pipe: add documentation and comments
[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 /*
89  * Check whether the contents of buf is OK to access. Since the content
90  * is a page cache page, IO may be in flight.
91  */
92 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
93                                        struct pipe_buffer *buf)
94 {
95         struct page *page = buf->page;
96         int err;
97
98         if (!PageUptodate(page)) {
99                 lock_page(page);
100
101                 /*
102                  * Page got truncated/unhashed. This will cause a 0-byte
103                  * splice, if this is the first page.
104                  */
105                 if (!page->mapping) {
106                         err = -ENODATA;
107                         goto error;
108                 }
109
110                 /*
111                  * Uh oh, read-error from disk.
112                  */
113                 if (!PageUptodate(page)) {
114                         err = -EIO;
115                         goto error;
116                 }
117
118                 /*
119                  * Page is ok afterall, we are done.
120                  */
121                 unlock_page(page);
122         }
123
124         return 0;
125 error:
126         unlock_page(page);
127         return err;
128 }
129
130 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
131         .can_merge = 0,
132         .map = generic_pipe_buf_map,
133         .unmap = generic_pipe_buf_unmap,
134         .confirm = page_cache_pipe_buf_confirm,
135         .release = page_cache_pipe_buf_release,
136         .steal = page_cache_pipe_buf_steal,
137         .get = generic_pipe_buf_get,
138 };
139
140 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
141                                     struct pipe_buffer *buf)
142 {
143         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
144                 return 1;
145
146         buf->flags |= PIPE_BUF_FLAG_LRU;
147         return generic_pipe_buf_steal(pipe, buf);
148 }
149
150 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
151         .can_merge = 0,
152         .map = generic_pipe_buf_map,
153         .unmap = generic_pipe_buf_unmap,
154         .confirm = generic_pipe_buf_confirm,
155         .release = page_cache_pipe_buf_release,
156         .steal = user_page_pipe_buf_steal,
157         .get = generic_pipe_buf_get,
158 };
159
160 /**
161  * splice_to_pipe - fill passed data into a pipe
162  * @pipe:       pipe to fill
163  * @spd:        data to fill
164  *
165  * Description:
166  *    @spd contains a map of pages and len/offset tupples, a long with
167  *    the struct pipe_buf_operations associated with these pages. This
168  *    function will link that data to the pipe.
169  *
170  */
171 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
172                        struct splice_pipe_desc *spd)
173 {
174         unsigned int spd_pages = spd->nr_pages;
175         int ret, do_wakeup, page_nr;
176
177         ret = 0;
178         do_wakeup = 0;
179         page_nr = 0;
180
181         if (pipe->inode)
182                 mutex_lock(&pipe->inode->i_mutex);
183
184         for (;;) {
185                 if (!pipe->readers) {
186                         send_sig(SIGPIPE, current, 0);
187                         if (!ret)
188                                 ret = -EPIPE;
189                         break;
190                 }
191
192                 if (pipe->nrbufs < PIPE_BUFFERS) {
193                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
194                         struct pipe_buffer *buf = pipe->bufs + newbuf;
195
196                         buf->page = spd->pages[page_nr];
197                         buf->offset = spd->partial[page_nr].offset;
198                         buf->len = spd->partial[page_nr].len;
199                         buf->private = spd->partial[page_nr].private;
200                         buf->ops = spd->ops;
201                         if (spd->flags & SPLICE_F_GIFT)
202                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
203
204                         pipe->nrbufs++;
205                         page_nr++;
206                         ret += buf->len;
207
208                         if (pipe->inode)
209                                 do_wakeup = 1;
210
211                         if (!--spd->nr_pages)
212                                 break;
213                         if (pipe->nrbufs < PIPE_BUFFERS)
214                                 continue;
215
216                         break;
217                 }
218
219                 if (spd->flags & SPLICE_F_NONBLOCK) {
220                         if (!ret)
221                                 ret = -EAGAIN;
222                         break;
223                 }
224
225                 if (signal_pending(current)) {
226                         if (!ret)
227                                 ret = -ERESTARTSYS;
228                         break;
229                 }
230
231                 if (do_wakeup) {
232                         smp_mb();
233                         if (waitqueue_active(&pipe->wait))
234                                 wake_up_interruptible_sync(&pipe->wait);
235                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
236                         do_wakeup = 0;
237                 }
238
239                 pipe->waiting_writers++;
240                 pipe_wait(pipe);
241                 pipe->waiting_writers--;
242         }
243
244         if (pipe->inode) {
245                 mutex_unlock(&pipe->inode->i_mutex);
246
247                 if (do_wakeup) {
248                         smp_mb();
249                         if (waitqueue_active(&pipe->wait))
250                                 wake_up_interruptible(&pipe->wait);
251                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
252                 }
253         }
254
255         while (page_nr < spd_pages)
256                 page_cache_release(spd->pages[page_nr++]);
257
258         return ret;
259 }
260
261 static int
262 __generic_file_splice_read(struct file *in, loff_t *ppos,
263                            struct pipe_inode_info *pipe, size_t len,
264                            unsigned int flags)
265 {
266         struct address_space *mapping = in->f_mapping;
267         unsigned int loff, nr_pages;
268         struct page *pages[PIPE_BUFFERS];
269         struct partial_page partial[PIPE_BUFFERS];
270         struct page *page;
271         pgoff_t index, end_index;
272         loff_t isize;
273         int error, page_nr;
274         struct splice_pipe_desc spd = {
275                 .pages = pages,
276                 .partial = partial,
277                 .flags = flags,
278                 .ops = &page_cache_pipe_buf_ops,
279         };
280
281         index = *ppos >> PAGE_CACHE_SHIFT;
282         loff = *ppos & ~PAGE_CACHE_MASK;
283         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
284
285         if (nr_pages > PIPE_BUFFERS)
286                 nr_pages = PIPE_BUFFERS;
287
288         /*
289          * Don't try to 2nd guess the read-ahead logic, call into
290          * page_cache_readahead() like the page cache reads would do.
291          */
292         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
293
294         /*
295          * Lookup the (hopefully) full range of pages we need.
296          */
297         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
298
299         /*
300          * If find_get_pages_contig() returned fewer pages than we needed,
301          * allocate the rest and fill in the holes.
302          */
303         error = 0;
304         index += spd.nr_pages;
305         while (spd.nr_pages < nr_pages) {
306                 /*
307                  * Page could be there, find_get_pages_contig() breaks on
308                  * the first hole.
309                  */
310                 page = find_get_page(mapping, index);
311                 if (!page) {
312                         /*
313                          * Make sure the read-ahead engine is notified
314                          * about this failure.
315                          */
316                         handle_ra_miss(mapping, &in->f_ra, index);
317
318                         /*
319                          * page didn't exist, allocate one.
320                          */
321                         page = page_cache_alloc_cold(mapping);
322                         if (!page)
323                                 break;
324
325                         error = add_to_page_cache_lru(page, mapping, index,
326                                               GFP_KERNEL);
327                         if (unlikely(error)) {
328                                 page_cache_release(page);
329                                 if (error == -EEXIST)
330                                         continue;
331                                 break;
332                         }
333                         /*
334                          * add_to_page_cache() locks the page, unlock it
335                          * to avoid convoluting the logic below even more.
336                          */
337                         unlock_page(page);
338                 }
339
340                 pages[spd.nr_pages++] = page;
341                 index++;
342         }
343
344         /*
345          * Now loop over the map and see if we need to start IO on any
346          * pages, fill in the partial map, etc.
347          */
348         index = *ppos >> PAGE_CACHE_SHIFT;
349         nr_pages = spd.nr_pages;
350         spd.nr_pages = 0;
351         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
352                 unsigned int this_len;
353
354                 if (!len)
355                         break;
356
357                 /*
358                  * this_len is the max we'll use from this page
359                  */
360                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
361                 page = pages[page_nr];
362
363                 /*
364                  * If the page isn't uptodate, we may need to start io on it
365                  */
366                 if (!PageUptodate(page)) {
367                         /*
368                          * If in nonblock mode then dont block on waiting
369                          * for an in-flight io page
370                          */
371                         if (flags & SPLICE_F_NONBLOCK) {
372                                 if (TestSetPageLocked(page))
373                                         break;
374                         } else
375                                 lock_page(page);
376
377                         /*
378                          * page was truncated, stop here. if this isn't the
379                          * first page, we'll just complete what we already
380                          * added
381                          */
382                         if (!page->mapping) {
383                                 unlock_page(page);
384                                 break;
385                         }
386                         /*
387                          * page was already under io and is now done, great
388                          */
389                         if (PageUptodate(page)) {
390                                 unlock_page(page);
391                                 goto fill_it;
392                         }
393
394                         /*
395                          * need to read in the page
396                          */
397                         error = mapping->a_ops->readpage(in, page);
398                         if (unlikely(error)) {
399                                 /*
400                                  * We really should re-lookup the page here,
401                                  * but it complicates things a lot. Instead
402                                  * lets just do what we already stored, and
403                                  * we'll get it the next time we are called.
404                                  */
405                                 if (error == AOP_TRUNCATED_PAGE)
406                                         error = 0;
407
408                                 break;
409                         }
410                 }
411 fill_it:
412                 /*
413                  * i_size must be checked after PageUptodate.
414                  */
415                 isize = i_size_read(mapping->host);
416                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
417                 if (unlikely(!isize || index > end_index))
418                         break;
419
420                 /*
421                  * if this is the last page, see if we need to shrink
422                  * the length and stop
423                  */
424                 if (end_index == index) {
425                         unsigned int plen;
426
427                         /*
428                          * max good bytes in this page
429                          */
430                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
431                         if (plen <= loff)
432                                 break;
433
434                         /*
435                          * force quit after adding this page
436                          */
437                         this_len = min(this_len, plen - loff);
438                         len = this_len;
439                 }
440
441                 partial[page_nr].offset = loff;
442                 partial[page_nr].len = this_len;
443                 len -= this_len;
444                 loff = 0;
445                 spd.nr_pages++;
446                 index++;
447         }
448
449         /*
450          * Release any pages at the end, if we quit early. 'page_nr' is how far
451          * we got, 'nr_pages' is how many pages are in the map.
452          */
453         while (page_nr < nr_pages)
454                 page_cache_release(pages[page_nr++]);
455
456         if (spd.nr_pages)
457                 return splice_to_pipe(pipe, &spd);
458
459         return error;
460 }
461
462 /**
463  * generic_file_splice_read - splice data from file to a pipe
464  * @in:         file to splice from
465  * @ppos:       position in @in
466  * @pipe:       pipe to splice to
467  * @len:        number of bytes to splice
468  * @flags:      splice modifier flags
469  *
470  * Description:
471  *    Will read pages from given file and fill them into a pipe. Can be
472  *    used as long as the address_space operations for the source implements
473  *    a readpage() hook.
474  *
475  */
476 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
477                                  struct pipe_inode_info *pipe, size_t len,
478                                  unsigned int flags)
479 {
480         ssize_t spliced;
481         int ret;
482         loff_t isize, left;
483
484         isize = i_size_read(in->f_mapping->host);
485         if (unlikely(*ppos >= isize))
486                 return 0;
487
488         left = isize - *ppos;
489         if (unlikely(left < len))
490                 len = left;
491
492         ret = 0;
493         spliced = 0;
494         while (len) {
495                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
496
497                 if (ret < 0)
498                         break;
499                 else if (!ret) {
500                         if (spliced)
501                                 break;
502                         if (flags & SPLICE_F_NONBLOCK) {
503                                 ret = -EAGAIN;
504                                 break;
505                         }
506                 }
507
508                 *ppos += ret;
509                 len -= ret;
510                 spliced += ret;
511         }
512
513         if (spliced)
514                 return spliced;
515
516         return ret;
517 }
518
519 EXPORT_SYMBOL(generic_file_splice_read);
520
521 /*
522  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
523  * using sendpage(). Return the number of bytes sent.
524  */
525 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
526                             struct pipe_buffer *buf, struct splice_desc *sd)
527 {
528         struct file *file = sd->u.file;
529         loff_t pos = sd->pos;
530         int ret, more;
531
532         ret = buf->ops->confirm(pipe, buf);
533         if (!ret) {
534                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
535
536                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
537                                            sd->len, &pos, more);
538         }
539
540         return ret;
541 }
542
543 /*
544  * This is a little more tricky than the file -> pipe splicing. There are
545  * basically three cases:
546  *
547  *      - Destination page already exists in the address space and there
548  *        are users of it. For that case we have no other option that
549  *        copying the data. Tough luck.
550  *      - Destination page already exists in the address space, but there
551  *        are no users of it. Make sure it's uptodate, then drop it. Fall
552  *        through to last case.
553  *      - Destination page does not exist, we can add the pipe page to
554  *        the page cache and avoid the copy.
555  *
556  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
557  * sd->flags), we attempt to migrate pages from the pipe to the output
558  * file address space page cache. This is possible if no one else has
559  * the pipe page referenced outside of the pipe and page cache. If
560  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
561  * a new page in the output file page cache and fill/dirty that.
562  */
563 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
564                         struct splice_desc *sd)
565 {
566         struct file *file = sd->u.file;
567         struct address_space *mapping = file->f_mapping;
568         unsigned int offset, this_len;
569         struct page *page;
570         pgoff_t index;
571         int ret;
572
573         /*
574          * make sure the data in this buffer is uptodate
575          */
576         ret = buf->ops->confirm(pipe, buf);
577         if (unlikely(ret))
578                 return ret;
579
580         index = sd->pos >> PAGE_CACHE_SHIFT;
581         offset = sd->pos & ~PAGE_CACHE_MASK;
582
583         this_len = sd->len;
584         if (this_len + offset > PAGE_CACHE_SIZE)
585                 this_len = PAGE_CACHE_SIZE - offset;
586
587 find_page:
588         page = find_lock_page(mapping, index);
589         if (!page) {
590                 ret = -ENOMEM;
591                 page = page_cache_alloc_cold(mapping);
592                 if (unlikely(!page))
593                         goto out_ret;
594
595                 /*
596                  * This will also lock the page
597                  */
598                 ret = add_to_page_cache_lru(page, mapping, index,
599                                             GFP_KERNEL);
600                 if (unlikely(ret))
601                         goto out;
602         }
603
604         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
605         if (unlikely(ret)) {
606                 loff_t isize = i_size_read(mapping->host);
607
608                 if (ret != AOP_TRUNCATED_PAGE)
609                         unlock_page(page);
610                 page_cache_release(page);
611                 if (ret == AOP_TRUNCATED_PAGE)
612                         goto find_page;
613
614                 /*
615                  * prepare_write() may have instantiated a few blocks
616                  * outside i_size.  Trim these off again.
617                  */
618                 if (sd->pos + this_len > isize)
619                         vmtruncate(mapping->host, isize);
620
621                 goto out_ret;
622         }
623
624         if (buf->page != page) {
625                 /*
626                  * Careful, ->map() uses KM_USER0!
627                  */
628                 char *src = buf->ops->map(pipe, buf, 1);
629                 char *dst = kmap_atomic(page, KM_USER1);
630
631                 memcpy(dst + offset, src + buf->offset, this_len);
632                 flush_dcache_page(page);
633                 kunmap_atomic(dst, KM_USER1);
634                 buf->ops->unmap(pipe, buf, src);
635         }
636
637         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
638         if (ret) {
639                 if (ret == AOP_TRUNCATED_PAGE) {
640                         page_cache_release(page);
641                         goto find_page;
642                 }
643                 if (ret < 0)
644                         goto out;
645                 /*
646                  * Partial write has happened, so 'ret' already initialized by
647                  * number of bytes written, Where is nothing we have to do here.
648                  */
649         } else
650                 ret = this_len;
651         /*
652          * Return the number of bytes written and mark page as
653          * accessed, we are now done!
654          */
655         mark_page_accessed(page);
656 out:
657         page_cache_release(page);
658         unlock_page(page);
659 out_ret:
660         return ret;
661 }
662
663 /**
664  * __splice_from_pipe - splice data from a pipe to given actor
665  * @pipe:       pipe to splice from
666  * @sd:         information to @actor
667  * @actor:      handler that splices the data
668  *
669  * Description:
670  *    This function does little more than loop over the pipe and call
671  *    @actor to do the actual moving of a single struct pipe_buffer to
672  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
673  *    pipe_to_user.
674  *
675  */
676 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
677                            splice_actor *actor)
678 {
679         int ret, do_wakeup, err;
680
681         ret = 0;
682         do_wakeup = 0;
683
684         for (;;) {
685                 if (pipe->nrbufs) {
686                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
687                         const struct pipe_buf_operations *ops = buf->ops;
688
689                         sd->len = buf->len;
690                         if (sd->len > sd->total_len)
691                                 sd->len = sd->total_len;
692
693                         err = actor(pipe, buf, sd);
694                         if (err <= 0) {
695                                 if (!ret && err != -ENODATA)
696                                         ret = err;
697
698                                 break;
699                         }
700
701                         ret += err;
702                         buf->offset += err;
703                         buf->len -= err;
704
705                         sd->len -= err;
706                         sd->pos += err;
707                         sd->total_len -= err;
708                         if (sd->len)
709                                 continue;
710
711                         if (!buf->len) {
712                                 buf->ops = NULL;
713                                 ops->release(pipe, buf);
714                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
715                                 pipe->nrbufs--;
716                                 if (pipe->inode)
717                                         do_wakeup = 1;
718                         }
719
720                         if (!sd->total_len)
721                                 break;
722                 }
723
724                 if (pipe->nrbufs)
725                         continue;
726                 if (!pipe->writers)
727                         break;
728                 if (!pipe->waiting_writers) {
729                         if (ret)
730                                 break;
731                 }
732
733                 if (sd->flags & SPLICE_F_NONBLOCK) {
734                         if (!ret)
735                                 ret = -EAGAIN;
736                         break;
737                 }
738
739                 if (signal_pending(current)) {
740                         if (!ret)
741                                 ret = -ERESTARTSYS;
742                         break;
743                 }
744
745                 if (do_wakeup) {
746                         smp_mb();
747                         if (waitqueue_active(&pipe->wait))
748                                 wake_up_interruptible_sync(&pipe->wait);
749                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
750                         do_wakeup = 0;
751                 }
752
753                 pipe_wait(pipe);
754         }
755
756         if (do_wakeup) {
757                 smp_mb();
758                 if (waitqueue_active(&pipe->wait))
759                         wake_up_interruptible(&pipe->wait);
760                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
761         }
762
763         return ret;
764 }
765 EXPORT_SYMBOL(__splice_from_pipe);
766
767 /**
768  * splice_from_pipe - splice data from a pipe to a file
769  * @pipe:       pipe to splice from
770  * @out:        file to splice to
771  * @ppos:       position in @out
772  * @len:        how many bytes to splice
773  * @flags:      splice modifier flags
774  * @actor:      handler that splices the data
775  *
776  * Description:
777  *    See __splice_from_pipe. This function locks the input and output inodes,
778  *    otherwise it's identical to __splice_from_pipe().
779  *
780  */
781 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
782                          loff_t *ppos, size_t len, unsigned int flags,
783                          splice_actor *actor)
784 {
785         ssize_t ret;
786         struct inode *inode = out->f_mapping->host;
787         struct splice_desc sd = {
788                 .total_len = len,
789                 .flags = flags,
790                 .pos = *ppos,
791                 .u.file = out,
792         };
793
794         /*
795          * The actor worker might be calling ->prepare_write and
796          * ->commit_write. Most of the time, these expect i_mutex to
797          * be held. Since this may result in an ABBA deadlock with
798          * pipe->inode, we have to order lock acquiry here.
799          */
800         inode_double_lock(inode, pipe->inode);
801         ret = __splice_from_pipe(pipe, &sd, actor);
802         inode_double_unlock(inode, pipe->inode);
803
804         return ret;
805 }
806
807 /**
808  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
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. The caller is responsible
818  *    for acquiring i_mutex on both inodes.
819  *
820  */
821 ssize_t
822 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
823                                  loff_t *ppos, size_t len, unsigned int flags)
824 {
825         struct address_space *mapping = out->f_mapping;
826         struct inode *inode = mapping->host;
827         struct splice_desc sd = {
828                 .total_len = len,
829                 .flags = flags,
830                 .pos = *ppos,
831                 .u.file = out,
832         };
833         ssize_t ret;
834         int err;
835
836         err = remove_suid(out->f_path.dentry);
837         if (unlikely(err))
838                 return err;
839
840         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
841         if (ret > 0) {
842                 unsigned long nr_pages;
843
844                 *ppos += ret;
845                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
846
847                 /*
848                  * If file or inode is SYNC and we actually wrote some data,
849                  * sync it.
850                  */
851                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
852                         err = generic_osync_inode(inode, mapping,
853                                                   OSYNC_METADATA|OSYNC_DATA);
854
855                         if (err)
856                                 ret = err;
857                 }
858                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
859         }
860
861         return ret;
862 }
863
864 EXPORT_SYMBOL(generic_file_splice_write_nolock);
865
866 /**
867  * generic_file_splice_write - splice data from a pipe to a file
868  * @pipe:       pipe info
869  * @out:        file to write to
870  * @ppos:       position in @out
871  * @len:        number of bytes to splice
872  * @flags:      splice modifier flags
873  *
874  * Description:
875  *    Will either move or copy pages (determined by @flags options) from
876  *    the given pipe inode to the given file.
877  *
878  */
879 ssize_t
880 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
881                           loff_t *ppos, size_t len, unsigned int flags)
882 {
883         struct address_space *mapping = out->f_mapping;
884         struct inode *inode = mapping->host;
885         ssize_t ret;
886         int err;
887
888         err = should_remove_suid(out->f_path.dentry);
889         if (unlikely(err)) {
890                 mutex_lock(&inode->i_mutex);
891                 err = __remove_suid(out->f_path.dentry, err);
892                 mutex_unlock(&inode->i_mutex);
893                 if (err)
894                         return err;
895         }
896
897         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
898         if (ret > 0) {
899                 unsigned long nr_pages;
900
901                 *ppos += ret;
902                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
903
904                 /*
905                  * If file or inode is SYNC and we actually wrote some data,
906                  * sync it.
907                  */
908                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
909                         mutex_lock(&inode->i_mutex);
910                         err = generic_osync_inode(inode, mapping,
911                                                   OSYNC_METADATA|OSYNC_DATA);
912                         mutex_unlock(&inode->i_mutex);
913
914                         if (err)
915                                 ret = err;
916                 }
917                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
918         }
919
920         return ret;
921 }
922
923 EXPORT_SYMBOL(generic_file_splice_write);
924
925 /**
926  * generic_splice_sendpage - splice data from a pipe to a socket
927  * @pipe:       pipe to splice from
928  * @out:        socket to write to
929  * @ppos:       position in @out
930  * @len:        number of bytes to splice
931  * @flags:      splice modifier flags
932  *
933  * Description:
934  *    Will send @len bytes from the pipe to a network socket. No data copying
935  *    is involved.
936  *
937  */
938 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
939                                 loff_t *ppos, size_t len, unsigned int flags)
940 {
941         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
942 }
943
944 EXPORT_SYMBOL(generic_splice_sendpage);
945
946 /*
947  * Attempt to initiate a splice from pipe to file.
948  */
949 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
950                            loff_t *ppos, size_t len, unsigned int flags)
951 {
952         int ret;
953
954         if (unlikely(!out->f_op || !out->f_op->splice_write))
955                 return -EINVAL;
956
957         if (unlikely(!(out->f_mode & FMODE_WRITE)))
958                 return -EBADF;
959
960         ret = rw_verify_area(WRITE, out, ppos, len);
961         if (unlikely(ret < 0))
962                 return ret;
963
964         return out->f_op->splice_write(pipe, out, ppos, len, flags);
965 }
966
967 /*
968  * Attempt to initiate a splice from a file to a pipe.
969  */
970 static long do_splice_to(struct file *in, loff_t *ppos,
971                          struct pipe_inode_info *pipe, size_t len,
972                          unsigned int flags)
973 {
974         int ret;
975
976         if (unlikely(!in->f_op || !in->f_op->splice_read))
977                 return -EINVAL;
978
979         if (unlikely(!(in->f_mode & FMODE_READ)))
980                 return -EBADF;
981
982         ret = rw_verify_area(READ, in, ppos, len);
983         if (unlikely(ret < 0))
984                 return ret;
985
986         return in->f_op->splice_read(in, ppos, pipe, len, flags);
987 }
988
989 /**
990  * splice_direct_to_actor - splices data directly between two non-pipes
991  * @in:         file to splice from
992  * @sd:         actor information on where to splice to
993  * @actor:      handles the data splicing
994  *
995  * Description:
996  *    This is a special case helper to splice directly between two
997  *    points, without requiring an explicit pipe. Internally an allocated
998  *    pipe is cached in the process, and reused during the life time of
999  *    that process.
1000  *
1001  */
1002 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1003                                splice_direct_actor *actor)
1004 {
1005         struct pipe_inode_info *pipe;
1006         long ret, bytes;
1007         umode_t i_mode;
1008         size_t len;
1009         int i, flags;
1010
1011         /*
1012          * We require the input being a regular file, as we don't want to
1013          * randomly drop data for eg socket -> socket splicing. Use the
1014          * piped splicing for that!
1015          */
1016         i_mode = in->f_path.dentry->d_inode->i_mode;
1017         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1018                 return -EINVAL;
1019
1020         /*
1021          * neither in nor out is a pipe, setup an internal pipe attached to
1022          * 'out' and transfer the wanted data from 'in' to 'out' through that
1023          */
1024         pipe = current->splice_pipe;
1025         if (unlikely(!pipe)) {
1026                 pipe = alloc_pipe_info(NULL);
1027                 if (!pipe)
1028                         return -ENOMEM;
1029
1030                 /*
1031                  * We don't have an immediate reader, but we'll read the stuff
1032                  * out of the pipe right after the splice_to_pipe(). So set
1033                  * PIPE_READERS appropriately.
1034                  */
1035                 pipe->readers = 1;
1036
1037                 current->splice_pipe = pipe;
1038         }
1039
1040         /*
1041          * Do the splice.
1042          */
1043         ret = 0;
1044         bytes = 0;
1045         len = sd->total_len;
1046         flags = sd->flags;
1047
1048         /*
1049          * Don't block on output, we have to drain the direct pipe.
1050          */
1051         sd->flags &= ~SPLICE_F_NONBLOCK;
1052
1053         while (len) {
1054                 size_t read_len, max_read_len;
1055
1056                 /*
1057                  * Do at most PIPE_BUFFERS pages worth of transfer:
1058                  */
1059                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1060
1061                 ret = do_splice_to(in, &sd->pos, pipe, max_read_len, flags);
1062                 if (unlikely(ret < 0))
1063                         goto out_release;
1064
1065                 read_len = ret;
1066                 sd->total_len = read_len;
1067
1068                 /*
1069                  * NOTE: nonblocking mode only applies to the input. We
1070                  * must not do the output in nonblocking mode as then we
1071                  * could get stuck data in the internal pipe:
1072                  */
1073                 ret = actor(pipe, sd);
1074                 if (unlikely(ret < 0))
1075                         goto out_release;
1076
1077                 bytes += ret;
1078                 len -= ret;
1079
1080                 /*
1081                  * In nonblocking mode, if we got back a short read then
1082                  * that was due to either an IO error or due to the
1083                  * pagecache entry not being there. In the IO error case
1084                  * the _next_ splice attempt will produce a clean IO error
1085                  * return value (not a short read), so in both cases it's
1086                  * correct to break out of the loop here:
1087                  */
1088                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1089                         break;
1090         }
1091
1092         pipe->nrbufs = pipe->curbuf = 0;
1093
1094         return bytes;
1095
1096 out_release:
1097         /*
1098          * If we did an incomplete transfer we must release
1099          * the pipe buffers in question:
1100          */
1101         for (i = 0; i < PIPE_BUFFERS; i++) {
1102                 struct pipe_buffer *buf = pipe->bufs + i;
1103
1104                 if (buf->ops) {
1105                         buf->ops->release(pipe, buf);
1106                         buf->ops = NULL;
1107                 }
1108         }
1109         pipe->nrbufs = pipe->curbuf = 0;
1110
1111         /*
1112          * If we transferred some data, return the number of bytes:
1113          */
1114         if (bytes > 0)
1115                 return bytes;
1116
1117         return ret;
1118
1119 }
1120 EXPORT_SYMBOL(splice_direct_to_actor);
1121
1122 static int direct_splice_actor(struct pipe_inode_info *pipe,
1123                                struct splice_desc *sd)
1124 {
1125         struct file *file = sd->u.file;
1126
1127         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1128 }
1129
1130 /**
1131  * do_splice_direct - splices data directly between two files
1132  * @in:         file to splice from
1133  * @ppos:       input file offset
1134  * @out:        file to splice to
1135  * @len:        number of bytes to splice
1136  * @flags:      splice modifier flags
1137  *
1138  * Description:
1139  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1140  *    doing it in the application would incur an extra system call
1141  *    (splice in + splice out, as compared to just sendfile()). So this helper
1142  *    can splice directly through a process-private pipe.
1143  *
1144  */
1145 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1146                       size_t len, unsigned int flags)
1147 {
1148         struct splice_desc sd = {
1149                 .len            = len,
1150                 .total_len      = len,
1151                 .flags          = flags,
1152                 .pos            = *ppos,
1153                 .u.file         = out,
1154         };
1155         size_t ret;
1156
1157         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1158         *ppos = sd.pos;
1159         return ret;
1160 }
1161
1162 /*
1163  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1164  * location, so checking ->i_pipe is not enough to verify that this is a
1165  * pipe.
1166  */
1167 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1168 {
1169         if (S_ISFIFO(inode->i_mode))
1170                 return inode->i_pipe;
1171
1172         return NULL;
1173 }
1174
1175 /*
1176  * Determine where to splice to/from.
1177  */
1178 static long do_splice(struct file *in, loff_t __user *off_in,
1179                       struct file *out, loff_t __user *off_out,
1180                       size_t len, unsigned int flags)
1181 {
1182         struct pipe_inode_info *pipe;
1183         loff_t offset, *off;
1184         long ret;
1185
1186         pipe = pipe_info(in->f_path.dentry->d_inode);
1187         if (pipe) {
1188                 if (off_in)
1189                         return -ESPIPE;
1190                 if (off_out) {
1191                         if (out->f_op->llseek == no_llseek)
1192                                 return -EINVAL;
1193                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1194                                 return -EFAULT;
1195                         off = &offset;
1196                 } else
1197                         off = &out->f_pos;
1198
1199                 ret = do_splice_from(pipe, out, off, len, flags);
1200
1201                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1202                         ret = -EFAULT;
1203
1204                 return ret;
1205         }
1206
1207         pipe = pipe_info(out->f_path.dentry->d_inode);
1208         if (pipe) {
1209                 if (off_out)
1210                         return -ESPIPE;
1211                 if (off_in) {
1212                         if (in->f_op->llseek == no_llseek)
1213                                 return -EINVAL;
1214                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1215                                 return -EFAULT;
1216                         off = &offset;
1217                 } else
1218                         off = &in->f_pos;
1219
1220                 ret = do_splice_to(in, off, pipe, len, flags);
1221
1222                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1223                         ret = -EFAULT;
1224
1225                 return ret;
1226         }
1227
1228         return -EINVAL;
1229 }
1230
1231 /*
1232  * Map an iov into an array of pages and offset/length tupples. With the
1233  * partial_page structure, we can map several non-contiguous ranges into
1234  * our ones pages[] map instead of splitting that operation into pieces.
1235  * Could easily be exported as a generic helper for other users, in which
1236  * case one would probably want to add a 'max_nr_pages' parameter as well.
1237  */
1238 static int get_iovec_page_array(const struct iovec __user *iov,
1239                                 unsigned int nr_vecs, struct page **pages,
1240                                 struct partial_page *partial, int aligned)
1241 {
1242         int buffers = 0, error = 0;
1243
1244         /*
1245          * It's ok to take the mmap_sem for reading, even
1246          * across a "get_user()".
1247          */
1248         down_read(&current->mm->mmap_sem);
1249
1250         while (nr_vecs) {
1251                 unsigned long off, npages;
1252                 void __user *base;
1253                 size_t len;
1254                 int i;
1255
1256                 /*
1257                  * Get user address base and length for this iovec.
1258                  */
1259                 error = get_user(base, &iov->iov_base);
1260                 if (unlikely(error))
1261                         break;
1262                 error = get_user(len, &iov->iov_len);
1263                 if (unlikely(error))
1264                         break;
1265
1266                 /*
1267                  * Sanity check this iovec. 0 read succeeds.
1268                  */
1269                 if (unlikely(!len))
1270                         break;
1271                 error = -EFAULT;
1272                 if (unlikely(!base))
1273                         break;
1274
1275                 /*
1276                  * Get this base offset and number of pages, then map
1277                  * in the user pages.
1278                  */
1279                 off = (unsigned long) base & ~PAGE_MASK;
1280
1281                 /*
1282                  * If asked for alignment, the offset must be zero and the
1283                  * length a multiple of the PAGE_SIZE.
1284                  */
1285                 error = -EINVAL;
1286                 if (aligned && (off || len & ~PAGE_MASK))
1287                         break;
1288
1289                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1290                 if (npages > PIPE_BUFFERS - buffers)
1291                         npages = PIPE_BUFFERS - buffers;
1292
1293                 error = get_user_pages(current, current->mm,
1294                                        (unsigned long) base, npages, 0, 0,
1295                                        &pages[buffers], NULL);
1296
1297                 if (unlikely(error <= 0))
1298                         break;
1299
1300                 /*
1301                  * Fill this contiguous range into the partial page map.
1302                  */
1303                 for (i = 0; i < error; i++) {
1304                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1305
1306                         partial[buffers].offset = off;
1307                         partial[buffers].len = plen;
1308
1309                         off = 0;
1310                         len -= plen;
1311                         buffers++;
1312                 }
1313
1314                 /*
1315                  * We didn't complete this iov, stop here since it probably
1316                  * means we have to move some of this into a pipe to
1317                  * be able to continue.
1318                  */
1319                 if (len)
1320                         break;
1321
1322                 /*
1323                  * Don't continue if we mapped fewer pages than we asked for,
1324                  * or if we mapped the max number of pages that we have
1325                  * room for.
1326                  */
1327                 if (error < npages || buffers == PIPE_BUFFERS)
1328                         break;
1329
1330                 nr_vecs--;
1331                 iov++;
1332         }
1333
1334         up_read(&current->mm->mmap_sem);
1335
1336         if (buffers)
1337                 return buffers;
1338
1339         return error;
1340 }
1341
1342 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1343                         struct splice_desc *sd)
1344 {
1345         char *src;
1346         int ret;
1347
1348         ret = buf->ops->confirm(pipe, buf);
1349         if (unlikely(ret))
1350                 return ret;
1351
1352         /*
1353          * See if we can use the atomic maps, by prefaulting in the
1354          * pages and doing an atomic copy
1355          */
1356         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1357                 src = buf->ops->map(pipe, buf, 1);
1358                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1359                                                         sd->len);
1360                 buf->ops->unmap(pipe, buf, src);
1361                 if (!ret) {
1362                         ret = sd->len;
1363                         goto out;
1364                 }
1365         }
1366
1367         /*
1368          * No dice, use slow non-atomic map and copy
1369          */
1370         src = buf->ops->map(pipe, buf, 0);
1371
1372         ret = sd->len;
1373         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1374                 ret = -EFAULT;
1375
1376 out:
1377         if (ret > 0)
1378                 sd->u.userptr += ret;
1379         buf->ops->unmap(pipe, buf, src);
1380         return ret;
1381 }
1382
1383 /*
1384  * For lack of a better implementation, implement vmsplice() to userspace
1385  * as a simple copy of the pipes pages to the user iov.
1386  */
1387 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1388                              unsigned long nr_segs, unsigned int flags)
1389 {
1390         struct pipe_inode_info *pipe;
1391         struct splice_desc sd;
1392         ssize_t size;
1393         int error;
1394         long ret;
1395
1396         pipe = pipe_info(file->f_path.dentry->d_inode);
1397         if (!pipe)
1398                 return -EBADF;
1399
1400         if (pipe->inode)
1401                 mutex_lock(&pipe->inode->i_mutex);
1402
1403         error = ret = 0;
1404         while (nr_segs) {
1405                 void __user *base;
1406                 size_t len;
1407
1408                 /*
1409                  * Get user address base and length for this iovec.
1410                  */
1411                 error = get_user(base, &iov->iov_base);
1412                 if (unlikely(error))
1413                         break;
1414                 error = get_user(len, &iov->iov_len);
1415                 if (unlikely(error))
1416                         break;
1417
1418                 /*
1419                  * Sanity check this iovec. 0 read succeeds.
1420                  */
1421                 if (unlikely(!len))
1422                         break;
1423                 if (unlikely(!base)) {
1424                         error = -EFAULT;
1425                         break;
1426                 }
1427
1428                 sd.len = 0;
1429                 sd.total_len = len;
1430                 sd.flags = flags;
1431                 sd.u.userptr = base;
1432                 sd.pos = 0;
1433
1434                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1435                 if (size < 0) {
1436                         if (!ret)
1437                                 ret = size;
1438
1439                         break;
1440                 }
1441
1442                 ret += size;
1443
1444                 if (size < len)
1445                         break;
1446
1447                 nr_segs--;
1448                 iov++;
1449         }
1450
1451         if (pipe->inode)
1452                 mutex_unlock(&pipe->inode->i_mutex);
1453
1454         if (!ret)
1455                 ret = error;
1456
1457         return ret;
1458 }
1459
1460 /*
1461  * vmsplice splices a user address range into a pipe. It can be thought of
1462  * as splice-from-memory, where the regular splice is splice-from-file (or
1463  * to file). In both cases the output is a pipe, naturally.
1464  */
1465 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1466                              unsigned long nr_segs, unsigned int flags)
1467 {
1468         struct pipe_inode_info *pipe;
1469         struct page *pages[PIPE_BUFFERS];
1470         struct partial_page partial[PIPE_BUFFERS];
1471         struct splice_pipe_desc spd = {
1472                 .pages = pages,
1473                 .partial = partial,
1474                 .flags = flags,
1475                 .ops = &user_page_pipe_buf_ops,
1476         };
1477
1478         pipe = pipe_info(file->f_path.dentry->d_inode);
1479         if (!pipe)
1480                 return -EBADF;
1481
1482         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1483                                             flags & SPLICE_F_GIFT);
1484         if (spd.nr_pages <= 0)
1485                 return spd.nr_pages;
1486
1487         return splice_to_pipe(pipe, &spd);
1488 }
1489
1490 /*
1491  * Note that vmsplice only really supports true splicing _from_ user memory
1492  * to a pipe, not the other way around. Splicing from user memory is a simple
1493  * operation that can be supported without any funky alignment restrictions
1494  * or nasty vm tricks. We simply map in the user memory and fill them into
1495  * a pipe. The reverse isn't quite as easy, though. There are two possible
1496  * solutions for that:
1497  *
1498  *      - memcpy() the data internally, at which point we might as well just
1499  *        do a regular read() on the buffer anyway.
1500  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1501  *        has restriction limitations on both ends of the pipe).
1502  *
1503  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1504  *
1505  */
1506 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1507                              unsigned long nr_segs, unsigned int flags)
1508 {
1509         struct file *file;
1510         long error;
1511         int fput;
1512
1513         if (unlikely(nr_segs > UIO_MAXIOV))
1514                 return -EINVAL;
1515         else if (unlikely(!nr_segs))
1516                 return 0;
1517
1518         error = -EBADF;
1519         file = fget_light(fd, &fput);
1520         if (file) {
1521                 if (file->f_mode & FMODE_WRITE)
1522                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1523                 else if (file->f_mode & FMODE_READ)
1524                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1525
1526                 fput_light(file, fput);
1527         }
1528
1529         return error;
1530 }
1531
1532 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1533                            int fd_out, loff_t __user *off_out,
1534                            size_t len, unsigned int flags)
1535 {
1536         long error;
1537         struct file *in, *out;
1538         int fput_in, fput_out;
1539
1540         if (unlikely(!len))
1541                 return 0;
1542
1543         error = -EBADF;
1544         in = fget_light(fd_in, &fput_in);
1545         if (in) {
1546                 if (in->f_mode & FMODE_READ) {
1547                         out = fget_light(fd_out, &fput_out);
1548                         if (out) {
1549                                 if (out->f_mode & FMODE_WRITE)
1550                                         error = do_splice(in, off_in,
1551                                                           out, off_out,
1552                                                           len, flags);
1553                                 fput_light(out, fput_out);
1554                         }
1555                 }
1556
1557                 fput_light(in, fput_in);
1558         }
1559
1560         return error;
1561 }
1562
1563 /*
1564  * Make sure there's data to read. Wait for input if we can, otherwise
1565  * return an appropriate error.
1566  */
1567 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1568 {
1569         int ret;
1570
1571         /*
1572          * Check ->nrbufs without the inode lock first. This function
1573          * is speculative anyways, so missing one is ok.
1574          */
1575         if (pipe->nrbufs)
1576                 return 0;
1577
1578         ret = 0;
1579         mutex_lock(&pipe->inode->i_mutex);
1580
1581         while (!pipe->nrbufs) {
1582                 if (signal_pending(current)) {
1583                         ret = -ERESTARTSYS;
1584                         break;
1585                 }
1586                 if (!pipe->writers)
1587                         break;
1588                 if (!pipe->waiting_writers) {
1589                         if (flags & SPLICE_F_NONBLOCK) {
1590                                 ret = -EAGAIN;
1591                                 break;
1592                         }
1593                 }
1594                 pipe_wait(pipe);
1595         }
1596
1597         mutex_unlock(&pipe->inode->i_mutex);
1598         return ret;
1599 }
1600
1601 /*
1602  * Make sure there's writeable room. Wait for room if we can, otherwise
1603  * return an appropriate error.
1604  */
1605 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1606 {
1607         int ret;
1608
1609         /*
1610          * Check ->nrbufs without the inode lock first. This function
1611          * is speculative anyways, so missing one is ok.
1612          */
1613         if (pipe->nrbufs < PIPE_BUFFERS)
1614                 return 0;
1615
1616         ret = 0;
1617         mutex_lock(&pipe->inode->i_mutex);
1618
1619         while (pipe->nrbufs >= PIPE_BUFFERS) {
1620                 if (!pipe->readers) {
1621                         send_sig(SIGPIPE, current, 0);
1622                         ret = -EPIPE;
1623                         break;
1624                 }
1625                 if (flags & SPLICE_F_NONBLOCK) {
1626                         ret = -EAGAIN;
1627                         break;
1628                 }
1629                 if (signal_pending(current)) {
1630                         ret = -ERESTARTSYS;
1631                         break;
1632                 }
1633                 pipe->waiting_writers++;
1634                 pipe_wait(pipe);
1635                 pipe->waiting_writers--;
1636         }
1637
1638         mutex_unlock(&pipe->inode->i_mutex);
1639         return ret;
1640 }
1641
1642 /*
1643  * Link contents of ipipe to opipe.
1644  */
1645 static int link_pipe(struct pipe_inode_info *ipipe,
1646                      struct pipe_inode_info *opipe,
1647                      size_t len, unsigned int flags)
1648 {
1649         struct pipe_buffer *ibuf, *obuf;
1650         int ret = 0, i = 0, nbuf;
1651
1652         /*
1653          * Potential ABBA deadlock, work around it by ordering lock
1654          * grabbing by inode address. Otherwise two different processes
1655          * could deadlock (one doing tee from A -> B, the other from B -> A).
1656          */
1657         inode_double_lock(ipipe->inode, opipe->inode);
1658
1659         do {
1660                 if (!opipe->readers) {
1661                         send_sig(SIGPIPE, current, 0);
1662                         if (!ret)
1663                                 ret = -EPIPE;
1664                         break;
1665                 }
1666
1667                 /*
1668                  * If we have iterated all input buffers or ran out of
1669                  * output room, break.
1670                  */
1671                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1672                         break;
1673
1674                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1675                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1676
1677                 /*
1678                  * Get a reference to this pipe buffer,
1679                  * so we can copy the contents over.
1680                  */
1681                 ibuf->ops->get(ipipe, ibuf);
1682
1683                 obuf = opipe->bufs + nbuf;
1684                 *obuf = *ibuf;
1685
1686                 /*
1687                  * Don't inherit the gift flag, we need to
1688                  * prevent multiple steals of this page.
1689                  */
1690                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1691
1692                 if (obuf->len > len)
1693                         obuf->len = len;
1694
1695                 opipe->nrbufs++;
1696                 ret += obuf->len;
1697                 len -= obuf->len;
1698                 i++;
1699         } while (len);
1700
1701         inode_double_unlock(ipipe->inode, opipe->inode);
1702
1703         /*
1704          * If we put data in the output pipe, wakeup any potential readers.
1705          */
1706         if (ret > 0) {
1707                 smp_mb();
1708                 if (waitqueue_active(&opipe->wait))
1709                         wake_up_interruptible(&opipe->wait);
1710                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1711         }
1712
1713         return ret;
1714 }
1715
1716 /*
1717  * This is a tee(1) implementation that works on pipes. It doesn't copy
1718  * any data, it simply references the 'in' pages on the 'out' pipe.
1719  * The 'flags' used are the SPLICE_F_* variants, currently the only
1720  * applicable one is SPLICE_F_NONBLOCK.
1721  */
1722 static long do_tee(struct file *in, struct file *out, size_t len,
1723                    unsigned int flags)
1724 {
1725         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1726         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1727         int ret = -EINVAL;
1728
1729         /*
1730          * Duplicate the contents of ipipe to opipe without actually
1731          * copying the data.
1732          */
1733         if (ipipe && opipe && ipipe != opipe) {
1734                 /*
1735                  * Keep going, unless we encounter an error. The ipipe/opipe
1736                  * ordering doesn't really matter.
1737                  */
1738                 ret = link_ipipe_prep(ipipe, flags);
1739                 if (!ret) {
1740                         ret = link_opipe_prep(opipe, flags);
1741                         if (!ret) {
1742                                 ret = link_pipe(ipipe, opipe, len, flags);
1743                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1744                                         ret = -EAGAIN;
1745                         }
1746                 }
1747         }
1748
1749         return ret;
1750 }
1751
1752 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1753 {
1754         struct file *in;
1755         int error, fput_in;
1756
1757         if (unlikely(!len))
1758                 return 0;
1759
1760         error = -EBADF;
1761         in = fget_light(fdin, &fput_in);
1762         if (in) {
1763                 if (in->f_mode & FMODE_READ) {
1764                         int fput_out;
1765                         struct file *out = fget_light(fdout, &fput_out);
1766
1767                         if (out) {
1768                                 if (out->f_mode & FMODE_WRITE)
1769                                         error = do_tee(in, out, len, flags);
1770                                 fput_light(out, fput_out);
1771                         }
1772                 }
1773                 fput_light(in, fput_in);
1774         }
1775
1776         return error;
1777 }