples, along with
168  *    the struct pipe_buf_operations associated with these pages. This
169  *    function will link that data to the pipe.
170  *
171  */
172 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
173                        struct splice_pipe_desc *spd)
174 {
175         unsigned int spd_pages = spd->nr_pages;
176         int ret, do_wakeup, page_nr;
177
178         ret = 0;
179         do_wakeup = 0;
180         page_nr = 0;
181
182         if (pipe->inode)
183                 mutex_lock(&pipe->inode->i_mutex);
184
185         for (;;) {
186                 if (!pipe->readers) {
187                         send_sig(SIGPIPE, current, 0);
188                         if (!ret)
189                                 ret = -EPIPE;
190                         break;
191                 }
192
193                 if (pipe->nrbufs < PIPE_BUFFERS) {
194                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
195                         struct pipe_buffer *buf = pipe->bufs + newbuf;
196
197                         buf->page = spd->pages[page_nr];
198                         buf->offset = spd->partial[page_nr].offset;
199                         buf->len = spd->partial[page_nr].len;
200                         buf->private = spd->partial[page_nr].private;
201                         buf->ops = spd->ops;
202                         if (spd->flags & SPLICE_F_GIFT)
203                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
204
205                         pipe->nrbufs++;
206                         page_nr++;
207                         ret += buf->len;
208
209                         if (pipe->inode)
210                                 do_wakeup = 1;
211
212                         if (!--spd->nr_pages)
213                                 break;
214                         if (pipe->nrbufs < PIPE_BUFFERS)
215                                 continue;
216
217                         break;
218                 }
219
220                 if (spd->flags & SPLICE_F_NONBLOCK) {
221                         if (!ret)
222                                 ret = -EAGAIN;
223                         break;
224                 }
225
226                 if (signal_pending(current)) {
227                         if (!ret)
228                                 ret = -ERESTARTSYS;
229                         break;
230                 }
231
232                 if (do_wakeup) {
233                         smp_mb();
234                         if (waitqueue_active(&pipe->wait))
235                                 wake_up_interruptible_sync(&pipe->wait);
236                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
237                         do_wakeup = 0;
238                 }
239
240                 pipe->waiting_writers++;
241                 pipe_wait(pipe);
242                 pipe->waiting_writers--;
243         }
244
245         if (pipe->inode) {
246                 mutex_unlock(&pipe->inode->i_mutex);
247
248                 if (do_wakeup) {
249                         smp_mb();
250                         if (waitqueue_active(&pipe->wait))
251                                 wake_up_interruptible(&pipe->wait);
252                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
253                 }
254         }
255
256         while (page_nr < spd_pages)
257                 spd->spd_release(spd, page_nr++);
258
259         return ret;
260 }
261
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
263 {
264         page_cache_release(spd->pages[i]);
265 }
266
267 static int
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269                            struct pipe_inode_info *pipe, size_t len,
270                            unsigned int flags)
271 {
272         struct address_space *mapping = in->f_mapping;
273         unsigned int loff, nr_pages, req_pages;
274         struct page *pages[PIPE_BUFFERS];
275         struct partial_page partial[PIPE_BUFFERS];
276         struct page *page;
277         pgoff_t index, end_index;
278         loff_t isize;
279         int error, page_nr;
280         struct splice_pipe_desc spd = {
281                 .pages = pages,
282                 .partial = partial,
283                 .flags = flags,
284                 .ops = &page_cache_pipe_buf_ops,
285                 .spd_release = spd_release_page,
286         };
287
288         index = *ppos >> PAGE_CACHE_SHIFT;
289         loff = *ppos & ~PAGE_CACHE_MASK;
290         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
292
293         /*
294          * Lookup the (hopefully) full range of pages we need.
295          */
296         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297         index += spd.nr_pages;
298
299         /*
300          * If find_get_pages_contig() returned fewer pages than we needed,
301          * readahead/allocate the rest and fill in the holes.
302          */
303         if (spd.nr_pages < nr_pages)
304                 page_cache_sync_readahead(mapping, &in->f_ra, in,
305                                 index, req_pages - spd.nr_pages);
306
307         error = 0;
308         while (spd.nr_pages < nr_pages) {
309                 /*
310                  * Page could be there, find_get_pages_contig() breaks on
311                  * the first hole.
312                  */
313                 page = find_get_page(mapping, index);
314                 if (!page) {
315                         /*
316                          * page didn't exist, allocate one.
317                          */
318                         page = page_cache_alloc_cold(mapping);
319                         if (!page)
320                                 break;
321
322                         error = add_to_page_cache_lru(page, mapping, index,
323                                                 mapping_gfp_mask(mapping));
324                         if (unlikely(error)) {
325                                 page_cache_release(page);
326                                 if (error == -EEXIST)
327                                         continue;
328                                 break;
329                         }
330                         /*
331                          * add_to_page_cache() locks the page, unlock it
332                          * to avoid convoluting the logic below even more.
333                          */
334                         unlock_page(page);
335                 }
336
337                 pages[spd.nr_pages++] = page;
338                 index++;
339         }
340
341         /*
342          * Now loop over the map and see if we need to start IO on any
343          * pages, fill in the partial map, etc.
344          */
345         index = *ppos >> PAGE_CACHE_SHIFT;
346         nr_pages = spd.nr_pages;
347         spd.nr_pages = 0;
348         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349                 unsigned int this_len;
350
351                 if (!len)
352                         break;
353
354                 /*
355                  * this_len is the max we'll use from this page
356                  */
357                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358                 page = pages[page_nr];
359
360                 if (PageReadahead(page))
361                         page_cache_async_readahead(mapping, &in->f_ra, in,
362                                         page, index, req_pages - page_nr);
363
364                 /*
365                  * If the page isn't uptodate, we may need to start io on it
366                  */
367                 if (!PageUptodate(page)) {
368                         /*
369                          * If in nonblock mode then dont block on waiting
370                          * for an in-flight io page
371                          */
372                         if (flags & SPLICE_F_NONBLOCK) {
373                                 if (TestSetPageLocked(page)) {
374                                         error = -EAGAIN;
375                                         break;
376                                 }
377                         } else
378                                 lock_page(page);
379
380                         /*
381                          * page was truncated, stop here. if this isn't the
382                          * first page, we'll just complete what we already
383                          * added
384                          */
385                         if (!page->mapping) {
386                                 unlock_page(page);
387                                 break;
388                         }
389                         /*
390                          * page was already under io and is now done, great
391                          */
392                         if (PageUptodate(page)) {
393                                 unlock_page(page);
394                                 goto fill_it;
395                         }
396
397                         /*
398                          * need to read in the page
399                          */
400                         error = mapping->a_ops->readpage(in, page);
401                         if (unlikely(error)) {
402                                 /*
403                                  * We really should re-lookup the page here,
404                                  * but it complicates things a lot. Instead
405                                  * lets just do what we already stored, and
406                                  * we'll get it the next time we are called.
407                                  */
408                                 if (error == AOP_TRUNCATED_PAGE)
409                                         error = 0;
410
411                                 break;
412                         }
413                 }
414 fill_it:
415                 /*
416                  * i_size must be checked after PageUptodate.
417                  */
418                 isize = i_size_read(mapping->host);
419                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
420                 if (unlikely(!isize || index > end_index))
421                         break;
422
423                 /*
424                  * if this is the last page, see if we need to shrink
425                  * the length and stop
426                  */
427                 if (end_index == index) {
428                         unsigned int plen;
429
430                         /*
431                          * max good bytes in this page
432                          */
433                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
434                         if (plen <= loff)
435                                 break;
436
437                         /*
438                          * force quit after adding this page
439                          */
440                         this_len = min(this_len, plen - loff);
441                         len = this_len;
442                 }
443
444                 partial[page_nr].offset = loff;
445                 partial[page_nr].len = this_len;
446                 len -= this_len;
447                 loff = 0;
448                 spd.nr_pages++;
449                 index++;
450         }
451
452         /*
453          * Release any pages at the end, if we quit early. 'page_nr' is how far
454          * we got, 'nr_pages' is how many pages are in the map.
455          */
456         while (page_nr < nr_pages)
457                 page_cache_release(pages[page_nr++]);
458         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
459
460         if (spd.nr_pages)
461                 return splice_to_pipe(pipe, &spd);
462
463         return error;
464 }
465
466 /**
467  * generic_file_splice_read - splice data from file to a pipe
468  * @in:         file to splice from
469  * @ppos:       position in @in
470  * @pipe:       pipe to splice to
471  * @len:        number of bytes to splice
472  * @flags:      splice modifier flags
473  *
474  * Description:
475  *    Will read pages from given file and fill them into a pipe. Can be
476  *    used as long as the address_space operations for the source implements
477  *    a readpage() hook.
478  *
479  */
480 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
481                                  struct pipe_inode_info *pipe, size_t len,
482                                  unsigned int flags)
483 {
484         loff_t isize, left;
485         int ret;
486
487         isize = i_size_read(in->f_mapping->host);
488         if (unlikely(*ppos >= isize))
489                 return 0;
490
491         left = isize - *ppos;
492         if (unlikely(left < len))
493                 len = left;
494
495         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
496         if (ret > 0)
497                 *ppos += ret;
498
499         return ret;
500 }
501
502 EXPORT_SYMBOL(generic_file_splice_read);
503
504 /*
505  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
506  * using sendpage(). Return the number of bytes sent.
507  */
508 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
509                             struct pipe_buffer *buf, struct splice_desc *sd)
510 {
511         struct file *file = sd->u.file;
512         loff_t pos = sd->pos;
513         int ret, more;
514
515         ret = buf->ops->confirm(pipe, buf);
516         if (!ret) {
517                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
518
519                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
520                                            sd->len, &pos, more);
521         }
522
523         return ret;
524 }
525
526 /*
527  * This is a little more tricky than the file -> pipe splicing. There are
528  * basically three cases:
529  *
530  *      - Destination page already exists in the address space and there
531  *        are users of it. For that case we have no other option that
532  *        copying the data. Tough luck.
533  *      - Destination page already exists in the address space, but there
534  *        are no users of it. Make sure it's uptodate, then drop it. Fall
535  *        through to last case.
536  *      - Destination page does not exist, we can add the pipe page to
537  *        the page cache and avoid the copy.
538  *
539  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
540  * sd->flags), we attempt to migrate pages from the pipe to the output
541  * file address space page cache. This is possible if no one else has
542  * the pipe page referenced outside of the pipe and page cache. If
543  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
544  * a new page in the output file page cache and fill/dirty that.
545  */
546 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
547                         struct splice_desc *sd)
548 {
549         struct file *file = sd->u.file;
550         struct address_space *mapping = file->f_mapping;
551         unsigned int offset, this_len;
552         struct page *page;
553         void *fsdata;
554         int ret;
555
556         /*
557          * make sure the data in this buffer is uptodate
558          */
559         ret = buf->ops->confirm(pipe, buf);
560         if (unlikely(ret))
561                 return ret;
562
563         offset = sd->pos & ~PAGE_CACHE_MASK;
564
565         this_len = sd->len;
566         if (this_len + offset > PAGE_CACHE_SIZE)
567                 this_len = PAGE_CACHE_SIZE - offset;
568
569         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
570                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
571         if (unlikely(ret))
572                 goto out;
573
574         if (buf->page != page) {
575                 /*
576                  * Careful, ->map() uses KM_USER0!
577                  */
578                 char *src = buf->ops->map(pipe, buf, 1);
579                 char *dst = kmap_atomic(page, KM_USER1);
580
581                 memcpy(dst + offset, src + buf->offset, this_len);
582                 flush_dcache_page(page);
583                 kunmap_atomic(dst, KM_USER1);
584                 buf->ops->unmap(pipe, buf, src);
585         }
586         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
587                                 page, fsdata);
588 out:
589         return ret;
590 }
591
592 /**
593  * __splice_from_pipe - splice data from a pipe to given actor
594  * @pipe:       pipe to splice from
595  * @sd:         information to @actor
596  * @actor:      handler that splices the data
597  *
598  * Description:
599  *    This function does little more than loop over the pipe and call
600  *    @actor to do the actual moving of a single struct pipe_buffer to
601  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
602  *    pipe_to_user.
603  *
604  */
605 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
606                            splice_actor *actor)
607 {
608         int ret, do_wakeup, err;
609
610         ret = 0;
611         do_wakeup = 0;
612
613         for (;;) {
614                 if (pipe->nrbufs) {
615                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
616                         const struct pipe_buf_operations *ops = buf->ops;
617
618                         sd->len = buf->len;
619                         if (sd->len > sd->total_len)
620                                 sd->len = sd->total_len;
621
622                         err = actor(pipe, buf, sd);
623                         if (err <= 0) {
624                                 if (!ret && err != -ENODATA)
625                                         ret = err;
626
627                                 break;
628                         }
629
630                         ret += err;
631                         buf->offset += err;
632                         buf->len -= err;
633
634                         sd->len -= err;
635                         sd->pos += err;
636                         sd->total_len -= err;
637                         if (sd->len)
638                                 continue;
639
640                         if (!buf->len) {
641                                 buf->ops = NULL;
642                                 ops->release(pipe, buf);
643                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
644                                 pipe->nrbufs--;
645                                 if (pipe->inode)
646                                         do_wakeup = 1;
647                         }
648
649                         if (!sd->total_len)
650                                 break;
651                 }
652
653                 if (pipe->nrbufs)
654                         continue;
655                 if (!pipe->writers)
656                         break;
657                 if (!pipe->waiting_writers) {
658                         if (ret)
659                                 break;
660                 }
661
662                 if (sd->flags & SPLICE_F_NONBLOCK) {
663                         if (!ret)
664                                 ret = -EAGAIN;
665                         break;
666                 }
667
668                 if (signal_pending(current)) {
669                         if (!ret)
670                                 ret = -ERESTARTSYS;
671                         break;
672                 }
673
674                 if (do_wakeup) {
675                         smp_mb();
676                         if (waitqueue_active(&pipe->wait))
677                                 wake_up_interruptible_sync(&pipe->wait);
678                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
679                         do_wakeup = 0;
680                 }
681
682                 pipe_wait(pipe);
683         }
684
685         if (do_wakeup) {
686                 smp_mb();
687                 if (waitqueue_active(&pipe->wait))
688                         wake_up_interruptible(&pipe->wait);
689                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
690         }
691
692         return ret;
693 }
694 EXPORT_SYMBOL(__splice_from_pipe);
695
696 /**
697  * splice_from_pipe - splice data from a pipe to a file
698  * @pipe:       pipe to splice from
699  * @out:        file to splice to
700  * @ppos:       position in @out
701  * @len:        how many bytes to splice
702  * @flags:      splice modifier flags
703  * @actor:      handler that splices the data
704  *
705  * Description:
706  *    See __splice_from_pipe. This function locks the input and output inodes,
707  *    otherwise it's identical to __splice_from_pipe().
708  *
709  */
710 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
711                          loff_t *ppos, size_t len, unsigned int flags,
712                          splice_actor *actor)
713 {
714         ssize_t ret;
715         struct inode *inode = out->f_mapping->host;
716         struct splice_desc sd = {
717                 .total_len = len,
718                 .flags = flags,
719                 .pos = *ppos,
720                 .u.file = out,
721         };
722
723         /*
724          * The actor worker might be calling ->prepare_write and
725          * ->commit_write. Most of the time, these expect i_mutex to
726          * be held. Since this may result in an ABBA deadlock with
727          * pipe->inode, we have to order lock acquiry here.
728          */
729         inode_double_lock(inode, pipe->inode);
730         ret = __splice_from_pipe(pipe, &sd, actor);
731         inode_double_unlock(inode, pipe->inode);
732
733         return ret;
734 }
735
736 /**
737  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
738  * @pipe:       pipe info
739  * @out:        file to write to
740  * @ppos:       position in @out
741  * @len:        number of bytes to splice
742  * @flags:      splice modifier flags
743  *
744  * Description:
745  *    Will either move or copy pages (determined by @flags options) from
746  *    the given pipe inode to the given file. The caller is responsible
747  *    for acquiring i_mutex on both inodes.
748  *
749  */
750 ssize_t
751 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
752                                  loff_t *ppos, size_t len, unsigned int flags)
753 {
754         struct address_space *mapping = out->f_mapping;
755         struct inode *inode = mapping->host;
756         struct splice_desc sd = {
757                 .total_len = len,
758                 .flags = flags,
759                 .pos = *ppos,
760                 .u.file = out,
761         };
762         ssize_t ret;
763         int err;
764
765         err = remove_suid(out->f_path.dentry);
766         if (unlikely(err))
767                 return err;
768
769         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
770         if (ret > 0) {
771                 unsigned long nr_pages;
772
773                 *ppos += ret;
774                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
775
776                 /*
777                  * If file or inode is SYNC and we actually wrote some data,
778                  * sync it.
779                  */
780                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
781                         err = generic_osync_inode(inode, mapping,
782                                                   OSYNC_METADATA|OSYNC_DATA);
783
784                         if (err)
785                                 ret = err;
786                 }
787                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
788         }
789
790         return ret;
791 }
792
793 EXPORT_SYMBOL(generic_file_splice_write_nolock);
794
795 /**
796  * generic_file_splice_write - splice data from a pipe to a file
797  * @pipe:       pipe info
798  * @out:        file to write to
799  * @ppos:       position in @out
800  * @len:        number of bytes to splice
801  * @flags:      splice modifier flags
802  *
803  * Description:
804  *    Will either move or copy pages (determined by @flags options) from
805  *    the given pipe inode to the given file.
806  *
807  */
808 ssize_t
809 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
810                           loff_t *ppos, size_t len, unsigned int flags)
811 {
812         struct address_space *mapping = out->f_mapping;
813         struct inode *inode = mapping->host;
814         struct splice_desc sd = {
815                 .total_len = len,
816                 .flags = flags,
817                 .pos = *ppos,
818                 .u.file = out,
819         };
820         ssize_t ret;
821
822         inode_double_lock(inode, pipe->inode);
823         ret = remove_suid(out->f_path.dentry);
824         if (likely(!ret))
825                 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
826         inode_double_unlock(inode, pipe->inode);
827         if (ret > 0) {
828                 unsigned long nr_pages;
829
830                 *ppos += ret;
831                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
832
833                 /*
834                  * If file or inode is SYNC and we actually wrote some data,
835                  * sync it.
836                  */
837                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
838                         int err;
839
840                         mutex_lock(&inode->i_mutex);
841                         err = generic_osync_inode(inode, mapping,
842                                                   OSYNC_METADATA|OSYNC_DATA);
843                         mutex_unlock(&inode->i_mutex);
844
845                         if (err)
846                                 ret = err;
847                 }
848                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
849         }
850
851         return ret;
852 }
853
854 EXPORT_SYMBOL(generic_file_splice_write);
855
856 /**
857  * generic_splice_sendpage - splice data from a pipe to a socket
858  * @pipe:       pipe to splice from
859  * @out:        socket to write to
860  * @ppos:       position in @out
861  * @len:        number of bytes to splice
862  * @flags:      splice modifier flags
863  *
864  * Description:
865  *    Will send @len bytes from the pipe to a network socket. No data copying
866  *    is involved.
867  *
868  */
869 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
870                                 loff_t *ppos, size_t len, unsigned int flags)
871 {
872         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
873 }
874
875 EXPORT_SYMBOL(generic_splice_sendpage);
876
877 /*
878  * Attempt to initiate a splice from pipe to file.
879  */
880 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
881                            loff_t *ppos, size_t len, unsigned int flags)
882 {
883         int ret;
884
885         if (unlikely(!out->f_op || !out->f_op->splice_write))
886                 return -EINVAL;
887
888         if (unlikely(!(out->f_mode & FMODE_WRITE)))
889                 return -EBADF;
890
891         ret = rw_verify_area(WRITE, out, ppos, len);
892         if (unlikely(ret < 0))
893                 return ret;
894
895         return out->f_op->splice_write(pipe, out, ppos, len, flags);
896 }
897
898 /*
899  * Attempt to initiate a splice from a file to a pipe.
900  */
901 static long do_splice_to(struct file *in, loff_t *ppos,
902                          struct pipe_inode_info *pipe, size_t len,
903                          unsigned int flags)
904 {
905         int ret;
906
907         if (unlikely(!in->f_op || !in->f_op->splice_read))
908                 return -EINVAL;
909
910         if (unlikely(!(in->f_mode & FMODE_READ)))
911                 return -EBADF;
912
913         ret = rw_verify_area(READ, in, ppos, len);
914         if (unlikely(ret < 0))
915                 return ret;
916
917         return in->f_op->splice_read(in, ppos, pipe, len, flags);
918 }
919
920 /**
921  * splice_direct_to_actor - splices data directly between two non-pipes
922  * @in:         file to splice from
923  * @sd:         actor information on where to splice to
924  * @actor:      handles the data splicing
925  *
926  * Description:
927  *    This is a special case helper to splice directly between two
928  *    points, without requiring an explicit pipe. Internally an allocated
929  *    pipe is cached in the process, and reused during the lifetime of
930  *    that process.
931  *
932  */
933 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
934                                splice_direct_actor *actor)
935 {
936         struct pipe_inode_info *pipe;
937         long ret, bytes;
938         umode_t i_mode;
939         size_t len;
940         int i, flags;
941
942         /*
943          * We require the input being a regular file, as we don't want to
944          * randomly drop data for eg socket -> socket splicing. Use the
945          * piped splicing for that!
946          */
947         i_mode = in->f_path.dentry->d_inode->i_mode;
948         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
949                 return -EINVAL;
950
951         /*
952          * neither in nor out is a pipe, setup an internal pipe attached to
953          * 'out' and transfer the wanted data from 'in' to 'out' through that
954          */
955         pipe = current->splice_pipe;
956         if (unlikely(!pipe)) {
957                 pipe = alloc_pipe_info(NULL);
958                 if (!pipe)
959                         return -ENOMEM;
960
961                 /*
962                  * We don't have an immediate reader, but we'll read the stuff
963                  * out of the pipe right after the splice_to_pipe(). So set
964                  * PIPE_READERS appropriately.
965                  */
966                 pipe->readers = 1;
967
968                 current->splice_pipe = pipe;
969         }
970
971         /*
972          * Do the splice.
973          */
974         ret = 0;
975         bytes = 0;
976         len = sd->total_len;
977         flags = sd->flags;
978
979         /*
980          * Don't block on output, we have to drain the direct pipe.
981          */
982         sd->flags &= ~SPLICE_F_NONBLOCK;
983
984         while (len) {
985                 size_t read_len;
986                 loff_t pos = sd->pos, prev_pos = pos;
987
988                 ret = do_splice_to(in, &pos, pipe, len, flags);
989                 if (unlikely(ret <= 0))
990                         goto out_release;
991
992                 read_len = ret;
993                 sd->total_len = read_len;
994
995                 /*
996                  * NOTE: nonblocking mode only applies to the input. We
997                  * must not do the output in nonblocking mode as then we
998                  * could get stuck data in the internal pipe:
999                  */
1000                 ret = actor(pipe, sd);
1001                 if (unlikely(ret <= 0)) {
1002                         sd->pos = prev_pos;
1003                         goto out_release;
1004                 }
1005
1006                 bytes += ret;
1007                 len -= ret;
1008                 sd->pos = pos;
1009
1010                 if (ret < read_len) {
1011                         sd->pos = prev_pos + ret;
1012                         goto out_release;
1013                 }
1014         }
1015
1016 done:
1017         pipe->nrbufs = pipe->curbuf = 0;
1018         file_accessed(in);
1019         return bytes;
1020
1021 out_release:
1022         /*
1023          * If we did an incomplete transfer we must release
1024          * the pipe buffers in question:
1025          */
1026         for (i = 0; i < PIPE_BUFFERS; i++) {
1027                 struct pipe_buffer *buf = pipe->bufs + i;
1028
1029                 if (buf->ops) {
1030                         buf->ops->release(pipe, buf);
1031                         buf->ops = NULL;
1032                 }
1033         }
1034
1035         if (!bytes)
1036                 bytes = ret;
1037
1038         goto done;
1039 }
1040 EXPORT_SYMBOL(splice_direct_to_actor);
1041
1042 static int direct_splice_actor(struct pipe_inode_info *pipe,
1043                                struct splice_desc *sd)
1044 {
1045         struct file *file = sd->u.file;
1046
1047         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1048 }
1049
1050 /**
1051  * do_splice_direct - splices data directly between two files
1052  * @in:         file to splice from
1053  * @ppos:       input file offset
1054  * @out:        file to splice to
1055  * @len:        number of bytes to splice
1056  * @flags:      splice modifier flags
1057  *
1058  * Description:
1059  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1060  *    doing it in the application would incur an extra system call
1061  *    (splice in + splice out, as compared to just sendfile()). So this helper
1062  *    can splice directly through a process-private pipe.
1063  *
1064  */
1065 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1066                       size_t len, unsigned int flags)
1067 {
1068         struct splice_desc sd = {
1069                 .len            = len,
1070                 .total_len      = len,
1071                 .flags          = flags,
1072                 .pos            = *ppos,
1073                 .u.file         = out,
1074         };
1075         long ret;
1076
1077         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1078         if (ret > 0)
1079                 *ppos = sd.pos;
1080
1081         return ret;
1082 }
1083
1084 /*
1085  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1086  * location, so checking ->i_pipe is not enough to verify that this is a
1087  * pipe.
1088  */
1089 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1090 {
1091         if (S_ISFIFO(inode->i_mode))
1092                 return inode->i_pipe;
1093
1094         return NULL;
1095 }
1096
1097 /*
1098  * Determine where to splice to/from.
1099  */
1100 static long do_splice(struct file *in, loff_t __user *off_in,
1101                       struct file *out, loff_t __user *off_out,
1102                       size_t len, unsigned int flags)
1103 {
1104         struct pipe_inode_info *pipe;
1105         loff_t offset, *off;
1106         long ret;
1107
1108         pipe = pipe_info(in->f_path.dentry->d_inode);
1109         if (pipe) {
1110                 if (off_in)
1111                         return -ESPIPE;
1112                 if (off_out) {
1113                         if (out->f_op->llseek == no_llseek)
1114                                 return -EINVAL;
1115                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1116                                 return -EFAULT;
1117                         off = &offset;
1118                 } else
1119                         off = &out->f_pos;
1120
1121                 ret = do_splice_from(pipe, out, off, len, flags);
1122
1123                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1124                         ret = -EFAULT;
1125
1126                 return ret;
1127         }
1128
1129         pipe = pipe_info(out->f_path.dentry->d_inode);
1130         if (pipe) {
1131                 if (off_out)
1132                         return -ESPIPE;
1133                 if (off_in) {
1134                         if (in->f_op->llseek == no_llseek)
1135                                 return -EINVAL;
1136                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1137                                 return -EFAULT;
1138                         off = &offset;
1139                 } else
1140                         off = &in->f_pos;
1141
1142                 ret = do_splice_to(in, off, pipe, len, flags);
1143
1144                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1145                         ret = -EFAULT;
1146
1147                 return ret;
1148         }
1149
1150         return -EINVAL;
1151 }
1152
1153 /*
1154  * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1155  * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1156  * for writing) and page faulting on the user memory pointed to by src.
1157  * This assumes that we will very rarely hit the partial != 0 path, or this
1158  * will not be a win.
1159  */
1160 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1161 {
1162         int partial;
1163
1164         if (!access_ok(VERIFY_READ, src, n))
1165                 return -EFAULT;
1166
1167         pagefault_disable();
1168         partial = __copy_from_user_inatomic(dst, src, n);
1169         pagefault_enable();
1170
1171         /*
1172          * Didn't copy everything, drop the mmap_sem and do a faulting copy
1173          */
1174         if (unlikely(partial)) {
1175                 up_read(&current->mm->mmap_sem);
1176                 partial = copy_from_user(dst, src, n);
1177                 down_read(&current->mm->mmap_sem);
1178         }
1179
1180         return partial;
1181 }
1182
1183 /*
1184  * Map an iov into an array of pages and offset/length tupples. With the
1185  * partial_page structure, we can map several non-contiguous ranges into
1186  * our ones pages[] map instead of splitting that operation into pieces.
1187  * Could easily be exported as a generic helper for other users, in which
1188  * case one would probably want to add a 'max_nr_pages' parameter as well.
1189  */
1190 static int get_iovec_page_array(const struct iovec __user *iov,
1191                                 unsigned int nr_vecs, struct page **pages,
1192                                 struct partial_page *partial, int aligned)
1193 {
1194         int buffers = 0, error = 0;
1195
1196         down_read(&current->mm->mmap_sem);
1197
1198         while (nr_vecs) {
1199                 unsigned long off, npages;
1200                 struct iovec entry;
1201                 void __user *base;
1202                 size_t len;
1203                 int i;
1204
1205                 error = -EFAULT;
1206                 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1207                         break;
1208
1209                 base = entry.iov_base;
1210                 len = entry.iov_len;
1211
1212                 /*
1213                  * Sanity check this iovec. 0 read succeeds.
1214                  */
1215                 error = 0;
1216                 if (unlikely(!len))
1217                         break;
1218                 error = -EFAULT;
1219                 if (!access_ok(VERIFY_READ, base, len))
1220                         break;
1221
1222                 /*
1223                  * Get this base offset and number of pages, then map
1224                  * in the user pages.
1225                  */
1226                 off = (unsigned long) base & ~PAGE_MASK;
1227
1228                 /*
1229                  * If asked for alignment, the offset must be zero and the
1230                  * length a multiple of the PAGE_SIZE.
1231                  */
1232                 error = -EINVAL;
1233                 if (aligned && (off || len & ~PAGE_MASK))
1234                         break;
1235
1236                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1237                 if (npages > PIPE_BUFFERS - buffers)
1238                         npages = PIPE_BUFFERS - buffers;
1239
1240                 error = get_user_pages(current, current->mm,
1241                                        (unsigned long) base, npages, 0, 0,
1242                                        &pages[buffers], NULL);
1243
1244                 if (unlikely(error <= 0))
1245                         break;
1246
1247                 /*
1248                  * Fill this contiguous range into the partial page map.
1249                  */
1250                 for (i = 0; i < error; i++) {
1251                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1252
1253                         partial[buffers].offset = off;
1254                         partial[buffers].len = plen;
1255
1256                         off = 0;
1257                         len -= plen;
1258                         buffers++;
1259                 }
1260
1261                 /*
1262                  * We didn't complete this iov, stop here since it probably
1263                  * means we have to move some of this into a pipe to
1264                  * be able to continue.
1265                  */
1266                 if (len)
1267                         break;
1268
1269                 /*
1270                  * Don't continue if we mapped fewer pages than we asked for,
1271                  * or if we mapped the max number of pages that we have
1272                  * room for.
1273                  */
1274                 if (error < npages || buffers == PIPE_BUFFERS)
1275                         break;
1276
1277                 nr_vecs--;
1278                 iov++;
1279         }
1280
1281         up_read(&current->mm->mmap_sem);
1282
1283         if (buffers)
1284                 return buffers;
1285
1286         return error;
1287 }
1288
1289 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1290                         struct splice_desc *sd)
1291 {
1292         char *src;
1293         int ret;
1294
1295         ret = buf->ops->confirm(pipe, buf);
1296         if (unlikely(ret))
1297                 return ret;
1298
1299         /*
1300          * See if we can use the atomic maps, by prefaulting in the
1301          * pages and doing an atomic copy
1302          */
1303         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1304                 src = buf->ops->map(pipe, buf, 1);
1305                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1306                                                         sd->len);
1307                 buf->ops->unmap(pipe, buf, src);
1308                 if (!ret) {
1309                         ret = sd->len;
1310                         goto out;
1311                 }
1312         }
1313
1314         /*
1315          * No dice, use slow non-atomic map and copy
1316          */
1317         src = buf->ops->map(pipe, buf, 0);
1318
1319         ret = sd->len;
1320         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1321                 ret = -EFAULT;
1322
1323         buf->ops->unmap(pipe, buf, src);
1324 out:
1325         if (ret > 0)
1326                 sd->u.userptr += ret;
1327         return ret;
1328 }
1329
1330 /*
1331  * For lack of a better implementation, implement vmsplice() to userspace
1332  * as a simple copy of the pipes pages to the user iov.
1333  */
1334 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1335                              unsigned long nr_segs, unsigned int flags)
1336 {
1337         struct pipe_inode_info *pipe;
1338         struct splice_desc sd;
1339         ssize_t size;
1340         int error;
1341         long ret;
1342
1343         pipe = pipe_info(file->f_path.dentry->d_inode);
1344         if (!pipe)
1345                 return -EBADF;
1346
1347         if (pipe->inode)
1348                 mutex_lock(&pipe->inode->i_mutex);
1349
1350         error = ret = 0;
1351         while (nr_segs) {
1352                 void __user *base;
1353                 size_t len;
1354
1355                 /*
1356                  * Get user address base and length for this iovec.
1357                  */
1358                 error = get_user(base, &iov->iov_base);
1359                 if (unlikely(error))
1360                         break;
1361                 error = get_user(len, &iov->iov_len);
1362                 if (unlikely(error))
1363                         break;
1364
1365                 /*
1366                  * Sanity check this iovec. 0 read succeeds.
1367                  */
1368                 if (unlikely(!len))
1369                         break;
1370                 if (unlikely(!base)) {
1371                         error = -EFAULT;
1372                         break;
1373                 }
1374
1375                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1376                         error = -EFAULT;
1377                         break;
1378                 }
1379
1380                 sd.len = 0;
1381                 sd.total_len = len;
1382                 sd.flags = flags;
1383                 sd.u.userptr = base;
1384                 sd.pos = 0;
1385
1386                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1387                 if (size < 0) {
1388                         if (!ret)
1389                                 ret = size;
1390
1391                         break;
1392                 }
1393
1394                 ret += size;
1395
1396                 if (size < len)
1397                         break;
1398
1399                 nr_segs--;
1400                 iov++;
1401         }
1402
1403         if (pipe->inode)
1404                 mutex_unlock(&pipe->inode->i_mutex);
1405
1406         if (!ret)
1407                 ret = error;
1408
1409         return ret;
1410 }
1411
1412 /*
1413  * vmsplice splices a user address range into a pipe. It can be thought of
1414  * as splice-from-memory, where the regular splice is splice-from-file (or
1415  * to file). In both cases the output is a pipe, naturally.
1416  */
1417 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1418                              unsigned long nr_segs, unsigned int flags)
1419 {
1420         struct pipe_inode_info *pipe;
1421         struct page *pages[PIPE_BUFFERS];
1422         struct partial_page partial[PIPE_BUFFERS];
1423         struct splice_pipe_desc spd = {
1424                 .pages = pages,
1425                 .partial = partial,
1426                 .flags = flags,
1427                 .ops = &user_page_pipe_buf_ops,
1428                 .spd_release = spd_release_page,
1429         };
1430
1431         pipe = pipe_info(file->f_path.dentry->d_inode);
1432         if (!pipe)
1433                 return -EBADF;
1434
1435         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1436                                             flags & SPLICE_F_GIFT);
1437         if (spd.nr_pages <= 0)
1438                 return spd.nr_pages;
1439
1440         return splice_to_pipe(pipe, &spd);
1441 }
1442
1443 /*
1444  * Note that vmsplice only really supports true splicing _from_ user memory
1445  * to a pipe, not the other way around. Splicing from user memory is a simple
1446  * operation that can be supported without any funky alignment restrictions
1447  * or nasty vm tricks. We simply map in the user memory and fill them into
1448  * a pipe. The reverse isn't quite as easy, though. There are two possible
1449  * solutions for that:
1450  *
1451  *      - memcpy() the data internally, at which point we might as well just
1452  *        do a regular read() on the buffer anyway.
1453  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1454  *        has restriction limitations on both ends of the pipe).
1455  *
1456  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1457  *
1458  */
1459 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1460                              unsigned long nr_segs, unsigned int flags)
1461 {
1462         struct file *file;
1463         long error;
1464         int fput;
1465
1466         if (unlikely(nr_segs > UIO_MAXIOV))
1467                 return -EINVAL;
1468         else if (unlikely(!nr_segs))
1469                 return 0;
1470
1471         error = -EBADF;
1472         file = fget_light(fd, &fput);
1473         if (file) {
1474                 if (file->f_mode & FMODE_WRITE)
1475                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1476                 else if (file->f_mode & FMODE_READ)
1477                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1478
1479                 fput_light(file, fput);
1480         }
1481
1482         return error;
1483 }
1484
1485 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1486                            int fd_out, loff_t __user *off_out,
1487                            size_t len, unsigned int flags)
1488 {
1489         long error;
1490         struct file *in, *out;
1491         int fput_in, fput_out;
1492
1493         if (unlikely(!len))
1494                 return 0;
1495
1496         error = -EBADF;
1497         in = fget_light(fd_in, &fput_in);
1498         if (in) {
1499                 if (in->f_mode & FMODE_READ) {
1500                         out = fget_light(fd_out, &fput_out);
1501                         if (out) {
1502                                 if (out->f_mode & FMODE_WRITE)
1503                                         error = do_splice(in, off_in,
1504                                                           out, off_out,
1505                                                           len, flags);
1506                                 fput_light(out, fput_out);
1507                         }
1508                 }
1509
1510                 fput_light(in, fput_in);
1511         }
1512
1513         return error;
1514 }
1515
1516 /*
1517  * Make sure there's data to read. Wait for input if we can, otherwise
1518  * return an appropriate error.
1519  */
1520 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1521 {
1522         int ret;
1523
1524         /*
1525          * Check ->nrbufs without the inode lock first. This function
1526          * is speculative anyways, so missing one is ok.
1527          */
1528         if (pipe->nrbufs)
1529                 return 0;
1530
1531         ret = 0;
1532         mutex_lock(&pipe->inode->i_mutex);
1533
1534         while (!pipe->nrbufs) {
1535                 if (signal_pending(current)) {
1536                         ret = -ERESTARTSYS;
1537                         break;
1538                 }
1539                 if (!pipe->writers)
1540                         break;
1541                 if (!pipe->waiting_writers) {
1542                         if (flags & SPLICE_F_NONBLOCK) {
1543                                 ret = -EAGAIN;
1544                                 break;
1545                         }
1546                 }
1547                 pipe_wait(pipe);
1548         }
1549
1550         mutex_unlock(&pipe->inode->i_mutex);
1551         return ret;
1552 }
1553
1554 /*
1555  * Make sure there's writeable room. Wait for room if we can, otherwise
1556  * return an appropriate error.
1557  */
1558 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1559 {
1560         int ret;
1561
1562         /*
1563          * Check ->nrbufs without the inode lock first. This function
1564          * is speculative anyways, so missing one is ok.
1565          */
1566         if (pipe->nrbufs < PIPE_BUFFERS)
1567                 return 0;
1568
1569         ret = 0;
1570         mutex_lock(&pipe->inode->i_mutex);
1571
1572         while (pipe->nrbufs >= PIPE_BUFFERS) {
1573                 if (!pipe->readers) {
1574                         send_sig(SIGPIPE, current, 0);
1575                         ret = -EPIPE;
1576                         break;
1577                 }
1578                 if (flags & SPLICE_F_NONBLOCK) {
1579                         ret = -EAGAIN;
1580                         break;
1581                 }
1582                 if (signal_pending(current)) {
1583                         ret = -ERESTARTSYS;
1584                         break;
1585                 }
1586                 pipe->waiting_writers++;
1587                 pipe_wait(pipe);
1588                 pipe->waiting_writers--;
1589         }
1590
1591         mutex_unlock(&pipe->inode->i_mutex);
1592         return ret;
1593 }
1594
1595 /*
1596  * Link contents of ipipe to opipe.
1597  */
1598 static int link_pipe(struct pipe_inode_info *ipipe,
1599                      struct pipe_inode_info *opipe,
1600                      size_t len, unsigned int flags)
1601 {
1602         struct pipe_buffer *ibuf, *obuf;
1603         int ret = 0, i = 0, nbuf;
1604
1605         /*
1606          * Potential ABBA deadlock, work around it by ordering lock
1607          * grabbing by inode address. Otherwise two different processes
1608          * could deadlock (one doing tee from A -> B, the other from B -> A).
1609          */
1610         inode_double_lock(ipipe->inode, opipe->inode);
1611
1612         do {
1613                 if (!opipe->readers) {
1614                         send_sig(SIGPIPE, current, 0);
1615                         if (!ret)
1616                                 ret = -EPIPE;
1617                         break;
1618                 }
1619
1620                 /*
1621                  * If we have iterated all input buffers or ran out of
1622                  * output room, break.
1623                  */
1624                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1625                         break;
1626
1627                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1628                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1629
1630                 /*
1631                  * Get a reference to this pipe buffer,
1632                  * so we can copy the contents over.
1633                  */
1634                 ibuf->ops->get(ipipe, ibuf);
1635
1636                 obuf = opipe->bufs + nbuf;
1637                 *obuf = *ibuf;
1638
1639                 /*
1640                  * Don't inherit the gift flag, we need to
1641                  * prevent multiple steals of this page.
1642                  */
1643                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1644
1645                 if (obuf->len > len)
1646                         obuf->len = len;
1647
1648                 opipe->nrbufs++;
1649                 ret += obuf->len;
1650                 len -= obuf->len;
1651                 i++;
1652         } while (len);
1653
1654         /*
1655          * return EAGAIN if we have the potential of some data in the
1656          * future, otherwise just return 0
1657          */
1658         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1659                 ret = -EAGAIN;
1660
1661         inode_double_unlock(ipipe->inode, opipe->inode);
1662
1663         /*
1664          * If we put data in the output pipe, wakeup any potential readers.
1665          */
1666         if (ret > 0) {
1667                 smp_mb();
1668                 if (waitqueue_active(&opipe->wait))
1669                         wake_up_interruptible(&opipe->wait);
1670                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1671         }
1672
1673         return ret;
1674 }
1675
1676 /*
1677  * This is a tee(1) implementation that works on pipes. It doesn't copy
1678  * any data, it simply references the 'in' pages on the 'out' pipe.
1679  * The 'flags' used are the SPLICE_F_* variants, currently the only
1680  * applicable one is SPLICE_F_NONBLOCK.
1681  */
1682 static long do_tee(struct file *in, struct file *out, size_t len,
1683                    unsigned int flags)
1684 {
1685         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1686         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1687         int ret = -EINVAL;
1688
1689         /*
1690          * Duplicate the contents of ipipe to opipe without actually
1691          * copying the data.
1692          */
1693         if (ipipe && opipe && ipipe != opipe) {
1694                 /*
1695                  * Keep going, unless we encounter an error. The ipipe/opipe
1696                  * ordering doesn't really matter.
1697                  */
1698                 ret = link_ipipe_prep(ipipe, flags);
1699                 if (!ret) {
1700                         ret = link_opipe_prep(opipe, flags);
1701                         if (!ret)
1702                                 ret = link_pipe(ipipe, opipe, len, flags);
1703                 }
1704         }
1705
1706         return ret;
1707 }
1708
1709 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1710 {
1711         struct file *in;
1712         int error, fput_in;
1713
1714         if (unlikely(!len))
1715                 return 0;
1716
1717         error = -EBADF;
1718         in = fget_light(fdin, &fput_in);
1719         if (in) {
1720                 if (in->f_mode & FMODE_READ) {
1721                         int fput_out;
1722                         struct file *out = fget_light(fdout, &fput_out);
1723
1724                         if (out) {
1725                                 if (out->f_mode & FMODE_WRITE)
1726                                         error = do_tee(in, out, len, flags);
1727                                 fput_light(out, fput_out);
1728                         }
1729                 }
1730                 fput_light(in, fput_in);
1731         }
1732
1733         return error;
1734 }