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