splice: completely document external interface with kerneldoc
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31
32 /*
33  * Attempt to steal a page from a pipe buffer. This should perhaps go into
34  * a vm helper function, it's already simplified quite a bit by the
35  * addition of remove_mapping(). If success is returned, the caller may
36  * attempt to reuse this page for another destination.
37  */
38 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
39                                      struct pipe_buffer *buf)
40 {
41         struct page *page = buf->page;
42         struct address_space *mapping;
43
44         lock_page(page);
45
46         mapping = page_mapping(page);
47         if (mapping) {
48                 WARN_ON(!PageUptodate(page));
49
50                 /*
51                  * At least for ext2 with nobh option, we need to wait on
52                  * writeback completing on this page, since we'll remove it
53                  * from the pagecache.  Otherwise truncate wont wait on the
54                  * page, allowing the disk blocks to be reused by someone else
55                  * before we actually wrote our data to them. fs corruption
56                  * ensues.
57                  */
58                 wait_on_page_writeback(page);
59
60                 if (PagePrivate(page))
61                         try_to_release_page(page, GFP_KERNEL);
62
63                 /*
64                  * If we succeeded in removing the mapping, set LRU flag
65                  * and return good.
66                  */
67                 if (remove_mapping(mapping, page)) {
68                         buf->flags |= PIPE_BUF_FLAG_LRU;
69                         return 0;
70                 }
71         }
72
73         /*
74          * Raced with truncate or failed to remove page from current
75          * address space, unlock and return failure.
76          */
77         unlock_page(page);
78         return 1;
79 }
80
81 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
82                                         struct pipe_buffer *buf)
83 {
84         page_cache_release(buf->page);
85         buf->flags &= ~PIPE_BUF_FLAG_LRU;
86 }
87
88 static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
89                                    struct pipe_buffer *buf)
90 {
91         struct page *page = buf->page;
92         int err;
93
94         if (!PageUptodate(page)) {
95                 lock_page(page);
96
97                 /*
98                  * Page got truncated/unhashed. This will cause a 0-byte
99                  * splice, if this is the first page.
100                  */
101                 if (!page->mapping) {
102                         err = -ENODATA;
103                         goto error;
104                 }
105
106                 /*
107                  * Uh oh, read-error from disk.
108                  */
109                 if (!PageUptodate(page)) {
110                         err = -EIO;
111                         goto error;
112                 }
113
114                 /*
115                  * Page is ok afterall, we are done.
116                  */
117                 unlock_page(page);
118         }
119
120         return 0;
121 error:
122         unlock_page(page);
123         return err;
124 }
125
126 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
127         .can_merge = 0,
128         .map = generic_pipe_buf_map,
129         .unmap = generic_pipe_buf_unmap,
130         .pin = page_cache_pipe_buf_pin,
131         .release = page_cache_pipe_buf_release,
132         .steal = page_cache_pipe_buf_steal,
133         .get = generic_pipe_buf_get,
134 };
135
136 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
137                                     struct pipe_buffer *buf)
138 {
139         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
140                 return 1;
141
142         buf->flags |= PIPE_BUF_FLAG_LRU;
143         return generic_pipe_buf_steal(pipe, buf);
144 }
145
146 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
147         .can_merge = 0,
148         .map = generic_pipe_buf_map,
149         .unmap = generic_pipe_buf_unmap,
150         .pin = generic_pipe_buf_pin,
151         .release = page_cache_pipe_buf_release,
152         .steal = user_page_pipe_buf_steal,
153         .get = generic_pipe_buf_get,
154 };
155
156 /**
157  * splice_to_pipe - fill passed data into a pipe
158  * @pipe:       pipe to fill
159  * @spd:        data to fill
160  *
161  * Description:
162  *    @spd contains a map of pages and len/offset tupples, a long with
163  *    the struct pipe_buf_operations associated with these pages. This
164  *    function will link that data to the pipe.
165  *
166  */
167 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
168                        struct splice_pipe_desc *spd)
169 {
170         unsigned int spd_pages = spd->nr_pages;
171         int ret, do_wakeup, page_nr;
172
173         ret = 0;
174         do_wakeup = 0;
175         page_nr = 0;
176
177         if (pipe->inode)
178                 mutex_lock(&pipe->inode->i_mutex);
179
180         for (;;) {
181                 if (!pipe->readers) {
182                         send_sig(SIGPIPE, current, 0);
183                         if (!ret)
184                                 ret = -EPIPE;
185                         break;
186                 }
187
188                 if (pipe->nrbufs < PIPE_BUFFERS) {
189                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
190                         struct pipe_buffer *buf = pipe->bufs + newbuf;
191
192                         buf->page = spd->pages[page_nr];
193                         buf->offset = spd->partial[page_nr].offset;
194                         buf->len = spd->partial[page_nr].len;
195                         buf->private = spd->partial[page_nr].private;
196                         buf->ops = spd->ops;
197                         if (spd->flags & SPLICE_F_GIFT)
198                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
199
200                         pipe->nrbufs++;
201                         page_nr++;
202                         ret += buf->len;
203
204                         if (pipe->inode)
205                                 do_wakeup = 1;
206
207                         if (!--spd->nr_pages)
208                                 break;
209                         if (pipe->nrbufs < PIPE_BUFFERS)
210                                 continue;
211
212                         break;
213                 }
214
215                 if (spd->flags & SPLICE_F_NONBLOCK) {
216                         if (!ret)
217                                 ret = -EAGAIN;
218                         break;
219                 }
220
221                 if (signal_pending(current)) {
222                         if (!ret)
223                                 ret = -ERESTARTSYS;
224                         break;
225                 }
226
227                 if (do_wakeup) {
228                         smp_mb();
229                         if (waitqueue_active(&pipe->wait))
230                                 wake_up_interruptible_sync(&pipe->wait);
231                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
232                         do_wakeup = 0;
233                 }
234
235                 pipe->waiting_writers++;
236                 pipe_wait(pipe);
237                 pipe->waiting_writers--;
238         }
239
240         if (pipe->inode) {
241                 mutex_unlock(&pipe->inode->i_mutex);
242
243                 if (do_wakeup) {
244                         smp_mb();
245                         if (waitqueue_active(&pipe->wait))
246                                 wake_up_interruptible(&pipe->wait);
247                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
248                 }
249         }
250
251         while (page_nr < spd_pages)
252                 page_cache_release(spd->pages[page_nr++]);
253
254         return ret;
255 }
256
257 static int
258 __generic_file_splice_read(struct file *in, loff_t *ppos,
259                            struct pipe_inode_info *pipe, size_t len,
260                            unsigned int flags)
261 {
262         struct address_space *mapping = in->f_mapping;
263         unsigned int loff, nr_pages;
264         struct page *pages[PIPE_BUFFERS];
265         struct partial_page partial[PIPE_BUFFERS];
266         struct page *page;
267         pgoff_t index, end_index;
268         loff_t isize;
269         int error, page_nr;
270         struct splice_pipe_desc spd = {
271                 .pages = pages,
272                 .partial = partial,
273                 .flags = flags,
274                 .ops = &page_cache_pipe_buf_ops,
275         };
276
277         index = *ppos >> PAGE_CACHE_SHIFT;
278         loff = *ppos & ~PAGE_CACHE_MASK;
279         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
280
281         if (nr_pages > PIPE_BUFFERS)
282                 nr_pages = PIPE_BUFFERS;
283
284         /*
285          * Don't try to 2nd guess the read-ahead logic, call into
286          * page_cache_readahead() like the page cache reads would do.
287          */
288         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
289
290         /*
291          * Lookup the (hopefully) full range of pages we need.
292          */
293         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
294
295         /*
296          * If find_get_pages_contig() returned fewer pages than we needed,
297          * allocate the rest and fill in the holes.
298          */
299         error = 0;
300         index += spd.nr_pages;
301         while (spd.nr_pages < nr_pages) {
302                 /*
303                  * Page could be there, find_get_pages_contig() breaks on
304                  * the first hole.
305                  */
306                 page = find_get_page(mapping, index);
307                 if (!page) {
308                         /*
309                          * Make sure the read-ahead engine is notified
310                          * about this failure.
311                          */
312                         handle_ra_miss(mapping, &in->f_ra, index);
313
314                         /*
315                          * page didn't exist, allocate one.
316                          */
317                         page = page_cache_alloc_cold(mapping);
318                         if (!page)
319                                 break;
320
321                         error = add_to_page_cache_lru(page, mapping, index,
322                                               GFP_KERNEL);
323                         if (unlikely(error)) {
324                                 page_cache_release(page);
325                                 if (error == -EEXIST)
326                                         continue;
327                                 break;
328                         }
329                         /*
330                          * add_to_page_cache() locks the page, unlock it
331                          * to avoid convoluting the logic below even more.
332                          */
333                         unlock_page(page);
334                 }
335
336                 pages[spd.nr_pages++] = page;
337                 index++;
338         }
339
340         /*
341          * Now loop over the map and see if we need to start IO on any
342          * pages, fill in the partial map, etc.
343          */
344         index = *ppos >> PAGE_CACHE_SHIFT;
345         nr_pages = spd.nr_pages;
346         spd.nr_pages = 0;
347         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
348                 unsigned int this_len;
349
350                 if (!len)
351                         break;
352
353                 /*
354                  * this_len is the max we'll use from this page
355                  */
356                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
357                 page = pages[page_nr];
358
359                 /*
360                  * If the page isn't uptodate, we may need to start io on it
361                  */
362                 if (!PageUptodate(page)) {
363                         /*
364                          * If in nonblock mode then dont block on waiting
365                          * for an in-flight io page
366                          */
367                         if (flags & SPLICE_F_NONBLOCK) {
368                                 if (TestSetPageLocked(page))
369                                         break;
370                         } else
371                                 lock_page(page);
372
373                         /*
374                          * page was truncated, stop here. if this isn't the
375                          * first page, we'll just complete what we already
376                          * added
377                          */
378                         if (!page->mapping) {
379                                 unlock_page(page);
380                                 break;
381                         }
382                         /*
383                          * page was already under io and is now done, great
384                          */
385                         if (PageUptodate(page)) {
386                                 unlock_page(page);
387                                 goto fill_it;
388                         }
389
390                         /*
391                          * need to read in the page
392                          */
393                         error = mapping->a_ops->readpage(in, page);
394                         if (unlikely(error)) {
395                                 /*
396                                  * We really should re-lookup the page here,
397                                  * but it complicates things a lot. Instead
398                                  * lets just do what we already stored, and
399                                  * we'll get it the next time we are called.
400                                  */
401                                 if (error == AOP_TRUNCATED_PAGE)
402                                         error = 0;
403
404                                 break;
405                         }
406                 }
407 fill_it:
408                 /*
409                  * i_size must be checked after PageUptodate.
410                  */
411                 isize = i_size_read(mapping->host);
412                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
413                 if (unlikely(!isize || index > end_index))
414                         break;
415
416                 /*
417                  * if this is the last page, see if we need to shrink
418                  * the length and stop
419                  */
420                 if (end_index == index) {
421                         unsigned int plen;
422
423                         /*
424                          * max good bytes in this page
425                          */
426                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
427                         if (plen <= loff)
428                                 break;
429
430                         /*
431                          * force quit after adding this page
432                          */
433                         this_len = min(this_len, plen - loff);
434                         len = this_len;
435                 }
436
437                 partial[page_nr].offset = loff;
438                 partial[page_nr].len = this_len;
439                 len -= this_len;
440                 loff = 0;
441                 spd.nr_pages++;
442                 index++;
443         }
444
445         /*
446          * Release any pages at the end, if we quit early. 'page_nr' is how far
447          * we got, 'nr_pages' is how many pages are in the map.
448          */
449         while (page_nr < nr_pages)
450                 page_cache_release(pages[page_nr++]);
451
452         if (spd.nr_pages)
453                 return splice_to_pipe(pipe, &spd);
454
455         return error;
456 }
457
458 /**
459  * generic_file_splice_read - splice data from file to a pipe
460  * @in:         file to splice from
461  * @ppos:       position in @in
462  * @pipe:       pipe to splice to
463  * @len:        number of bytes to splice
464  * @flags:      splice modifier flags
465  *
466  * Description:
467  *    Will read pages from given file and fill them into a pipe. Can be
468  *    used as long as the address_space operations for the source implements
469  *    a readpage() hook.
470  *
471  */
472 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
473                                  struct pipe_inode_info *pipe, size_t len,
474                                  unsigned int flags)
475 {
476         ssize_t spliced;
477         int ret;
478         loff_t isize, left;
479
480         isize = i_size_read(in->f_mapping->host);
481         if (unlikely(*ppos >= isize))
482                 return 0;
483
484         left = isize - *ppos;
485         if (unlikely(left < len))
486                 len = left;
487
488         ret = 0;
489         spliced = 0;
490         while (len) {
491                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
492
493                 if (ret < 0)
494                         break;
495                 else if (!ret) {
496                         if (spliced)
497                                 break;
498                         if (flags & SPLICE_F_NONBLOCK) {
499                                 ret = -EAGAIN;
500                                 break;
501                         }
502                 }
503
504                 *ppos += ret;
505                 len -= ret;
506                 spliced += ret;
507         }
508
509         if (spliced)
510                 return spliced;
511
512         return ret;
513 }
514
515 EXPORT_SYMBOL(generic_file_splice_read);
516
517 /*
518  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
519  * using sendpage(). Return the number of bytes sent.
520  */
521 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
522                             struct pipe_buffer *buf, struct splice_desc *sd)
523 {
524         struct file *file = sd->u.file;
525         loff_t pos = sd->pos;
526         int ret, more;
527
528         ret = buf->ops->pin(pipe, buf);
529         if (!ret) {
530                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
531
532                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
533                                            sd->len, &pos, more);
534         }
535
536         return ret;
537 }
538
539 /*
540  * This is a little more tricky than the file -> pipe splicing. There are
541  * basically three cases:
542  *
543  *      - Destination page already exists in the address space and there
544  *        are users of it. For that case we have no other option that
545  *        copying the data. Tough luck.
546  *      - Destination page already exists in the address space, but there
547  *        are no users of it. Make sure it's uptodate, then drop it. Fall
548  *        through to last case.
549  *      - Destination page does not exist, we can add the pipe page to
550  *        the page cache and avoid the copy.
551  *
552  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
553  * sd->flags), we attempt to migrate pages from the pipe to the output
554  * file address space page cache. This is possible if no one else has
555  * the pipe page referenced outside of the pipe and page cache. If
556  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
557  * a new page in the output file page cache and fill/dirty that.
558  */
559 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
560                         struct splice_desc *sd)
561 {
562         struct file *file = sd->u.file;
563         struct address_space *mapping = file->f_mapping;
564         unsigned int offset, this_len;
565         struct page *page;
566         pgoff_t index;
567         int ret;
568
569         /*
570          * make sure the data in this buffer is uptodate
571          */
572         ret = buf->ops->pin(pipe, buf);
573         if (unlikely(ret))
574                 return ret;
575
576         index = sd->pos >> PAGE_CACHE_SHIFT;
577         offset = sd->pos & ~PAGE_CACHE_MASK;
578
579         this_len = sd->len;
580         if (this_len + offset > PAGE_CACHE_SIZE)
581                 this_len = PAGE_CACHE_SIZE - offset;
582
583 find_page:
584         page = find_lock_page(mapping, index);
585         if (!page) {
586                 ret = -ENOMEM;
587                 page = page_cache_alloc_cold(mapping);
588                 if (unlikely(!page))
589                         goto out_ret;
590
591                 /*
592                  * This will also lock the page
593                  */
594                 ret = add_to_page_cache_lru(page, mapping, index,
595                                             GFP_KERNEL);
596                 if (unlikely(ret))
597                         goto out;
598         }
599
600         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
601         if (unlikely(ret)) {
602                 loff_t isize = i_size_read(mapping->host);
603
604                 if (ret != AOP_TRUNCATED_PAGE)
605                         unlock_page(page);
606                 page_cache_release(page);
607                 if (ret == AOP_TRUNCATED_PAGE)
608                         goto find_page;
609
610                 /*
611                  * prepare_write() may have instantiated a few blocks
612                  * outside i_size.  Trim these off again.
613                  */
614                 if (sd->pos + this_len > isize)
615                         vmtruncate(mapping->host, isize);
616
617                 goto out_ret;
618         }
619
620         if (buf->page != page) {
621                 /*
622                  * Careful, ->map() uses KM_USER0!
623                  */
624                 char *src = buf->ops->map(pipe, buf, 1);
625                 char *dst = kmap_atomic(page, KM_USER1);
626
627                 memcpy(dst + offset, src + buf->offset, this_len);
628                 flush_dcache_page(page);
629                 kunmap_atomic(dst, KM_USER1);
630                 buf->ops->unmap(pipe, buf, src);
631         }
632
633         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
634         if (ret) {
635                 if (ret == AOP_TRUNCATED_PAGE) {
636                         page_cache_release(page);
637                         goto find_page;
638                 }
639                 if (ret < 0)
640                         goto out;
641                 /*
642                  * Partial write has happened, so 'ret' already initialized by
643                  * number of bytes written, Where is nothing we have to do here.
644                  */
645         } else
646                 ret = this_len;
647         /*
648          * Return the number of bytes written and mark page as
649          * accessed, we are now done!
650          */
651         mark_page_accessed(page);
652 out:
653         page_cache_release(page);
654         unlock_page(page);
655 out_ret:
656         return ret;
657 }
658
659 /**
660  * __splice_from_pipe - splice data from a pipe to given actor
661  * @pipe:       pipe to splice from
662  * @sd:         information to @actor
663  * @actor:      handler that splices the data
664  *
665  * Description:
666  *    This function does little more than loop over the pipe and call
667  *    @actor to do the actual moving of a single struct pipe_buffer to
668  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
669  *    pipe_to_user.
670  *
671  */
672 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
673                            splice_actor *actor)
674 {
675         int ret, do_wakeup, err;
676
677         ret = 0;
678         do_wakeup = 0;
679
680         for (;;) {
681                 if (pipe->nrbufs) {
682                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
683                         const struct pipe_buf_operations *ops = buf->ops;
684
685                         sd->len = buf->len;
686                         if (sd->len > sd->total_len)
687                                 sd->len = sd->total_len;
688
689                         err = actor(pipe, buf, sd);
690                         if (err <= 0) {
691                                 if (!ret && err != -ENODATA)
692                                         ret = err;
693
694                                 break;
695                         }
696
697                         ret += err;
698                         buf->offset += err;
699                         buf->len -= err;
700
701                         sd->len -= err;
702                         sd->pos += err;
703                         sd->total_len -= err;
704                         if (sd->len)
705                                 continue;
706
707                         if (!buf->len) {
708                                 buf->ops = NULL;
709                                 ops->release(pipe, buf);
710                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
711                                 pipe->nrbufs--;
712                                 if (pipe->inode)
713                                         do_wakeup = 1;
714                         }
715
716                         if (!sd->total_len)
717                                 break;
718                 }
719
720                 if (pipe->nrbufs)
721                         continue;
722                 if (!pipe->writers)
723                         break;
724                 if (!pipe->waiting_writers) {
725                         if (ret)
726                                 break;
727                 }
728
729                 if (sd->flags & SPLICE_F_NONBLOCK) {
730                         if (!ret)
731                                 ret = -EAGAIN;
732                         break;
733                 }
734
735                 if (signal_pending(current)) {
736                         if (!ret)
737                                 ret = -ERESTARTSYS;
738                         break;
739                 }
740
741                 if (do_wakeup) {
742                         smp_mb();
743                         if (waitqueue_active(&pipe->wait))
744                                 wake_up_interruptible_sync(&pipe->wait);
745                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
746                         do_wakeup = 0;
747                 }
748
749                 pipe_wait(pipe);
750         }
751
752         if (do_wakeup) {
753                 smp_mb();
754                 if (waitqueue_active(&pipe->wait))
755                         wake_up_interruptible(&pipe->wait);
756                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
757         }
758
759         return ret;
760 }
761 EXPORT_SYMBOL(__splice_from_pipe);
762
763 /**
764  * splice_from_pipe - splice data from a pipe to a file
765  * @pipe:       pipe to splice from
766  * @out:        file to splice to
767  * @ppos:       position in @out
768  * @len:        how many bytes to splice
769  * @flags:      splice modifier flags
770  * @actor:      handler that splices the data
771  *
772  * Description:
773  *    See __splice_from_pipe. This function locks the input and output inodes,
774  *    otherwise it's identical to __splice_from_pipe().
775  *
776  */
777 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
778                          loff_t *ppos, size_t len, unsigned int flags,
779                          splice_actor *actor)
780 {
781         ssize_t ret;
782         struct inode *inode = out->f_mapping->host;
783         struct splice_desc sd = {
784                 .total_len = len,
785                 .flags = flags,
786                 .pos = *ppos,
787                 .u.file = out,
788         };
789
790         /*
791          * The actor worker might be calling ->prepare_write and
792          * ->commit_write. Most of the time, these expect i_mutex to
793          * be held. Since this may result in an ABBA deadlock with
794          * pipe->inode, we have to order lock acquiry here.
795          */
796         inode_double_lock(inode, pipe->inode);
797         ret = __splice_from_pipe(pipe, &sd, actor);
798         inode_double_unlock(inode, pipe->inode);
799
800         return ret;
801 }
802
803 /**
804  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
805  * @pipe:       pipe info
806  * @out:        file to write to
807  * @ppos:       position in @out
808  * @len:        number of bytes to splice
809  * @flags:      splice modifier flags
810  *
811  * Description:
812  *    Will either move or copy pages (determined by @flags options) from
813  *    the given pipe inode to the given file. The caller is responsible
814  *    for acquiring i_mutex on both inodes.
815  *
816  */
817 ssize_t
818 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
819                                  loff_t *ppos, size_t len, unsigned int flags)
820 {
821         struct address_space *mapping = out->f_mapping;
822         struct inode *inode = mapping->host;
823         struct splice_desc sd = {
824                 .total_len = len,
825                 .flags = flags,
826                 .pos = *ppos,
827                 .u.file = out,
828         };
829         ssize_t ret;
830         int err;
831
832         err = remove_suid(out->f_path.dentry);
833         if (unlikely(err))
834                 return err;
835
836         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
837         if (ret > 0) {
838                 unsigned long nr_pages;
839
840                 *ppos += ret;
841                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
842
843                 /*
844                  * If file or inode is SYNC and we actually wrote some data,
845                  * sync it.
846                  */
847                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
848                         err = generic_osync_inode(inode, mapping,
849                                                   OSYNC_METADATA|OSYNC_DATA);
850
851                         if (err)
852                                 ret = err;
853                 }
854                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
855         }
856
857         return ret;
858 }
859
860 EXPORT_SYMBOL(generic_file_splice_write_nolock);
861
862 /**
863  * generic_file_splice_write - splice data from a pipe to a file
864  * @pipe:       pipe info
865  * @out:        file to write to
866  * @ppos:       position in @out
867  * @len:        number of bytes to splice
868  * @flags:      splice modifier flags
869  *
870  * Description:
871  *    Will either move or copy pages (determined by @flags options) from
872  *    the given pipe inode to the given file.
873  *
874  */
875 ssize_t
876 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
877                           loff_t *ppos, size_t len, unsigned int flags)
878 {
879         struct address_space *mapping = out->f_mapping;
880         struct inode *inode = mapping->host;
881         ssize_t ret;
882         int err;
883
884         err = should_remove_suid(out->f_path.dentry);
885         if (unlikely(err)) {
886                 mutex_lock(&inode->i_mutex);
887                 err = __remove_suid(out->f_path.dentry, err);
888                 mutex_unlock(&inode->i_mutex);
889                 if (err)
890                         return err;
891         }
892
893         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
894         if (ret > 0) {
895                 unsigned long nr_pages;
896
897                 *ppos += ret;
898                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
899
900                 /*
901                  * If file or inode is SYNC and we actually wrote some data,
902                  * sync it.
903                  */
904                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
905                         mutex_lock(&inode->i_mutex);
906                         err = generic_osync_inode(inode, mapping,
907                                                   OSYNC_METADATA|OSYNC_DATA);
908                         mutex_unlock(&inode->i_mutex);
909
910                         if (err)
911                                 ret = err;
912                 }
913                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
914         }
915
916         return ret;
917 }
918
919 EXPORT_SYMBOL(generic_file_splice_write);
920
921 /**
922  * generic_splice_sendpage - splice data from a pipe to a socket
923  * @pipe:       pipe to splice from
924  * @out:        socket to write to
925  * @ppos:       position in @out
926  * @len:        number of bytes to splice
927  * @flags:      splice modifier flags
928  *
929  * Description:
930  *    Will send @len bytes from the pipe to a network socket. No data copying
931  *    is involved.
932  *
933  */
934 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
935                                 loff_t *ppos, size_t len, unsigned int flags)
936 {
937         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
938 }
939
940 EXPORT_SYMBOL(generic_splice_sendpage);
941
942 /*
943  * Attempt to initiate a splice from pipe to file.
944  */
945 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
946                            loff_t *ppos, size_t len, unsigned int flags)
947 {
948         int ret;
949
950         if (unlikely(!out->f_op || !out->f_op->splice_write))
951                 return -EINVAL;
952
953         if (unlikely(!(out->f_mode & FMODE_WRITE)))
954                 return -EBADF;
955
956         ret = rw_verify_area(WRITE, out, ppos, len);
957         if (unlikely(ret < 0))
958                 return ret;
959
960         return out->f_op->splice_write(pipe, out, ppos, len, flags);
961 }
962
963 /*
964  * Attempt to initiate a splice from a file to a pipe.
965  */
966 static long do_splice_to(struct file *in, loff_t *ppos,
967                          struct pipe_inode_info *pipe, size_t len,
968                          unsigned int flags)
969 {
970         int ret;
971
972         if (unlikely(!in->f_op || !in->f_op->splice_read))
973                 return -EINVAL;
974
975         if (unlikely(!(in->f_mode & FMODE_READ)))
976                 return -EBADF;
977
978         ret = rw_verify_area(READ, in, ppos, len);
979         if (unlikely(ret < 0))
980                 return ret;
981
982         return in->f_op->splice_read(in, ppos, pipe, len, flags);
983 }
984
985 /**
986  * splice_direct_to_actor - splices data directly between two non-pipes
987  * @in:         file to splice from
988  * @sd:         actor information on where to splice to
989  * @actor:      handles the data splicing
990  *
991  * Description:
992  *    This is a special case helper to splice directly between two
993  *    points, without requiring an explicit pipe. Internally an allocated
994  *    pipe is cached in the process, and reused during the life time of
995  *    that process.
996  *
997  */
998 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
999                                splice_direct_actor *actor)
1000 {
1001         struct pipe_inode_info *pipe;
1002         long ret, bytes;
1003         umode_t i_mode;
1004         size_t len;
1005         int i, flags;
1006
1007         /*
1008          * We require the input being a regular file, as we don't want to
1009          * randomly drop data for eg socket -> socket splicing. Use the
1010          * piped splicing for that!
1011          */
1012         i_mode = in->f_path.dentry->d_inode->i_mode;
1013         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1014                 return -EINVAL;
1015
1016         /*
1017          * neither in nor out is a pipe, setup an internal pipe attached to
1018          * 'out' and transfer the wanted data from 'in' to 'out' through that
1019          */
1020         pipe = current->splice_pipe;
1021         if (unlikely(!pipe)) {
1022                 pipe = alloc_pipe_info(NULL);
1023                 if (!pipe)
1024                         return -ENOMEM;
1025
1026                 /*
1027                  * We don't have an immediate reader, but we'll read the stuff
1028                  * out of the pipe right after the splice_to_pipe(). So set
1029                  * PIPE_READERS appropriately.
1030                  */
1031                 pipe->readers = 1;
1032
1033                 current->splice_pipe = pipe;
1034         }
1035
1036         /*
1037          * Do the splice.
1038          */
1039         ret = 0;
1040         bytes = 0;
1041         len = sd->total_len;
1042         flags = sd->flags;
1043
1044         /*
1045          * Don't block on output, we have to drain the direct pipe.
1046          */
1047         sd->flags &= ~SPLICE_F_NONBLOCK;
1048
1049         while (len) {
1050                 size_t read_len, max_read_len;
1051
1052                 /*
1053                  * Do at most PIPE_BUFFERS pages worth of transfer:
1054                  */
1055                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1056
1057                 ret = do_splice_to(in, &sd->pos, pipe, max_read_len, flags);
1058                 if (unlikely(ret < 0))
1059                         goto out_release;
1060
1061                 read_len = ret;
1062                 sd->total_len = read_len;
1063
1064                 /*
1065                  * NOTE: nonblocking mode only applies to the input. We
1066                  * must not do the output in nonblocking mode as then we
1067                  * could get stuck data in the internal pipe:
1068                  */
1069                 ret = actor(pipe, sd);
1070                 if (unlikely(ret < 0))
1071                         goto out_release;
1072
1073                 bytes += ret;
1074                 len -= ret;
1075
1076                 /*
1077                  * In nonblocking mode, if we got back a short read then
1078                  * that was due to either an IO error or due to the
1079                  * pagecache entry not being there. In the IO error case
1080                  * the _next_ splice attempt will produce a clean IO error
1081                  * return value (not a short read), so in both cases it's
1082                  * correct to break out of the loop here:
1083                  */
1084                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1085                         break;
1086         }
1087
1088         pipe->nrbufs = pipe->curbuf = 0;
1089
1090         return bytes;
1091
1092 out_release:
1093         /*
1094          * If we did an incomplete transfer we must release
1095          * the pipe buffers in question:
1096          */
1097         for (i = 0; i < PIPE_BUFFERS; i++) {
1098                 struct pipe_buffer *buf = pipe->bufs + i;
1099
1100                 if (buf->ops) {
1101                         buf->ops->release(pipe, buf);
1102                         buf->ops = NULL;
1103                 }
1104         }
1105         pipe->nrbufs = pipe->curbuf = 0;
1106
1107         /*
1108          * If we transferred some data, return the number of bytes:
1109          */
1110         if (bytes > 0)
1111                 return bytes;
1112
1113         return ret;
1114
1115 }
1116 EXPORT_SYMBOL(splice_direct_to_actor);
1117
1118 static int direct_splice_actor(struct pipe_inode_info *pipe,
1119                                struct splice_desc *sd)
1120 {
1121         struct file *file = sd->u.file;
1122
1123         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1124 }
1125
1126 /**
1127  * do_splice_direct - splices data directly between two files
1128  * @in:         file to splice from
1129  * @ppos:       input file offset
1130  * @out:        file to splice to
1131  * @len:        number of bytes to splice
1132  * @flags:      splice modifier flags
1133  *
1134  * Description:
1135  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1136  *    doing it in the application would incur an extra system call
1137  *    (splice in + splice out, as compared to just sendfile()). So this helper
1138  *    can splice directly through a process-private pipe.
1139  *
1140  */
1141 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1142                       size_t len, unsigned int flags)
1143 {
1144         struct splice_desc sd = {
1145                 .len            = len,
1146                 .total_len      = len,
1147                 .flags          = flags,
1148                 .pos            = *ppos,
1149                 .u.file         = out,
1150         };
1151         size_t ret;
1152
1153         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1154         *ppos = sd.pos;
1155         return ret;
1156 }
1157
1158 /*
1159  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1160  * location, so checking ->i_pipe is not enough to verify that this is a
1161  * pipe.
1162  */
1163 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1164 {
1165         if (S_ISFIFO(inode->i_mode))
1166                 return inode->i_pipe;
1167
1168         return NULL;
1169 }
1170
1171 /*
1172  * Determine where to splice to/from.
1173  */
1174 static long do_splice(struct file *in, loff_t __user *off_in,
1175                       struct file *out, loff_t __user *off_out,
1176                       size_t len, unsigned int flags)
1177 {
1178         struct pipe_inode_info *pipe;
1179         loff_t offset, *off;
1180         long ret;
1181
1182         pipe = pipe_info(in->f_path.dentry->d_inode);
1183         if (pipe) {
1184                 if (off_in)
1185                         return -ESPIPE;
1186                 if (off_out) {
1187                         if (out->f_op->llseek == no_llseek)
1188                                 return -EINVAL;
1189                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1190                                 return -EFAULT;
1191                         off = &offset;
1192                 } else
1193                         off = &out->f_pos;
1194
1195                 ret = do_splice_from(pipe, out, off, len, flags);
1196
1197                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1198                         ret = -EFAULT;
1199
1200                 return ret;
1201         }
1202
1203         pipe = pipe_info(out->f_path.dentry->d_inode);
1204         if (pipe) {
1205                 if (off_out)
1206                         return -ESPIPE;
1207                 if (off_in) {
1208                         if (in->f_op->llseek == no_llseek)
1209                                 return -EINVAL;
1210                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1211                                 return -EFAULT;
1212                         off = &offset;
1213                 } else
1214                         off = &in->f_pos;
1215
1216                 ret = do_splice_to(in, off, pipe, len, flags);
1217
1218                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1219                         ret = -EFAULT;
1220
1221                 return ret;
1222         }
1223
1224         return -EINVAL;
1225 }
1226
1227 /*
1228  * Map an iov into an array of pages and offset/length tupples. With the
1229  * partial_page structure, we can map several non-contiguous ranges into
1230  * our ones pages[] map instead of splitting that operation into pieces.
1231  * Could easily be exported as a generic helper for other users, in which
1232  * case one would probably want to add a 'max_nr_pages' parameter as well.
1233  */
1234 static int get_iovec_page_array(const struct iovec __user *iov,
1235                                 unsigned int nr_vecs, struct page **pages,
1236                                 struct partial_page *partial, int aligned)
1237 {
1238         int buffers = 0, error = 0;
1239
1240         /*
1241          * It's ok to take the mmap_sem for reading, even
1242          * across a "get_user()".
1243          */
1244         down_read(&current->mm->mmap_sem);
1245
1246         while (nr_vecs) {
1247                 unsigned long off, npages;
1248                 void __user *base;
1249                 size_t len;
1250                 int i;
1251
1252                 /*
1253                  * Get user address base and length for this iovec.
1254                  */
1255                 error = get_user(base, &iov->iov_base);
1256                 if (unlikely(error))
1257                         break;
1258                 error = get_user(len, &iov->iov_len);
1259                 if (unlikely(error))
1260                         break;
1261
1262                 /*
1263                  * Sanity check this iovec. 0 read succeeds.
1264                  */
1265                 if (unlikely(!len))
1266                         break;
1267                 error = -EFAULT;
1268                 if (unlikely(!base))
1269                         break;
1270
1271                 /*
1272                  * Get this base offset and number of pages, then map
1273                  * in the user pages.
1274                  */
1275                 off = (unsigned long) base & ~PAGE_MASK;
1276
1277                 /*
1278                  * If asked for alignment, the offset must be zero and the
1279                  * length a multiple of the PAGE_SIZE.
1280                  */
1281                 error = -EINVAL;
1282                 if (aligned && (off || len & ~PAGE_MASK))
1283                         break;
1284
1285                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1286                 if (npages > PIPE_BUFFERS - buffers)
1287                         npages = PIPE_BUFFERS - buffers;
1288
1289                 error = get_user_pages(current, current->mm,
1290                                        (unsigned long) base, npages, 0, 0,
1291                                        &pages[buffers], NULL);
1292
1293                 if (unlikely(error <= 0))
1294                         break;
1295
1296                 /*
1297                  * Fill this contiguous range into the partial page map.
1298                  */
1299                 for (i = 0; i < error; i++) {
1300                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1301
1302                         partial[buffers].offset = off;
1303                         partial[buffers].len = plen;
1304
1305                         off = 0;
1306                         len -= plen;
1307                         buffers++;
1308                 }
1309
1310                 /*
1311                  * We didn't complete this iov, stop here since it probably
1312                  * means we have to move some of this into a pipe to
1313                  * be able to continue.
1314                  */
1315                 if (len)
1316                         break;
1317
1318                 /*
1319                  * Don't continue if we mapped fewer pages than we asked for,
1320                  * or if we mapped the max number of pages that we have
1321                  * room for.
1322                  */
1323                 if (error < npages || buffers == PIPE_BUFFERS)
1324                         break;
1325
1326                 nr_vecs--;
1327                 iov++;
1328         }
1329
1330         up_read(&current->mm->mmap_sem);
1331
1332         if (buffers)
1333                 return buffers;
1334
1335         return error;
1336 }
1337
1338 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1339                         struct splice_desc *sd)
1340 {
1341         char *src;
1342         int ret;
1343
1344         ret = buf->ops->pin(pipe, buf);
1345         if (unlikely(ret))
1346                 return ret;
1347
1348         /*
1349          * See if we can use the atomic maps, by prefaulting in the
1350          * pages and doing an atomic copy
1351          */
1352         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1353                 src = buf->ops->map(pipe, buf, 1);
1354                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1355                                                         sd->len);
1356                 buf->ops->unmap(pipe, buf, src);
1357                 if (!ret) {
1358                         ret = sd->len;
1359                         goto out;
1360                 }
1361         }
1362
1363         /*
1364          * No dice, use slow non-atomic map and copy
1365          */
1366         src = buf->ops->map(pipe, buf, 0);
1367
1368         ret = sd->len;
1369         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1370                 ret = -EFAULT;
1371
1372 out:
1373         if (ret > 0)
1374                 sd->u.userptr += ret;
1375         buf->ops->unmap(pipe, buf, src);
1376         return ret;
1377 }
1378
1379 /*
1380  * For lack of a better implementation, implement vmsplice() to userspace
1381  * as a simple copy of the pipes pages to the user iov.
1382  */
1383 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1384                              unsigned long nr_segs, unsigned int flags)
1385 {
1386         struct pipe_inode_info *pipe;
1387         struct splice_desc sd;
1388         ssize_t size;
1389         int error;
1390         long ret;
1391
1392         pipe = pipe_info(file->f_path.dentry->d_inode);
1393         if (!pipe)
1394                 return -EBADF;
1395
1396         if (pipe->inode)
1397                 mutex_lock(&pipe->inode->i_mutex);
1398
1399         error = ret = 0;
1400         while (nr_segs) {
1401                 void __user *base;
1402                 size_t len;
1403
1404                 /*
1405                  * Get user address base and length for this iovec.
1406                  */
1407                 error = get_user(base, &iov->iov_base);
1408                 if (unlikely(error))
1409                         break;
1410                 error = get_user(len, &iov->iov_len);
1411                 if (unlikely(error))
1412                         break;
1413
1414                 /*
1415                  * Sanity check this iovec. 0 read succeeds.
1416                  */
1417                 if (unlikely(!len))
1418                         break;
1419                 if (unlikely(!base)) {
1420                         error = -EFAULT;
1421                         break;
1422                 }
1423
1424                 sd.len = 0;
1425                 sd.total_len = len;
1426                 sd.flags = flags;
1427                 sd.u.userptr = base;
1428                 sd.pos = 0;
1429
1430                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1431                 if (size < 0) {
1432                         if (!ret)
1433                                 ret = size;
1434
1435                         break;
1436                 }
1437
1438                 ret += size;
1439
1440                 if (size < len)
1441                         break;
1442
1443                 nr_segs--;
1444                 iov++;
1445         }
1446
1447         if (pipe->inode)
1448                 mutex_unlock(&pipe->inode->i_mutex);
1449
1450         if (!ret)
1451                 ret = error;
1452
1453         return ret;
1454 }
1455
1456 /*
1457  * vmsplice splices a user address range into a pipe. It can be thought of
1458  * as splice-from-memory, where the regular splice is splice-from-file (or
1459  * to file). In both cases the output is a pipe, naturally.
1460  */
1461 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1462                              unsigned long nr_segs, unsigned int flags)
1463 {
1464         struct pipe_inode_info *pipe;
1465         struct page *pages[PIPE_BUFFERS];
1466         struct partial_page partial[PIPE_BUFFERS];
1467         struct splice_pipe_desc spd = {
1468                 .pages = pages,
1469                 .partial = partial,
1470                 .flags = flags,
1471                 .ops = &user_page_pipe_buf_ops,
1472         };
1473
1474         pipe = pipe_info(file->f_path.dentry->d_inode);
1475         if (!pipe)
1476                 return -EBADF;
1477
1478         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1479                                             flags & SPLICE_F_GIFT);
1480         if (spd.nr_pages <= 0)
1481                 return spd.nr_pages;
1482
1483         return splice_to_pipe(pipe, &spd);
1484 }
1485
1486 /*
1487  * Note that vmsplice only really supports true splicing _from_ user memory
1488  * to a pipe, not the other way around. Splicing from user memory is a simple
1489  * operation that can be supported without any funky alignment restrictions
1490  * or nasty vm tricks. We simply map in the user memory and fill them into
1491  * a pipe. The reverse isn't quite as easy, though. There are two possible
1492  * solutions for that:
1493  *
1494  *      - memcpy() the data internally, at which point we might as well just
1495  *        do a regular read() on the buffer anyway.
1496  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1497  *        has restriction limitations on both ends of the pipe).
1498  *
1499  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1500  *
1501  */
1502 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1503                              unsigned long nr_segs, unsigned int flags)
1504 {
1505         struct file *file;
1506         long error;
1507         int fput;
1508
1509         if (unlikely(nr_segs > UIO_MAXIOV))
1510                 return -EINVAL;
1511         else if (unlikely(!nr_segs))
1512                 return 0;
1513
1514         error = -EBADF;
1515         file = fget_light(fd, &fput);
1516         if (file) {
1517                 if (file->f_mode & FMODE_WRITE)
1518                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1519                 else if (file->f_mode & FMODE_READ)
1520                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1521
1522                 fput_light(file, fput);
1523         }
1524
1525         return error;
1526 }
1527
1528 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1529                            int fd_out, loff_t __user *off_out,
1530                            size_t len, unsigned int flags)
1531 {
1532         long error;
1533         struct file *in, *out;
1534         int fput_in, fput_out;
1535
1536         if (unlikely(!len))
1537                 return 0;
1538
1539         error = -EBADF;
1540         in = fget_light(fd_in, &fput_in);
1541         if (in) {
1542                 if (in->f_mode & FMODE_READ) {
1543                         out = fget_light(fd_out, &fput_out);
1544                         if (out) {
1545                                 if (out->f_mode & FMODE_WRITE)
1546                                         error = do_splice(in, off_in,
1547                                                           out, off_out,
1548                                                           len, flags);
1549                                 fput_light(out, fput_out);
1550                         }
1551                 }
1552
1553                 fput_light(in, fput_in);
1554         }
1555
1556         return error;
1557 }
1558
1559 /*
1560  * Make sure there's data to read. Wait for input if we can, otherwise
1561  * return an appropriate error.
1562  */
1563 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1564 {
1565         int ret;
1566
1567         /*
1568          * Check ->nrbufs without the inode lock first. This function
1569          * is speculative anyways, so missing one is ok.
1570          */
1571         if (pipe->nrbufs)
1572                 return 0;
1573
1574         ret = 0;
1575         mutex_lock(&pipe->inode->i_mutex);
1576
1577         while (!pipe->nrbufs) {
1578                 if (signal_pending(current)) {
1579                         ret = -ERESTARTSYS;
1580                         break;
1581                 }
1582                 if (!pipe->writers)
1583                         break;
1584                 if (!pipe->waiting_writers) {
1585                         if (flags & SPLICE_F_NONBLOCK) {
1586                                 ret = -EAGAIN;
1587                                 break;
1588                         }
1589                 }
1590                 pipe_wait(pipe);
1591         }
1592
1593         mutex_unlock(&pipe->inode->i_mutex);
1594         return ret;
1595 }
1596
1597 /*
1598  * Make sure there's writeable room. Wait for room if we can, otherwise
1599  * return an appropriate error.
1600  */
1601 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1602 {
1603         int ret;
1604
1605         /*
1606          * Check ->nrbufs without the inode lock first. This function
1607          * is speculative anyways, so missing one is ok.
1608          */
1609         if (pipe->nrbufs < PIPE_BUFFERS)
1610                 return 0;
1611
1612         ret = 0;
1613         mutex_lock(&pipe->inode->i_mutex);
1614
1615         while (pipe->nrbufs >= PIPE_BUFFERS) {
1616                 if (!pipe->readers) {
1617                         send_sig(SIGPIPE, current, 0);
1618                         ret = -EPIPE;
1619                         break;
1620                 }
1621                 if (flags & SPLICE_F_NONBLOCK) {
1622                         ret = -EAGAIN;
1623                         break;
1624                 }
1625                 if (signal_pending(current)) {
1626                         ret = -ERESTARTSYS;
1627                         break;
1628                 }
1629                 pipe->waiting_writers++;
1630                 pipe_wait(pipe);
1631                 pipe->waiting_writers--;
1632         }
1633
1634         mutex_unlock(&pipe->inode->i_mutex);
1635         return ret;
1636 }
1637
1638 /*
1639  * Link contents of ipipe to opipe.
1640  */
1641 static int link_pipe(struct pipe_inode_info *ipipe,
1642                      struct pipe_inode_info *opipe,
1643                      size_t len, unsigned int flags)
1644 {
1645         struct pipe_buffer *ibuf, *obuf;
1646         int ret = 0, i = 0, nbuf;
1647
1648         /*
1649          * Potential ABBA deadlock, work around it by ordering lock
1650          * grabbing by inode address. Otherwise two different processes
1651          * could deadlock (one doing tee from A -> B, the other from B -> A).
1652          */
1653         inode_double_lock(ipipe->inode, opipe->inode);
1654
1655         do {
1656                 if (!opipe->readers) {
1657                         send_sig(SIGPIPE, current, 0);
1658                         if (!ret)
1659                                 ret = -EPIPE;
1660                         break;
1661                 }
1662
1663                 /*
1664                  * If we have iterated all input buffers or ran out of
1665                  * output room, break.
1666                  */
1667                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1668                         break;
1669
1670                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1671                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1672
1673                 /*
1674                  * Get a reference to this pipe buffer,
1675                  * so we can copy the contents over.
1676                  */
1677                 ibuf->ops->get(ipipe, ibuf);
1678
1679                 obuf = opipe->bufs + nbuf;
1680                 *obuf = *ibuf;
1681
1682                 /*
1683                  * Don't inherit the gift flag, we need to
1684                  * prevent multiple steals of this page.
1685                  */
1686                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1687
1688                 if (obuf->len > len)
1689                         obuf->len = len;
1690
1691                 opipe->nrbufs++;
1692                 ret += obuf->len;
1693                 len -= obuf->len;
1694                 i++;
1695         } while (len);
1696
1697         inode_double_unlock(ipipe->inode, opipe->inode);
1698
1699         /*
1700          * If we put data in the output pipe, wakeup any potential readers.
1701          */
1702         if (ret > 0) {
1703                 smp_mb();
1704                 if (waitqueue_active(&opipe->wait))
1705                         wake_up_interruptible(&opipe->wait);
1706                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1707         }
1708
1709         return ret;
1710 }
1711
1712 /*
1713  * This is a tee(1) implementation that works on pipes. It doesn't copy
1714  * any data, it simply references the 'in' pages on the 'out' pipe.
1715  * The 'flags' used are the SPLICE_F_* variants, currently the only
1716  * applicable one is SPLICE_F_NONBLOCK.
1717  */
1718 static long do_tee(struct file *in, struct file *out, size_t len,
1719                    unsigned int flags)
1720 {
1721         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1722         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1723         int ret = -EINVAL;
1724
1725         /*
1726          * Duplicate the contents of ipipe to opipe without actually
1727          * copying the data.
1728          */
1729         if (ipipe && opipe && ipipe != opipe) {
1730                 /*
1731                  * Keep going, unless we encounter an error. The ipipe/opipe
1732                  * ordering doesn't really matter.
1733                  */
1734                 ret = link_ipipe_prep(ipipe, flags);
1735                 if (!ret) {
1736                         ret = link_opipe_prep(opipe, flags);
1737                         if (!ret) {
1738                                 ret = link_pipe(ipipe, opipe, len, flags);
1739                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1740                                         ret = -EAGAIN;
1741                         }
1742                 }
1743         }
1744
1745         return ret;
1746 }
1747
1748 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1749 {
1750         struct file *in;
1751         int error, fput_in;
1752
1753         if (unlikely(!len))
1754                 return 0;
1755
1756         error = -EBADF;
1757         in = fget_light(fdin, &fput_in);
1758         if (in) {
1759                 if (in->f_mode & FMODE_READ) {
1760                         int fput_out;
1761                         struct file *out = fget_light(fdout, &fput_out);
1762
1763                         if (out) {
1764                                 if (out->f_mode & FMODE_WRITE)
1765                                         error = do_tee(in, out, len, flags);
1766                                 fput_light(out, fput_out);
1767                         }
1768                 }
1769                 fput_light(in, fput_in);
1770         }
1771
1772         return error;
1773 }