[PATCH] Take i_mutex in splice_from_pipe()
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/pipe_fs_i.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31
32 struct partial_page {
33         unsigned int offset;
34         unsigned int len;
35 };
36
37 /*
38  * Passed to splice_to_pipe
39  */
40 struct splice_pipe_desc {
41         struct page **pages;            /* page map */
42         struct partial_page *partial;   /* pages[] may not be contig */
43         int nr_pages;                   /* number of pages in map */
44         unsigned int flags;             /* splice flags */
45         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_ret;
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_ret;
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_ret:
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         struct inode *inode = out->f_mapping->host;
717
718         ret = 0;
719         do_wakeup = 0;
720
721         sd.total_len = len;
722         sd.flags = flags;
723         sd.file = out;
724         sd.pos = *ppos;
725
726         /*
727          * The actor worker might be calling ->prepare_write and
728          * ->commit_write. Most of the time, these expect i_mutex to
729          * be held. Since this may result in an ABBA deadlock with
730          * pipe->inode, we have to order lock acquiry here.
731          */
732         inode_double_lock(inode, pipe->inode);
733
734         for (;;) {
735                 if (pipe->nrbufs) {
736                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
737                         struct pipe_buf_operations *ops = buf->ops;
738
739                         sd.len = buf->len;
740                         if (sd.len > sd.total_len)
741                                 sd.len = sd.total_len;
742
743                         err = actor(pipe, buf, &sd);
744                         if (err <= 0) {
745                                 if (!ret && err != -ENODATA)
746                                         ret = err;
747
748                                 break;
749                         }
750
751                         ret += err;
752                         buf->offset += err;
753                         buf->len -= err;
754
755                         sd.len -= err;
756                         sd.pos += err;
757                         sd.total_len -= err;
758                         if (sd.len)
759                                 continue;
760
761                         if (!buf->len) {
762                                 buf->ops = NULL;
763                                 ops->release(pipe, buf);
764                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
765                                 pipe->nrbufs--;
766                                 if (pipe->inode)
767                                         do_wakeup = 1;
768                         }
769
770                         if (!sd.total_len)
771                                 break;
772                 }
773
774                 if (pipe->nrbufs)
775                         continue;
776                 if (!pipe->writers)
777                         break;
778                 if (!pipe->waiting_writers) {
779                         if (ret)
780                                 break;
781                 }
782
783                 if (flags & SPLICE_F_NONBLOCK) {
784                         if (!ret)
785                                 ret = -EAGAIN;
786                         break;
787                 }
788
789                 if (signal_pending(current)) {
790                         if (!ret)
791                                 ret = -ERESTARTSYS;
792                         break;
793                 }
794
795                 if (do_wakeup) {
796                         smp_mb();
797                         if (waitqueue_active(&pipe->wait))
798                                 wake_up_interruptible_sync(&pipe->wait);
799                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
800                         do_wakeup = 0;
801                 }
802
803                 pipe_wait(pipe);
804         }
805
806         inode_double_unlock(inode, pipe->inode);
807
808         if (do_wakeup) {
809                 smp_mb();
810                 if (waitqueue_active(&pipe->wait))
811                         wake_up_interruptible(&pipe->wait);
812                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
813         }
814
815         return ret;
816 }
817
818 /**
819  * generic_file_splice_write - splice data from a pipe to a file
820  * @pipe:       pipe info
821  * @out:        file to write to
822  * @len:        number of bytes to splice
823  * @flags:      splice modifier flags
824  *
825  * Will either move or copy pages (determined by @flags options) from
826  * the given pipe inode to the given file.
827  *
828  */
829 ssize_t
830 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
831                           loff_t *ppos, size_t len, unsigned int flags)
832 {
833         struct address_space *mapping = out->f_mapping;
834         ssize_t ret;
835
836         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
837         if (ret > 0) {
838                 struct inode *inode = mapping->host;
839
840                 *ppos += ret;
841
842                 /*
843                  * If file or inode is SYNC and we actually wrote some data,
844                  * sync it.
845                  */
846                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
847                         int err;
848
849                         mutex_lock(&inode->i_mutex);
850                         err = generic_osync_inode(inode, mapping,
851                                                   OSYNC_METADATA|OSYNC_DATA);
852                         mutex_unlock(&inode->i_mutex);
853
854                         if (err)
855                                 ret = err;
856                 }
857         }
858
859         return ret;
860 }
861
862 EXPORT_SYMBOL(generic_file_splice_write);
863
864 /**
865  * generic_splice_sendpage - splice data from a pipe to a socket
866  * @inode:      pipe inode
867  * @out:        socket to write to
868  * @len:        number of bytes to splice
869  * @flags:      splice modifier flags
870  *
871  * Will send @len bytes from the pipe to a network socket. No data copying
872  * is involved.
873  *
874  */
875 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
876                                 loff_t *ppos, size_t len, unsigned int flags)
877 {
878         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
879 }
880
881 EXPORT_SYMBOL(generic_splice_sendpage);
882
883 /*
884  * Attempt to initiate a splice from pipe to file.
885  */
886 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
887                            loff_t *ppos, size_t len, unsigned int flags)
888 {
889         int ret;
890
891         if (unlikely(!out->f_op || !out->f_op->splice_write))
892                 return -EINVAL;
893
894         if (unlikely(!(out->f_mode & FMODE_WRITE)))
895                 return -EBADF;
896
897         ret = rw_verify_area(WRITE, out, ppos, len);
898         if (unlikely(ret < 0))
899                 return ret;
900
901         return out->f_op->splice_write(pipe, out, ppos, len, flags);
902 }
903
904 /*
905  * Attempt to initiate a splice from a file to a pipe.
906  */
907 static long do_splice_to(struct file *in, loff_t *ppos,
908                          struct pipe_inode_info *pipe, size_t len,
909                          unsigned int flags)
910 {
911         loff_t isize, left;
912         int ret;
913
914         if (unlikely(!in->f_op || !in->f_op->splice_read))
915                 return -EINVAL;
916
917         if (unlikely(!(in->f_mode & FMODE_READ)))
918                 return -EBADF;
919
920         ret = rw_verify_area(READ, in, ppos, len);
921         if (unlikely(ret < 0))
922                 return ret;
923
924         isize = i_size_read(in->f_mapping->host);
925         if (unlikely(*ppos >= isize))
926                 return 0;
927         
928         left = isize - *ppos;
929         if (unlikely(left < len))
930                 len = left;
931
932         return in->f_op->splice_read(in, ppos, pipe, len, flags);
933 }
934
935 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
936                       size_t len, unsigned int flags)
937 {
938         struct pipe_inode_info *pipe;
939         long ret, bytes;
940         loff_t out_off;
941         umode_t i_mode;
942         int i;
943
944         /*
945          * We require the input being a regular file, as we don't want to
946          * randomly drop data for eg socket -> socket splicing. Use the
947          * piped splicing for that!
948          */
949         i_mode = in->f_dentry->d_inode->i_mode;
950         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
951                 return -EINVAL;
952
953         /*
954          * neither in nor out is a pipe, setup an internal pipe attached to
955          * 'out' and transfer the wanted data from 'in' to 'out' through that
956          */
957         pipe = current->splice_pipe;
958         if (unlikely(!pipe)) {
959                 pipe = alloc_pipe_info(NULL);
960                 if (!pipe)
961                         return -ENOMEM;
962
963                 /*
964                  * We don't have an immediate reader, but we'll read the stuff
965                  * out of the pipe right after the splice_to_pipe(). So set
966                  * PIPE_READERS appropriately.
967                  */
968                 pipe->readers = 1;
969
970                 current->splice_pipe = pipe;
971         }
972
973         /*
974          * Do the splice.
975          */
976         ret = 0;
977         bytes = 0;
978         out_off = 0;
979
980         while (len) {
981                 size_t read_len, max_read_len;
982
983                 /*
984                  * Do at most PIPE_BUFFERS pages worth of transfer:
985                  */
986                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
987
988                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
989                 if (unlikely(ret < 0))
990                         goto out_release;
991
992                 read_len = ret;
993
994                 /*
995                  * NOTE: nonblocking mode only applies to the input. We
996                  * must not do the output in nonblocking mode as then we
997                  * could get stuck data in the internal pipe:
998                  */
999                 ret = do_splice_from(pipe, out, &out_off, read_len,
1000                                      flags & ~SPLICE_F_NONBLOCK);
1001                 if (unlikely(ret < 0))
1002                         goto out_release;
1003
1004                 bytes += ret;
1005                 len -= ret;
1006
1007                 /*
1008                  * In nonblocking mode, if we got back a short read then
1009                  * that was due to either an IO error or due to the
1010                  * pagecache entry not being there. In the IO error case
1011                  * the _next_ splice attempt will produce a clean IO error
1012                  * return value (not a short read), so in both cases it's
1013                  * correct to break out of the loop here:
1014                  */
1015                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1016                         break;
1017         }
1018
1019         pipe->nrbufs = pipe->curbuf = 0;
1020
1021         return bytes;
1022
1023 out_release:
1024         /*
1025          * If we did an incomplete transfer we must release
1026          * the pipe buffers in question:
1027          */
1028         for (i = 0; i < PIPE_BUFFERS; i++) {
1029                 struct pipe_buffer *buf = pipe->bufs + i;
1030
1031                 if (buf->ops) {
1032                         buf->ops->release(pipe, buf);
1033                         buf->ops = NULL;
1034                 }
1035         }
1036         pipe->nrbufs = pipe->curbuf = 0;
1037
1038         /*
1039          * If we transferred some data, return the number of bytes:
1040          */
1041         if (bytes > 0)
1042                 return bytes;
1043
1044         return ret;
1045 }
1046
1047 EXPORT_SYMBOL(do_splice_direct);
1048
1049 /*
1050  * Determine where to splice to/from.
1051  */
1052 static long do_splice(struct file *in, loff_t __user *off_in,
1053                       struct file *out, loff_t __user *off_out,
1054                       size_t len, unsigned int flags)
1055 {
1056         struct pipe_inode_info *pipe;
1057         loff_t offset, *off;
1058         long ret;
1059
1060         pipe = in->f_dentry->d_inode->i_pipe;
1061         if (pipe) {
1062                 if (off_in)
1063                         return -ESPIPE;
1064                 if (off_out) {
1065                         if (out->f_op->llseek == no_llseek)
1066                                 return -EINVAL;
1067                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1068                                 return -EFAULT;
1069                         off = &offset;
1070                 } else
1071                         off = &out->f_pos;
1072
1073                 ret = do_splice_from(pipe, out, off, len, flags);
1074
1075                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1076                         ret = -EFAULT;
1077
1078                 return ret;
1079         }
1080
1081         pipe = out->f_dentry->d_inode->i_pipe;
1082         if (pipe) {
1083                 if (off_out)
1084                         return -ESPIPE;
1085                 if (off_in) {
1086                         if (in->f_op->llseek == no_llseek)
1087                                 return -EINVAL;
1088                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1089                                 return -EFAULT;
1090                         off = &offset;
1091                 } else
1092                         off = &in->f_pos;
1093
1094                 ret = do_splice_to(in, off, pipe, len, flags);
1095
1096                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1097                         ret = -EFAULT;
1098
1099                 return ret;
1100         }
1101
1102         return -EINVAL;
1103 }
1104
1105 /*
1106  * Map an iov into an array of pages and offset/length tupples. With the
1107  * partial_page structure, we can map several non-contiguous ranges into
1108  * our ones pages[] map instead of splitting that operation into pieces.
1109  * Could easily be exported as a generic helper for other users, in which
1110  * case one would probably want to add a 'max_nr_pages' parameter as well.
1111  */
1112 static int get_iovec_page_array(const struct iovec __user *iov,
1113                                 unsigned int nr_vecs, struct page **pages,
1114                                 struct partial_page *partial, int aligned)
1115 {
1116         int buffers = 0, error = 0;
1117
1118         /*
1119          * It's ok to take the mmap_sem for reading, even
1120          * across a "get_user()".
1121          */
1122         down_read(&current->mm->mmap_sem);
1123
1124         while (nr_vecs) {
1125                 unsigned long off, npages;
1126                 void __user *base;
1127                 size_t len;
1128                 int i;
1129
1130                 /*
1131                  * Get user address base and length for this iovec.
1132                  */
1133                 error = get_user(base, &iov->iov_base);
1134                 if (unlikely(error))
1135                         break;
1136                 error = get_user(len, &iov->iov_len);
1137                 if (unlikely(error))
1138                         break;
1139
1140                 /*
1141                  * Sanity check this iovec. 0 read succeeds.
1142                  */
1143                 if (unlikely(!len))
1144                         break;
1145                 error = -EFAULT;
1146                 if (unlikely(!base))
1147                         break;
1148
1149                 /*
1150                  * Get this base offset and number of pages, then map
1151                  * in the user pages.
1152                  */
1153                 off = (unsigned long) base & ~PAGE_MASK;
1154
1155                 /*
1156                  * If asked for alignment, the offset must be zero and the
1157                  * length a multiple of the PAGE_SIZE.
1158                  */
1159                 error = -EINVAL;
1160                 if (aligned && (off || len & ~PAGE_MASK))
1161                         break;
1162
1163                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1164                 if (npages > PIPE_BUFFERS - buffers)
1165                         npages = PIPE_BUFFERS - buffers;
1166
1167                 error = get_user_pages(current, current->mm,
1168                                        (unsigned long) base, npages, 0, 0,
1169                                        &pages[buffers], NULL);
1170
1171                 if (unlikely(error <= 0))
1172                         break;
1173
1174                 /*
1175                  * Fill this contiguous range into the partial page map.
1176                  */
1177                 for (i = 0; i < error; i++) {
1178                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1179
1180                         partial[buffers].offset = off;
1181                         partial[buffers].len = plen;
1182
1183                         off = 0;
1184                         len -= plen;
1185                         buffers++;
1186                 }
1187
1188                 /*
1189                  * We didn't complete this iov, stop here since it probably
1190                  * means we have to move some of this into a pipe to
1191                  * be able to continue.
1192                  */
1193                 if (len)
1194                         break;
1195
1196                 /*
1197                  * Don't continue if we mapped fewer pages than we asked for,
1198                  * or if we mapped the max number of pages that we have
1199                  * room for.
1200                  */
1201                 if (error < npages || buffers == PIPE_BUFFERS)
1202                         break;
1203
1204                 nr_vecs--;
1205                 iov++;
1206         }
1207
1208         up_read(&current->mm->mmap_sem);
1209
1210         if (buffers)
1211                 return buffers;
1212
1213         return error;
1214 }
1215
1216 /*
1217  * vmsplice splices a user address range into a pipe. It can be thought of
1218  * as splice-from-memory, where the regular splice is splice-from-file (or
1219  * to file). In both cases the output is a pipe, naturally.
1220  *
1221  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1222  * not the other way around. Splicing from user memory is a simple operation
1223  * that can be supported without any funky alignment restrictions or nasty
1224  * vm tricks. We simply map in the user memory and fill them into a pipe.
1225  * The reverse isn't quite as easy, though. There are two possible solutions
1226  * for that:
1227  *
1228  *      - memcpy() the data internally, at which point we might as well just
1229  *        do a regular read() on the buffer anyway.
1230  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1231  *        has restriction limitations on both ends of the pipe).
1232  *
1233  * Alas, it isn't here.
1234  *
1235  */
1236 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1237                         unsigned long nr_segs, unsigned int flags)
1238 {
1239         struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1240         struct page *pages[PIPE_BUFFERS];
1241         struct partial_page partial[PIPE_BUFFERS];
1242         struct splice_pipe_desc spd = {
1243                 .pages = pages,
1244                 .partial = partial,
1245                 .flags = flags,
1246                 .ops = &user_page_pipe_buf_ops,
1247         };
1248
1249         if (unlikely(!pipe))
1250                 return -EBADF;
1251         if (unlikely(nr_segs > UIO_MAXIOV))
1252                 return -EINVAL;
1253         else if (unlikely(!nr_segs))
1254                 return 0;
1255
1256         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1257                                             flags & SPLICE_F_GIFT);
1258         if (spd.nr_pages <= 0)
1259                 return spd.nr_pages;
1260
1261         return splice_to_pipe(pipe, &spd);
1262 }
1263
1264 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1265                              unsigned long nr_segs, unsigned int flags)
1266 {
1267         struct file *file;
1268         long error;
1269         int fput;
1270
1271         error = -EBADF;
1272         file = fget_light(fd, &fput);
1273         if (file) {
1274                 if (file->f_mode & FMODE_WRITE)
1275                         error = do_vmsplice(file, iov, nr_segs, flags);
1276
1277                 fput_light(file, fput);
1278         }
1279
1280         return error;
1281 }
1282
1283 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1284                            int fd_out, loff_t __user *off_out,
1285                            size_t len, unsigned int flags)
1286 {
1287         long error;
1288         struct file *in, *out;
1289         int fput_in, fput_out;
1290
1291         if (unlikely(!len))
1292                 return 0;
1293
1294         error = -EBADF;
1295         in = fget_light(fd_in, &fput_in);
1296         if (in) {
1297                 if (in->f_mode & FMODE_READ) {
1298                         out = fget_light(fd_out, &fput_out);
1299                         if (out) {
1300                                 if (out->f_mode & FMODE_WRITE)
1301                                         error = do_splice(in, off_in,
1302                                                           out, off_out,
1303                                                           len, flags);
1304                                 fput_light(out, fput_out);
1305                         }
1306                 }
1307
1308                 fput_light(in, fput_in);
1309         }
1310
1311         return error;
1312 }
1313
1314 /*
1315  * Make sure there's data to read. Wait for input if we can, otherwise
1316  * return an appropriate error.
1317  */
1318 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1319 {
1320         int ret;
1321
1322         /*
1323          * Check ->nrbufs without the inode lock first. This function
1324          * is speculative anyways, so missing one is ok.
1325          */
1326         if (pipe->nrbufs)
1327                 return 0;
1328
1329         ret = 0;
1330         mutex_lock(&pipe->inode->i_mutex);
1331
1332         while (!pipe->nrbufs) {
1333                 if (signal_pending(current)) {
1334                         ret = -ERESTARTSYS;
1335                         break;
1336                 }
1337                 if (!pipe->writers)
1338                         break;
1339                 if (!pipe->waiting_writers) {
1340                         if (flags & SPLICE_F_NONBLOCK) {
1341                                 ret = -EAGAIN;
1342                                 break;
1343                         }
1344                 }
1345                 pipe_wait(pipe);
1346         }
1347
1348         mutex_unlock(&pipe->inode->i_mutex);
1349         return ret;
1350 }
1351
1352 /*
1353  * Make sure there's writeable room. Wait for room if we can, otherwise
1354  * return an appropriate error.
1355  */
1356 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1357 {
1358         int ret;
1359
1360         /*
1361          * Check ->nrbufs without the inode lock first. This function
1362          * is speculative anyways, so missing one is ok.
1363          */
1364         if (pipe->nrbufs < PIPE_BUFFERS)
1365                 return 0;
1366
1367         ret = 0;
1368         mutex_lock(&pipe->inode->i_mutex);
1369
1370         while (pipe->nrbufs >= PIPE_BUFFERS) {
1371                 if (!pipe->readers) {
1372                         send_sig(SIGPIPE, current, 0);
1373                         ret = -EPIPE;
1374                         break;
1375                 }
1376                 if (flags & SPLICE_F_NONBLOCK) {
1377                         ret = -EAGAIN;
1378                         break;
1379                 }
1380                 if (signal_pending(current)) {
1381                         ret = -ERESTARTSYS;
1382                         break;
1383                 }
1384                 pipe->waiting_writers++;
1385                 pipe_wait(pipe);
1386                 pipe->waiting_writers--;
1387         }
1388
1389         mutex_unlock(&pipe->inode->i_mutex);
1390         return ret;
1391 }
1392
1393 /*
1394  * Link contents of ipipe to opipe.
1395  */
1396 static int link_pipe(struct pipe_inode_info *ipipe,
1397                      struct pipe_inode_info *opipe,
1398                      size_t len, unsigned int flags)
1399 {
1400         struct pipe_buffer *ibuf, *obuf;
1401         int ret = 0, i = 0, nbuf;
1402
1403         /*
1404          * Potential ABBA deadlock, work around it by ordering lock
1405          * grabbing by inode address. Otherwise two different processes
1406          * could deadlock (one doing tee from A -> B, the other from B -> A).
1407          */
1408         inode_double_lock(ipipe->inode, opipe->inode);
1409
1410         do {
1411                 if (!opipe->readers) {
1412                         send_sig(SIGPIPE, current, 0);
1413                         if (!ret)
1414                                 ret = -EPIPE;
1415                         break;
1416                 }
1417
1418                 /*
1419                  * If we have iterated all input buffers or ran out of
1420                  * output room, break.
1421                  */
1422                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1423                         break;
1424
1425                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1426                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1427
1428                 /*
1429                  * Get a reference to this pipe buffer,
1430                  * so we can copy the contents over.
1431                  */
1432                 ibuf->ops->get(ipipe, ibuf);
1433
1434                 obuf = opipe->bufs + nbuf;
1435                 *obuf = *ibuf;
1436
1437                 /*
1438                  * Don't inherit the gift flag, we need to
1439                  * prevent multiple steals of this page.
1440                  */
1441                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1442
1443                 if (obuf->len > len)
1444                         obuf->len = len;
1445
1446                 opipe->nrbufs++;
1447                 ret += obuf->len;
1448                 len -= obuf->len;
1449                 i++;
1450         } while (len);
1451
1452         inode_double_unlock(ipipe->inode, opipe->inode);
1453
1454         /*
1455          * If we put data in the output pipe, wakeup any potential readers.
1456          */
1457         if (ret > 0) {
1458                 smp_mb();
1459                 if (waitqueue_active(&opipe->wait))
1460                         wake_up_interruptible(&opipe->wait);
1461                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1462         }
1463
1464         return ret;
1465 }
1466
1467 /*
1468  * This is a tee(1) implementation that works on pipes. It doesn't copy
1469  * any data, it simply references the 'in' pages on the 'out' pipe.
1470  * The 'flags' used are the SPLICE_F_* variants, currently the only
1471  * applicable one is SPLICE_F_NONBLOCK.
1472  */
1473 static long do_tee(struct file *in, struct file *out, size_t len,
1474                    unsigned int flags)
1475 {
1476         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1477         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1478         int ret = -EINVAL;
1479
1480         /*
1481          * Duplicate the contents of ipipe to opipe without actually
1482          * copying the data.
1483          */
1484         if (ipipe && opipe && ipipe != opipe) {
1485                 /*
1486                  * Keep going, unless we encounter an error. The ipipe/opipe
1487                  * ordering doesn't really matter.
1488                  */
1489                 ret = link_ipipe_prep(ipipe, flags);
1490                 if (!ret) {
1491                         ret = link_opipe_prep(opipe, flags);
1492                         if (!ret) {
1493                                 ret = link_pipe(ipipe, opipe, len, flags);
1494                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1495                                         ret = -EAGAIN;
1496                         }
1497                 }
1498         }
1499
1500         return ret;
1501 }
1502
1503 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1504 {
1505         struct file *in;
1506         int error, fput_in;
1507
1508         if (unlikely(!len))
1509                 return 0;
1510
1511         error = -EBADF;
1512         in = fget_light(fdin, &fput_in);
1513         if (in) {
1514                 if (in->f_mode & FMODE_READ) {
1515                         int fput_out;
1516                         struct file *out = fget_light(fdout, &fput_out);
1517
1518                         if (out) {
1519                                 if (out->f_mode & FMODE_WRITE)
1520                                         error = do_tee(in, out, len, flags);
1521                                 fput_light(out, fput_out);
1522                         }
1523                 }
1524                 fput_light(in, fput_in);
1525         }
1526
1527         return error;
1528 }