splice: add helpers for locking pipe inode
[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/memcontrol.h>
25 #include <linux/mm_inline.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
33
34 /*
35  * Attempt to steal a page from a pipe buffer. This should perhaps go into
36  * a vm helper function, it's already simplified quite a bit by the
37  * addition of remove_mapping(). If success is returned, the caller may
38  * attempt to reuse this page for another destination.
39  */
40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
41                                      struct pipe_buffer *buf)
42 {
43         struct page *page = buf->page;
44         struct address_space *mapping;
45
46         lock_page(page);
47
48         mapping = page_mapping(page);
49         if (mapping) {
50                 WARN_ON(!PageUptodate(page));
51
52                 /*
53                  * At least for ext2 with nobh option, we need to wait on
54                  * writeback completing on this page, since we'll remove it
55                  * from the pagecache.  Otherwise truncate wont wait on the
56                  * page, allowing the disk blocks to be reused by someone else
57                  * before we actually wrote our data to them. fs corruption
58                  * ensues.
59                  */
60                 wait_on_page_writeback(page);
61
62                 if (page_has_private(page) &&
63                     !try_to_release_page(page, GFP_KERNEL))
64                         goto out_unlock;
65
66                 /*
67                  * If we succeeded in removing the mapping, set LRU flag
68                  * and return good.
69                  */
70                 if (remove_mapping(mapping, page)) {
71                         buf->flags |= PIPE_BUF_FLAG_LRU;
72                         return 0;
73                 }
74         }
75
76         /*
77          * Raced with truncate or failed to remove page from current
78          * address space, unlock and return failure.
79          */
80 out_unlock:
81         unlock_page(page);
82         return 1;
83 }
84
85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
86                                         struct pipe_buffer *buf)
87 {
88         page_cache_release(buf->page);
89         buf->flags &= ~PIPE_BUF_FLAG_LRU;
90 }
91
92 /*
93  * Check whether the contents of buf is OK to access. Since the content
94  * is a page cache page, IO may be in flight.
95  */
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
97                                        struct pipe_buffer *buf)
98 {
99         struct page *page = buf->page;
100         int err;
101
102         if (!PageUptodate(page)) {
103                 lock_page(page);
104
105                 /*
106                  * Page got truncated/unhashed. This will cause a 0-byte
107                  * splice, if this is the first page.
108                  */
109                 if (!page->mapping) {
110                         err = -ENODATA;
111                         goto error;
112                 }
113
114                 /*
115                  * Uh oh, read-error from disk.
116                  */
117                 if (!PageUptodate(page)) {
118                         err = -EIO;
119                         goto error;
120                 }
121
122                 /*
123                  * Page is ok afterall, we are done.
124                  */
125                 unlock_page(page);
126         }
127
128         return 0;
129 error:
130         unlock_page(page);
131         return err;
132 }
133
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
135         .can_merge = 0,
136         .map = generic_pipe_buf_map,
137         .unmap = generic_pipe_buf_unmap,
138         .confirm = page_cache_pipe_buf_confirm,
139         .release = page_cache_pipe_buf_release,
140         .steal = page_cache_pipe_buf_steal,
141         .get = generic_pipe_buf_get,
142 };
143
144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
145                                     struct pipe_buffer *buf)
146 {
147         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
148                 return 1;
149
150         buf->flags |= PIPE_BUF_FLAG_LRU;
151         return generic_pipe_buf_steal(pipe, buf);
152 }
153
154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
155         .can_merge = 0,
156         .map = generic_pipe_buf_map,
157         .unmap = generic_pipe_buf_unmap,
158         .confirm = generic_pipe_buf_confirm,
159         .release = page_cache_pipe_buf_release,
160         .steal = user_page_pipe_buf_steal,
161         .get = generic_pipe_buf_get,
162 };
163
164 /**
165  * splice_to_pipe - fill passed data into a pipe
166  * @pipe:       pipe to fill
167  * @spd:        data to fill
168  *
169  * Description:
170  *    @spd contains a map of pages and len/offset tuples, along with
171  *    the struct pipe_buf_operations associated with these pages. This
172  *    function will link that data to the pipe.
173  *
174  */
175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
176                        struct splice_pipe_desc *spd)
177 {
178         unsigned int spd_pages = spd->nr_pages;
179         int ret, do_wakeup, page_nr;
180
181         ret = 0;
182         do_wakeup = 0;
183         page_nr = 0;
184
185         pipe_lock(pipe);
186
187         for (;;) {
188                 if (!pipe->readers) {
189                         send_sig(SIGPIPE, current, 0);
190                         if (!ret)
191                                 ret = -EPIPE;
192                         break;
193                 }
194
195                 if (pipe->nrbufs < PIPE_BUFFERS) {
196                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
197                         struct pipe_buffer *buf = pipe->bufs + newbuf;
198
199                         buf->page = spd->pages[page_nr];
200                         buf->offset = spd->partial[page_nr].offset;
201                         buf->len = spd->partial[page_nr].len;
202                         buf->private = spd->partial[page_nr].private;
203                         buf->ops = spd->ops;
204                         if (spd->flags & SPLICE_F_GIFT)
205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
206
207                         pipe->nrbufs++;
208                         page_nr++;
209                         ret += buf->len;
210
211                         if (pipe->inode)
212                                 do_wakeup = 1;
213
214                         if (!--spd->nr_pages)
215                                 break;
216                         if (pipe->nrbufs < PIPE_BUFFERS)
217                                 continue;
218
219                         break;
220                 }
221
222                 if (spd->flags & SPLICE_F_NONBLOCK) {
223                         if (!ret)
224                                 ret = -EAGAIN;
225                         break;
226                 }
227
228                 if (signal_pending(current)) {
229                         if (!ret)
230                                 ret = -ERESTARTSYS;
231                         break;
232                 }
233
234                 if (do_wakeup) {
235                         smp_mb();
236                         if (waitqueue_active(&pipe->wait))
237                                 wake_up_interruptible_sync(&pipe->wait);
238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
239                         do_wakeup = 0;
240                 }
241
242                 pipe->waiting_writers++;
243                 pipe_wait(pipe);
244                 pipe->waiting_writers--;
245         }
246
247         pipe_unlock(pipe);
248
249         if (do_wakeup) {
250                 smp_mb();
251                 if (waitqueue_active(&pipe->wait))
252                         wake_up_interruptible(&pipe->wait);
253                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
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 (!trylock_page(page)) {
374                                         error = -EAGAIN;
375                                         break;
376                                 }
377                         } else
378                                 lock_page(page);
379
380                         /*
381                          * Page was truncated, or invalidated by the
382                          * filesystem.  Redo the find/create, but this time the
383                          * page is kept locked, so there's no chance of another
384                          * race with truncate/invalidate.
385                          */
386                         if (!page->mapping) {
387                                 unlock_page(page);
388                                 page = find_or_create_page(mapping, index,
389                                                 mapping_gfp_mask(mapping));
390
391                                 if (!page) {
392                                         error = -ENOMEM;
393                                         break;
394                                 }
395                                 page_cache_release(pages[page_nr]);
396                                 pages[page_nr] = page;
397                         }
398                         /*
399                          * page was already under io and is now done, great
400                          */
401                         if (PageUptodate(page)) {
402                                 unlock_page(page);
403                                 goto fill_it;
404                         }
405
406                         /*
407                          * need to read in the page
408                          */
409                         error = mapping->a_ops->readpage(in, page);
410                         if (unlikely(error)) {
411                                 /*
412                                  * We really should re-lookup the page here,
413                                  * but it complicates things a lot. Instead
414                                  * lets just do what we already stored, and
415                                  * we'll get it the next time we are called.
416                                  */
417                                 if (error == AOP_TRUNCATED_PAGE)
418                                         error = 0;
419
420                                 break;
421                         }
422                 }
423 fill_it:
424                 /*
425                  * i_size must be checked after PageUptodate.
426                  */
427                 isize = i_size_read(mapping->host);
428                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
429                 if (unlikely(!isize || index > end_index))
430                         break;
431
432                 /*
433                  * if this is the last page, see if we need to shrink
434                  * the length and stop
435                  */
436                 if (end_index == index) {
437                         unsigned int plen;
438
439                         /*
440                          * max good bytes in this page
441                          */
442                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
443                         if (plen <= loff)
444                                 break;
445
446                         /*
447                          * force quit after adding this page
448                          */
449                         this_len = min(this_len, plen - loff);
450                         len = this_len;
451                 }
452
453                 partial[page_nr].offset = loff;
454                 partial[page_nr].len = this_len;
455                 len -= this_len;
456                 loff = 0;
457                 spd.nr_pages++;
458                 index++;
459         }
460
461         /*
462          * Release any pages at the end, if we quit early. 'page_nr' is how far
463          * we got, 'nr_pages' is how many pages are in the map.
464          */
465         while (page_nr < nr_pages)
466                 page_cache_release(pages[page_nr++]);
467         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
468
469         if (spd.nr_pages)
470                 return splice_to_pipe(pipe, &spd);
471
472         return error;
473 }
474
475 /**
476  * generic_file_splice_read - splice data from file to a pipe
477  * @in:         file to splice from
478  * @ppos:       position in @in
479  * @pipe:       pipe to splice to
480  * @len:        number of bytes to splice
481  * @flags:      splice modifier flags
482  *
483  * Description:
484  *    Will read pages from given file and fill them into a pipe. Can be
485  *    used as long as the address_space operations for the source implements
486  *    a readpage() hook.
487  *
488  */
489 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
490                                  struct pipe_inode_info *pipe, size_t len,
491                                  unsigned int flags)
492 {
493         loff_t isize, left;
494         int ret;
495
496         isize = i_size_read(in->f_mapping->host);
497         if (unlikely(*ppos >= isize))
498                 return 0;
499
500         left = isize - *ppos;
501         if (unlikely(left < len))
502                 len = left;
503
504         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
505         if (ret > 0)
506                 *ppos += ret;
507
508         return ret;
509 }
510
511 EXPORT_SYMBOL(generic_file_splice_read);
512
513 /*
514  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
515  * using sendpage(). Return the number of bytes sent.
516  */
517 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
518                             struct pipe_buffer *buf, struct splice_desc *sd)
519 {
520         struct file *file = sd->u.file;
521         loff_t pos = sd->pos;
522         int ret, more;
523
524         ret = buf->ops->confirm(pipe, buf);
525         if (!ret) {
526                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
527
528                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
529                                            sd->len, &pos, more);
530         }
531
532         return ret;
533 }
534
535 /*
536  * This is a little more tricky than the file -> pipe splicing. There are
537  * basically three cases:
538  *
539  *      - Destination page already exists in the address space and there
540  *        are users of it. For that case we have no other option that
541  *        copying the data. Tough luck.
542  *      - Destination page already exists in the address space, but there
543  *        are no users of it. Make sure it's uptodate, then drop it. Fall
544  *        through to last case.
545  *      - Destination page does not exist, we can add the pipe page to
546  *        the page cache and avoid the copy.
547  *
548  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
549  * sd->flags), we attempt to migrate pages from the pipe to the output
550  * file address space page cache. This is possible if no one else has
551  * the pipe page referenced outside of the pipe and page cache. If
552  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
553  * a new page in the output file page cache and fill/dirty that.
554  */
555 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
556                  struct splice_desc *sd)
557 {
558         struct file *file = sd->u.file;
559         struct address_space *mapping = file->f_mapping;
560         unsigned int offset, this_len;
561         struct page *page;
562         void *fsdata;
563         int ret;
564
565         /*
566          * make sure the data in this buffer is uptodate
567          */
568         ret = buf->ops->confirm(pipe, buf);
569         if (unlikely(ret))
570                 return ret;
571
572         offset = sd->pos & ~PAGE_CACHE_MASK;
573
574         this_len = sd->len;
575         if (this_len + offset > PAGE_CACHE_SIZE)
576                 this_len = PAGE_CACHE_SIZE - offset;
577
578         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
579                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
580         if (unlikely(ret))
581                 goto out;
582
583         if (buf->page != page) {
584                 /*
585                  * Careful, ->map() uses KM_USER0!
586                  */
587                 char *src = buf->ops->map(pipe, buf, 1);
588                 char *dst = kmap_atomic(page, KM_USER1);
589
590                 memcpy(dst + offset, src + buf->offset, this_len);
591                 flush_dcache_page(page);
592                 kunmap_atomic(dst, KM_USER1);
593                 buf->ops->unmap(pipe, buf, src);
594         }
595         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
596                                 page, fsdata);
597 out:
598         return ret;
599 }
600 EXPORT_SYMBOL(pipe_to_file);
601
602 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
603 {
604         smp_mb();
605         if (waitqueue_active(&pipe->wait))
606                 wake_up_interruptible(&pipe->wait);
607         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
608 }
609
610 /**
611  * splice_from_pipe_feed - feed available data from a pipe to a file
612  * @pipe:       pipe to splice from
613  * @sd:         information to @actor
614  * @actor:      handler that splices the data
615  *
616  * Description:
617
618  *    This function loops over the pipe and calls @actor to do the
619  *    actual moving of a single struct pipe_buffer to the desired
620  *    destination.  It returns when there's no more buffers left in
621  *    the pipe or if the requested number of bytes (@sd->total_len)
622  *    have been copied.  It returns a positive number (one) if the
623  *    pipe needs to be filled with more data, zero if the required
624  *    number of bytes have been copied and -errno on error.
625  *
626  *    This, together with splice_from_pipe_{begin,end,next}, may be
627  *    used to implement the functionality of __splice_from_pipe() when
628  *    locking is required around copying the pipe buffers to the
629  *    destination.
630  */
631 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
632                           splice_actor *actor)
633 {
634         int ret;
635
636         while (pipe->nrbufs) {
637                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
638                 const struct pipe_buf_operations *ops = buf->ops;
639
640                 sd->len = buf->len;
641                 if (sd->len > sd->total_len)
642                         sd->len = sd->total_len;
643
644                 ret = actor(pipe, buf, sd);
645                 if (ret <= 0) {
646                         if (ret == -ENODATA)
647                                 ret = 0;
648                         return ret;
649                 }
650                 buf->offset += ret;
651                 buf->len -= ret;
652
653                 sd->num_spliced += ret;
654                 sd->len -= ret;
655                 sd->pos += ret;
656                 sd->total_len -= ret;
657
658                 if (!buf->len) {
659                         buf->ops = NULL;
660                         ops->release(pipe, buf);
661                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
662                         pipe->nrbufs--;
663                         if (pipe->inode)
664                                 sd->need_wakeup = true;
665                 }
666
667                 if (!sd->total_len)
668                         return 0;
669         }
670
671         return 1;
672 }
673 EXPORT_SYMBOL(splice_from_pipe_feed);
674
675 /**
676  * splice_from_pipe_next - wait for some data to splice from
677  * @pipe:       pipe to splice from
678  * @sd:         information about the splice operation
679  *
680  * Description:
681  *    This function will wait for some data and return a positive
682  *    value (one) if pipe buffers are available.  It will return zero
683  *    or -errno if no more data needs to be spliced.
684  */
685 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
686 {
687         while (!pipe->nrbufs) {
688                 if (!pipe->writers)
689                         return 0;
690
691                 if (!pipe->waiting_writers && sd->num_spliced)
692                         return 0;
693
694                 if (sd->flags & SPLICE_F_NONBLOCK)
695                         return -EAGAIN;
696
697                 if (signal_pending(current))
698                         return -ERESTARTSYS;
699
700                 if (sd->need_wakeup) {
701                         wakeup_pipe_writers(pipe);
702                         sd->need_wakeup = false;
703                 }
704
705                 pipe_wait(pipe);
706         }
707
708         return 1;
709 }
710 EXPORT_SYMBOL(splice_from_pipe_next);
711
712 /**
713  * splice_from_pipe_begin - start splicing from pipe
714  * @pipe:       pipe to splice from
715  *
716  * Description:
717  *    This function should be called before a loop containing
718  *    splice_from_pipe_next() and splice_from_pipe_feed() to
719  *    initialize the necessary fields of @sd.
720  */
721 void splice_from_pipe_begin(struct splice_desc *sd)
722 {
723         sd->num_spliced = 0;
724         sd->need_wakeup = false;
725 }
726 EXPORT_SYMBOL(splice_from_pipe_begin);
727
728 /**
729  * splice_from_pipe_end - finish splicing from pipe
730  * @pipe:       pipe to splice from
731  * @sd:         information about the splice operation
732  *
733  * Description:
734  *    This function will wake up pipe writers if necessary.  It should
735  *    be called after a loop containing splice_from_pipe_next() and
736  *    splice_from_pipe_feed().
737  */
738 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
739 {
740         if (sd->need_wakeup)
741                 wakeup_pipe_writers(pipe);
742 }
743 EXPORT_SYMBOL(splice_from_pipe_end);
744
745 /**
746  * __splice_from_pipe - splice data from a pipe to given actor
747  * @pipe:       pipe to splice from
748  * @sd:         information to @actor
749  * @actor:      handler that splices the data
750  *
751  * Description:
752  *    This function does little more than loop over the pipe and call
753  *    @actor to do the actual moving of a single struct pipe_buffer to
754  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
755  *    pipe_to_user.
756  *
757  */
758 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
759                            splice_actor *actor)
760 {
761         int ret;
762
763         splice_from_pipe_begin(sd);
764         do {
765                 ret = splice_from_pipe_next(pipe, sd);
766                 if (ret > 0)
767                         ret = splice_from_pipe_feed(pipe, sd, actor);
768         } while (ret > 0);
769         splice_from_pipe_end(pipe, sd);
770
771         return sd->num_spliced ? sd->num_spliced : ret;
772 }
773 EXPORT_SYMBOL(__splice_from_pipe);
774
775 /**
776  * splice_from_pipe - splice data from a pipe to a file
777  * @pipe:       pipe to splice from
778  * @out:        file to splice to
779  * @ppos:       position in @out
780  * @len:        how many bytes to splice
781  * @flags:      splice modifier flags
782  * @actor:      handler that splices the data
783  *
784  * Description:
785  *    See __splice_from_pipe. This function locks the pipe inode,
786  *    otherwise it's identical to __splice_from_pipe().
787  *
788  */
789 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
790                          loff_t *ppos, size_t len, unsigned int flags,
791                          splice_actor *actor)
792 {
793         ssize_t ret;
794         struct splice_desc sd = {
795                 .total_len = len,
796                 .flags = flags,
797                 .pos = *ppos,
798                 .u.file = out,
799         };
800
801         pipe_lock(pipe);
802         ret = __splice_from_pipe(pipe, &sd, actor);
803         pipe_unlock(pipe);
804
805         return ret;
806 }
807
808 /**
809  * generic_file_splice_write - splice data from a pipe to a file
810  * @pipe:       pipe info
811  * @out:        file to write to
812  * @ppos:       position in @out
813  * @len:        number of bytes to splice
814  * @flags:      splice modifier flags
815  *
816  * Description:
817  *    Will either move or copy pages (determined by @flags options) from
818  *    the given pipe inode to the given file.
819  *
820  */
821 ssize_t
822 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
823                           loff_t *ppos, size_t len, unsigned int flags)
824 {
825         struct address_space *mapping = out->f_mapping;
826         struct inode *inode = mapping->host;
827         struct splice_desc sd = {
828                 .total_len = len,
829                 .flags = flags,
830                 .pos = *ppos,
831                 .u.file = out,
832         };
833         ssize_t ret;
834
835         pipe_lock(pipe);
836
837         splice_from_pipe_begin(&sd);
838         do {
839                 ret = splice_from_pipe_next(pipe, &sd);
840                 if (ret <= 0)
841                         break;
842
843                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
844                 ret = file_remove_suid(out);
845                 if (!ret)
846                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
847                 mutex_unlock(&inode->i_mutex);
848         } while (ret > 0);
849         splice_from_pipe_end(pipe, &sd);
850
851         pipe_unlock(pipe);
852
853         if (sd.num_spliced)
854                 ret = sd.num_spliced;
855
856         if (ret > 0) {
857                 unsigned long nr_pages;
858
859                 *ppos += ret;
860                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
861
862                 /*
863                  * If file or inode is SYNC and we actually wrote some data,
864                  * sync it.
865                  */
866                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
867                         int err;
868
869                         mutex_lock(&inode->i_mutex);
870                         err = generic_osync_inode(inode, mapping,
871                                                   OSYNC_METADATA|OSYNC_DATA);
872                         mutex_unlock(&inode->i_mutex);
873
874                         if (err)
875                                 ret = err;
876                 }
877                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
878         }
879
880         return ret;
881 }
882
883 EXPORT_SYMBOL(generic_file_splice_write);
884
885 /**
886  * generic_splice_sendpage - splice data from a pipe to a socket
887  * @pipe:       pipe to splice from
888  * @out:        socket to write to
889  * @ppos:       position in @out
890  * @len:        number of bytes to splice
891  * @flags:      splice modifier flags
892  *
893  * Description:
894  *    Will send @len bytes from the pipe to a network socket. No data copying
895  *    is involved.
896  *
897  */
898 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
899                                 loff_t *ppos, size_t len, unsigned int flags)
900 {
901         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
902 }
903
904 EXPORT_SYMBOL(generic_splice_sendpage);
905
906 /*
907  * Attempt to initiate a splice from pipe to file.
908  */
909 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
910                            loff_t *ppos, size_t len, unsigned int flags)
911 {
912         int ret;
913
914         if (unlikely(!out->f_op || !out->f_op->splice_write))
915                 return -EINVAL;
916
917         if (unlikely(!(out->f_mode & FMODE_WRITE)))
918                 return -EBADF;
919
920         if (unlikely(out->f_flags & O_APPEND))
921                 return -EINVAL;
922
923         ret = rw_verify_area(WRITE, out, ppos, len);
924         if (unlikely(ret < 0))
925                 return ret;
926
927         return out->f_op->splice_write(pipe, out, ppos, len, flags);
928 }
929
930 /*
931  * Attempt to initiate a splice from a file to a pipe.
932  */
933 static long do_splice_to(struct file *in, loff_t *ppos,
934                          struct pipe_inode_info *pipe, size_t len,
935                          unsigned int flags)
936 {
937         int ret;
938
939         if (unlikely(!in->f_op || !in->f_op->splice_read))
940                 return -EINVAL;
941
942         if (unlikely(!(in->f_mode & FMODE_READ)))
943                 return -EBADF;
944
945         ret = rw_verify_area(READ, in, ppos, len);
946         if (unlikely(ret < 0))
947                 return ret;
948
949         return in->f_op->splice_read(in, ppos, pipe, len, flags);
950 }
951
952 /**
953  * splice_direct_to_actor - splices data directly between two non-pipes
954  * @in:         file to splice from
955  * @sd:         actor information on where to splice to
956  * @actor:      handles the data splicing
957  *
958  * Description:
959  *    This is a special case helper to splice directly between two
960  *    points, without requiring an explicit pipe. Internally an allocated
961  *    pipe is cached in the process, and reused during the lifetime of
962  *    that process.
963  *
964  */
965 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
966                                splice_direct_actor *actor)
967 {
968         struct pipe_inode_info *pipe;
969         long ret, bytes;
970         umode_t i_mode;
971         size_t len;
972         int i, flags;
973
974         /*
975          * We require the input being a regular file, as we don't want to
976          * randomly drop data for eg socket -> socket splicing. Use the
977          * piped splicing for that!
978          */
979         i_mode = in->f_path.dentry->d_inode->i_mode;
980         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
981                 return -EINVAL;
982
983         /*
984          * neither in nor out is a pipe, setup an internal pipe attached to
985          * 'out' and transfer the wanted data from 'in' to 'out' through that
986          */
987         pipe = current->splice_pipe;
988         if (unlikely(!pipe)) {
989                 pipe = alloc_pipe_info(NULL);
990                 if (!pipe)
991                         return -ENOMEM;
992
993                 /*
994                  * We don't have an immediate reader, but we'll read the stuff
995                  * out of the pipe right after the splice_to_pipe(). So set
996                  * PIPE_READERS appropriately.
997                  */
998                 pipe->readers = 1;
999
1000                 current->splice_pipe = pipe;
1001         }
1002
1003         /*
1004          * Do the splice.
1005          */
1006         ret = 0;
1007         bytes = 0;
1008         len = sd->total_len;
1009         flags = sd->flags;
1010
1011         /*
1012          * Don't block on output, we have to drain the direct pipe.
1013          */
1014         sd->flags &= ~SPLICE_F_NONBLOCK;
1015
1016         while (len) {
1017                 size_t read_len;
1018                 loff_t pos = sd->pos, prev_pos = pos;
1019
1020                 ret = do_splice_to(in, &pos, pipe, len, flags);
1021                 if (unlikely(ret <= 0))
1022                         goto out_release;
1023
1024                 read_len = ret;
1025                 sd->total_len = read_len;
1026
1027                 /*
1028                  * NOTE: nonblocking mode only applies to the input. We
1029                  * must not do the output in nonblocking mode as then we
1030                  * could get stuck data in the internal pipe:
1031                  */
1032                 ret = actor(pipe, sd);
1033                 if (unlikely(ret <= 0)) {
1034                         sd->pos = prev_pos;
1035                         goto out_release;
1036                 }
1037
1038                 bytes += ret;
1039                 len -= ret;
1040                 sd->pos = pos;
1041
1042                 if (ret < read_len) {
1043                         sd->pos = prev_pos + ret;
1044                         goto out_release;
1045                 }
1046         }
1047
1048 done:
1049         pipe->nrbufs = pipe->curbuf = 0;
1050         file_accessed(in);
1051         return bytes;
1052
1053 out_release:
1054         /*
1055          * If we did an incomplete transfer we must release
1056          * the pipe buffers in question:
1057          */
1058         for (i = 0; i < PIPE_BUFFERS; i++) {
1059                 struct pipe_buffer *buf = pipe->bufs + i;
1060
1061                 if (buf->ops) {
1062                         buf->ops->release(pipe, buf);
1063                         buf->ops = NULL;
1064                 }
1065         }
1066
1067         if (!bytes)
1068                 bytes = ret;
1069
1070         goto done;
1071 }
1072 EXPORT_SYMBOL(splice_direct_to_actor);
1073
1074 static int direct_splice_actor(struct pipe_inode_info *pipe,
1075                                struct splice_desc *sd)
1076 {
1077         struct file *file = sd->u.file;
1078
1079         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1080 }
1081
1082 /**
1083  * do_splice_direct - splices data directly between two files
1084  * @in:         file to splice from
1085  * @ppos:       input file offset
1086  * @out:        file to splice to
1087  * @len:        number of bytes to splice
1088  * @flags:      splice modifier flags
1089  *
1090  * Description:
1091  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1092  *    doing it in the application would incur an extra system call
1093  *    (splice in + splice out, as compared to just sendfile()). So this helper
1094  *    can splice directly through a process-private pipe.
1095  *
1096  */
1097 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1098                       size_t len, unsigned int flags)
1099 {
1100         struct splice_desc sd = {
1101                 .len            = len,
1102                 .total_len      = len,
1103                 .flags          = flags,
1104                 .pos            = *ppos,
1105                 .u.file         = out,
1106         };
1107         long ret;
1108
1109         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1110         if (ret > 0)
1111                 *ppos = sd.pos;
1112
1113         return ret;
1114 }
1115
1116 /*
1117  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1118  * location, so checking ->i_pipe is not enough to verify that this is a
1119  * pipe.
1120  */
1121 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1122 {
1123         if (S_ISFIFO(inode->i_mode))
1124                 return inode->i_pipe;
1125
1126         return NULL;
1127 }
1128
1129 /*
1130  * Determine where to splice to/from.
1131  */
1132 static long do_splice(struct file *in, loff_t __user *off_in,
1133                       struct file *out, loff_t __user *off_out,
1134                       size_t len, unsigned int flags)
1135 {
1136         struct pipe_inode_info *pipe;
1137         loff_t offset, *off;
1138         long ret;
1139
1140         pipe = pipe_info(in->f_path.dentry->d_inode);
1141         if (pipe) {
1142                 if (off_in)
1143                         return -ESPIPE;
1144                 if (off_out) {
1145                         if (out->f_op->llseek == no_llseek)
1146                                 return -EINVAL;
1147                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1148                                 return -EFAULT;
1149                         off = &offset;
1150                 } else
1151                         off = &out->f_pos;
1152
1153                 ret = do_splice_from(pipe, out, off, len, flags);
1154
1155                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1156                         ret = -EFAULT;
1157
1158                 return ret;
1159         }
1160
1161         pipe = pipe_info(out->f_path.dentry->d_inode);
1162         if (pipe) {
1163                 if (off_out)
1164                         return -ESPIPE;
1165                 if (off_in) {
1166                         if (in->f_op->llseek == no_llseek)
1167                                 return -EINVAL;
1168                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1169                                 return -EFAULT;
1170                         off = &offset;
1171                 } else
1172                         off = &in->f_pos;
1173
1174                 ret = do_splice_to(in, off, pipe, len, flags);
1175
1176                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1177                         ret = -EFAULT;
1178
1179                 return ret;
1180         }
1181
1182         return -EINVAL;
1183 }
1184
1185 /*
1186  * Map an iov into an array of pages and offset/length tupples. With the
1187  * partial_page structure, we can map several non-contiguous ranges into
1188  * our ones pages[] map instead of splitting that operation into pieces.
1189  * Could easily be exported as a generic helper for other users, in which
1190  * case one would probably want to add a 'max_nr_pages' parameter as well.
1191  */
1192 static int get_iovec_page_array(const struct iovec __user *iov,
1193                                 unsigned int nr_vecs, struct page **pages,
1194                                 struct partial_page *partial, int aligned)
1195 {
1196         int buffers = 0, error = 0;
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(&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_fast((unsigned long)base, npages,
1241                                         0, &pages[buffers]);
1242
1243                 if (unlikely(error <= 0))
1244                         break;
1245
1246                 /*
1247                  * Fill this contiguous range into the partial page map.
1248                  */
1249                 for (i = 0; i < error; i++) {
1250                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1251
1252                         partial[buffers].offset = off;
1253                         partial[buffers].len = plen;
1254
1255                         off = 0;
1256                         len -= plen;
1257                         buffers++;
1258                 }
1259
1260                 /*
1261                  * We didn't complete this iov, stop here since it probably
1262                  * means we have to move some of this into a pipe to
1263                  * be able to continue.
1264                  */
1265                 if (len)
1266                         break;
1267
1268                 /*
1269                  * Don't continue if we mapped fewer pages than we asked for,
1270                  * or if we mapped the max number of pages that we have
1271                  * room for.
1272                  */
1273                 if (error < npages || buffers == PIPE_BUFFERS)
1274                         break;
1275
1276                 nr_vecs--;
1277                 iov++;
1278         }
1279
1280         if (buffers)
1281                 return buffers;
1282
1283         return error;
1284 }
1285
1286 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1287                         struct splice_desc *sd)
1288 {
1289         char *src;
1290         int ret;
1291
1292         ret = buf->ops->confirm(pipe, buf);
1293         if (unlikely(ret))
1294                 return ret;
1295
1296         /*
1297          * See if we can use the atomic maps, by prefaulting in the
1298          * pages and doing an atomic copy
1299          */
1300         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1301                 src = buf->ops->map(pipe, buf, 1);
1302                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1303                                                         sd->len);
1304                 buf->ops->unmap(pipe, buf, src);
1305                 if (!ret) {
1306                         ret = sd->len;
1307                         goto out;
1308                 }
1309         }
1310
1311         /*
1312          * No dice, use slow non-atomic map and copy
1313          */
1314         src = buf->ops->map(pipe, buf, 0);
1315
1316         ret = sd->len;
1317         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1318                 ret = -EFAULT;
1319
1320         buf->ops->unmap(pipe, buf, src);
1321 out:
1322         if (ret > 0)
1323                 sd->u.userptr += ret;
1324         return ret;
1325 }
1326
1327 /*
1328  * For lack of a better implementation, implement vmsplice() to userspace
1329  * as a simple copy of the pipes pages to the user iov.
1330  */
1331 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1332                              unsigned long nr_segs, unsigned int flags)
1333 {
1334         struct pipe_inode_info *pipe;
1335         struct splice_desc sd;
1336         ssize_t size;
1337         int error;
1338         long ret;
1339
1340         pipe = pipe_info(file->f_path.dentry->d_inode);
1341         if (!pipe)
1342                 return -EBADF;
1343
1344         pipe_lock(pipe);
1345
1346         error = ret = 0;
1347         while (nr_segs) {
1348                 void __user *base;
1349                 size_t len;
1350
1351                 /*
1352                  * Get user address base and length for this iovec.
1353                  */
1354                 error = get_user(base, &iov->iov_base);
1355                 if (unlikely(error))
1356                         break;
1357                 error = get_user(len, &iov->iov_len);
1358                 if (unlikely(error))
1359                         break;
1360
1361                 /*
1362                  * Sanity check this iovec. 0 read succeeds.
1363                  */
1364                 if (unlikely(!len))
1365                         break;
1366                 if (unlikely(!base)) {
1367                         error = -EFAULT;
1368                         break;
1369                 }
1370
1371                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1372                         error = -EFAULT;
1373                         break;
1374                 }
1375
1376                 sd.len = 0;
1377                 sd.total_len = len;
1378                 sd.flags = flags;
1379                 sd.u.userptr = base;
1380                 sd.pos = 0;
1381
1382                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1383                 if (size < 0) {
1384                         if (!ret)
1385                                 ret = size;
1386
1387                         break;
1388                 }
1389
1390                 ret += size;
1391
1392                 if (size < len)
1393                         break;
1394
1395                 nr_segs--;
1396                 iov++;
1397         }
1398
1399         pipe_unlock(pipe);
1400
1401         if (!ret)
1402                 ret = error;
1403
1404         return ret;
1405 }
1406
1407 /*
1408  * vmsplice splices a user address range into a pipe. It can be thought of
1409  * as splice-from-memory, where the regular splice is splice-from-file (or
1410  * to file). In both cases the output is a pipe, naturally.
1411  */
1412 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1413                              unsigned long nr_segs, unsigned int flags)
1414 {
1415         struct pipe_inode_info *pipe;
1416         struct page *pages[PIPE_BUFFERS];
1417         struct partial_page partial[PIPE_BUFFERS];
1418         struct splice_pipe_desc spd = {
1419                 .pages = pages,
1420                 .partial = partial,
1421                 .flags = flags,
1422                 .ops = &user_page_pipe_buf_ops,
1423                 .spd_release = spd_release_page,
1424         };
1425
1426         pipe = pipe_info(file->f_path.dentry->d_inode);
1427         if (!pipe)
1428                 return -EBADF;
1429
1430         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1431                                             flags & SPLICE_F_GIFT);
1432         if (spd.nr_pages <= 0)
1433                 return spd.nr_pages;
1434
1435         return splice_to_pipe(pipe, &spd);
1436 }
1437
1438 /*
1439  * Note that vmsplice only really supports true splicing _from_ user memory
1440  * to a pipe, not the other way around. Splicing from user memory is a simple
1441  * operation that can be supported without any funky alignment restrictions
1442  * or nasty vm tricks. We simply map in the user memory and fill them into
1443  * a pipe. The reverse isn't quite as easy, though. There are two possible
1444  * solutions for that:
1445  *
1446  *      - memcpy() the data internally, at which point we might as well just
1447  *        do a regular read() on the buffer anyway.
1448  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1449  *        has restriction limitations on both ends of the pipe).
1450  *
1451  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1452  *
1453  */
1454 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1455                 unsigned long, nr_segs, unsigned int, flags)
1456 {
1457         struct file *file;
1458         long error;
1459         int fput;
1460
1461         if (unlikely(nr_segs > UIO_MAXIOV))
1462                 return -EINVAL;
1463         else if (unlikely(!nr_segs))
1464                 return 0;
1465
1466         error = -EBADF;
1467         file = fget_light(fd, &fput);
1468         if (file) {
1469                 if (file->f_mode & FMODE_WRITE)
1470                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1471                 else if (file->f_mode & FMODE_READ)
1472                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1473
1474                 fput_light(file, fput);
1475         }
1476
1477         return error;
1478 }
1479
1480 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1481                 int, fd_out, loff_t __user *, off_out,
1482                 size_t, len, unsigned int, flags)
1483 {
1484         long error;
1485         struct file *in, *out;
1486         int fput_in, fput_out;
1487
1488         if (unlikely(!len))
1489                 return 0;
1490
1491         error = -EBADF;
1492         in = fget_light(fd_in, &fput_in);
1493         if (in) {
1494                 if (in->f_mode & FMODE_READ) {
1495                         out = fget_light(fd_out, &fput_out);
1496                         if (out) {
1497                                 if (out->f_mode & FMODE_WRITE)
1498                                         error = do_splice(in, off_in,
1499                                                           out, off_out,
1500                                                           len, flags);
1501                                 fput_light(out, fput_out);
1502                         }
1503                 }
1504
1505                 fput_light(in, fput_in);
1506         }
1507
1508         return error;
1509 }
1510
1511 /*
1512  * Make sure there's data to read. Wait for input if we can, otherwise
1513  * return an appropriate error.
1514  */
1515 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1516 {
1517         int ret;
1518
1519         /*
1520          * Check ->nrbufs without the inode lock first. This function
1521          * is speculative anyways, so missing one is ok.
1522          */
1523         if (pipe->nrbufs)
1524                 return 0;
1525
1526         ret = 0;
1527         pipe_lock(pipe);
1528
1529         while (!pipe->nrbufs) {
1530                 if (signal_pending(current)) {
1531                         ret = -ERESTARTSYS;
1532                         break;
1533                 }
1534                 if (!pipe->writers)
1535                         break;
1536                 if (!pipe->waiting_writers) {
1537                         if (flags & SPLICE_F_NONBLOCK) {
1538                                 ret = -EAGAIN;
1539                                 break;
1540                         }
1541                 }
1542                 pipe_wait(pipe);
1543         }
1544
1545         pipe_unlock(pipe);
1546         return ret;
1547 }
1548
1549 /*
1550  * Make sure there's writeable room. Wait for room if we can, otherwise
1551  * return an appropriate error.
1552  */
1553 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1554 {
1555         int ret;
1556
1557         /*
1558          * Check ->nrbufs without the inode lock first. This function
1559          * is speculative anyways, so missing one is ok.
1560          */
1561         if (pipe->nrbufs < PIPE_BUFFERS)
1562                 return 0;
1563
1564         ret = 0;
1565         pipe_lock(pipe);
1566
1567         while (pipe->nrbufs >= PIPE_BUFFERS) {
1568                 if (!pipe->readers) {
1569                         send_sig(SIGPIPE, current, 0);
1570                         ret = -EPIPE;
1571                         break;
1572                 }
1573                 if (flags & SPLICE_F_NONBLOCK) {
1574                         ret = -EAGAIN;
1575                         break;
1576                 }
1577                 if (signal_pending(current)) {
1578                         ret = -ERESTARTSYS;
1579                         break;
1580                 }
1581                 pipe->waiting_writers++;
1582                 pipe_wait(pipe);
1583                 pipe->waiting_writers--;
1584         }
1585
1586         pipe_unlock(pipe);
1587         return ret;
1588 }
1589
1590 /*
1591  * Link contents of ipipe to opipe.
1592  */
1593 static int link_pipe(struct pipe_inode_info *ipipe,
1594                      struct pipe_inode_info *opipe,
1595                      size_t len, unsigned int flags)
1596 {
1597         struct pipe_buffer *ibuf, *obuf;
1598         int ret = 0, i = 0, nbuf;
1599
1600         /*
1601          * Potential ABBA deadlock, work around it by ordering lock
1602          * grabbing by pipe info address. Otherwise two different processes
1603          * could deadlock (one doing tee from A -> B, the other from B -> A).
1604          */
1605         pipe_double_lock(ipipe, opipe);
1606
1607         do {
1608                 if (!opipe->readers) {
1609                         send_sig(SIGPIPE, current, 0);
1610                         if (!ret)
1611                                 ret = -EPIPE;
1612                         break;
1613                 }
1614
1615                 /*
1616                  * If we have iterated all input buffers or ran out of
1617                  * output room, break.
1618                  */
1619                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1620                         break;
1621
1622                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1623                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1624
1625                 /*
1626                  * Get a reference to this pipe buffer,
1627                  * so we can copy the contents over.
1628                  */
1629                 ibuf->ops->get(ipipe, ibuf);
1630
1631                 obuf = opipe->bufs + nbuf;
1632                 *obuf = *ibuf;
1633
1634                 /*
1635                  * Don't inherit the gift flag, we need to
1636                  * prevent multiple steals of this page.
1637                  */
1638                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1639
1640                 if (obuf->len > len)
1641                         obuf->len = len;
1642
1643                 opipe->nrbufs++;
1644                 ret += obuf->len;
1645                 len -= obuf->len;
1646                 i++;
1647         } while (len);
1648
1649         /*
1650          * return EAGAIN if we have the potential of some data in the
1651          * future, otherwise just return 0
1652          */
1653         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1654                 ret = -EAGAIN;
1655
1656         pipe_unlock(ipipe);
1657         pipe_unlock(opipe);
1658
1659         /*
1660          * If we put data in the output pipe, wakeup any potential readers.
1661          */
1662         if (ret > 0) {
1663                 smp_mb();
1664                 if (waitqueue_active(&opipe->wait))
1665                         wake_up_interruptible(&opipe->wait);
1666                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1667         }
1668
1669         return ret;
1670 }
1671
1672 /*
1673  * This is a tee(1) implementation that works on pipes. It doesn't copy
1674  * any data, it simply references the 'in' pages on the 'out' pipe.
1675  * The 'flags' used are the SPLICE_F_* variants, currently the only
1676  * applicable one is SPLICE_F_NONBLOCK.
1677  */
1678 static long do_tee(struct file *in, struct file *out, size_t len,
1679                    unsigned int flags)
1680 {
1681         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1682         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1683         int ret = -EINVAL;
1684
1685         /*
1686          * Duplicate the contents of ipipe to opipe without actually
1687          * copying the data.
1688          */
1689         if (ipipe && opipe && ipipe != opipe) {
1690                 /*
1691                  * Keep going, unless we encounter an error. The ipipe/opipe
1692                  * ordering doesn't really matter.
1693                  */
1694                 ret = link_ipipe_prep(ipipe, flags);
1695                 if (!ret) {
1696                         ret = link_opipe_prep(opipe, flags);
1697                         if (!ret)
1698                                 ret = link_pipe(ipipe, opipe, len, flags);
1699                 }
1700         }
1701
1702         return ret;
1703 }
1704
1705 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1706 {
1707         struct file *in;
1708         int error, fput_in;
1709
1710         if (unlikely(!len))
1711                 return 0;
1712
1713         error = -EBADF;
1714         in = fget_light(fdin, &fput_in);
1715         if (in) {
1716                 if (in->f_mode & FMODE_READ) {
1717                         int fput_out;
1718                         struct file *out = fget_light(fdout, &fput_out);
1719
1720                         if (out) {
1721                                 if (out->f_mode & FMODE_WRITE)
1722                                         error = do_tee(in, out, len, flags);
1723                                 fput_light(out, fput_out);
1724                         }
1725                 }
1726                 fput_light(in, fput_in);
1727         }
1728
1729         return error;
1730 }