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