splice: implement pipe to pipe splicing
[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  *    This function loops over the pipe and calls @actor to do the
618  *    actual moving of a single struct pipe_buffer to the desired
619  *    destination.  It returns when there's no more buffers left in
620  *    the pipe or if the requested number of bytes (@sd->total_len)
621  *    have been copied.  It returns a positive number (one) if the
622  *    pipe needs to be filled with more data, zero if the required
623  *    number of bytes have been copied and -errno on error.
624  *
625  *    This, together with splice_from_pipe_{begin,end,next}, may be
626  *    used to implement the functionality of __splice_from_pipe() when
627  *    locking is required around copying the pipe buffers to the
628  *    destination.
629  */
630 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
631                           splice_actor *actor)
632 {
633         int ret;
634
635         while (pipe->nrbufs) {
636                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
637                 const struct pipe_buf_operations *ops = buf->ops;
638
639                 sd->len = buf->len;
640                 if (sd->len > sd->total_len)
641                         sd->len = sd->total_len;
642
643                 ret = actor(pipe, buf, sd);
644                 if (ret <= 0) {
645                         if (ret == -ENODATA)
646                                 ret = 0;
647                         return ret;
648                 }
649                 buf->offset += ret;
650                 buf->len -= ret;
651
652                 sd->num_spliced += ret;
653                 sd->len -= ret;
654                 sd->pos += ret;
655                 sd->total_len -= ret;
656
657                 if (!buf->len) {
658                         buf->ops = NULL;
659                         ops->release(pipe, buf);
660                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
661                         pipe->nrbufs--;
662                         if (pipe->inode)
663                                 sd->need_wakeup = true;
664                 }
665
666                 if (!sd->total_len)
667                         return 0;
668         }
669
670         return 1;
671 }
672 EXPORT_SYMBOL(splice_from_pipe_feed);
673
674 /**
675  * splice_from_pipe_next - wait for some data to splice from
676  * @pipe:       pipe to splice from
677  * @sd:         information about the splice operation
678  *
679  * Description:
680  *    This function will wait for some data and return a positive
681  *    value (one) if pipe buffers are available.  It will return zero
682  *    or -errno if no more data needs to be spliced.
683  */
684 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
685 {
686         while (!pipe->nrbufs) {
687                 if (!pipe->writers)
688                         return 0;
689
690                 if (!pipe->waiting_writers && sd->num_spliced)
691                         return 0;
692
693                 if (sd->flags & SPLICE_F_NONBLOCK)
694                         return -EAGAIN;
695
696                 if (signal_pending(current))
697                         return -ERESTARTSYS;
698
699                 if (sd->need_wakeup) {
700                         wakeup_pipe_writers(pipe);
701                         sd->need_wakeup = false;
702                 }
703
704                 pipe_wait(pipe);
705         }
706
707         return 1;
708 }
709 EXPORT_SYMBOL(splice_from_pipe_next);
710
711 /**
712  * splice_from_pipe_begin - start splicing from pipe
713  * @sd:         information about the splice operation
714  *
715  * Description:
716  *    This function should be called before a loop containing
717  *    splice_from_pipe_next() and splice_from_pipe_feed() to
718  *    initialize the necessary fields of @sd.
719  */
720 void splice_from_pipe_begin(struct splice_desc *sd)
721 {
722         sd->num_spliced = 0;
723         sd->need_wakeup = false;
724 }
725 EXPORT_SYMBOL(splice_from_pipe_begin);
726
727 /**
728  * splice_from_pipe_end - finish splicing from pipe
729  * @pipe:       pipe to splice from
730  * @sd:         information about the splice operation
731  *
732  * Description:
733  *    This function will wake up pipe writers if necessary.  It should
734  *    be called after a loop containing splice_from_pipe_next() and
735  *    splice_from_pipe_feed().
736  */
737 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
738 {
739         if (sd->need_wakeup)
740                 wakeup_pipe_writers(pipe);
741 }
742 EXPORT_SYMBOL(splice_from_pipe_end);
743
744 /**
745  * __splice_from_pipe - splice data from a pipe to given actor
746  * @pipe:       pipe to splice from
747  * @sd:         information to @actor
748  * @actor:      handler that splices the data
749  *
750  * Description:
751  *    This function does little more than loop over the pipe and call
752  *    @actor to do the actual moving of a single struct pipe_buffer to
753  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
754  *    pipe_to_user.
755  *
756  */
757 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
758                            splice_actor *actor)
759 {
760         int ret;
761
762         splice_from_pipe_begin(sd);
763         do {
764                 ret = splice_from_pipe_next(pipe, sd);
765                 if (ret > 0)
766                         ret = splice_from_pipe_feed(pipe, sd, actor);
767         } while (ret > 0);
768         splice_from_pipe_end(pipe, sd);
769
770         return sd->num_spliced ? sd->num_spliced : ret;
771 }
772 EXPORT_SYMBOL(__splice_from_pipe);
773
774 /**
775  * splice_from_pipe - splice data from a pipe to a file
776  * @pipe:       pipe to splice from
777  * @out:        file to splice to
778  * @ppos:       position in @out
779  * @len:        how many bytes to splice
780  * @flags:      splice modifier flags
781  * @actor:      handler that splices the data
782  *
783  * Description:
784  *    See __splice_from_pipe. This function locks the pipe inode,
785  *    otherwise it's identical to __splice_from_pipe().
786  *
787  */
788 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
789                          loff_t *ppos, size_t len, unsigned int flags,
790                          splice_actor *actor)
791 {
792         ssize_t ret;
793         struct splice_desc sd = {
794                 .total_len = len,
795                 .flags = flags,
796                 .pos = *ppos,
797                 .u.file = out,
798         };
799
800         pipe_lock(pipe);
801         ret = __splice_from_pipe(pipe, &sd, actor);
802         pipe_unlock(pipe);
803
804         return ret;
805 }
806
807 /**
808  * generic_file_splice_write - splice data from a pipe to a file
809  * @pipe:       pipe info
810  * @out:        file to write to
811  * @ppos:       position in @out
812  * @len:        number of bytes to splice
813  * @flags:      splice modifier flags
814  *
815  * Description:
816  *    Will either move or copy pages (determined by @flags options) from
817  *    the given pipe inode to the given file.
818  *
819  */
820 ssize_t
821 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
822                           loff_t *ppos, size_t len, unsigned int flags)
823 {
824         struct address_space *mapping = out->f_mapping;
825         struct inode *inode = mapping->host;
826         struct splice_desc sd = {
827                 .total_len = len,
828                 .flags = flags,
829                 .pos = *ppos,
830                 .u.file = out,
831         };
832         ssize_t ret;
833
834         pipe_lock(pipe);
835
836         splice_from_pipe_begin(&sd);
837         do {
838                 ret = splice_from_pipe_next(pipe, &sd);
839                 if (ret <= 0)
840                         break;
841
842                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
843                 ret = file_remove_suid(out);
844                 if (!ret)
845                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
846                 mutex_unlock(&inode->i_mutex);
847         } while (ret > 0);
848         splice_from_pipe_end(pipe, &sd);
849
850         pipe_unlock(pipe);
851
852         if (sd.num_spliced)
853                 ret = sd.num_spliced;
854
855         if (ret > 0) {
856                 unsigned long nr_pages;
857
858                 *ppos += ret;
859                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
860
861                 /*
862                  * If file or inode is SYNC and we actually wrote some data,
863                  * sync it.
864                  */
865                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
866                         int err;
867
868                         mutex_lock(&inode->i_mutex);
869                         err = generic_osync_inode(inode, mapping,
870                                                   OSYNC_METADATA|OSYNC_DATA);
871                         mutex_unlock(&inode->i_mutex);
872
873                         if (err)
874                                 ret = err;
875                 }
876                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
877         }
878
879         return ret;
880 }
881
882 EXPORT_SYMBOL(generic_file_splice_write);
883
884 /**
885  * generic_splice_sendpage - splice data from a pipe to a socket
886  * @pipe:       pipe to splice from
887  * @out:        socket to write to
888  * @ppos:       position in @out
889  * @len:        number of bytes to splice
890  * @flags:      splice modifier flags
891  *
892  * Description:
893  *    Will send @len bytes from the pipe to a network socket. No data copying
894  *    is involved.
895  *
896  */
897 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
898                                 loff_t *ppos, size_t len, unsigned int flags)
899 {
900         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
901 }
902
903 EXPORT_SYMBOL(generic_splice_sendpage);
904
905 /*
906  * Attempt to initiate a splice from pipe to file.
907  */
908 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
909                            loff_t *ppos, size_t len, unsigned int flags)
910 {
911         int ret;
912
913         if (unlikely(!out->f_op || !out->f_op->splice_write))
914                 return -EINVAL;
915
916         if (unlikely(!(out->f_mode & FMODE_WRITE)))
917                 return -EBADF;
918
919         if (unlikely(out->f_flags & O_APPEND))
920                 return -EINVAL;
921
922         ret = rw_verify_area(WRITE, out, ppos, len);
923         if (unlikely(ret < 0))
924                 return ret;
925
926         return out->f_op->splice_write(pipe, out, ppos, len, flags);
927 }
928
929 /*
930  * Attempt to initiate a splice from a file to a pipe.
931  */
932 static long do_splice_to(struct file *in, loff_t *ppos,
933                          struct pipe_inode_info *pipe, size_t len,
934                          unsigned int flags)
935 {
936         int ret;
937
938         if (unlikely(!in->f_op || !in->f_op->splice_read))
939                 return -EINVAL;
940
941         if (unlikely(!(in->f_mode & FMODE_READ)))
942                 return -EBADF;
943
944         ret = rw_verify_area(READ, in, ppos, len);
945         if (unlikely(ret < 0))
946                 return ret;
947
948         return in->f_op->splice_read(in, ppos, pipe, len, flags);
949 }
950
951 /**
952  * splice_direct_to_actor - splices data directly between two non-pipes
953  * @in:         file to splice from
954  * @sd:         actor information on where to splice to
955  * @actor:      handles the data splicing
956  *
957  * Description:
958  *    This is a special case helper to splice directly between two
959  *    points, without requiring an explicit pipe. Internally an allocated
960  *    pipe is cached in the process, and reused during the lifetime of
961  *    that process.
962  *
963  */
964 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
965                                splice_direct_actor *actor)
966 {
967         struct pipe_inode_info *pipe;
968         long ret, bytes;
969         umode_t i_mode;
970         size_t len;
971         int i, flags;
972
973         /*
974          * We require the input being a regular file, as we don't want to
975          * randomly drop data for eg socket -> socket splicing. Use the
976          * piped splicing for that!
977          */
978         i_mode = in->f_path.dentry->d_inode->i_mode;
979         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
980                 return -EINVAL;
981
982         /*
983          * neither in nor out is a pipe, setup an internal pipe attached to
984          * 'out' and transfer the wanted data from 'in' to 'out' through that
985          */
986         pipe = current->splice_pipe;
987         if (unlikely(!pipe)) {
988                 pipe = alloc_pipe_info(NULL);
989                 if (!pipe)
990                         return -ENOMEM;
991
992                 /*
993                  * We don't have an immediate reader, but we'll read the stuff
994                  * out of the pipe right after the splice_to_pipe(). So set
995                  * PIPE_READERS appropriately.
996                  */
997                 pipe->readers = 1;
998
999                 current->splice_pipe = pipe;
1000         }
1001
1002         /*
1003          * Do the splice.
1004          */
1005         ret = 0;
1006         bytes = 0;
1007         len = sd->total_len;
1008         flags = sd->flags;
1009
1010         /*
1011          * Don't block on output, we have to drain the direct pipe.
1012          */
1013         sd->flags &= ~SPLICE_F_NONBLOCK;
1014
1015         while (len) {
1016                 size_t read_len;
1017                 loff_t pos = sd->pos, prev_pos = pos;
1018
1019                 ret = do_splice_to(in, &pos, pipe, len, flags);
1020                 if (unlikely(ret <= 0))
1021                         goto out_release;
1022
1023                 read_len = ret;
1024                 sd->total_len = read_len;
1025
1026                 /*
1027                  * NOTE: nonblocking mode only applies to the input. We
1028                  * must not do the output in nonblocking mode as then we
1029                  * could get stuck data in the internal pipe:
1030                  */
1031                 ret = actor(pipe, sd);
1032                 if (unlikely(ret <= 0)) {
1033                         sd->pos = prev_pos;
1034                         goto out_release;
1035                 }
1036
1037                 bytes += ret;
1038                 len -= ret;
1039                 sd->pos = pos;
1040
1041                 if (ret < read_len) {
1042                         sd->pos = prev_pos + ret;
1043                         goto out_release;
1044                 }
1045         }
1046
1047 done:
1048         pipe->nrbufs = pipe->curbuf = 0;
1049         file_accessed(in);
1050         return bytes;
1051
1052 out_release:
1053         /*
1054          * If we did an incomplete transfer we must release
1055          * the pipe buffers in question:
1056          */
1057         for (i = 0; i < PIPE_BUFFERS; i++) {
1058                 struct pipe_buffer *buf = pipe->bufs + i;
1059
1060                 if (buf->ops) {
1061                         buf->ops->release(pipe, buf);
1062                         buf->ops = NULL;
1063                 }
1064         }
1065
1066         if (!bytes)
1067                 bytes = ret;
1068
1069         goto done;
1070 }
1071 EXPORT_SYMBOL(splice_direct_to_actor);
1072
1073 static int direct_splice_actor(struct pipe_inode_info *pipe,
1074                                struct splice_desc *sd)
1075 {
1076         struct file *file = sd->u.file;
1077
1078         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1079 }
1080
1081 /**
1082  * do_splice_direct - splices data directly between two files
1083  * @in:         file to splice from
1084  * @ppos:       input file offset
1085  * @out:        file to splice to
1086  * @len:        number of bytes to splice
1087  * @flags:      splice modifier flags
1088  *
1089  * Description:
1090  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1091  *    doing it in the application would incur an extra system call
1092  *    (splice in + splice out, as compared to just sendfile()). So this helper
1093  *    can splice directly through a process-private pipe.
1094  *
1095  */
1096 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1097                       size_t len, unsigned int flags)
1098 {
1099         struct splice_desc sd = {
1100                 .len            = len,
1101                 .total_len      = len,
1102                 .flags          = flags,
1103                 .pos            = *ppos,
1104                 .u.file         = out,
1105         };
1106         long ret;
1107
1108         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1109         if (ret > 0)
1110                 *ppos = sd.pos;
1111
1112         return ret;
1113 }
1114
1115 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1116                                struct pipe_inode_info *opipe,
1117                                size_t len, unsigned int flags);
1118 /*
1119  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1120  * location, so checking ->i_pipe is not enough to verify that this is a
1121  * pipe.
1122  */
1123 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1124 {
1125         if (S_ISFIFO(inode->i_mode))
1126                 return inode->i_pipe;
1127
1128         return NULL;
1129 }
1130
1131 /*
1132  * Determine where to splice to/from.
1133  */
1134 static long do_splice(struct file *in, loff_t __user *off_in,
1135                       struct file *out, loff_t __user *off_out,
1136                       size_t len, unsigned int flags)
1137 {
1138         struct pipe_inode_info *ipipe;
1139         struct pipe_inode_info *opipe;
1140         loff_t offset, *off;
1141         long ret;
1142
1143         ipipe = pipe_info(in->f_path.dentry->d_inode);
1144         opipe = pipe_info(out->f_path.dentry->d_inode);
1145
1146         if (ipipe && opipe) {
1147                 if (off_in || off_out)
1148                         return -ESPIPE;
1149
1150                 if (!(in->f_mode & FMODE_READ))
1151                         return -EBADF;
1152
1153                 if (!(out->f_mode & FMODE_WRITE))
1154                         return -EBADF;
1155
1156                 /* Splicing to self would be fun, but... */
1157                 if (ipipe == opipe)
1158                         return -EINVAL;
1159
1160                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1161         }
1162
1163         if (ipipe) {
1164                 if (off_in)
1165                         return -ESPIPE;
1166                 if (off_out) {
1167                         if (out->f_op->llseek == no_llseek)
1168                                 return -EINVAL;
1169                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1170                                 return -EFAULT;
1171                         off = &offset;
1172                 } else
1173                         off = &out->f_pos;
1174
1175                 ret = do_splice_from(ipipe, out, off, len, flags);
1176
1177                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1178                         ret = -EFAULT;
1179
1180                 return ret;
1181         }
1182
1183         if (opipe) {
1184                 if (off_out)
1185                         return -ESPIPE;
1186                 if (off_in) {
1187                         if (in->f_op->llseek == no_llseek)
1188                                 return -EINVAL;
1189                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1190                                 return -EFAULT;
1191                         off = &offset;
1192                 } else
1193                         off = &in->f_pos;
1194
1195                 ret = do_splice_to(in, off, opipe, len, flags);
1196
1197                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1198                         ret = -EFAULT;
1199
1200                 return ret;
1201         }
1202
1203         return -EINVAL;
1204 }
1205
1206 /*
1207  * Map an iov into an array of pages and offset/length tupples. With the
1208  * partial_page structure, we can map several non-contiguous ranges into
1209  * our ones pages[] map instead of splitting that operation into pieces.
1210  * Could easily be exported as a generic helper for other users, in which
1211  * case one would probably want to add a 'max_nr_pages' parameter as well.
1212  */
1213 static int get_iovec_page_array(const struct iovec __user *iov,
1214                                 unsigned int nr_vecs, struct page **pages,
1215                                 struct partial_page *partial, int aligned)
1216 {
1217         int buffers = 0, error = 0;
1218
1219         while (nr_vecs) {
1220                 unsigned long off, npages;
1221                 struct iovec entry;
1222                 void __user *base;
1223                 size_t len;
1224                 int i;
1225
1226                 error = -EFAULT;
1227                 if (copy_from_user(&entry, iov, sizeof(entry)))
1228                         break;
1229
1230                 base = entry.iov_base;
1231                 len = entry.iov_len;
1232
1233                 /*
1234                  * Sanity check this iovec. 0 read succeeds.
1235                  */
1236                 error = 0;
1237                 if (unlikely(!len))
1238                         break;
1239                 error = -EFAULT;
1240                 if (!access_ok(VERIFY_READ, base, len))
1241                         break;
1242
1243                 /*
1244                  * Get this base offset and number of pages, then map
1245                  * in the user pages.
1246                  */
1247                 off = (unsigned long) base & ~PAGE_MASK;
1248
1249                 /*
1250                  * If asked for alignment, the offset must be zero and the
1251                  * length a multiple of the PAGE_SIZE.
1252                  */
1253                 error = -EINVAL;
1254                 if (aligned && (off || len & ~PAGE_MASK))
1255                         break;
1256
1257                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1258                 if (npages > PIPE_BUFFERS - buffers)
1259                         npages = PIPE_BUFFERS - buffers;
1260
1261                 error = get_user_pages_fast((unsigned long)base, npages,
1262                                         0, &pages[buffers]);
1263
1264                 if (unlikely(error <= 0))
1265                         break;
1266
1267                 /*
1268                  * Fill this contiguous range into the partial page map.
1269                  */
1270                 for (i = 0; i < error; i++) {
1271                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1272
1273                         partial[buffers].offset = off;
1274                         partial[buffers].len = plen;
1275
1276                         off = 0;
1277                         len -= plen;
1278                         buffers++;
1279                 }
1280
1281                 /*
1282                  * We didn't complete this iov, stop here since it probably
1283                  * means we have to move some of this into a pipe to
1284                  * be able to continue.
1285                  */
1286                 if (len)
1287                         break;
1288
1289                 /*
1290                  * Don't continue if we mapped fewer pages than we asked for,
1291                  * or if we mapped the max number of pages that we have
1292                  * room for.
1293                  */
1294                 if (error < npages || buffers == PIPE_BUFFERS)
1295                         break;
1296
1297                 nr_vecs--;
1298                 iov++;
1299         }
1300
1301         if (buffers)
1302                 return buffers;
1303
1304         return error;
1305 }
1306
1307 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1308                         struct splice_desc *sd)
1309 {
1310         char *src;
1311         int ret;
1312
1313         ret = buf->ops->confirm(pipe, buf);
1314         if (unlikely(ret))
1315                 return ret;
1316
1317         /*
1318          * See if we can use the atomic maps, by prefaulting in the
1319          * pages and doing an atomic copy
1320          */
1321         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1322                 src = buf->ops->map(pipe, buf, 1);
1323                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1324                                                         sd->len);
1325                 buf->ops->unmap(pipe, buf, src);
1326                 if (!ret) {
1327                         ret = sd->len;
1328                         goto out;
1329                 }
1330         }
1331
1332         /*
1333          * No dice, use slow non-atomic map and copy
1334          */
1335         src = buf->ops->map(pipe, buf, 0);
1336
1337         ret = sd->len;
1338         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1339                 ret = -EFAULT;
1340
1341         buf->ops->unmap(pipe, buf, src);
1342 out:
1343         if (ret > 0)
1344                 sd->u.userptr += ret;
1345         return ret;
1346 }
1347
1348 /*
1349  * For lack of a better implementation, implement vmsplice() to userspace
1350  * as a simple copy of the pipes pages to the user iov.
1351  */
1352 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1353                              unsigned long nr_segs, unsigned int flags)
1354 {
1355         struct pipe_inode_info *pipe;
1356         struct splice_desc sd;
1357         ssize_t size;
1358         int error;
1359         long ret;
1360
1361         pipe = pipe_info(file->f_path.dentry->d_inode);
1362         if (!pipe)
1363                 return -EBADF;
1364
1365         pipe_lock(pipe);
1366
1367         error = ret = 0;
1368         while (nr_segs) {
1369                 void __user *base;
1370                 size_t len;
1371
1372                 /*
1373                  * Get user address base and length for this iovec.
1374                  */
1375                 error = get_user(base, &iov->iov_base);
1376                 if (unlikely(error))
1377                         break;
1378                 error = get_user(len, &iov->iov_len);
1379                 if (unlikely(error))
1380                         break;
1381
1382                 /*
1383                  * Sanity check this iovec. 0 read succeeds.
1384                  */
1385                 if (unlikely(!len))
1386                         break;
1387                 if (unlikely(!base)) {
1388                         error = -EFAULT;
1389                         break;
1390                 }
1391
1392                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1393                         error = -EFAULT;
1394                         break;
1395                 }
1396
1397                 sd.len = 0;
1398                 sd.total_len = len;
1399                 sd.flags = flags;
1400                 sd.u.userptr = base;
1401                 sd.pos = 0;
1402
1403                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1404                 if (size < 0) {
1405                         if (!ret)
1406                                 ret = size;
1407
1408                         break;
1409                 }
1410
1411                 ret += size;
1412
1413                 if (size < len)
1414                         break;
1415
1416                 nr_segs--;
1417                 iov++;
1418         }
1419
1420         pipe_unlock(pipe);
1421
1422         if (!ret)
1423                 ret = error;
1424
1425         return ret;
1426 }
1427
1428 /*
1429  * vmsplice splices a user address range into a pipe. It can be thought of
1430  * as splice-from-memory, where the regular splice is splice-from-file (or
1431  * to file). In both cases the output is a pipe, naturally.
1432  */
1433 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1434                              unsigned long nr_segs, unsigned int flags)
1435 {
1436         struct pipe_inode_info *pipe;
1437         struct page *pages[PIPE_BUFFERS];
1438         struct partial_page partial[PIPE_BUFFERS];
1439         struct splice_pipe_desc spd = {
1440                 .pages = pages,
1441                 .partial = partial,
1442                 .flags = flags,
1443                 .ops = &user_page_pipe_buf_ops,
1444                 .spd_release = spd_release_page,
1445         };
1446
1447         pipe = pipe_info(file->f_path.dentry->d_inode);
1448         if (!pipe)
1449                 return -EBADF;
1450
1451         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1452                                             flags & SPLICE_F_GIFT);
1453         if (spd.nr_pages <= 0)
1454                 return spd.nr_pages;
1455
1456         return splice_to_pipe(pipe, &spd);
1457 }
1458
1459 /*
1460  * Note that vmsplice only really supports true splicing _from_ user memory
1461  * to a pipe, not the other way around. Splicing from user memory is a simple
1462  * operation that can be supported without any funky alignment restrictions
1463  * or nasty vm tricks. We simply map in the user memory and fill them into
1464  * a pipe. The reverse isn't quite as easy, though. There are two possible
1465  * solutions for that:
1466  *
1467  *      - memcpy() the data internally, at which point we might as well just
1468  *        do a regular read() on the buffer anyway.
1469  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1470  *        has restriction limitations on both ends of the pipe).
1471  *
1472  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1473  *
1474  */
1475 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1476                 unsigned long, nr_segs, unsigned int, flags)
1477 {
1478         struct file *file;
1479         long error;
1480         int fput;
1481
1482         if (unlikely(nr_segs > UIO_MAXIOV))
1483                 return -EINVAL;
1484         else if (unlikely(!nr_segs))
1485                 return 0;
1486
1487         error = -EBADF;
1488         file = fget_light(fd, &fput);
1489         if (file) {
1490                 if (file->f_mode & FMODE_WRITE)
1491                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1492                 else if (file->f_mode & FMODE_READ)
1493                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1494
1495                 fput_light(file, fput);
1496         }
1497
1498         return error;
1499 }
1500
1501 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1502                 int, fd_out, loff_t __user *, off_out,
1503                 size_t, len, unsigned int, flags)
1504 {
1505         long error;
1506         struct file *in, *out;
1507         int fput_in, fput_out;
1508
1509         if (unlikely(!len))
1510                 return 0;
1511
1512         error = -EBADF;
1513         in = fget_light(fd_in, &fput_in);
1514         if (in) {
1515                 if (in->f_mode & FMODE_READ) {
1516                         out = fget_light(fd_out, &fput_out);
1517                         if (out) {
1518                                 if (out->f_mode & FMODE_WRITE)
1519                                         error = do_splice(in, off_in,
1520                                                           out, off_out,
1521                                                           len, flags);
1522                                 fput_light(out, fput_out);
1523                         }
1524                 }
1525
1526                 fput_light(in, fput_in);
1527         }
1528
1529         return error;
1530 }
1531
1532 /*
1533  * Make sure there's data to read. Wait for input if we can, otherwise
1534  * return an appropriate error.
1535  */
1536 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1537 {
1538         int ret;
1539
1540         /*
1541          * Check ->nrbufs without the inode lock first. This function
1542          * is speculative anyways, so missing one is ok.
1543          */
1544         if (pipe->nrbufs)
1545                 return 0;
1546
1547         ret = 0;
1548         pipe_lock(pipe);
1549
1550         while (!pipe->nrbufs) {
1551                 if (signal_pending(current)) {
1552                         ret = -ERESTARTSYS;
1553                         break;
1554                 }
1555                 if (!pipe->writers)
1556                         break;
1557                 if (!pipe->waiting_writers) {
1558                         if (flags & SPLICE_F_NONBLOCK) {
1559                                 ret = -EAGAIN;
1560                                 break;
1561                         }
1562                 }
1563                 pipe_wait(pipe);
1564         }
1565
1566         pipe_unlock(pipe);
1567         return ret;
1568 }
1569
1570 /*
1571  * Make sure there's writeable room. Wait for room if we can, otherwise
1572  * return an appropriate error.
1573  */
1574 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1575 {
1576         int ret;
1577
1578         /*
1579          * Check ->nrbufs without the inode lock first. This function
1580          * is speculative anyways, so missing one is ok.
1581          */
1582         if (pipe->nrbufs < PIPE_BUFFERS)
1583                 return 0;
1584
1585         ret = 0;
1586         pipe_lock(pipe);
1587
1588         while (pipe->nrbufs >= PIPE_BUFFERS) {
1589                 if (!pipe->readers) {
1590                         send_sig(SIGPIPE, current, 0);
1591                         ret = -EPIPE;
1592                         break;
1593                 }
1594                 if (flags & SPLICE_F_NONBLOCK) {
1595                         ret = -EAGAIN;
1596                         break;
1597                 }
1598                 if (signal_pending(current)) {
1599                         ret = -ERESTARTSYS;
1600                         break;
1601                 }
1602                 pipe->waiting_writers++;
1603                 pipe_wait(pipe);
1604                 pipe->waiting_writers--;
1605         }
1606
1607         pipe_unlock(pipe);
1608         return ret;
1609 }
1610
1611 /*
1612  * Splice contents of ipipe to opipe.
1613  */
1614 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1615                                struct pipe_inode_info *opipe,
1616                                size_t len, unsigned int flags)
1617 {
1618         struct pipe_buffer *ibuf, *obuf;
1619         int ret = 0, nbuf;
1620         bool input_wakeup = false;
1621
1622
1623 retry:
1624         ret = ipipe_prep(ipipe, flags);
1625         if (ret)
1626                 return ret;
1627
1628         ret = opipe_prep(opipe, flags);
1629         if (ret)
1630                 return ret;
1631
1632         /*
1633          * Potential ABBA deadlock, work around it by ordering lock
1634          * grabbing by pipe info address. Otherwise two different processes
1635          * could deadlock (one doing tee from A -> B, the other from B -> A).
1636          */
1637         pipe_double_lock(ipipe, opipe);
1638
1639         do {
1640                 if (!opipe->readers) {
1641                         send_sig(SIGPIPE, current, 0);
1642                         if (!ret)
1643                                 ret = -EPIPE;
1644                         break;
1645                 }
1646
1647                 if (!ipipe->nrbufs && !ipipe->writers)
1648                         break;
1649
1650                 /*
1651                  * Cannot make any progress, because either the input
1652                  * pipe is empty or the output pipe is full.
1653                  */
1654                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1655                         /* Already processed some buffers, break */
1656                         if (ret)
1657                                 break;
1658
1659                         if (flags & SPLICE_F_NONBLOCK) {
1660                                 ret = -EAGAIN;
1661                                 break;
1662                         }
1663
1664                         /*
1665                          * We raced with another reader/writer and haven't
1666                          * managed to process any buffers.  A zero return
1667                          * value means EOF, so retry instead.
1668                          */
1669                         pipe_unlock(ipipe);
1670                         pipe_unlock(opipe);
1671                         goto retry;
1672                 }
1673
1674                 ibuf = ipipe->bufs + ipipe->curbuf;
1675                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1676                 obuf = opipe->bufs + nbuf;
1677
1678                 if (len >= ibuf->len) {
1679                         /*
1680                          * Simply move the whole buffer from ipipe to opipe
1681                          */
1682                         *obuf = *ibuf;
1683                         ibuf->ops = NULL;
1684                         opipe->nrbufs++;
1685                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1686                         ipipe->nrbufs--;
1687                         input_wakeup = true;
1688                 } else {
1689                         /*
1690                          * Get a reference to this pipe buffer,
1691                          * so we can copy the contents over.
1692                          */
1693                         ibuf->ops->get(ipipe, ibuf);
1694                         *obuf = *ibuf;
1695
1696                         /*
1697                          * Don't inherit the gift flag, we need to
1698                          * prevent multiple steals of this page.
1699                          */
1700                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1701
1702                         obuf->len = len;
1703                         opipe->nrbufs++;
1704                         ibuf->offset += obuf->len;
1705                         ibuf->len -= obuf->len;
1706                 }
1707                 ret += obuf->len;
1708                 len -= obuf->len;
1709         } while (len);
1710
1711         pipe_unlock(ipipe);
1712         pipe_unlock(opipe);
1713
1714         /*
1715          * If we put data in the output pipe, wakeup any potential readers.
1716          */
1717         if (ret > 0) {
1718                 smp_mb();
1719                 if (waitqueue_active(&opipe->wait))
1720                         wake_up_interruptible(&opipe->wait);
1721                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1722         }
1723         if (input_wakeup)
1724                 wakeup_pipe_writers(ipipe);
1725
1726         return ret;
1727 }
1728
1729 /*
1730  * Link contents of ipipe to opipe.
1731  */
1732 static int link_pipe(struct pipe_inode_info *ipipe,
1733                      struct pipe_inode_info *opipe,
1734                      size_t len, unsigned int flags)
1735 {
1736         struct pipe_buffer *ibuf, *obuf;
1737         int ret = 0, i = 0, nbuf;
1738
1739         /*
1740          * Potential ABBA deadlock, work around it by ordering lock
1741          * grabbing by pipe info address. Otherwise two different processes
1742          * could deadlock (one doing tee from A -> B, the other from B -> A).
1743          */
1744         pipe_double_lock(ipipe, opipe);
1745
1746         do {
1747                 if (!opipe->readers) {
1748                         send_sig(SIGPIPE, current, 0);
1749                         if (!ret)
1750                                 ret = -EPIPE;
1751                         break;
1752                 }
1753
1754                 /*
1755                  * If we have iterated all input buffers or ran out of
1756                  * output room, break.
1757                  */
1758                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1759                         break;
1760
1761                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1762                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1763
1764                 /*
1765                  * Get a reference to this pipe buffer,
1766                  * so we can copy the contents over.
1767                  */
1768                 ibuf->ops->get(ipipe, ibuf);
1769
1770                 obuf = opipe->bufs + nbuf;
1771                 *obuf = *ibuf;
1772
1773                 /*
1774                  * Don't inherit the gift flag, we need to
1775                  * prevent multiple steals of this page.
1776                  */
1777                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1778
1779                 if (obuf->len > len)
1780                         obuf->len = len;
1781
1782                 opipe->nrbufs++;
1783                 ret += obuf->len;
1784                 len -= obuf->len;
1785                 i++;
1786         } while (len);
1787
1788         /*
1789          * return EAGAIN if we have the potential of some data in the
1790          * future, otherwise just return 0
1791          */
1792         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1793                 ret = -EAGAIN;
1794
1795         pipe_unlock(ipipe);
1796         pipe_unlock(opipe);
1797
1798         /*
1799          * If we put data in the output pipe, wakeup any potential readers.
1800          */
1801         if (ret > 0) {
1802                 smp_mb();
1803                 if (waitqueue_active(&opipe->wait))
1804                         wake_up_interruptible(&opipe->wait);
1805                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1806         }
1807
1808         return ret;
1809 }
1810
1811 /*
1812  * This is a tee(1) implementation that works on pipes. It doesn't copy
1813  * any data, it simply references the 'in' pages on the 'out' pipe.
1814  * The 'flags' used are the SPLICE_F_* variants, currently the only
1815  * applicable one is SPLICE_F_NONBLOCK.
1816  */
1817 static long do_tee(struct file *in, struct file *out, size_t len,
1818                    unsigned int flags)
1819 {
1820         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1821         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1822         int ret = -EINVAL;
1823
1824         /*
1825          * Duplicate the contents of ipipe to opipe without actually
1826          * copying the data.
1827          */
1828         if (ipipe && opipe && ipipe != opipe) {
1829                 /*
1830                  * Keep going, unless we encounter an error. The ipipe/opipe
1831                  * ordering doesn't really matter.
1832                  */
1833                 ret = ipipe_prep(ipipe, flags);
1834                 if (!ret) {
1835                         ret = opipe_prep(opipe, flags);
1836                         if (!ret)
1837                                 ret = link_pipe(ipipe, opipe, len, flags);
1838                 }
1839         }
1840
1841         return ret;
1842 }
1843
1844 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1845 {
1846         struct file *in;
1847         int error, fput_in;
1848
1849         if (unlikely(!len))
1850                 return 0;
1851
1852         error = -EBADF;
1853         in = fget_light(fdin, &fput_in);
1854         if (in) {
1855                 if (in->f_mode & FMODE_READ) {
1856                         int fput_out;
1857                         struct file *out = fget_light(fdout, &fput_out);
1858
1859                         if (out) {
1860                                 if (out->f_mode & FMODE_WRITE)
1861                                         error = do_tee(in, out, len, flags);
1862                                 fput_light(out, fput_out);
1863                         }
1864                 }
1865                 fput_light(in, fput_in);
1866         }
1867
1868         return error;
1869 }