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