splice: fix double kunmap() in vmsplice copy path
[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 tuples, along 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, req_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         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
285         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
286
287         /*
288          * Lookup the (hopefully) full range of pages we need.
289          */
290         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
291         index += spd.nr_pages;
292
293         /*
294          * If find_get_pages_contig() returned fewer pages than we needed,
295          * readahead/allocate the rest and fill in the holes.
296          */
297         if (spd.nr_pages < nr_pages)
298                 page_cache_sync_readahead(mapping, &in->f_ra, in,
299                                 index, req_pages - spd.nr_pages);
300
301         error = 0;
302         while (spd.nr_pages < nr_pages) {
303                 /*
304                  * Page could be there, find_get_pages_contig() breaks on
305                  * the first hole.
306                  */
307                 page = find_get_page(mapping, index);
308                 if (!page) {
309                         /*
310                          * page didn't exist, allocate one.
311                          */
312                         page = page_cache_alloc_cold(mapping);
313                         if (!page)
314                                 break;
315
316                         error = add_to_page_cache_lru(page, mapping, index,
317                                               GFP_KERNEL);
318                         if (unlikely(error)) {
319                                 page_cache_release(page);
320                                 if (error == -EEXIST)
321                                         continue;
322                                 break;
323                         }
324                         /*
325                          * add_to_page_cache() locks the page, unlock it
326                          * to avoid convoluting the logic below even more.
327                          */
328                         unlock_page(page);
329                 }
330
331                 pages[spd.nr_pages++] = page;
332                 index++;
333         }
334
335         /*
336          * Now loop over the map and see if we need to start IO on any
337          * pages, fill in the partial map, etc.
338          */
339         index = *ppos >> PAGE_CACHE_SHIFT;
340         nr_pages = spd.nr_pages;
341         spd.nr_pages = 0;
342         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
343                 unsigned int this_len;
344
345                 if (!len)
346                         break;
347
348                 /*
349                  * this_len is the max we'll use from this page
350                  */
351                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
352                 page = pages[page_nr];
353
354                 if (PageReadahead(page))
355                         page_cache_async_readahead(mapping, &in->f_ra, in,
356                                         page, index, req_pages - page_nr);
357
358                 /*
359                  * If the page isn't uptodate, we may need to start io on it
360                  */
361                 if (!PageUptodate(page)) {
362                         /*
363                          * If in nonblock mode then dont block on waiting
364                          * for an in-flight io page
365                          */
366                         if (flags & SPLICE_F_NONBLOCK) {
367                                 if (TestSetPageLocked(page))
368                                         break;
369                         } else
370                                 lock_page(page);
371
372                         /*
373                          * page was truncated, stop here. if this isn't the
374                          * first page, we'll just complete what we already
375                          * added
376                          */
377                         if (!page->mapping) {
378                                 unlock_page(page);
379                                 break;
380                         }
381                         /*
382                          * page was already under io and is now done, great
383                          */
384                         if (PageUptodate(page)) {
385                                 unlock_page(page);
386                                 goto fill_it;
387                         }
388
389                         /*
390                          * need to read in the page
391                          */
392                         error = mapping->a_ops->readpage(in, page);
393                         if (unlikely(error)) {
394                                 /*
395                                  * We really should re-lookup the page here,
396                                  * but it complicates things a lot. Instead
397                                  * lets just do what we already stored, and
398                                  * we'll get it the next time we are called.
399                                  */
400                                 if (error == AOP_TRUNCATED_PAGE)
401                                         error = 0;
402
403                                 break;
404                         }
405                 }
406 fill_it:
407                 /*
408                  * i_size must be checked after PageUptodate.
409                  */
410                 isize = i_size_read(mapping->host);
411                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
412                 if (unlikely(!isize || index > end_index))
413                         break;
414
415                 /*
416                  * if this is the last page, see if we need to shrink
417                  * the length and stop
418                  */
419                 if (end_index == index) {
420                         unsigned int plen;
421
422                         /*
423                          * max good bytes in this page
424                          */
425                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
426                         if (plen <= loff)
427                                 break;
428
429                         /*
430                          * force quit after adding this page
431                          */
432                         this_len = min(this_len, plen - loff);
433                         len = this_len;
434                 }
435
436                 partial[page_nr].offset = loff;
437                 partial[page_nr].len = this_len;
438                 len -= this_len;
439                 loff = 0;
440                 spd.nr_pages++;
441                 index++;
442         }
443
444         /*
445          * Release any pages at the end, if we quit early. 'page_nr' is how far
446          * we got, 'nr_pages' is how many pages are in the map.
447          */
448         while (page_nr < nr_pages)
449                 page_cache_release(pages[page_nr++]);
450         in->f_ra.prev_index = index;
451
452         if (spd.nr_pages)
453                 return splice_to_pipe(pipe, &spd);
454
455         return error;
456 }
457
458 /**
459  * generic_file_splice_read - splice data from file to a pipe
460  * @in:         file to splice from
461  * @ppos:       position in @in
462  * @pipe:       pipe to splice to
463  * @len:        number of bytes to splice
464  * @flags:      splice modifier flags
465  *
466  * Description:
467  *    Will read pages from given file and fill them into a pipe. Can be
468  *    used as long as the address_space operations for the source implements
469  *    a readpage() hook.
470  *
471  */
472 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
473                                  struct pipe_inode_info *pipe, size_t len,
474                                  unsigned int flags)
475 {
476         ssize_t spliced;
477         int ret;
478         loff_t isize, left;
479
480         isize = i_size_read(in->f_mapping->host);
481         if (unlikely(*ppos >= isize))
482                 return 0;
483
484         left = isize - *ppos;
485         if (unlikely(left < len))
486                 len = left;
487
488         ret = 0;
489         spliced = 0;
490         while (len && !spliced) {
491                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
492
493                 if (ret < 0)
494                         break;
495                 else if (!ret) {
496                         if (spliced)
497                                 break;
498                         if (flags & SPLICE_F_NONBLOCK) {
499                                 ret = -EAGAIN;
500                                 break;
501                         }
502                 }
503
504                 *ppos += ret;
505                 len -= ret;
506                 spliced += ret;
507         }
508
509         if (spliced)
510                 return spliced;
511
512         return ret;
513 }
514
515 EXPORT_SYMBOL(generic_file_splice_read);
516
517 /*
518  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
519  * using sendpage(). Return the number of bytes sent.
520  */
521 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
522                             struct pipe_buffer *buf, struct splice_desc *sd)
523 {
524         struct file *file = sd->u.file;
525         loff_t pos = sd->pos;
526         int ret, more;
527
528         ret = buf->ops->confirm(pipe, buf);
529         if (!ret) {
530                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
531
532                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
533                                            sd->len, &pos, more);
534         }
535
536         return ret;
537 }
538
539 /*
540  * This is a little more tricky than the file -> pipe splicing. There are
541  * basically three cases:
542  *
543  *      - Destination page already exists in the address space and there
544  *        are users of it. For that case we have no other option that
545  *        copying the data. Tough luck.
546  *      - Destination page already exists in the address space, but there
547  *        are no users of it. Make sure it's uptodate, then drop it. Fall
548  *        through to last case.
549  *      - Destination page does not exist, we can add the pipe page to
550  *        the page cache and avoid the copy.
551  *
552  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
553  * sd->flags), we attempt to migrate pages from the pipe to the output
554  * file address space page cache. This is possible if no one else has
555  * the pipe page referenced outside of the pipe and page cache. If
556  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
557  * a new page in the output file page cache and fill/dirty that.
558  */
559 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
560                         struct splice_desc *sd)
561 {
562         struct file *file = sd->u.file;
563         struct address_space *mapping = file->f_mapping;
564         unsigned int offset, this_len;
565         struct page *page;
566         pgoff_t index;
567         int ret;
568
569         /*
570          * make sure the data in this buffer is uptodate
571          */
572         ret = buf->ops->confirm(pipe, buf);
573         if (unlikely(ret))
574                 return ret;
575
576         index = sd->pos >> PAGE_CACHE_SHIFT;
577         offset = sd->pos & ~PAGE_CACHE_MASK;
578
579         this_len = sd->len;
580         if (this_len + offset > PAGE_CACHE_SIZE)
581                 this_len = PAGE_CACHE_SIZE - offset;
582
583 find_page:
584         page = find_lock_page(mapping, index);
585         if (!page) {
586                 ret = -ENOMEM;
587                 page = page_cache_alloc_cold(mapping);
588                 if (unlikely(!page))
589                         goto out_ret;
590
591                 /*
592                  * This will also lock the page
593                  */
594                 ret = add_to_page_cache_lru(page, mapping, index,
595                                             GFP_KERNEL);
596                 if (unlikely(ret))
597                         goto out_release;
598         }
599
600         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
601         if (unlikely(ret)) {
602                 loff_t isize = i_size_read(mapping->host);
603
604                 if (ret != AOP_TRUNCATED_PAGE)
605                         unlock_page(page);
606                 page_cache_release(page);
607                 if (ret == AOP_TRUNCATED_PAGE)
608                         goto find_page;
609
610                 /*
611                  * prepare_write() may have instantiated a few blocks
612                  * outside i_size.  Trim these off again.
613                  */
614                 if (sd->pos + this_len > isize)
615                         vmtruncate(mapping->host, isize);
616
617                 goto out_ret;
618         }
619
620         if (buf->page != page) {
621                 /*
622                  * Careful, ->map() uses KM_USER0!
623                  */
624                 char *src = buf->ops->map(pipe, buf, 1);
625                 char *dst = kmap_atomic(page, KM_USER1);
626
627                 memcpy(dst + offset, src + buf->offset, this_len);
628                 flush_dcache_page(page);
629                 kunmap_atomic(dst, KM_USER1);
630                 buf->ops->unmap(pipe, buf, src);
631         }
632
633         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
634         if (ret) {
635                 if (ret == AOP_TRUNCATED_PAGE) {
636                         page_cache_release(page);
637                         goto find_page;
638                 }
639                 if (ret < 0)
640                         goto out;
641                 /*
642                  * Partial write has happened, so 'ret' already initialized by
643                  * number of bytes written, Where is nothing we have to do here.
644                  */
645         } else
646                 ret = this_len;
647         /*
648          * Return the number of bytes written and mark page as
649          * accessed, we are now done!
650          */
651         mark_page_accessed(page);
652 out:
653         unlock_page(page);
654 out_release:
655         page_cache_release(page);
656 out_ret:
657         return ret;
658 }
659
660 /**
661  * __splice_from_pipe - splice data from a pipe to given actor
662  * @pipe:       pipe to splice from
663  * @sd:         information to @actor
664  * @actor:      handler that splices the data
665  *
666  * Description:
667  *    This function does little more than loop over the pipe and call
668  *    @actor to do the actual moving of a single struct pipe_buffer to
669  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
670  *    pipe_to_user.
671  *
672  */
673 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
674                            splice_actor *actor)
675 {
676         int ret, do_wakeup, err;
677
678         ret = 0;
679         do_wakeup = 0;
680
681         for (;;) {
682                 if (pipe->nrbufs) {
683                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
684                         const struct pipe_buf_operations *ops = buf->ops;
685
686                         sd->len = buf->len;
687                         if (sd->len > sd->total_len)
688                                 sd->len = sd->total_len;
689
690                         err = actor(pipe, buf, sd);
691                         if (err <= 0) {
692                                 if (!ret && err != -ENODATA)
693                                         ret = err;
694
695                                 break;
696                         }
697
698                         ret += err;
699                         buf->offset += err;
700                         buf->len -= err;
701
702                         sd->len -= err;
703                         sd->pos += err;
704                         sd->total_len -= err;
705                         if (sd->len)
706                                 continue;
707
708                         if (!buf->len) {
709                                 buf->ops = NULL;
710                                 ops->release(pipe, buf);
711                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
712                                 pipe->nrbufs--;
713                                 if (pipe->inode)
714                                         do_wakeup = 1;
715                         }
716
717                         if (!sd->total_len)
718                                 break;
719                 }
720
721                 if (pipe->nrbufs)
722                         continue;
723                 if (!pipe->writers)
724                         break;
725                 if (!pipe->waiting_writers) {
726                         if (ret)
727                                 break;
728                 }
729
730                 if (sd->flags & SPLICE_F_NONBLOCK) {
731                         if (!ret)
732                                 ret = -EAGAIN;
733                         break;
734                 }
735
736                 if (signal_pending(current)) {
737                         if (!ret)
738                                 ret = -ERESTARTSYS;
739                         break;
740                 }
741
742                 if (do_wakeup) {
743                         smp_mb();
744                         if (waitqueue_active(&pipe->wait))
745                                 wake_up_interruptible_sync(&pipe->wait);
746                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
747                         do_wakeup = 0;
748                 }
749
750                 pipe_wait(pipe);
751         }
752
753         if (do_wakeup) {
754                 smp_mb();
755                 if (waitqueue_active(&pipe->wait))
756                         wake_up_interruptible(&pipe->wait);
757                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
758         }
759
760         return ret;
761 }
762 EXPORT_SYMBOL(__splice_from_pipe);
763
764 /**
765  * splice_from_pipe - splice data from a pipe to a file
766  * @pipe:       pipe to splice from
767  * @out:        file to splice to
768  * @ppos:       position in @out
769  * @len:        how many bytes to splice
770  * @flags:      splice modifier flags
771  * @actor:      handler that splices the data
772  *
773  * Description:
774  *    See __splice_from_pipe. This function locks the input and output inodes,
775  *    otherwise it's identical to __splice_from_pipe().
776  *
777  */
778 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
779                          loff_t *ppos, size_t len, unsigned int flags,
780                          splice_actor *actor)
781 {
782         ssize_t ret;
783         struct inode *inode = out->f_mapping->host;
784         struct splice_desc sd = {
785                 .total_len = len,
786                 .flags = flags,
787                 .pos = *ppos,
788                 .u.file = out,
789         };
790
791         /*
792          * The actor worker might be calling ->prepare_write and
793          * ->commit_write. Most of the time, these expect i_mutex to
794          * be held. Since this may result in an ABBA deadlock with
795          * pipe->inode, we have to order lock acquiry here.
796          */
797         inode_double_lock(inode, pipe->inode);
798         ret = __splice_from_pipe(pipe, &sd, actor);
799         inode_double_unlock(inode, pipe->inode);
800
801         return ret;
802 }
803
804 /**
805  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
806  * @pipe:       pipe info
807  * @out:        file to write to
808  * @ppos:       position in @out
809  * @len:        number of bytes to splice
810  * @flags:      splice modifier flags
811  *
812  * Description:
813  *    Will either move or copy pages (determined by @flags options) from
814  *    the given pipe inode to the given file. The caller is responsible
815  *    for acquiring i_mutex on both inodes.
816  *
817  */
818 ssize_t
819 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
820                                  loff_t *ppos, size_t len, unsigned int flags)
821 {
822         struct address_space *mapping = out->f_mapping;
823         struct inode *inode = mapping->host;
824         struct splice_desc sd = {
825                 .total_len = len,
826                 .flags = flags,
827                 .pos = *ppos,
828                 .u.file = out,
829         };
830         ssize_t ret;
831         int err;
832
833         err = remove_suid(out->f_path.dentry);
834         if (unlikely(err))
835                 return err;
836
837         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
838         if (ret > 0) {
839                 unsigned long nr_pages;
840
841                 *ppos += ret;
842                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
843
844                 /*
845                  * If file or inode is SYNC and we actually wrote some data,
846                  * sync it.
847                  */
848                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
849                         err = generic_osync_inode(inode, mapping,
850                                                   OSYNC_METADATA|OSYNC_DATA);
851
852                         if (err)
853                                 ret = err;
854                 }
855                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
856         }
857
858         return ret;
859 }
860
861 EXPORT_SYMBOL(generic_file_splice_write_nolock);
862
863 /**
864  * generic_file_splice_write - splice data from a pipe to a file
865  * @pipe:       pipe info
866  * @out:        file to write to
867  * @ppos:       position in @out
868  * @len:        number of bytes to splice
869  * @flags:      splice modifier flags
870  *
871  * Description:
872  *    Will either move or copy pages (determined by @flags options) from
873  *    the given pipe inode to the given file.
874  *
875  */
876 ssize_t
877 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
878                           loff_t *ppos, size_t len, unsigned int flags)
879 {
880         struct address_space *mapping = out->f_mapping;
881         struct inode *inode = mapping->host;
882         ssize_t ret;
883         int err;
884
885         err = should_remove_suid(out->f_path.dentry);
886         if (unlikely(err)) {
887                 mutex_lock(&inode->i_mutex);
888                 err = __remove_suid(out->f_path.dentry, err);
889                 mutex_unlock(&inode->i_mutex);
890                 if (err)
891                         return err;
892         }
893
894         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
895         if (ret > 0) {
896                 unsigned long nr_pages;
897
898                 *ppos += ret;
899                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
900
901                 /*
902                  * If file or inode is SYNC and we actually wrote some data,
903                  * sync it.
904                  */
905                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
906                         mutex_lock(&inode->i_mutex);
907                         err = generic_osync_inode(inode, mapping,
908                                                   OSYNC_METADATA|OSYNC_DATA);
909                         mutex_unlock(&inode->i_mutex);
910
911                         if (err)
912                                 ret = err;
913                 }
914                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
915         }
916
917         return ret;
918 }
919
920 EXPORT_SYMBOL(generic_file_splice_write);
921
922 /**
923  * generic_splice_sendpage - splice data from a pipe to a socket
924  * @pipe:       pipe to splice from
925  * @out:        socket to write to
926  * @ppos:       position in @out
927  * @len:        number of bytes to splice
928  * @flags:      splice modifier flags
929  *
930  * Description:
931  *    Will send @len bytes from the pipe to a network socket. No data copying
932  *    is involved.
933  *
934  */
935 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
936                                 loff_t *ppos, size_t len, unsigned int flags)
937 {
938         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
939 }
940
941 EXPORT_SYMBOL(generic_splice_sendpage);
942
943 /*
944  * Attempt to initiate a splice from pipe to file.
945  */
946 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
947                            loff_t *ppos, size_t len, unsigned int flags)
948 {
949         int ret;
950
951         if (unlikely(!out->f_op || !out->f_op->splice_write))
952                 return -EINVAL;
953
954         if (unlikely(!(out->f_mode & FMODE_WRITE)))
955                 return -EBADF;
956
957         ret = rw_verify_area(WRITE, out, ppos, len);
958         if (unlikely(ret < 0))
959                 return ret;
960
961         ret = security_file_permission(out, MAY_WRITE);
962         if (unlikely(ret < 0))
963                 return ret;
964
965         return out->f_op->splice_write(pipe, out, ppos, len, flags);
966 }
967
968 /*
969  * Attempt to initiate a splice from a file to a pipe.
970  */
971 static long do_splice_to(struct file *in, loff_t *ppos,
972                          struct pipe_inode_info *pipe, size_t len,
973                          unsigned int flags)
974 {
975         int ret;
976
977         if (unlikely(!in->f_op || !in->f_op->splice_read))
978                 return -EINVAL;
979
980         if (unlikely(!(in->f_mode & FMODE_READ)))
981                 return -EBADF;
982
983         ret = rw_verify_area(READ, in, ppos, len);
984         if (unlikely(ret < 0))
985                 return ret;
986
987         ret = security_file_permission(in, MAY_READ);
988         if (unlikely(ret < 0))
989                 return ret;
990
991         return in->f_op->splice_read(in, ppos, pipe, len, flags);
992 }
993
994 /**
995  * splice_direct_to_actor - splices data directly between two non-pipes
996  * @in:         file to splice from
997  * @sd:         actor information on where to splice to
998  * @actor:      handles the data splicing
999  *
1000  * Description:
1001  *    This is a special case helper to splice directly between two
1002  *    points, without requiring an explicit pipe. Internally an allocated
1003  *    pipe is cached in the process, and reused during the lifetime of
1004  *    that process.
1005  *
1006  */
1007 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1008                                splice_direct_actor *actor)
1009 {
1010         struct pipe_inode_info *pipe;
1011         long ret, bytes;
1012         umode_t i_mode;
1013         size_t len;
1014         int i, flags;
1015
1016         /*
1017          * We require the input being a regular file, as we don't want to
1018          * randomly drop data for eg socket -> socket splicing. Use the
1019          * piped splicing for that!
1020          */
1021         i_mode = in->f_path.dentry->d_inode->i_mode;
1022         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1023                 return -EINVAL;
1024
1025         /*
1026          * neither in nor out is a pipe, setup an internal pipe attached to
1027          * 'out' and transfer the wanted data from 'in' to 'out' through that
1028          */
1029         pipe = current->splice_pipe;
1030         if (unlikely(!pipe)) {
1031                 pipe = alloc_pipe_info(NULL);
1032                 if (!pipe)
1033                         return -ENOMEM;
1034
1035                 /*
1036                  * We don't have an immediate reader, but we'll read the stuff
1037                  * out of the pipe right after the splice_to_pipe(). So set
1038                  * PIPE_READERS appropriately.
1039                  */
1040                 pipe->readers = 1;
1041
1042                 current->splice_pipe = pipe;
1043         }
1044
1045         /*
1046          * Do the splice.
1047          */
1048         ret = 0;
1049         bytes = 0;
1050         len = sd->total_len;
1051         flags = sd->flags;
1052
1053         /*
1054          * Don't block on output, we have to drain the direct pipe.
1055          */
1056         sd->flags &= ~SPLICE_F_NONBLOCK;
1057
1058         while (len) {
1059                 size_t read_len;
1060                 loff_t pos = sd->pos;
1061
1062                 ret = do_splice_to(in, &pos, pipe, len, flags);
1063                 if (unlikely(ret <= 0))
1064                         goto out_release;
1065
1066                 read_len = ret;
1067                 sd->total_len = read_len;
1068
1069                 /*
1070                  * NOTE: nonblocking mode only applies to the input. We
1071                  * must not do the output in nonblocking mode as then we
1072                  * could get stuck data in the internal pipe:
1073                  */
1074                 ret = actor(pipe, sd);
1075                 if (unlikely(ret <= 0))
1076                         goto out_release;
1077
1078                 bytes += ret;
1079                 len -= ret;
1080                 sd->pos = pos;
1081
1082                 if (ret < read_len)
1083                         goto out_release;
1084         }
1085
1086         pipe->nrbufs = pipe->curbuf = 0;
1087         return bytes;
1088
1089 out_release:
1090         /*
1091          * If we did an incomplete transfer we must release
1092          * the pipe buffers in question:
1093          */
1094         for (i = 0; i < PIPE_BUFFERS; i++) {
1095                 struct pipe_buffer *buf = pipe->bufs + i;
1096
1097                 if (buf->ops) {
1098                         buf->ops->release(pipe, buf);
1099                         buf->ops = NULL;
1100                 }
1101         }
1102         pipe->nrbufs = pipe->curbuf = 0;
1103
1104         /*
1105          * If we transferred some data, return the number of bytes:
1106          */
1107         if (bytes > 0)
1108                 return bytes;
1109
1110         return ret;
1111
1112 }
1113 EXPORT_SYMBOL(splice_direct_to_actor);
1114
1115 static int direct_splice_actor(struct pipe_inode_info *pipe,
1116                                struct splice_desc *sd)
1117 {
1118         struct file *file = sd->u.file;
1119
1120         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1121 }
1122
1123 /**
1124  * do_splice_direct - splices data directly between two files
1125  * @in:         file to splice from
1126  * @ppos:       input file offset
1127  * @out:        file to splice to
1128  * @len:        number of bytes to splice
1129  * @flags:      splice modifier flags
1130  *
1131  * Description:
1132  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1133  *    doing it in the application would incur an extra system call
1134  *    (splice in + splice out, as compared to just sendfile()). So this helper
1135  *    can splice directly through a process-private pipe.
1136  *
1137  */
1138 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1139                       size_t len, unsigned int flags)
1140 {
1141         struct splice_desc sd = {
1142                 .len            = len,
1143                 .total_len      = len,
1144                 .flags          = flags,
1145                 .pos            = *ppos,
1146                 .u.file         = out,
1147         };
1148         long ret;
1149
1150         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1151         if (ret > 0)
1152                 *ppos += ret;
1153
1154         return ret;
1155 }
1156
1157 /*
1158  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1159  * location, so checking ->i_pipe is not enough to verify that this is a
1160  * pipe.
1161  */
1162 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1163 {
1164         if (S_ISFIFO(inode->i_mode))
1165                 return inode->i_pipe;
1166
1167         return NULL;
1168 }
1169
1170 /*
1171  * Determine where to splice to/from.
1172  */
1173 static long do_splice(struct file *in, loff_t __user *off_in,
1174                       struct file *out, loff_t __user *off_out,
1175                       size_t len, unsigned int flags)
1176 {
1177         struct pipe_inode_info *pipe;
1178         loff_t offset, *off;
1179         long ret;
1180
1181         pipe = pipe_info(in->f_path.dentry->d_inode);
1182         if (pipe) {
1183                 if (off_in)
1184                         return -ESPIPE;
1185                 if (off_out) {
1186                         if (out->f_op->llseek == no_llseek)
1187                                 return -EINVAL;
1188                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1189                                 return -EFAULT;
1190                         off = &offset;
1191                 } else
1192                         off = &out->f_pos;
1193
1194                 ret = do_splice_from(pipe, out, off, len, flags);
1195
1196                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1197                         ret = -EFAULT;
1198
1199                 return ret;
1200         }
1201
1202         pipe = pipe_info(out->f_path.dentry->d_inode);
1203         if (pipe) {
1204                 if (off_out)
1205                         return -ESPIPE;
1206                 if (off_in) {
1207                         if (in->f_op->llseek == no_llseek)
1208                                 return -EINVAL;
1209                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1210                                 return -EFAULT;
1211                         off = &offset;
1212                 } else
1213                         off = &in->f_pos;
1214
1215                 ret = do_splice_to(in, off, pipe, len, flags);
1216
1217                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1218                         ret = -EFAULT;
1219
1220                 return ret;
1221         }
1222
1223         return -EINVAL;
1224 }
1225
1226 /*
1227  * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1228  * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1229  * for writing) and page faulting on the user memory pointed to by src.
1230  * This assumes that we will very rarely hit the partial != 0 path, or this
1231  * will not be a win.
1232  */
1233 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1234 {
1235         int partial;
1236
1237         pagefault_disable();
1238         partial = __copy_from_user_inatomic(dst, src, n);
1239         pagefault_enable();
1240
1241         /*
1242          * Didn't copy everything, drop the mmap_sem and do a faulting copy
1243          */
1244         if (unlikely(partial)) {
1245                 up_read(&current->mm->mmap_sem);
1246                 partial = copy_from_user(dst, src, n);
1247                 down_read(&current->mm->mmap_sem);
1248         }
1249
1250         return partial;
1251 }
1252
1253 /*
1254  * Map an iov into an array of pages and offset/length tupples. With the
1255  * partial_page structure, we can map several non-contiguous ranges into
1256  * our ones pages[] map instead of splitting that operation into pieces.
1257  * Could easily be exported as a generic helper for other users, in which
1258  * case one would probably want to add a 'max_nr_pages' parameter as well.
1259  */
1260 static int get_iovec_page_array(const struct iovec __user *iov,
1261                                 unsigned int nr_vecs, struct page **pages,
1262                                 struct partial_page *partial, int aligned)
1263 {
1264         int buffers = 0, error = 0;
1265
1266         down_read(&current->mm->mmap_sem);
1267
1268         while (nr_vecs) {
1269                 unsigned long off, npages;
1270                 struct iovec entry;
1271                 void __user *base;
1272                 size_t len;
1273                 int i;
1274
1275                 error = -EFAULT;
1276                 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1277                         break;
1278
1279                 base = entry.iov_base;
1280                 len = entry.iov_len;
1281
1282                 /*
1283                  * Sanity check this iovec. 0 read succeeds.
1284                  */
1285                 error = 0;
1286                 if (unlikely(!len))
1287                         break;
1288                 error = -EFAULT;
1289                 if (unlikely(!base))
1290                         break;
1291
1292                 /*
1293                  * Get this base offset and number of pages, then map
1294                  * in the user pages.
1295                  */
1296                 off = (unsigned long) base & ~PAGE_MASK;
1297
1298                 /*
1299                  * If asked for alignment, the offset must be zero and the
1300                  * length a multiple of the PAGE_SIZE.
1301                  */
1302                 error = -EINVAL;
1303                 if (aligned && (off || len & ~PAGE_MASK))
1304                         break;
1305
1306                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1307                 if (npages > PIPE_BUFFERS - buffers)
1308                         npages = PIPE_BUFFERS - buffers;
1309
1310                 error = get_user_pages(current, current->mm,
1311                                        (unsigned long) base, npages, 0, 0,
1312                                        &pages[buffers], NULL);
1313
1314                 if (unlikely(error <= 0))
1315                         break;
1316
1317                 /*
1318                  * Fill this contiguous range into the partial page map.
1319                  */
1320                 for (i = 0; i < error; i++) {
1321                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1322
1323                         partial[buffers].offset = off;
1324                         partial[buffers].len = plen;
1325
1326                         off = 0;
1327                         len -= plen;
1328                         buffers++;
1329                 }
1330
1331                 /*
1332                  * We didn't complete this iov, stop here since it probably
1333                  * means we have to move some of this into a pipe to
1334                  * be able to continue.
1335                  */
1336                 if (len)
1337                         break;
1338
1339                 /*
1340                  * Don't continue if we mapped fewer pages than we asked for,
1341                  * or if we mapped the max number of pages that we have
1342                  * room for.
1343                  */
1344                 if (error < npages || buffers == PIPE_BUFFERS)
1345                         break;
1346
1347                 nr_vecs--;
1348                 iov++;
1349         }
1350
1351         up_read(&current->mm->mmap_sem);
1352
1353         if (buffers)
1354                 return buffers;
1355
1356         return error;
1357 }
1358
1359 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1360                         struct splice_desc *sd)
1361 {
1362         char *src;
1363         int ret;
1364
1365         ret = buf->ops->confirm(pipe, buf);
1366         if (unlikely(ret))
1367                 return ret;
1368
1369         /*
1370          * See if we can use the atomic maps, by prefaulting in the
1371          * pages and doing an atomic copy
1372          */
1373         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1374                 src = buf->ops->map(pipe, buf, 1);
1375                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1376                                                         sd->len);
1377                 buf->ops->unmap(pipe, buf, src);
1378                 if (!ret) {
1379                         ret = sd->len;
1380                         goto out;
1381                 }
1382         }
1383
1384         /*
1385          * No dice, use slow non-atomic map and copy
1386          */
1387         src = buf->ops->map(pipe, buf, 0);
1388
1389         ret = sd->len;
1390         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1391                 ret = -EFAULT;
1392
1393         buf->ops->unmap(pipe, buf, src);
1394 out:
1395         if (ret > 0)
1396                 sd->u.userptr += ret;
1397         return ret;
1398 }
1399
1400 /*
1401  * For lack of a better implementation, implement vmsplice() to userspace
1402  * as a simple copy of the pipes pages to the user iov.
1403  */
1404 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1405                              unsigned long nr_segs, unsigned int flags)
1406 {
1407         struct pipe_inode_info *pipe;
1408         struct splice_desc sd;
1409         ssize_t size;
1410         int error;
1411         long ret;
1412
1413         pipe = pipe_info(file->f_path.dentry->d_inode);
1414         if (!pipe)
1415                 return -EBADF;
1416
1417         if (pipe->inode)
1418                 mutex_lock(&pipe->inode->i_mutex);
1419
1420         error = ret = 0;
1421         while (nr_segs) {
1422                 void __user *base;
1423                 size_t len;
1424
1425                 /*
1426                  * Get user address base and length for this iovec.
1427                  */
1428                 error = get_user(base, &iov->iov_base);
1429                 if (unlikely(error))
1430                         break;
1431                 error = get_user(len, &iov->iov_len);
1432                 if (unlikely(error))
1433                         break;
1434
1435                 /*
1436                  * Sanity check this iovec. 0 read succeeds.
1437                  */
1438                 if (unlikely(!len))
1439                         break;
1440                 if (unlikely(!base)) {
1441                         error = -EFAULT;
1442                         break;
1443                 }
1444
1445                 sd.len = 0;
1446                 sd.total_len = len;
1447                 sd.flags = flags;
1448                 sd.u.userptr = base;
1449                 sd.pos = 0;
1450
1451                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1452                 if (size < 0) {
1453                         if (!ret)
1454                                 ret = size;
1455
1456                         break;
1457                 }
1458
1459                 ret += size;
1460
1461                 if (size < len)
1462                         break;
1463
1464                 nr_segs--;
1465                 iov++;
1466         }
1467
1468         if (pipe->inode)
1469                 mutex_unlock(&pipe->inode->i_mutex);
1470
1471         if (!ret)
1472                 ret = error;
1473
1474         return ret;
1475 }
1476
1477 /*
1478  * vmsplice splices a user address range into a pipe. It can be thought of
1479  * as splice-from-memory, where the regular splice is splice-from-file (or
1480  * to file). In both cases the output is a pipe, naturally.
1481  */
1482 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1483                              unsigned long nr_segs, unsigned int flags)
1484 {
1485         struct pipe_inode_info *pipe;
1486         struct page *pages[PIPE_BUFFERS];
1487         struct partial_page partial[PIPE_BUFFERS];
1488         struct splice_pipe_desc spd = {
1489                 .pages = pages,
1490                 .partial = partial,
1491                 .flags = flags,
1492                 .ops = &user_page_pipe_buf_ops,
1493         };
1494
1495         pipe = pipe_info(file->f_path.dentry->d_inode);
1496         if (!pipe)
1497                 return -EBADF;
1498
1499         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1500                                             flags & SPLICE_F_GIFT);
1501         if (spd.nr_pages <= 0)
1502                 return spd.nr_pages;
1503
1504         return splice_to_pipe(pipe, &spd);
1505 }
1506
1507 /*
1508  * Note that vmsplice only really supports true splicing _from_ user memory
1509  * to a pipe, not the other way around. Splicing from user memory is a simple
1510  * operation that can be supported without any funky alignment restrictions
1511  * or nasty vm tricks. We simply map in the user memory and fill them into
1512  * a pipe. The reverse isn't quite as easy, though. There are two possible
1513  * solutions for that:
1514  *
1515  *      - memcpy() the data internally, at which point we might as well just
1516  *        do a regular read() on the buffer anyway.
1517  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1518  *        has restriction limitations on both ends of the pipe).
1519  *
1520  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1521  *
1522  */
1523 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1524                              unsigned long nr_segs, unsigned int flags)
1525 {
1526         struct file *file;
1527         long error;
1528         int fput;
1529
1530         if (unlikely(nr_segs > UIO_MAXIOV))
1531                 return -EINVAL;
1532         else if (unlikely(!nr_segs))
1533                 return 0;
1534
1535         error = -EBADF;
1536         file = fget_light(fd, &fput);
1537         if (file) {
1538                 if (file->f_mode & FMODE_WRITE)
1539                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1540                 else if (file->f_mode & FMODE_READ)
1541                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1542
1543                 fput_light(file, fput);
1544         }
1545
1546         return error;
1547 }
1548
1549 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1550                            int fd_out, loff_t __user *off_out,
1551                            size_t len, unsigned int flags)
1552 {
1553         long error;
1554         struct file *in, *out;
1555         int fput_in, fput_out;
1556
1557         if (unlikely(!len))
1558                 return 0;
1559
1560         error = -EBADF;
1561         in = fget_light(fd_in, &fput_in);
1562         if (in) {
1563                 if (in->f_mode & FMODE_READ) {
1564                         out = fget_light(fd_out, &fput_out);
1565                         if (out) {
1566                                 if (out->f_mode & FMODE_WRITE)
1567                                         error = do_splice(in, off_in,
1568                                                           out, off_out,
1569                                                           len, flags);
1570                                 fput_light(out, fput_out);
1571                         }
1572                 }
1573
1574                 fput_light(in, fput_in);
1575         }
1576
1577         return error;
1578 }
1579
1580 /*
1581  * Make sure there's data to read. Wait for input if we can, otherwise
1582  * return an appropriate error.
1583  */
1584 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1585 {
1586         int ret;
1587
1588         /*
1589          * Check ->nrbufs without the inode lock first. This function
1590          * is speculative anyways, so missing one is ok.
1591          */
1592         if (pipe->nrbufs)
1593                 return 0;
1594
1595         ret = 0;
1596         mutex_lock(&pipe->inode->i_mutex);
1597
1598         while (!pipe->nrbufs) {
1599                 if (signal_pending(current)) {
1600                         ret = -ERESTARTSYS;
1601                         break;
1602                 }
1603                 if (!pipe->writers)
1604                         break;
1605                 if (!pipe->waiting_writers) {
1606                         if (flags & SPLICE_F_NONBLOCK) {
1607                                 ret = -EAGAIN;
1608                                 break;
1609                         }
1610                 }
1611                 pipe_wait(pipe);
1612         }
1613
1614         mutex_unlock(&pipe->inode->i_mutex);
1615         return ret;
1616 }
1617
1618 /*
1619  * Make sure there's writeable room. Wait for room if we can, otherwise
1620  * return an appropriate error.
1621  */
1622 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1623 {
1624         int ret;
1625
1626         /*
1627          * Check ->nrbufs without the inode lock first. This function
1628          * is speculative anyways, so missing one is ok.
1629          */
1630         if (pipe->nrbufs < PIPE_BUFFERS)
1631                 return 0;
1632
1633         ret = 0;
1634         mutex_lock(&pipe->inode->i_mutex);
1635
1636         while (pipe->nrbufs >= PIPE_BUFFERS) {
1637                 if (!pipe->readers) {
1638                         send_sig(SIGPIPE, current, 0);
1639                         ret = -EPIPE;
1640                         break;
1641                 }
1642                 if (flags & SPLICE_F_NONBLOCK) {
1643                         ret = -EAGAIN;
1644                         break;
1645                 }
1646                 if (signal_pending(current)) {
1647                         ret = -ERESTARTSYS;
1648                         break;
1649                 }
1650                 pipe->waiting_writers++;
1651                 pipe_wait(pipe);
1652                 pipe->waiting_writers--;
1653         }
1654
1655         mutex_unlock(&pipe->inode->i_mutex);
1656         return ret;
1657 }
1658
1659 /*
1660  * Link contents of ipipe to opipe.
1661  */
1662 static int link_pipe(struct pipe_inode_info *ipipe,
1663                      struct pipe_inode_info *opipe,
1664                      size_t len, unsigned int flags)
1665 {
1666         struct pipe_buffer *ibuf, *obuf;
1667         int ret = 0, i = 0, nbuf;
1668
1669         /*
1670          * Potential ABBA deadlock, work around it by ordering lock
1671          * grabbing by inode address. Otherwise two different processes
1672          * could deadlock (one doing tee from A -> B, the other from B -> A).
1673          */
1674         inode_double_lock(ipipe->inode, opipe->inode);
1675
1676         do {
1677                 if (!opipe->readers) {
1678                         send_sig(SIGPIPE, current, 0);
1679                         if (!ret)
1680                                 ret = -EPIPE;
1681                         break;
1682                 }
1683
1684                 /*
1685                  * If we have iterated all input buffers or ran out of
1686                  * output room, break.
1687                  */
1688                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1689                         break;
1690
1691                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1692                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1693
1694                 /*
1695                  * Get a reference to this pipe buffer,
1696                  * so we can copy the contents over.
1697                  */
1698                 ibuf->ops->get(ipipe, ibuf);
1699
1700                 obuf = opipe->bufs + nbuf;
1701                 *obuf = *ibuf;
1702
1703                 /*
1704                  * Don't inherit the gift flag, we need to
1705                  * prevent multiple steals of this page.
1706                  */
1707                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1708
1709                 if (obuf->len > len)
1710                         obuf->len = len;
1711
1712                 opipe->nrbufs++;
1713                 ret += obuf->len;
1714                 len -= obuf->len;
1715                 i++;
1716         } while (len);
1717
1718         inode_double_unlock(ipipe->inode, opipe->inode);
1719
1720         /*
1721          * If we put data in the output pipe, wakeup any potential readers.
1722          */
1723         if (ret > 0) {
1724                 smp_mb();
1725                 if (waitqueue_active(&opipe->wait))
1726                         wake_up_interruptible(&opipe->wait);
1727                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1728         }
1729
1730         return ret;
1731 }
1732
1733 /*
1734  * This is a tee(1) implementation that works on pipes. It doesn't copy
1735  * any data, it simply references the 'in' pages on the 'out' pipe.
1736  * The 'flags' used are the SPLICE_F_* variants, currently the only
1737  * applicable one is SPLICE_F_NONBLOCK.
1738  */
1739 static long do_tee(struct file *in, struct file *out, size_t len,
1740                    unsigned int flags)
1741 {
1742         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1743         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1744         int ret = -EINVAL;
1745
1746         /*
1747          * Duplicate the contents of ipipe to opipe without actually
1748          * copying the data.
1749          */
1750         if (ipipe && opipe && ipipe != opipe) {
1751                 /*
1752                  * Keep going, unless we encounter an error. The ipipe/opipe
1753                  * ordering doesn't really matter.
1754                  */
1755                 ret = link_ipipe_prep(ipipe, flags);
1756                 if (!ret) {
1757                         ret = link_opipe_prep(opipe, flags);
1758                         if (!ret) {
1759                                 ret = link_pipe(ipipe, opipe, len, flags);
1760                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1761                                         ret = -EAGAIN;
1762                         }
1763                 }
1764         }
1765
1766         return ret;
1767 }
1768
1769 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1770 {
1771         struct file *in;
1772         int error, fput_in;
1773
1774         if (unlikely(!len))
1775                 return 0;
1776
1777         error = -EBADF;
1778         in = fget_light(fdin, &fput_in);
1779         if (in) {
1780                 if (in->f_mode & FMODE_READ) {
1781                         int fput_out;
1782                         struct file *out = fget_light(fdout, &fput_out);
1783
1784                         if (out) {
1785                                 if (out->f_mode & FMODE_WRITE)
1786                                         error = do_tee(in, out, len, flags);
1787                                 fput_light(out, fput_out);
1788                         }
1789                 }
1790                 fput_light(in, fput_in);
1791         }
1792
1793         return error;
1794 }