splice: fix offset mangling with direct splicing (sendfile)
[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 #include <linux/security.h>
32
33 /*
34  * Attempt to steal a page from a pipe buffer. This should perhaps go into
35  * a vm helper function, it's already simplified quite a bit by the
36  * addition of remove_mapping(). If success is returned, the caller may
37  * attempt to reuse this page for another destination.
38  */
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
40                                      struct pipe_buffer *buf)
41 {
42         struct page *page = buf->page;
43         struct address_space *mapping;
44
45         lock_page(page);
46
47         mapping = page_mapping(page);
48         if (mapping) {
49                 WARN_ON(!PageUptodate(page));
50
51                 /*
52                  * At least for ext2 with nobh option, we need to wait on
53                  * writeback completing on this page, since we'll remove it
54                  * from the pagecache.  Otherwise truncate wont wait on the
55                  * page, allowing the disk blocks to be reused by someone else
56                  * before we actually wrote our data to them. fs corruption
57                  * ensues.
58                  */
59                 wait_on_page_writeback(page);
60
61                 if (PagePrivate(page))
62                         try_to_release_page(page, GFP_KERNEL);
63
64                 /*
65                  * If we succeeded in removing the mapping, set LRU flag
66                  * and return good.
67                  */
68                 if (remove_mapping(mapping, page)) {
69                         buf->flags |= PIPE_BUF_FLAG_LRU;
70                         return 0;
71                 }
72         }
73
74         /*
75          * Raced with truncate or failed to remove page from current
76          * address space, unlock and return failure.
77          */
78         unlock_page(page);
79         return 1;
80 }
81
82 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
83                                         struct pipe_buffer *buf)
84 {
85         page_cache_release(buf->page);
86         buf->flags &= ~PIPE_BUF_FLAG_LRU;
87 }
88
89 /*
90  * Check whether the contents of buf is OK to access. Since the content
91  * is a page cache page, IO may be in flight.
92  */
93 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
94                                        struct pipe_buffer *buf)
95 {
96         struct page *page = buf->page;
97         int err;
98
99         if (!PageUptodate(page)) {
100                 lock_page(page);
101
102                 /*
103                  * Page got truncated/unhashed. This will cause a 0-byte
104                  * splice, if this is the first page.
105                  */
106                 if (!page->mapping) {
107                         err = -ENODATA;
108                         goto error;
109                 }
110
111                 /*
112                  * Uh oh, read-error from disk.
113                  */
114                 if (!PageUptodate(page)) {
115                         err = -EIO;
116                         goto error;
117                 }
118
119                 /*
120                  * Page is ok afterall, we are done.
121                  */
122                 unlock_page(page);
123         }
124
125         return 0;
126 error:
127         unlock_page(page);
128         return err;
129 }
130
131 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
132         .can_merge = 0,
133         .map = generic_pipe_buf_map,
134         .unmap = generic_pipe_buf_unmap,
135         .confirm = page_cache_pipe_buf_confirm,
136         .release = page_cache_pipe_buf_release,
137         .steal = page_cache_pipe_buf_steal,
138         .get = generic_pipe_buf_get,
139 };
140
141 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
142                                     struct pipe_buffer *buf)
143 {
144         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
145                 return 1;
146
147         buf->flags |= PIPE_BUF_FLAG_LRU;
148         return generic_pipe_buf_steal(pipe, buf);
149 }
150
151 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
152         .can_merge = 0,
153         .map = generic_pipe_buf_map,
154         .unmap = generic_pipe_buf_unmap,
155         .confirm = generic_pipe_buf_confirm,
156         .release = page_cache_pipe_buf_release,
157         .steal = user_page_pipe_buf_steal,
158         .get = generic_pipe_buf_get,
159 };
160
161 /**
162  * splice_to_pipe - fill passed data into a pipe
163  * @pipe:       pipe to fill
164  * @spd:        data to fill
165  *
166  * Description:
167  *    @spd contains a map of pages and len/offset tupples, a long with
168  *    the struct pipe_buf_operations associated with these pages. This
169  *    function will link that data to the pipe.
170  *
171  */
172 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
173                        struct splice_pipe_desc *spd)
174 {
175         unsigned int spd_pages = spd->nr_pages;
176         int ret, do_wakeup, page_nr;
177
178         ret = 0;
179         do_wakeup = 0;
180         page_nr = 0;
181
182         if (pipe->inode)
183                 mutex_lock(&pipe->inode->i_mutex);
184
185         for (;;) {
186                 if (!pipe->readers) {
187                         send_sig(SIGPIPE, current, 0);
188                         if (!ret)
189                                 ret = -EPIPE;
190                         break;
191                 }
192
193                 if (pipe->nrbufs < PIPE_BUFFERS) {
194                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
195                         struct pipe_buffer *buf = pipe->bufs + newbuf;
196
197                         buf->page = spd->pages[page_nr];
198                         buf->offset = spd->partial[page_nr].offset;
199                         buf->len = spd->partial[page_nr].len;
200                         buf->private = spd->partial[page_nr].private;
201                         buf->ops = spd->ops;
202                         if (spd->flags & SPLICE_F_GIFT)
203                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
204
205                         pipe->nrbufs++;
206                         page_nr++;
207                         ret += buf->len;
208
209                         if (pipe->inode)
210                                 do_wakeup = 1;
211
212                         if (!--spd->nr_pages)
213                                 break;
214                         if (pipe->nrbufs < PIPE_BUFFERS)
215                                 continue;
216
217                         break;
218                 }
219
220                 if (spd->flags & SPLICE_F_NONBLOCK) {
221                         if (!ret)
222                                 ret = -EAGAIN;
223                         break;
224                 }
225
226                 if (signal_pending(current)) {
227                         if (!ret)
228                                 ret = -ERESTARTSYS;
229                         break;
230                 }
231
232                 if (do_wakeup) {
233                         smp_mb();
234                         if (waitqueue_active(&pipe->wait))
235                                 wake_up_interruptible_sync(&pipe->wait);
236                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
237                         do_wakeup = 0;
238                 }
239
240                 pipe->waiting_writers++;
241                 pipe_wait(pipe);
242                 pipe->waiting_writers--;
243         }
244
245         if (pipe->inode) {
246                 mutex_unlock(&pipe->inode->i_mutex);
247
248                 if (do_wakeup) {
249                         smp_mb();
250                         if (waitqueue_active(&pipe->wait))
251                                 wake_up_interruptible(&pipe->wait);
252                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
253                 }
254         }
255
256         while (page_nr < spd_pages)
257                 page_cache_release(spd->pages[page_nr++]);
258
259         return ret;
260 }
261
262 static int
263 __generic_file_splice_read(struct file *in, loff_t *ppos,
264                            struct pipe_inode_info *pipe, size_t len,
265                            unsigned int flags)
266 {
267         struct address_space *mapping = in->f_mapping;
268         unsigned int loff, nr_pages;
269         struct page *pages[PIPE_BUFFERS];
270         struct partial_page partial[PIPE_BUFFERS];
271         struct page *page;
272         pgoff_t index, end_index;
273         loff_t isize;
274         int error, page_nr;
275         struct splice_pipe_desc spd = {
276                 .pages = pages,
277                 .partial = partial,
278                 .flags = flags,
279                 .ops = &page_cache_pipe_buf_ops,
280         };
281
282         index = *ppos >> PAGE_CACHE_SHIFT;
283         loff = *ppos & ~PAGE_CACHE_MASK;
284         nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
285
286         if (nr_pages > PIPE_BUFFERS)
287                 nr_pages = PIPE_BUFFERS;
288
289         /*
290          * Don't try to 2nd guess the read-ahead logic, call into
291          * page_cache_readahead() like the page cache reads would do.
292          */
293         page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
294
295         /*
296          * Lookup the (hopefully) full range of pages we need.
297          */
298         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
299
300         /*
301          * If find_get_pages_contig() returned fewer pages than we needed,
302          * allocate the rest and fill in the holes.
303          */
304         error = 0;
305         index += spd.nr_pages;
306         while (spd.nr_pages < nr_pages) {
307                 /*
308                  * Page could be there, find_get_pages_contig() breaks on
309                  * the first hole.
310                  */
311                 page = find_get_page(mapping, index);
312                 if (!page) {
313                         /*
314                          * Make sure the read-ahead engine is notified
315                          * about this failure.
316                          */
317                         handle_ra_miss(mapping, &in->f_ra, index);
318
319                         /*
320                          * page didn't exist, allocate one.
321                          */
322                         page = page_cache_alloc_cold(mapping);
323                         if (!page)
324                                 break;
325
326                         error = add_to_page_cache_lru(page, mapping, index,
327                                               GFP_KERNEL);
328                         if (unlikely(error)) {
329                                 page_cache_release(page);
330                                 if (error == -EEXIST)
331                                         continue;
332                                 break;
333                         }
334                         /*
335                          * add_to_page_cache() locks the page, unlock it
336                          * to avoid convoluting the logic below even more.
337                          */
338                         unlock_page(page);
339                 }
340
341                 pages[spd.nr_pages++] = page;
342                 index++;
343         }
344
345         /*
346          * Now loop over the map and see if we need to start IO on any
347          * pages, fill in the partial map, etc.
348          */
349         index = *ppos >> PAGE_CACHE_SHIFT;
350         nr_pages = spd.nr_pages;
351         spd.nr_pages = 0;
352         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
353                 unsigned int this_len;
354
355                 if (!len)
356                         break;
357
358                 /*
359                  * this_len is the max we'll use from this page
360                  */
361                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
362                 page = pages[page_nr];
363
364                 /*
365                  * If the page isn't uptodate, we may need to start io on it
366                  */
367                 if (!PageUptodate(page)) {
368                         /*
369                          * If in nonblock mode then dont block on waiting
370                          * for an in-flight io page
371                          */
372                         if (flags & SPLICE_F_NONBLOCK) {
373                                 if (TestSetPageLocked(page))
374                                         break;
375                         } else
376                                 lock_page(page);
377
378                         /*
379                          * page was truncated, stop here. if this isn't the
380                          * first page, we'll just complete what we already
381                          * added
382                          */
383                         if (!page->mapping) {
384                                 unlock_page(page);
385                                 break;
386                         }
387                         /*
388                          * page was already under io and is now done, great
389                          */
390                         if (PageUptodate(page)) {
391                                 unlock_page(page);
392                                 goto fill_it;
393                         }
394
395                         /*
396                          * need to read in the page
397                          */
398                         error = mapping->a_ops->readpage(in, page);
399                         if (unlikely(error)) {
400                                 /*
401                                  * We really should re-lookup the page here,
402                                  * but it complicates things a lot. Instead
403                                  * lets just do what we already stored, and
404                                  * we'll get it the next time we are called.
405                                  */
406                                 if (error == AOP_TRUNCATED_PAGE)
407                                         error = 0;
408
409                                 break;
410                         }
411                 }
412 fill_it:
413                 /*
414                  * i_size must be checked after PageUptodate.
415                  */
416                 isize = i_size_read(mapping->host);
417                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
418                 if (unlikely(!isize || index > end_index))
419                         break;
420
421                 /*
422                  * if this is the last page, see if we need to shrink
423                  * the length and stop
424                  */
425                 if (end_index == index) {
426                         unsigned int plen;
427
428                         /*
429                          * max good bytes in this page
430                          */
431                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
432                         if (plen <= loff)
433                                 break;
434
435                         /*
436                          * force quit after adding this page
437                          */
438                         this_len = min(this_len, plen - loff);
439                         len = this_len;
440                 }
441
442                 partial[page_nr].offset = loff;
443                 partial[page_nr].len = this_len;
444                 len -= this_len;
445                 loff = 0;
446                 spd.nr_pages++;
447                 index++;
448         }
449
450         /*
451          * Release any pages at the end, if we quit early. 'page_nr' is how far
452          * we got, 'nr_pages' is how many pages are in the map.
453          */
454         while (page_nr < nr_pages)
455                 page_cache_release(pages[page_nr++]);
456
457         if (spd.nr_pages)
458                 return splice_to_pipe(pipe, &spd);
459
460         return error;
461 }
462
463 /**
464  * generic_file_splice_read - splice data from file to a pipe
465  * @in:         file to splice from
466  * @ppos:       position in @in
467  * @pipe:       pipe to splice to
468  * @len:        number of bytes to splice
469  * @flags:      splice modifier flags
470  *
471  * Description:
472  *    Will read pages from given file and fill them into a pipe. Can be
473  *    used as long as the address_space operations for the source implements
474  *    a readpage() hook.
475  *
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 && !spliced) {
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->u.file;
530         loff_t pos = sd->pos;
531         int ret, more;
532
533         ret = buf->ops->confirm(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->u.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->confirm(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  * __splice_from_pipe - splice data from a pipe to given actor
666  * @pipe:       pipe to splice from
667  * @sd:         information to @actor
668  * @actor:      handler that splices the data
669  *
670  * Description:
671  *    This function does little more than loop over the pipe and call
672  *    @actor to do the actual moving of a single struct pipe_buffer to
673  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
674  *    pipe_to_user.
675  *
676  */
677 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
678                            splice_actor *actor)
679 {
680         int ret, do_wakeup, err;
681
682         ret = 0;
683         do_wakeup = 0;
684
685         for (;;) {
686                 if (pipe->nrbufs) {
687                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
688                         const struct pipe_buf_operations *ops = buf->ops;
689
690                         sd->len = buf->len;
691                         if (sd->len > sd->total_len)
692                                 sd->len = sd->total_len;
693
694                         err = actor(pipe, buf, sd);
695                         if (err <= 0) {
696                                 if (!ret && err != -ENODATA)
697                                         ret = err;
698
699                                 break;
700                         }
701
702                         ret += err;
703                         buf->offset += err;
704                         buf->len -= err;
705
706                         sd->len -= err;
707                         sd->pos += err;
708                         sd->total_len -= err;
709                         if (sd->len)
710                                 continue;
711
712                         if (!buf->len) {
713                                 buf->ops = NULL;
714                                 ops->release(pipe, buf);
715                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
716                                 pipe->nrbufs--;
717                                 if (pipe->inode)
718                                         do_wakeup = 1;
719                         }
720
721                         if (!sd->total_len)
722                                 break;
723                 }
724
725                 if (pipe->nrbufs)
726                         continue;
727                 if (!pipe->writers)
728                         break;
729                 if (!pipe->waiting_writers) {
730                         if (ret)
731                                 break;
732                 }
733
734                 if (sd->flags & SPLICE_F_NONBLOCK) {
735                         if (!ret)
736                                 ret = -EAGAIN;
737                         break;
738                 }
739
740                 if (signal_pending(current)) {
741                         if (!ret)
742                                 ret = -ERESTARTSYS;
743                         break;
744                 }
745
746                 if (do_wakeup) {
747                         smp_mb();
748                         if (waitqueue_active(&pipe->wait))
749                                 wake_up_interruptible_sync(&pipe->wait);
750                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
751                         do_wakeup = 0;
752                 }
753
754                 pipe_wait(pipe);
755         }
756
757         if (do_wakeup) {
758                 smp_mb();
759                 if (waitqueue_active(&pipe->wait))
760                         wake_up_interruptible(&pipe->wait);
761                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
762         }
763
764         return ret;
765 }
766 EXPORT_SYMBOL(__splice_from_pipe);
767
768 /**
769  * splice_from_pipe - splice data from a pipe to a file
770  * @pipe:       pipe to splice from
771  * @out:        file to splice to
772  * @ppos:       position in @out
773  * @len:        how many bytes to splice
774  * @flags:      splice modifier flags
775  * @actor:      handler that splices the data
776  *
777  * Description:
778  *    See __splice_from_pipe. This function locks the input and output inodes,
779  *    otherwise it's identical to __splice_from_pipe().
780  *
781  */
782 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
783                          loff_t *ppos, size_t len, unsigned int flags,
784                          splice_actor *actor)
785 {
786         ssize_t ret;
787         struct inode *inode = out->f_mapping->host;
788         struct splice_desc sd = {
789                 .total_len = len,
790                 .flags = flags,
791                 .pos = *ppos,
792                 .u.file = out,
793         };
794
795         /*
796          * The actor worker might be calling ->prepare_write and
797          * ->commit_write. Most of the time, these expect i_mutex to
798          * be held. Since this may result in an ABBA deadlock with
799          * pipe->inode, we have to order lock acquiry here.
800          */
801         inode_double_lock(inode, pipe->inode);
802         ret = __splice_from_pipe(pipe, &sd, actor);
803         inode_double_unlock(inode, pipe->inode);
804
805         return ret;
806 }
807
808 /**
809  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
810  * @pipe:       pipe info
811  * @out:        file to write to
812  * @ppos:       position in @out
813  * @len:        number of bytes to splice
814  * @flags:      splice modifier flags
815  *
816  * Description:
817  *    Will either move or copy pages (determined by @flags options) from
818  *    the given pipe inode to the given file. The caller is responsible
819  *    for acquiring i_mutex on both inodes.
820  *
821  */
822 ssize_t
823 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
824                                  loff_t *ppos, size_t len, unsigned int flags)
825 {
826         struct address_space *mapping = out->f_mapping;
827         struct inode *inode = mapping->host;
828         struct splice_desc sd = {
829                 .total_len = len,
830                 .flags = flags,
831                 .pos = *ppos,
832                 .u.file = out,
833         };
834         ssize_t ret;
835         int err;
836
837         err = remove_suid(out->f_path.dentry);
838         if (unlikely(err))
839                 return err;
840
841         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
842         if (ret > 0) {
843                 unsigned long nr_pages;
844
845                 *ppos += ret;
846                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
847
848                 /*
849                  * If file or inode is SYNC and we actually wrote some data,
850                  * sync it.
851                  */
852                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
853                         err = generic_osync_inode(inode, mapping,
854                                                   OSYNC_METADATA|OSYNC_DATA);
855
856                         if (err)
857                                 ret = err;
858                 }
859                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
860         }
861
862         return ret;
863 }
864
865 EXPORT_SYMBOL(generic_file_splice_write_nolock);
866
867 /**
868  * generic_file_splice_write - splice data from a pipe to a file
869  * @pipe:       pipe info
870  * @out:        file to write to
871  * @ppos:       position in @out
872  * @len:        number of bytes to splice
873  * @flags:      splice modifier flags
874  *
875  * Description:
876  *    Will either move or copy pages (determined by @flags options) from
877  *    the given pipe inode to the given file.
878  *
879  */
880 ssize_t
881 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
882                           loff_t *ppos, size_t len, unsigned int flags)
883 {
884         struct address_space *mapping = out->f_mapping;
885         struct inode *inode = mapping->host;
886         ssize_t ret;
887         int err;
888
889         err = should_remove_suid(out->f_path.dentry);
890         if (unlikely(err)) {
891                 mutex_lock(&inode->i_mutex);
892                 err = __remove_suid(out->f_path.dentry, err);
893                 mutex_unlock(&inode->i_mutex);
894                 if (err)
895                         return err;
896         }
897
898         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
899         if (ret > 0) {
900                 unsigned long nr_pages;
901
902                 *ppos += ret;
903                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
904
905                 /*
906                  * If file or inode is SYNC and we actually wrote some data,
907                  * sync it.
908                  */
909                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
910                         mutex_lock(&inode->i_mutex);
911                         err = generic_osync_inode(inode, mapping,
912                                                   OSYNC_METADATA|OSYNC_DATA);
913                         mutex_unlock(&inode->i_mutex);
914
915                         if (err)
916                                 ret = err;
917                 }
918                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
919         }
920
921         return ret;
922 }
923
924 EXPORT_SYMBOL(generic_file_splice_write);
925
926 /**
927  * generic_splice_sendpage - splice data from a pipe to a socket
928  * @pipe:       pipe to splice from
929  * @out:        socket to write to
930  * @ppos:       position in @out
931  * @len:        number of bytes to splice
932  * @flags:      splice modifier flags
933  *
934  * Description:
935  *    Will send @len bytes from the pipe to a network socket. No data copying
936  *    is involved.
937  *
938  */
939 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
940                                 loff_t *ppos, size_t len, unsigned int flags)
941 {
942         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
943 }
944
945 EXPORT_SYMBOL(generic_splice_sendpage);
946
947 /*
948  * Attempt to initiate a splice from pipe to file.
949  */
950 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
951                            loff_t *ppos, size_t len, unsigned int flags)
952 {
953         int ret;
954
955         if (unlikely(!out->f_op || !out->f_op->splice_write))
956                 return -EINVAL;
957
958         if (unlikely(!(out->f_mode & FMODE_WRITE)))
959                 return -EBADF;
960
961         ret = rw_verify_area(WRITE, out, ppos, len);
962         if (unlikely(ret < 0))
963                 return ret;
964
965         ret = security_file_permission(out, MAY_WRITE);
966         if (unlikely(ret < 0))
967                 return ret;
968
969         return out->f_op->splice_write(pipe, out, ppos, len, flags);
970 }
971
972 /*
973  * Attempt to initiate a splice from a file to a pipe.
974  */
975 static long do_splice_to(struct file *in, loff_t *ppos,
976                          struct pipe_inode_info *pipe, size_t len,
977                          unsigned int flags)
978 {
979         int ret;
980
981         if (unlikely(!in->f_op || !in->f_op->splice_read))
982                 return -EINVAL;
983
984         if (unlikely(!(in->f_mode & FMODE_READ)))
985                 return -EBADF;
986
987         ret = rw_verify_area(READ, in, ppos, len);
988         if (unlikely(ret < 0))
989                 return ret;
990
991         ret = security_file_permission(in, MAY_READ);
992         if (unlikely(ret < 0))
993                 return ret;
994
995         return in->f_op->splice_read(in, ppos, pipe, len, flags);
996 }
997
998 /**
999  * splice_direct_to_actor - splices data directly between two non-pipes
1000  * @in:         file to splice from
1001  * @sd:         actor information on where to splice to
1002  * @actor:      handles the data splicing
1003  *
1004  * Description:
1005  *    This is a special case helper to splice directly between two
1006  *    points, without requiring an explicit pipe. Internally an allocated
1007  *    pipe is cached in the process, and reused during the life time of
1008  *    that process.
1009  *
1010  */
1011 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1012                                splice_direct_actor *actor)
1013 {
1014         struct pipe_inode_info *pipe;
1015         long ret, bytes;
1016         umode_t i_mode;
1017         size_t len;
1018         int i, flags;
1019
1020         /*
1021          * We require the input being a regular file, as we don't want to
1022          * randomly drop data for eg socket -> socket splicing. Use the
1023          * piped splicing for that!
1024          */
1025         i_mode = in->f_path.dentry->d_inode->i_mode;
1026         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1027                 return -EINVAL;
1028
1029         /*
1030          * neither in nor out is a pipe, setup an internal pipe attached to
1031          * 'out' and transfer the wanted data from 'in' to 'out' through that
1032          */
1033         pipe = current->splice_pipe;
1034         if (unlikely(!pipe)) {
1035                 pipe = alloc_pipe_info(NULL);
1036                 if (!pipe)
1037                         return -ENOMEM;
1038
1039                 /*
1040                  * We don't have an immediate reader, but we'll read the stuff
1041                  * out of the pipe right after the splice_to_pipe(). So set
1042                  * PIPE_READERS appropriately.
1043                  */
1044                 pipe->readers = 1;
1045
1046                 current->splice_pipe = pipe;
1047         }
1048
1049         /*
1050          * Do the splice.
1051          */
1052         ret = 0;
1053         bytes = 0;
1054         len = sd->total_len;
1055         flags = sd->flags;
1056
1057         /*
1058          * Don't block on output, we have to drain the direct pipe.
1059          */
1060         sd->flags &= ~SPLICE_F_NONBLOCK;
1061
1062         while (len) {
1063                 size_t read_len;
1064
1065                 ret = do_splice_to(in, &sd->pos, pipe, len, flags);
1066                 if (unlikely(ret <= 0))
1067                         goto out_release;
1068
1069                 read_len = ret;
1070                 sd->total_len = read_len;
1071
1072                 /*
1073                  * NOTE: nonblocking mode only applies to the input. We
1074                  * must not do the output in nonblocking mode as then we
1075                  * could get stuck data in the internal pipe:
1076                  */
1077                 ret = actor(pipe, sd);
1078                 if (unlikely(ret <= 0))
1079                         goto out_release;
1080
1081                 bytes += ret;
1082                 len -= ret;
1083
1084                 if (ret < read_len)
1085                         goto out_release;
1086         }
1087
1088         pipe->nrbufs = pipe->curbuf = 0;
1089         return bytes;
1090
1091 out_release:
1092         /*
1093          * If we did an incomplete transfer we must release
1094          * the pipe buffers in question:
1095          */
1096         for (i = 0; i < PIPE_BUFFERS; i++) {
1097                 struct pipe_buffer *buf = pipe->bufs + i;
1098
1099                 if (buf->ops) {
1100                         buf->ops->release(pipe, buf);
1101                         buf->ops = NULL;
1102                 }
1103         }
1104         pipe->nrbufs = pipe->curbuf = 0;
1105
1106         /*
1107          * If we transferred some data, return the number of bytes:
1108          */
1109         if (bytes > 0)
1110                 return bytes;
1111
1112         return ret;
1113
1114 }
1115 EXPORT_SYMBOL(splice_direct_to_actor);
1116
1117 static int direct_splice_actor(struct pipe_inode_info *pipe,
1118                                struct splice_desc *sd)
1119 {
1120         struct file *file = sd->u.file;
1121
1122         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1123 }
1124
1125 /**
1126  * do_splice_direct - splices data directly between two files
1127  * @in:         file to splice from
1128  * @ppos:       input file offset
1129  * @out:        file to splice to
1130  * @len:        number of bytes to splice
1131  * @flags:      splice modifier flags
1132  *
1133  * Description:
1134  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1135  *    doing it in the application would incur an extra system call
1136  *    (splice in + splice out, as compared to just sendfile()). So this helper
1137  *    can splice directly through a process-private pipe.
1138  *
1139  */
1140 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1141                       size_t len, unsigned int flags)
1142 {
1143         struct splice_desc sd = {
1144                 .len            = len,
1145                 .total_len      = len,
1146                 .flags          = flags,
1147                 .pos            = *ppos,
1148                 .u.file         = out,
1149         };
1150         long ret;
1151
1152         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1153         if (ret > 0)
1154                 *ppos += ret;
1155
1156         return ret;
1157 }
1158
1159 /*
1160  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1161  * location, so checking ->i_pipe is not enough to verify that this is a
1162  * pipe.
1163  */
1164 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1165 {
1166         if (S_ISFIFO(inode->i_mode))
1167                 return inode->i_pipe;
1168
1169         return NULL;
1170 }
1171
1172 /*
1173  * Determine where to splice to/from.
1174  */
1175 static long do_splice(struct file *in, loff_t __user *off_in,
1176                       struct file *out, loff_t __user *off_out,
1177                       size_t len, unsigned int flags)
1178 {
1179         struct pipe_inode_info *pipe;
1180         loff_t offset, *off;
1181         long ret;
1182
1183         pipe = pipe_info(in->f_path.dentry->d_inode);
1184         if (pipe) {
1185                 if (off_in)
1186                         return -ESPIPE;
1187                 if (off_out) {
1188                         if (out->f_op->llseek == no_llseek)
1189                                 return -EINVAL;
1190                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1191                                 return -EFAULT;
1192                         off = &offset;
1193                 } else
1194                         off = &out->f_pos;
1195
1196                 ret = do_splice_from(pipe, out, off, len, flags);
1197
1198                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1199                         ret = -EFAULT;
1200
1201                 return ret;
1202         }
1203
1204         pipe = pipe_info(out->f_path.dentry->d_inode);
1205         if (pipe) {
1206                 if (off_out)
1207                         return -ESPIPE;
1208                 if (off_in) {
1209                         if (in->f_op->llseek == no_llseek)
1210                                 return -EINVAL;
1211                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1212                                 return -EFAULT;
1213                         off = &offset;
1214                 } else
1215                         off = &in->f_pos;
1216
1217                 ret = do_splice_to(in, off, pipe, len, flags);
1218
1219                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1220                         ret = -EFAULT;
1221
1222                 return ret;
1223         }
1224
1225         return -EINVAL;
1226 }
1227
1228 /*
1229  * Map an iov into an array of pages and offset/length tupples. With the
1230  * partial_page structure, we can map several non-contiguous ranges into
1231  * our ones pages[] map instead of splitting that operation into pieces.
1232  * Could easily be exported as a generic helper for other users, in which
1233  * case one would probably want to add a 'max_nr_pages' parameter as well.
1234  */
1235 static int get_iovec_page_array(const struct iovec __user *iov,
1236                                 unsigned int nr_vecs, struct page **pages,
1237                                 struct partial_page *partial, int aligned)
1238 {
1239         int buffers = 0, error = 0;
1240
1241         /*
1242          * It's ok to take the mmap_sem for reading, even
1243          * across a "get_user()".
1244          */
1245         down_read(&current->mm->mmap_sem);
1246
1247         while (nr_vecs) {
1248                 unsigned long off, npages;
1249                 void __user *base;
1250                 size_t len;
1251                 int i;
1252
1253                 /*
1254                  * Get user address base and length for this iovec.
1255                  */
1256                 error = get_user(base, &iov->iov_base);
1257                 if (unlikely(error))
1258                         break;
1259                 error = get_user(len, &iov->iov_len);
1260                 if (unlikely(error))
1261                         break;
1262
1263                 /*
1264                  * Sanity check this iovec. 0 read succeeds.
1265                  */
1266                 if (unlikely(!len))
1267                         break;
1268                 error = -EFAULT;
1269                 if (unlikely(!base))
1270                         break;
1271
1272                 /*
1273                  * Get this base offset and number of pages, then map
1274                  * in the user pages.
1275                  */
1276                 off = (unsigned long) base & ~PAGE_MASK;
1277
1278                 /*
1279                  * If asked for alignment, the offset must be zero and the
1280                  * length a multiple of the PAGE_SIZE.
1281                  */
1282                 error = -EINVAL;
1283                 if (aligned && (off || len & ~PAGE_MASK))
1284                         break;
1285
1286                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1287                 if (npages > PIPE_BUFFERS - buffers)
1288                         npages = PIPE_BUFFERS - buffers;
1289
1290                 error = get_user_pages(current, current->mm,
1291                                        (unsigned long) base, npages, 0, 0,
1292                                        &pages[buffers], NULL);
1293
1294                 if (unlikely(error <= 0))
1295                         break;
1296
1297                 /*
1298                  * Fill this contiguous range into the partial page map.
1299                  */
1300                 for (i = 0; i < error; i++) {
1301                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1302
1303                         partial[buffers].offset = off;
1304                         partial[buffers].len = plen;
1305
1306                         off = 0;
1307                         len -= plen;
1308                         buffers++;
1309                 }
1310
1311                 /*
1312                  * We didn't complete this iov, stop here since it probably
1313                  * means we have to move some of this into a pipe to
1314                  * be able to continue.
1315                  */
1316                 if (len)
1317                         break;
1318
1319                 /*
1320                  * Don't continue if we mapped fewer pages than we asked for,
1321                  * or if we mapped the max number of pages that we have
1322                  * room for.
1323                  */
1324                 if (error < npages || buffers == PIPE_BUFFERS)
1325                         break;
1326
1327                 nr_vecs--;
1328                 iov++;
1329         }
1330
1331         up_read(&current->mm->mmap_sem);
1332
1333         if (buffers)
1334                 return buffers;
1335
1336         return error;
1337 }
1338
1339 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1340                         struct splice_desc *sd)
1341 {
1342         char *src;
1343         int ret;
1344
1345         ret = buf->ops->confirm(pipe, buf);
1346         if (unlikely(ret))
1347                 return ret;
1348
1349         /*
1350          * See if we can use the atomic maps, by prefaulting in the
1351          * pages and doing an atomic copy
1352          */
1353         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1354                 src = buf->ops->map(pipe, buf, 1);
1355                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1356                                                         sd->len);
1357                 buf->ops->unmap(pipe, buf, src);
1358                 if (!ret) {
1359                         ret = sd->len;
1360                         goto out;
1361                 }
1362         }
1363
1364         /*
1365          * No dice, use slow non-atomic map and copy
1366          */
1367         src = buf->ops->map(pipe, buf, 0);
1368
1369         ret = sd->len;
1370         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1371                 ret = -EFAULT;
1372
1373 out:
1374         if (ret > 0)
1375                 sd->u.userptr += ret;
1376         buf->ops->unmap(pipe, buf, src);
1377         return ret;
1378 }
1379
1380 /*
1381  * For lack of a better implementation, implement vmsplice() to userspace
1382  * as a simple copy of the pipes pages to the user iov.
1383  */
1384 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1385                              unsigned long nr_segs, unsigned int flags)
1386 {
1387         struct pipe_inode_info *pipe;
1388         struct splice_desc sd;
1389         ssize_t size;
1390         int error;
1391         long ret;
1392
1393         pipe = pipe_info(file->f_path.dentry->d_inode);
1394         if (!pipe)
1395                 return -EBADF;
1396
1397         if (pipe->inode)
1398                 mutex_lock(&pipe->inode->i_mutex);
1399
1400         error = ret = 0;
1401         while (nr_segs) {
1402                 void __user *base;
1403                 size_t len;
1404
1405                 /*
1406                  * Get user address base and length for this iovec.
1407                  */
1408                 error = get_user(base, &iov->iov_base);
1409                 if (unlikely(error))
1410                         break;
1411                 error = get_user(len, &iov->iov_len);
1412                 if (unlikely(error))
1413                         break;
1414
1415                 /*
1416                  * Sanity check this iovec. 0 read succeeds.
1417                  */
1418                 if (unlikely(!len))
1419                         break;
1420                 if (unlikely(!base)) {
1421                         error = -EFAULT;
1422                         break;
1423                 }
1424
1425                 sd.len = 0;
1426                 sd.total_len = len;
1427                 sd.flags = flags;
1428                 sd.u.userptr = base;
1429                 sd.pos = 0;
1430
1431                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1432                 if (size < 0) {
1433                         if (!ret)
1434                                 ret = size;
1435
1436                         break;
1437                 }
1438
1439                 ret += size;
1440
1441                 if (size < len)
1442                         break;
1443
1444                 nr_segs--;
1445                 iov++;
1446         }
1447
1448         if (pipe->inode)
1449                 mutex_unlock(&pipe->inode->i_mutex);
1450
1451         if (!ret)
1452                 ret = error;
1453
1454         return ret;
1455 }
1456
1457 /*
1458  * vmsplice splices a user address range into a pipe. It can be thought of
1459  * as splice-from-memory, where the regular splice is splice-from-file (or
1460  * to file). In both cases the output is a pipe, naturally.
1461  */
1462 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1463                              unsigned long nr_segs, unsigned int flags)
1464 {
1465         struct pipe_inode_info *pipe;
1466         struct page *pages[PIPE_BUFFERS];
1467         struct partial_page partial[PIPE_BUFFERS];
1468         struct splice_pipe_desc spd = {
1469                 .pages = pages,
1470                 .partial = partial,
1471                 .flags = flags,
1472                 .ops = &user_page_pipe_buf_ops,
1473         };
1474
1475         pipe = pipe_info(file->f_path.dentry->d_inode);
1476         if (!pipe)
1477                 return -EBADF;
1478
1479         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1480                                             flags & SPLICE_F_GIFT);
1481         if (spd.nr_pages <= 0)
1482                 return spd.nr_pages;
1483
1484         return splice_to_pipe(pipe, &spd);
1485 }
1486
1487 /*
1488  * Note that vmsplice only really supports true splicing _from_ user memory
1489  * to a pipe, not the other way around. Splicing from user memory is a simple
1490  * operation that can be supported without any funky alignment restrictions
1491  * or nasty vm tricks. We simply map in the user memory and fill them into
1492  * a pipe. The reverse isn't quite as easy, though. There are two possible
1493  * solutions for that:
1494  *
1495  *      - memcpy() the data internally, at which point we might as well just
1496  *        do a regular read() on the buffer anyway.
1497  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1498  *        has restriction limitations on both ends of the pipe).
1499  *
1500  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1501  *
1502  */
1503 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1504                              unsigned long nr_segs, unsigned int flags)
1505 {
1506         struct file *file;
1507         long error;
1508         int fput;
1509
1510         if (unlikely(nr_segs > UIO_MAXIOV))
1511                 return -EINVAL;
1512         else if (unlikely(!nr_segs))
1513                 return 0;
1514
1515         error = -EBADF;
1516         file = fget_light(fd, &fput);
1517         if (file) {
1518                 if (file->f_mode & FMODE_WRITE)
1519                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1520                 else if (file->f_mode & FMODE_READ)
1521                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1522
1523                 fput_light(file, fput);
1524         }
1525
1526         return error;
1527 }
1528
1529 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1530                            int fd_out, loff_t __user *off_out,
1531                            size_t len, unsigned int flags)
1532 {
1533         long error;
1534         struct file *in, *out;
1535         int fput_in, fput_out;
1536
1537         if (unlikely(!len))
1538                 return 0;
1539
1540         error = -EBADF;
1541         in = fget_light(fd_in, &fput_in);
1542         if (in) {
1543                 if (in->f_mode & FMODE_READ) {
1544                         out = fget_light(fd_out, &fput_out);
1545                         if (out) {
1546                                 if (out->f_mode & FMODE_WRITE)
1547                                         error = do_splice(in, off_in,
1548                                                           out, off_out,
1549                                                           len, flags);
1550                                 fput_light(out, fput_out);
1551                         }
1552                 }
1553
1554                 fput_light(in, fput_in);
1555         }
1556
1557         return error;
1558 }
1559
1560 /*
1561  * Make sure there's data to read. Wait for input if we can, otherwise
1562  * return an appropriate error.
1563  */
1564 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1565 {
1566         int ret;
1567
1568         /*
1569          * Check ->nrbufs without the inode lock first. This function
1570          * is speculative anyways, so missing one is ok.
1571          */
1572         if (pipe->nrbufs)
1573                 return 0;
1574
1575         ret = 0;
1576         mutex_lock(&pipe->inode->i_mutex);
1577
1578         while (!pipe->nrbufs) {
1579                 if (signal_pending(current)) {
1580                         ret = -ERESTARTSYS;
1581                         break;
1582                 }
1583                 if (!pipe->writers)
1584                         break;
1585                 if (!pipe->waiting_writers) {
1586                         if (flags & SPLICE_F_NONBLOCK) {
1587                                 ret = -EAGAIN;
1588                                 break;
1589                         }
1590                 }
1591                 pipe_wait(pipe);
1592         }
1593
1594         mutex_unlock(&pipe->inode->i_mutex);
1595         return ret;
1596 }
1597
1598 /*
1599  * Make sure there's writeable room. Wait for room if we can, otherwise
1600  * return an appropriate error.
1601  */
1602 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1603 {
1604         int ret;
1605
1606         /*
1607          * Check ->nrbufs without the inode lock first. This function
1608          * is speculative anyways, so missing one is ok.
1609          */
1610         if (pipe->nrbufs < PIPE_BUFFERS)
1611                 return 0;
1612
1613         ret = 0;
1614         mutex_lock(&pipe->inode->i_mutex);
1615
1616         while (pipe->nrbufs >= PIPE_BUFFERS) {
1617                 if (!pipe->readers) {
1618                         send_sig(SIGPIPE, current, 0);
1619                         ret = -EPIPE;
1620                         break;
1621                 }
1622                 if (flags & SPLICE_F_NONBLOCK) {
1623                         ret = -EAGAIN;
1624                         break;
1625                 }
1626                 if (signal_pending(current)) {
1627                         ret = -ERESTARTSYS;
1628                         break;
1629                 }
1630                 pipe->waiting_writers++;
1631                 pipe_wait(pipe);
1632                 pipe->waiting_writers--;
1633         }
1634
1635         mutex_unlock(&pipe->inode->i_mutex);
1636         return ret;
1637 }
1638
1639 /*
1640  * Link contents of ipipe to opipe.
1641  */
1642 static int link_pipe(struct pipe_inode_info *ipipe,
1643                      struct pipe_inode_info *opipe,
1644                      size_t len, unsigned int flags)
1645 {
1646         struct pipe_buffer *ibuf, *obuf;
1647         int ret = 0, i = 0, nbuf;
1648
1649         /*
1650          * Potential ABBA deadlock, work around it by ordering lock
1651          * grabbing by inode address. Otherwise two different processes
1652          * could deadlock (one doing tee from A -> B, the other from B -> A).
1653          */
1654         inode_double_lock(ipipe->inode, opipe->inode);
1655
1656         do {
1657                 if (!opipe->readers) {
1658                         send_sig(SIGPIPE, current, 0);
1659                         if (!ret)
1660                                 ret = -EPIPE;
1661                         break;
1662                 }
1663
1664                 /*
1665                  * If we have iterated all input buffers or ran out of
1666                  * output room, break.
1667                  */
1668                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1669                         break;
1670
1671                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1672                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1673
1674                 /*
1675                  * Get a reference to this pipe buffer,
1676                  * so we can copy the contents over.
1677                  */
1678                 ibuf->ops->get(ipipe, ibuf);
1679
1680                 obuf = opipe->bufs + nbuf;
1681                 *obuf = *ibuf;
1682
1683                 /*
1684                  * Don't inherit the gift flag, we need to
1685                  * prevent multiple steals of this page.
1686                  */
1687                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1688
1689                 if (obuf->len > len)
1690                         obuf->len = len;
1691
1692                 opipe->nrbufs++;
1693                 ret += obuf->len;
1694                 len -= obuf->len;
1695                 i++;
1696         } while (len);
1697
1698         inode_double_unlock(ipipe->inode, opipe->inode);
1699
1700         /*
1701          * If we put data in the output pipe, wakeup any potential readers.
1702          */
1703         if (ret > 0) {
1704                 smp_mb();
1705                 if (waitqueue_active(&opipe->wait))
1706                         wake_up_interruptible(&opipe->wait);
1707                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1708         }
1709
1710         return ret;
1711 }
1712
1713 /*
1714  * This is a tee(1) implementation that works on pipes. It doesn't copy
1715  * any data, it simply references the 'in' pages on the 'out' pipe.
1716  * The 'flags' used are the SPLICE_F_* variants, currently the only
1717  * applicable one is SPLICE_F_NONBLOCK.
1718  */
1719 static long do_tee(struct file *in, struct file *out, size_t len,
1720                    unsigned int flags)
1721 {
1722         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1723         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1724         int ret = -EINVAL;
1725
1726         /*
1727          * Duplicate the contents of ipipe to opipe without actually
1728          * copying the data.
1729          */
1730         if (ipipe && opipe && ipipe != opipe) {
1731                 /*
1732                  * Keep going, unless we encounter an error. The ipipe/opipe
1733                  * ordering doesn't really matter.
1734                  */
1735                 ret = link_ipipe_prep(ipipe, flags);
1736                 if (!ret) {
1737                         ret = link_opipe_prep(opipe, flags);
1738                         if (!ret) {
1739                                 ret = link_pipe(ipipe, opipe, len, flags);
1740                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1741                                         ret = -EAGAIN;
1742                         }
1743                 }
1744         }
1745
1746         return ret;
1747 }
1748
1749 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1750 {
1751         struct file *in;
1752         int error, fput_in;
1753
1754         if (unlikely(!len))
1755                 return 0;
1756
1757         error = -EBADF;
1758         in = fget_light(fdin, &fput_in);
1759         if (in) {
1760                 if (in->f_mode & FMODE_READ) {
1761                         int fput_out;
1762                         struct file *out = fget_light(fdout, &fput_out);
1763
1764                         if (out) {
1765                                 if (out->f_mode & FMODE_WRITE)
1766                                         error = do_tee(in, out, len, flags);
1767                                 fput_light(out, fput_out);
1768                         }
1769                 }
1770                 fput_light(in, fput_in);
1771         }
1772
1773         return error;
1774 }