9df28d30efa0ede1eb5f149fd342d3f1c5b6cd8e
[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@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, page_nr;
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 || nr_pages > 1)
303                 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
304
305         /*
306          * Now fill in the holes:
307          */
308         error = 0;
309         total_len = 0;
310
311         /*
312          * Lookup the (hopefully) full range of pages we need.
313          */
314         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
315
316         /*
317          * If find_get_pages_contig() returned fewer pages than we needed,
318          * allocate the rest.
319          */
320         index += spd.nr_pages;
321         while (spd.nr_pages < nr_pages) {
322                 /*
323                  * Page could be there, find_get_pages_contig() breaks on
324                  * the first hole.
325                  */
326                 page = find_get_page(mapping, index);
327                 if (!page) {
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                                 break;
340                         }
341                         /*
342                          * add_to_page_cache() locks the page, unlock it
343                          * to avoid convoluting the logic below even more.
344                          */
345                         unlock_page(page);
346                 }
347
348                 pages[spd.nr_pages++] = page;
349                 index++;
350         }
351
352         /*
353          * Now loop over the map and see if we need to start IO on any
354          * pages, fill in the partial map, etc.
355          */
356         index = *ppos >> PAGE_CACHE_SHIFT;
357         nr_pages = spd.nr_pages;
358         spd.nr_pages = 0;
359         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
360                 unsigned int this_len;
361
362                 if (!len)
363                         break;
364
365                 /*
366                  * this_len is the max we'll use from this page
367                  */
368                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
369                 page = pages[page_nr];
370
371                 /*
372                  * If the page isn't uptodate, we may need to start io on it
373                  */
374                 if (!PageUptodate(page)) {
375                         /*
376                          * If in nonblock mode then dont block on waiting
377                          * for an in-flight io page
378                          */
379                         if (flags & SPLICE_F_NONBLOCK)
380                                 break;
381
382                         lock_page(page);
383
384                         /*
385                          * page was truncated, stop here. if this isn't the
386                          * first page, we'll just complete what we already
387                          * added
388                          */
389                         if (!page->mapping) {
390                                 unlock_page(page);
391                                 break;
392                         }
393                         /*
394                          * page was already under io and is now done, great
395                          */
396                         if (PageUptodate(page)) {
397                                 unlock_page(page);
398                                 goto fill_it;
399                         }
400
401                         /*
402                          * need to read in the page
403                          */
404                         error = mapping->a_ops->readpage(in, page);
405                         if (unlikely(error)) {
406                                 /*
407                                  * We really should re-lookup the page here,
408                                  * but it complicates things a lot. Instead
409                                  * lets just do what we already stored, and
410                                  * we'll get it the next time we are called.
411                                  */
412                                 if (error == AOP_TRUNCATED_PAGE)
413                                         error = 0;
414
415                                 break;
416                         }
417
418                         /*
419                          * i_size must be checked after ->readpage().
420                          */
421                         isize = i_size_read(mapping->host);
422                         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
423                         if (unlikely(!isize || index > end_index))
424                                 break;
425
426                         /*
427                          * if this is the last page, see if we need to shrink
428                          * the length and stop
429                          */
430                         if (end_index == index) {
431                                 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
432                                 if (total_len + loff > isize)
433                                         break;
434                                 /*
435                                  * force quit after adding this page
436                                  */
437                                 len = this_len;
438                                 this_len = min(this_len, loff);
439                                 loff = 0;
440                         }
441                 }
442 fill_it:
443                 partial[page_nr].offset = loff;
444                 partial[page_nr].len = this_len;
445                 len -= this_len;
446                 total_len += this_len;
447                 loff = 0;
448                 spd.nr_pages++;
449                 index++;
450         }
451
452         /*
453          * Release any pages at the end, if we quit early. 'i' is how far
454          * we got, 'nr_pages' is how many pages are in the map.
455          */
456         while (page_nr < nr_pages)
457                 page_cache_release(pages[page_nr++]);
458
459         if (spd.nr_pages)
460                 return splice_to_pipe(pipe, &spd);
461
462         return error;
463 }
464
465 /**
466  * generic_file_splice_read - splice data from file to a pipe
467  * @in:         file to splice from
468  * @pipe:       pipe to splice to
469  * @len:        number of bytes to splice
470  * @flags:      splice modifier flags
471  *
472  * Will read pages from given file and fill them into a pipe.
473  */
474 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
475                                  struct pipe_inode_info *pipe, size_t len,
476                                  unsigned int flags)
477 {
478         ssize_t spliced;
479         int ret;
480
481         ret = 0;
482         spliced = 0;
483
484         while (len) {
485                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
486
487                 if (ret < 0)
488                         break;
489                 else if (!ret) {
490                         if (spliced)
491                                 break;
492                         if (flags & SPLICE_F_NONBLOCK) {
493                                 ret = -EAGAIN;
494                                 break;
495                         }
496                 }
497
498                 *ppos += ret;
499                 len -= ret;
500                 spliced += ret;
501         }
502
503         if (spliced)
504                 return spliced;
505
506         return ret;
507 }
508
509 EXPORT_SYMBOL(generic_file_splice_read);
510
511 /*
512  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
513  * using sendpage(). Return the number of bytes sent.
514  */
515 static int pipe_to_sendpage(struct pipe_inode_info *info,
516                             struct pipe_buffer *buf, struct splice_desc *sd)
517 {
518         struct file *file = sd->file;
519         loff_t pos = sd->pos;
520         ssize_t ret;
521         void *ptr;
522         int more;
523
524         /*
525          * Sub-optimal, but we are limited by the pipe ->map. We don't
526          * need a kmap'ed buffer here, we just want to make sure we
527          * have the page pinned if the pipe page originates from the
528          * page cache.
529          */
530         ptr = buf->ops->map(file, info, buf);
531         if (IS_ERR(ptr))
532                 return PTR_ERR(ptr);
533
534         more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
535
536         ret = file->f_op->sendpage(file, buf->page, buf->offset, sd->len,
537                                    &pos, more);
538
539         buf->ops->unmap(info, buf);
540         return ret;
541 }
542
543 /*
544  * This is a little more tricky than the file -> pipe splicing. There are
545  * basically three cases:
546  *
547  *      - Destination page already exists in the address space and there
548  *        are users of it. For that case we have no other option that
549  *        copying the data. Tough luck.
550  *      - Destination page already exists in the address space, but there
551  *        are no users of it. Make sure it's uptodate, then drop it. Fall
552  *        through to last case.
553  *      - Destination page does not exist, we can add the pipe page to
554  *        the page cache and avoid the copy.
555  *
556  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
557  * sd->flags), we attempt to migrate pages from the pipe to the output
558  * file address space page cache. This is possible if no one else has
559  * the pipe page referenced outside of the pipe and page cache. If
560  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
561  * a new page in the output file page cache and fill/dirty that.
562  */
563 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
564                         struct splice_desc *sd)
565 {
566         struct file *file = sd->file;
567         struct address_space *mapping = file->f_mapping;
568         gfp_t gfp_mask = mapping_gfp_mask(mapping);
569         unsigned int offset, this_len;
570         struct page *page;
571         pgoff_t index;
572         char *src;
573         int ret;
574
575         /*
576          * make sure the data in this buffer is uptodate
577          */
578         src = buf->ops->map(file, info, buf);
579         if (IS_ERR(src))
580                 return PTR_ERR(src);
581
582         index = sd->pos >> PAGE_CACHE_SHIFT;
583         offset = sd->pos & ~PAGE_CACHE_MASK;
584
585         this_len = sd->len;
586         if (this_len + offset > PAGE_CACHE_SIZE)
587                 this_len = PAGE_CACHE_SIZE - offset;
588
589         /*
590          * Reuse buf page, if SPLICE_F_MOVE is set.
591          */
592         if (sd->flags & SPLICE_F_MOVE) {
593                 /*
594                  * If steal succeeds, buf->page is now pruned from the vm
595                  * side (LRU and page cache) and we can reuse it. The page
596                  * will also be looked on successful return.
597                  */
598                 if (buf->ops->steal(info, buf))
599                         goto find_page;
600
601                 page = buf->page;
602                 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
603                         unlock_page(page);
604                         goto find_page;
605                 }
606
607                 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
608                         lru_cache_add(page);
609         } else {
610 find_page:
611                 page = find_lock_page(mapping, index);
612                 if (!page) {
613                         ret = -ENOMEM;
614                         page = page_cache_alloc_cold(mapping);
615                         if (unlikely(!page))
616                                 goto out_nomem;
617
618                         /*
619                          * This will also lock the page
620                          */
621                         ret = add_to_page_cache_lru(page, mapping, index,
622                                                     gfp_mask);
623                         if (unlikely(ret))
624                                 goto out;
625                 }
626
627                 /*
628                  * We get here with the page locked. If the page is also
629                  * uptodate, we don't need to do more. If it isn't, we
630                  * may need to bring it in if we are not going to overwrite
631                  * the full page.
632                  */
633                 if (!PageUptodate(page)) {
634                         if (this_len < PAGE_CACHE_SIZE) {
635                                 ret = mapping->a_ops->readpage(file, page);
636                                 if (unlikely(ret))
637                                         goto out;
638
639                                 lock_page(page);
640
641                                 if (!PageUptodate(page)) {
642                                         /*
643                                          * Page got invalidated, repeat.
644                                          */
645                                         if (!page->mapping) {
646                                                 unlock_page(page);
647                                                 page_cache_release(page);
648                                                 goto find_page;
649                                         }
650                                         ret = -EIO;
651                                         goto out;
652                                 }
653                         } else
654                                 SetPageUptodate(page);
655                 }
656         }
657
658         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
659         if (ret == AOP_TRUNCATED_PAGE) {
660                 page_cache_release(page);
661                 goto find_page;
662         } else if (ret)
663                 goto out;
664
665         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
666                 char *dst = kmap_atomic(page, KM_USER0);
667
668                 memcpy(dst + offset, src + buf->offset, this_len);
669                 flush_dcache_page(page);
670                 kunmap_atomic(dst, KM_USER0);
671         }
672
673         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
674         if (ret == AOP_TRUNCATED_PAGE) {
675                 page_cache_release(page);
676                 goto find_page;
677         } else if (ret)
678                 goto out;
679
680         /*
681          * Return the number of bytes written.
682          */
683         ret = this_len;
684         mark_page_accessed(page);
685         balance_dirty_pages_ratelimited(mapping);
686 out:
687         if (!(buf->flags & PIPE_BUF_FLAG_STOLEN))
688                 page_cache_release(page);
689
690         unlock_page(page);
691 out_nomem:
692         buf->ops->unmap(info, buf);
693         return ret;
694 }
695
696 /*
697  * Pipe input worker. Most of this logic works like a regular pipe, the
698  * key here is the 'actor' worker passed in that actually moves the data
699  * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
700  */
701 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
702                          loff_t *ppos, size_t len, unsigned int flags,
703                          splice_actor *actor)
704 {
705         int ret, do_wakeup, err;
706         struct splice_desc sd;
707
708         ret = 0;
709         do_wakeup = 0;
710
711         sd.total_len = len;
712         sd.flags = flags;
713         sd.file = out;
714         sd.pos = *ppos;
715
716         if (pipe->inode)
717                 mutex_lock(&pipe->inode->i_mutex);
718
719         for (;;) {
720                 if (pipe->nrbufs) {
721                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
722                         struct pipe_buf_operations *ops = buf->ops;
723
724                         sd.len = buf->len;
725                         if (sd.len > sd.total_len)
726                                 sd.len = sd.total_len;
727
728                         err = actor(pipe, buf, &sd);
729                         if (err <= 0) {
730                                 if (!ret && err != -ENODATA)
731                                         ret = err;
732
733                                 break;
734                         }
735
736                         ret += err;
737                         buf->offset += err;
738                         buf->len -= err;
739
740                         sd.len -= err;
741                         sd.pos += err;
742                         sd.total_len -= err;
743                         if (sd.len)
744                                 continue;
745
746                         if (!buf->len) {
747                                 buf->ops = NULL;
748                                 ops->release(pipe, buf);
749                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
750                                 pipe->nrbufs--;
751                                 if (pipe->inode)
752                                         do_wakeup = 1;
753                         }
754
755                         if (!sd.total_len)
756                                 break;
757                 }
758
759                 if (pipe->nrbufs)
760                         continue;
761                 if (!pipe->writers)
762                         break;
763                 if (!pipe->waiting_writers) {
764                         if (ret)
765                                 break;
766                 }
767
768                 if (flags & SPLICE_F_NONBLOCK) {
769                         if (!ret)
770                                 ret = -EAGAIN;
771                         break;
772                 }
773
774                 if (signal_pending(current)) {
775                         if (!ret)
776                                 ret = -ERESTARTSYS;
777                         break;
778                 }
779
780                 if (do_wakeup) {
781                         smp_mb();
782                         if (waitqueue_active(&pipe->wait))
783                                 wake_up_interruptible_sync(&pipe->wait);
784                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
785                         do_wakeup = 0;
786                 }
787
788                 pipe_wait(pipe);
789         }
790
791         if (pipe->inode)
792                 mutex_unlock(&pipe->inode->i_mutex);
793
794         if (do_wakeup) {
795                 smp_mb();
796                 if (waitqueue_active(&pipe->wait))
797                         wake_up_interruptible(&pipe->wait);
798                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
799         }
800
801         return ret;
802 }
803
804 /**
805  * generic_file_splice_write - splice data from a pipe to a file
806  * @pipe:       pipe info
807  * @out:        file to write to
808  * @len:        number of bytes to splice
809  * @flags:      splice modifier flags
810  *
811  * Will either move or copy pages (determined by @flags options) from
812  * the given pipe inode to the given file.
813  *
814  */
815 ssize_t
816 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
817                           loff_t *ppos, size_t len, unsigned int flags)
818 {
819         struct address_space *mapping = out->f_mapping;
820         ssize_t ret;
821
822         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
823         if (ret > 0) {
824                 struct inode *inode = mapping->host;
825
826                 *ppos += ret;
827
828                 /*
829                  * If file or inode is SYNC and we actually wrote some data,
830                  * sync it.
831                  */
832                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
833                         int err;
834
835                         mutex_lock(&inode->i_mutex);
836                         err = generic_osync_inode(inode, mapping,
837                                                   OSYNC_METADATA|OSYNC_DATA);
838                         mutex_unlock(&inode->i_mutex);
839
840                         if (err)
841                                 ret = err;
842                 }
843         }
844
845         return ret;
846 }
847
848 EXPORT_SYMBOL(generic_file_splice_write);
849
850 /**
851  * generic_splice_sendpage - splice data from a pipe to a socket
852  * @inode:      pipe inode
853  * @out:        socket to write to
854  * @len:        number of bytes to splice
855  * @flags:      splice modifier flags
856  *
857  * Will send @len bytes from the pipe to a network socket. No data copying
858  * is involved.
859  *
860  */
861 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
862                                 loff_t *ppos, size_t len, unsigned int flags)
863 {
864         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
865 }
866
867 EXPORT_SYMBOL(generic_splice_sendpage);
868
869 /*
870  * Attempt to initiate a splice from pipe to file.
871  */
872 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
873                            loff_t *ppos, size_t len, unsigned int flags)
874 {
875         int ret;
876
877         if (unlikely(!out->f_op || !out->f_op->splice_write))
878                 return -EINVAL;
879
880         if (unlikely(!(out->f_mode & FMODE_WRITE)))
881                 return -EBADF;
882
883         ret = rw_verify_area(WRITE, out, ppos, len);
884         if (unlikely(ret < 0))
885                 return ret;
886
887         return out->f_op->splice_write(pipe, out, ppos, len, flags);
888 }
889
890 /*
891  * Attempt to initiate a splice from a file to a pipe.
892  */
893 static long do_splice_to(struct file *in, loff_t *ppos,
894                          struct pipe_inode_info *pipe, size_t len,
895                          unsigned int flags)
896 {
897         loff_t isize, left;
898         int ret;
899
900         if (unlikely(!in->f_op || !in->f_op->splice_read))
901                 return -EINVAL;
902
903         if (unlikely(!(in->f_mode & FMODE_READ)))
904                 return -EBADF;
905
906         ret = rw_verify_area(READ, in, ppos, len);
907         if (unlikely(ret < 0))
908                 return ret;
909
910         isize = i_size_read(in->f_mapping->host);
911         if (unlikely(*ppos >= isize))
912                 return 0;
913         
914         left = isize - *ppos;
915         if (unlikely(left < len))
916                 len = left;
917
918         return in->f_op->splice_read(in, ppos, pipe, len, flags);
919 }
920
921 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
922                       size_t len, unsigned int flags)
923 {
924         struct pipe_inode_info *pipe;
925         long ret, bytes;
926         loff_t out_off;
927         umode_t i_mode;
928         int i;
929
930         /*
931          * We require the input being a regular file, as we don't want to
932          * randomly drop data for eg socket -> socket splicing. Use the
933          * piped splicing for that!
934          */
935         i_mode = in->f_dentry->d_inode->i_mode;
936         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
937                 return -EINVAL;
938
939         /*
940          * neither in nor out is a pipe, setup an internal pipe attached to
941          * 'out' and transfer the wanted data from 'in' to 'out' through that
942          */
943         pipe = current->splice_pipe;
944         if (unlikely(!pipe)) {
945                 pipe = alloc_pipe_info(NULL);
946                 if (!pipe)
947                         return -ENOMEM;
948
949                 /*
950                  * We don't have an immediate reader, but we'll read the stuff
951                  * out of the pipe right after the splice_to_pipe(). So set
952                  * PIPE_READERS appropriately.
953                  */
954                 pipe->readers = 1;
955
956                 current->splice_pipe = pipe;
957         }
958
959         /*
960          * Do the splice.
961          */
962         ret = 0;
963         bytes = 0;
964         out_off = 0;
965
966         while (len) {
967                 size_t read_len, max_read_len;
968
969                 /*
970                  * Do at most PIPE_BUFFERS pages worth of transfer:
971                  */
972                 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
973
974                 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
975                 if (unlikely(ret < 0))
976                         goto out_release;
977
978                 read_len = ret;
979
980                 /*
981                  * NOTE: nonblocking mode only applies to the input. We
982                  * must not do the output in nonblocking mode as then we
983                  * could get stuck data in the internal pipe:
984                  */
985                 ret = do_splice_from(pipe, out, &out_off, read_len,
986                                      flags & ~SPLICE_F_NONBLOCK);
987                 if (unlikely(ret < 0))
988                         goto out_release;
989
990                 bytes += ret;
991                 len -= ret;
992
993                 /*
994                  * In nonblocking mode, if we got back a short read then
995                  * that was due to either an IO error or due to the
996                  * pagecache entry not being there. In the IO error case
997                  * the _next_ splice attempt will produce a clean IO error
998                  * return value (not a short read), so in both cases it's
999                  * correct to break out of the loop here:
1000                  */
1001                 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1002                         break;
1003         }
1004
1005         pipe->nrbufs = pipe->curbuf = 0;
1006
1007         return bytes;
1008
1009 out_release:
1010         /*
1011          * If we did an incomplete transfer we must release
1012          * the pipe buffers in question:
1013          */
1014         for (i = 0; i < PIPE_BUFFERS; i++) {
1015                 struct pipe_buffer *buf = pipe->bufs + i;
1016
1017                 if (buf->ops) {
1018                         buf->ops->release(pipe, buf);
1019                         buf->ops = NULL;
1020                 }
1021         }
1022         pipe->nrbufs = pipe->curbuf = 0;
1023
1024         /*
1025          * If we transferred some data, return the number of bytes:
1026          */
1027         if (bytes > 0)
1028                 return bytes;
1029
1030         return ret;
1031 }
1032
1033 EXPORT_SYMBOL(do_splice_direct);
1034
1035 /*
1036  * Determine where to splice to/from.
1037  */
1038 static long do_splice(struct file *in, loff_t __user *off_in,
1039                       struct file *out, loff_t __user *off_out,
1040                       size_t len, unsigned int flags)
1041 {
1042         struct pipe_inode_info *pipe;
1043         loff_t offset, *off;
1044         long ret;
1045
1046         pipe = in->f_dentry->d_inode->i_pipe;
1047         if (pipe) {
1048                 if (off_in)
1049                         return -ESPIPE;
1050                 if (off_out) {
1051                         if (out->f_op->llseek == no_llseek)
1052                                 return -EINVAL;
1053                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1054                                 return -EFAULT;
1055                         off = &offset;
1056                 } else
1057                         off = &out->f_pos;
1058
1059                 ret = do_splice_from(pipe, out, off, len, flags);
1060
1061                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1062                         ret = -EFAULT;
1063
1064                 return ret;
1065         }
1066
1067         pipe = out->f_dentry->d_inode->i_pipe;
1068         if (pipe) {
1069                 if (off_out)
1070                         return -ESPIPE;
1071                 if (off_in) {
1072                         if (in->f_op->llseek == no_llseek)
1073                                 return -EINVAL;
1074                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1075                                 return -EFAULT;
1076                         off = &offset;
1077                 } else
1078                         off = &in->f_pos;
1079
1080                 ret = do_splice_to(in, off, pipe, len, flags);
1081
1082                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1083                         ret = -EFAULT;
1084
1085                 return ret;
1086         }
1087
1088         return -EINVAL;
1089 }
1090
1091 /*
1092  * Map an iov into an array of pages and offset/length tupples. With the
1093  * partial_page structure, we can map several non-contiguous ranges into
1094  * our ones pages[] map instead of splitting that operation into pieces.
1095  * Could easily be exported as a generic helper for other users, in which
1096  * case one would probably want to add a 'max_nr_pages' parameter as well.
1097  */
1098 static int get_iovec_page_array(const struct iovec __user *iov,
1099                                 unsigned int nr_vecs, struct page **pages,
1100                                 struct partial_page *partial)
1101 {
1102         int buffers = 0, error = 0;
1103
1104         /*
1105          * It's ok to take the mmap_sem for reading, even
1106          * across a "get_user()".
1107          */
1108         down_read(&current->mm->mmap_sem);
1109
1110         while (nr_vecs) {
1111                 unsigned long off, npages;
1112                 void __user *base;
1113                 size_t len;
1114                 int i;
1115
1116                 /*
1117                  * Get user address base and length for this iovec.
1118                  */
1119                 error = get_user(base, &iov->iov_base);
1120                 if (unlikely(error))
1121                         break;
1122                 error = get_user(len, &iov->iov_len);
1123                 if (unlikely(error))
1124                         break;
1125
1126                 /*
1127                  * Sanity check this iovec. 0 read succeeds.
1128                  */
1129                 if (unlikely(!len))
1130                         break;
1131                 error = -EFAULT;
1132                 if (unlikely(!base))
1133                         break;
1134
1135                 /*
1136                  * Get this base offset and number of pages, then map
1137                  * in the user pages.
1138                  */
1139                 off = (unsigned long) base & ~PAGE_MASK;
1140                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1141                 if (npages > PIPE_BUFFERS - buffers)
1142                         npages = PIPE_BUFFERS - buffers;
1143
1144                 error = get_user_pages(current, current->mm,
1145                                        (unsigned long) base, npages, 0, 0,
1146                                        &pages[buffers], NULL);
1147
1148                 if (unlikely(error <= 0))
1149                         break;
1150
1151                 /*
1152                  * Fill this contiguous range into the partial page map.
1153                  */
1154                 for (i = 0; i < error; i++) {
1155                         const int plen = min_t(size_t, len, PAGE_SIZE) - off;
1156
1157                         partial[buffers].offset = off;
1158                         partial[buffers].len = plen;
1159
1160                         off = 0;
1161                         len -= plen;
1162                         buffers++;
1163                 }
1164
1165                 /*
1166                  * We didn't complete this iov, stop here since it probably
1167                  * means we have to move some of this into a pipe to
1168                  * be able to continue.
1169                  */
1170                 if (len)
1171                         break;
1172
1173                 /*
1174                  * Don't continue if we mapped fewer pages than we asked for,
1175                  * or if we mapped the max number of pages that we have
1176                  * room for.
1177                  */
1178                 if (error < npages || buffers == PIPE_BUFFERS)
1179                         break;
1180
1181                 nr_vecs--;
1182                 iov++;
1183         }
1184
1185         up_read(&current->mm->mmap_sem);
1186
1187         if (buffers)
1188                 return buffers;
1189
1190         return error;
1191 }
1192
1193 /*
1194  * vmsplice splices a user address range into a pipe. It can be thought of
1195  * as splice-from-memory, where the regular splice is splice-from-file (or
1196  * to file). In both cases the output is a pipe, naturally.
1197  *
1198  * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1199  * not the other way around. Splicing from user memory is a simple operation
1200  * that can be supported without any funky alignment restrictions or nasty
1201  * vm tricks. We simply map in the user memory and fill them into a pipe.
1202  * The reverse isn't quite as easy, though. There are two possible solutions
1203  * for that:
1204  *
1205  *      - memcpy() the data internally, at which point we might as well just
1206  *        do a regular read() on the buffer anyway.
1207  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1208  *        has restriction limitations on both ends of the pipe).
1209  *
1210  * Alas, it isn't here.
1211  *
1212  */
1213 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1214                         unsigned long nr_segs, unsigned int flags)
1215 {
1216         struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1217         struct page *pages[PIPE_BUFFERS];
1218         struct partial_page partial[PIPE_BUFFERS];
1219         struct splice_pipe_desc spd = {
1220                 .pages = pages,
1221                 .partial = partial,
1222                 .flags = flags,
1223                 .ops = &user_page_pipe_buf_ops,
1224         };
1225
1226         if (unlikely(!pipe))
1227                 return -EBADF;
1228         if (unlikely(nr_segs > UIO_MAXIOV))
1229                 return -EINVAL;
1230         else if (unlikely(!nr_segs))
1231                 return 0;
1232
1233         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial);
1234         if (spd.nr_pages <= 0)
1235                 return spd.nr_pages;
1236
1237         return splice_to_pipe(pipe, &spd);
1238 }
1239
1240 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1241                              unsigned long nr_segs, unsigned int flags)
1242 {
1243         struct file *file;
1244         long error;
1245         int fput;
1246
1247         error = -EBADF;
1248         file = fget_light(fd, &fput);
1249         if (file) {
1250                 if (file->f_mode & FMODE_WRITE)
1251                         error = do_vmsplice(file, iov, nr_segs, flags);
1252
1253                 fput_light(file, fput);
1254         }
1255
1256         return error;
1257 }
1258
1259 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1260                            int fd_out, loff_t __user *off_out,
1261                            size_t len, unsigned int flags)
1262 {
1263         long error;
1264         struct file *in, *out;
1265         int fput_in, fput_out;
1266
1267         if (unlikely(!len))
1268                 return 0;
1269
1270         error = -EBADF;
1271         in = fget_light(fd_in, &fput_in);
1272         if (in) {
1273                 if (in->f_mode & FMODE_READ) {
1274                         out = fget_light(fd_out, &fput_out);
1275                         if (out) {
1276                                 if (out->f_mode & FMODE_WRITE)
1277                                         error = do_splice(in, off_in,
1278                                                           out, off_out,
1279                                                           len, flags);
1280                                 fput_light(out, fput_out);
1281                         }
1282                 }
1283
1284                 fput_light(in, fput_in);
1285         }
1286
1287         return error;
1288 }
1289
1290 /*
1291  * Link contents of ipipe to opipe.
1292  */
1293 static int link_pipe(struct pipe_inode_info *ipipe,
1294                      struct pipe_inode_info *opipe,
1295                      size_t len, unsigned int flags)
1296 {
1297         struct pipe_buffer *ibuf, *obuf;
1298         int ret, do_wakeup, i, ipipe_first;
1299
1300         ret = do_wakeup = ipipe_first = 0;
1301
1302         /*
1303          * Potential ABBA deadlock, work around it by ordering lock
1304          * grabbing by inode address. Otherwise two different processes
1305          * could deadlock (one doing tee from A -> B, the other from B -> A).
1306          */
1307         if (ipipe->inode < opipe->inode) {
1308                 ipipe_first = 1;
1309                 mutex_lock(&ipipe->inode->i_mutex);
1310                 mutex_lock(&opipe->inode->i_mutex);
1311         } else {
1312                 mutex_lock(&opipe->inode->i_mutex);
1313                 mutex_lock(&ipipe->inode->i_mutex);
1314         }
1315
1316         for (i = 0;; i++) {
1317                 if (!opipe->readers) {
1318                         send_sig(SIGPIPE, current, 0);
1319                         if (!ret)
1320                                 ret = -EPIPE;
1321                         break;
1322                 }
1323                 if (ipipe->nrbufs - i) {
1324                         ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1325
1326                         /*
1327                          * If we have room, fill this buffer
1328                          */
1329                         if (opipe->nrbufs < PIPE_BUFFERS) {
1330                                 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1331
1332                                 /*
1333                                  * Get a reference to this pipe buffer,
1334                                  * so we can copy the contents over.
1335                                  */
1336                                 ibuf->ops->get(ipipe, ibuf);
1337
1338                                 obuf = opipe->bufs + nbuf;
1339                                 *obuf = *ibuf;
1340
1341                                 if (obuf->len > len)
1342                                         obuf->len = len;
1343
1344                                 opipe->nrbufs++;
1345                                 do_wakeup = 1;
1346                                 ret += obuf->len;
1347                                 len -= obuf->len;
1348
1349                                 if (!len)
1350                                         break;
1351                                 if (opipe->nrbufs < PIPE_BUFFERS)
1352                                         continue;
1353                         }
1354
1355                         /*
1356                          * We have input available, but no output room.
1357                          * If we already copied data, return that. If we
1358                          * need to drop the opipe lock, it must be ordered
1359                          * last to avoid deadlocks.
1360                          */
1361                         if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
1362                                 if (!ret)
1363                                         ret = -EAGAIN;
1364                                 break;
1365                         }
1366                         if (signal_pending(current)) {
1367                                 if (!ret)
1368                                         ret = -ERESTARTSYS;
1369                                 break;
1370                         }
1371                         if (do_wakeup) {
1372                                 smp_mb();
1373                                 if (waitqueue_active(&opipe->wait))
1374                                         wake_up_interruptible(&opipe->wait);
1375                                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1376                                 do_wakeup = 0;
1377                         }
1378
1379                         opipe->waiting_writers++;
1380                         pipe_wait(opipe);
1381                         opipe->waiting_writers--;
1382                         continue;
1383                 }
1384
1385                 /*
1386                  * No input buffers, do the usual checks for available
1387                  * writers and blocking and wait if necessary
1388                  */
1389                 if (!ipipe->writers)
1390                         break;
1391                 if (!ipipe->waiting_writers) {
1392                         if (ret)
1393                                 break;
1394                 }
1395                 /*
1396                  * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1397                  * with another process, we can only safely do that if
1398                  * the ipipe lock is ordered last.
1399                  */
1400                 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
1401                         if (!ret)
1402                                 ret = -EAGAIN;
1403                         break;
1404                 }
1405                 if (signal_pending(current)) {
1406                         if (!ret)
1407                                 ret = -ERESTARTSYS;
1408                         break;
1409                 }
1410
1411                 if (waitqueue_active(&ipipe->wait))
1412                         wake_up_interruptible_sync(&ipipe->wait);
1413                 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1414
1415                 pipe_wait(ipipe);
1416         }
1417
1418         mutex_unlock(&ipipe->inode->i_mutex);
1419         mutex_unlock(&opipe->inode->i_mutex);
1420
1421         if (do_wakeup) {
1422                 smp_mb();
1423                 if (waitqueue_active(&opipe->wait))
1424                         wake_up_interruptible(&opipe->wait);
1425                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1426         }
1427
1428         return ret;
1429 }
1430
1431 /*
1432  * This is a tee(1) implementation that works on pipes. It doesn't copy
1433  * any data, it simply references the 'in' pages on the 'out' pipe.
1434  * The 'flags' used are the SPLICE_F_* variants, currently the only
1435  * applicable one is SPLICE_F_NONBLOCK.
1436  */
1437 static long do_tee(struct file *in, struct file *out, size_t len,
1438                    unsigned int flags)
1439 {
1440         struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1441         struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1442
1443         /*
1444          * Link ipipe to the two output pipes, consuming as we go along.
1445          */
1446         if (ipipe && opipe)
1447                 return link_pipe(ipipe, opipe, len, flags);
1448
1449         return -EINVAL;
1450 }
1451
1452 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1453 {
1454         struct file *in;
1455         int error, fput_in;
1456
1457         if (unlikely(!len))
1458                 return 0;
1459
1460         error = -EBADF;
1461         in = fget_light(fdin, &fput_in);
1462         if (in) {
1463                 if (in->f_mode & FMODE_READ) {
1464                         int fput_out;
1465                         struct file *out = fget_light(fdout, &fput_out);
1466
1467                         if (out) {
1468                                 if (out->f_mode & FMODE_WRITE)
1469                                         error = do_tee(in, out, len, flags);
1470                                 fput_light(out, fput_out);
1471                         }
1472                 }
1473                 fput_light(in, fput_in);
1474         }
1475
1476         return error;
1477 }