splice: remove i_mutex locking in splice_from_pipe()
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/memcontrol.h>
25 #include <linux/mm_inline.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
33
34 /*
35  * Attempt to steal a page from a pipe buffer. This should perhaps go into
36  * a vm helper function, it's already simplified quite a bit by the
37  * addition of remove_mapping(). If success is returned, the caller may
38  * attempt to reuse this page for another destination.
39  */
40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
41                                      struct pipe_buffer *buf)
42 {
43         struct page *page = buf->page;
44         struct address_space *mapping;
45
46         lock_page(page);
47
48         mapping = page_mapping(page);
49         if (mapping) {
50                 WARN_ON(!PageUptodate(page));
51
52                 /*
53                  * At least for ext2 with nobh option, we need to wait on
54                  * writeback completing on this page, since we'll remove it
55                  * from the pagecache.  Otherwise truncate wont wait on the
56                  * page, allowing the disk blocks to be reused by someone else
57                  * before we actually wrote our data to them. fs corruption
58                  * ensues.
59                  */
60                 wait_on_page_writeback(page);
61
62                 if (page_has_private(page) &&
63                     !try_to_release_page(page, GFP_KERNEL))
64                         goto out_unlock;
65
66                 /*
67                  * If we succeeded in removing the mapping, set LRU flag
68                  * and return good.
69                  */
70                 if (remove_mapping(mapping, page)) {
71                         buf->flags |= PIPE_BUF_FLAG_LRU;
72                         return 0;
73                 }
74         }
75
76         /*
77          * Raced with truncate or failed to remove page from current
78          * address space, unlock and return failure.
79          */
80 out_unlock:
81         unlock_page(page);
82         return 1;
83 }
84
85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
86                                         struct pipe_buffer *buf)
87 {
88         page_cache_release(buf->page);
89         buf->flags &= ~PIPE_BUF_FLAG_LRU;
90 }
91
92 /*
93  * Check whether the contents of buf is OK to access. Since the content
94  * is a page cache page, IO may be in flight.
95  */
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
97                                        struct pipe_buffer *buf)
98 {
99         struct page *page = buf->page;
100         int err;
101
102         if (!PageUptodate(page)) {
103                 lock_page(page);
104
105                 /*
106                  * Page got truncated/unhashed. This will cause a 0-byte
107                  * splice, if this is the first page.
108                  */
109                 if (!page->mapping) {
110                         err = -ENODATA;
111                         goto error;
112                 }
113
114                 /*
115                  * Uh oh, read-error from disk.
116                  */
117                 if (!PageUptodate(page)) {
118                         err = -EIO;
119                         goto error;
120                 }
121
122                 /*
123                  * Page is ok afterall, we are done.
124                  */
125                 unlock_page(page);
126         }
127
128         return 0;
129 error:
130         unlock_page(page);
131         return err;
132 }
133
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
135         .can_merge = 0,
136         .map = generic_pipe_buf_map,
137         .unmap = generic_pipe_buf_unmap,
138         .confirm = page_cache_pipe_buf_confirm,
139         .release = page_cache_pipe_buf_release,
140         .steal = page_cache_pipe_buf_steal,
141         .get = generic_pipe_buf_get,
142 };
143
144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
145                                     struct pipe_buffer *buf)
146 {
147         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
148                 return 1;
149
150         buf->flags |= PIPE_BUF_FLAG_LRU;
151         return generic_pipe_buf_steal(pipe, buf);
152 }
153
154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
155         .can_merge = 0,
156         .map = generic_pipe_buf_map,
157         .unmap = generic_pipe_buf_unmap,
158         .confirm = generic_pipe_buf_confirm,
159         .release = page_cache_pipe_buf_release,
160         .steal = user_page_pipe_buf_steal,
161         .get = generic_pipe_buf_get,
162 };
163
164 /**
165  * splice_to_pipe - fill passed data into a pipe
166  * @pipe:       pipe to fill
167  * @spd:        data to fill
168  *
169  * Description:
170  *    @spd contains a map of pages and len/offset tuples, along with
171  *    the struct pipe_buf_operations associated with these pages. This
172  *    function will link that data to the pipe.
173  *
174  */
175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
176                        struct splice_pipe_desc *spd)
177 {
178         unsigned int spd_pages = spd->nr_pages;
179         int ret, do_wakeup, page_nr;
180
181         ret = 0;
182         do_wakeup = 0;
183         page_nr = 0;
184
185         if (pipe->inode)
186                 mutex_lock(&pipe->inode->i_mutex);
187
188         for (;;) {
189                 if (!pipe->readers) {
190                         send_sig(SIGPIPE, current, 0);
191                         if (!ret)
192                                 ret = -EPIPE;
193                         break;
194                 }
195
196                 if (pipe->nrbufs < PIPE_BUFFERS) {
197                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
198                         struct pipe_buffer *buf = pipe->bufs + newbuf;
199
200                         buf->page = spd->pages[page_nr];
201                         buf->offset = spd->partial[page_nr].offset;
202                         buf->len = spd->partial[page_nr].len;
203                         buf->private = spd->partial[page_nr].private;
204                         buf->ops = spd->ops;
205                         if (spd->flags & SPLICE_F_GIFT)
206                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
207
208                         pipe->nrbufs++;
209                         page_nr++;
210                         ret += buf->len;
211
212                         if (pipe->inode)
213                                 do_wakeup = 1;
214
215                         if (!--spd->nr_pages)
216                                 break;
217                         if (pipe->nrbufs < PIPE_BUFFERS)
218                                 continue;
219
220                         break;
221                 }
222
223                 if (spd->flags & SPLICE_F_NONBLOCK) {
224                         if (!ret)
225                                 ret = -EAGAIN;
226                         break;
227                 }
228
229                 if (signal_pending(current)) {
230                         if (!ret)
231                                 ret = -ERESTARTSYS;
232                         break;
233                 }
234
235                 if (do_wakeup) {
236                         smp_mb();
237                         if (waitqueue_active(&pipe->wait))
238                                 wake_up_interruptible_sync(&pipe->wait);
239                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
240                         do_wakeup = 0;
241                 }
242
243                 pipe->waiting_writers++;
244                 pipe_wait(pipe);
245                 pipe->waiting_writers--;
246         }
247
248         if (pipe->inode) {
249                 mutex_unlock(&pipe->inode->i_mutex);
250
251                 if (do_wakeup) {
252                         smp_mb();
253                         if (waitqueue_active(&pipe->wait))
254                                 wake_up_interruptible(&pipe->wait);
255                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
256                 }
257         }
258
259         while (page_nr < spd_pages)
260                 spd->spd_release(spd, page_nr++);
261
262         return ret;
263 }
264
265 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
266 {
267         page_cache_release(spd->pages[i]);
268 }
269
270 static int
271 __generic_file_splice_read(struct file *in, loff_t *ppos,
272                            struct pipe_inode_info *pipe, size_t len,
273                            unsigned int flags)
274 {
275         struct address_space *mapping = in->f_mapping;
276         unsigned int loff, nr_pages, req_pages;
277         struct page *pages[PIPE_BUFFERS];
278         struct partial_page partial[PIPE_BUFFERS];
279         struct page *page;
280         pgoff_t index, end_index;
281         loff_t isize;
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                 .spd_release = spd_release_page,
289         };
290
291         index = *ppos >> PAGE_CACHE_SHIFT;
292         loff = *ppos & ~PAGE_CACHE_MASK;
293         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
294         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
295
296         /*
297          * Lookup the (hopefully) full range of pages we need.
298          */
299         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
300         index += spd.nr_pages;
301
302         /*
303          * If find_get_pages_contig() returned fewer pages than we needed,
304          * readahead/allocate the rest and fill in the holes.
305          */
306         if (spd.nr_pages < nr_pages)
307                 page_cache_sync_readahead(mapping, &in->f_ra, in,
308                                 index, req_pages - spd.nr_pages);
309
310         error = 0;
311         while (spd.nr_pages < nr_pages) {
312                 /*
313                  * Page could be there, find_get_pages_contig() breaks on
314                  * the first hole.
315                  */
316                 page = find_get_page(mapping, index);
317                 if (!page) {
318                         /*
319                          * page didn't exist, allocate one.
320                          */
321                         page = page_cache_alloc_cold(mapping);
322                         if (!page)
323                                 break;
324
325                         error = add_to_page_cache_lru(page, mapping, index,
326                                                 mapping_gfp_mask(mapping));
327                         if (unlikely(error)) {
328                                 page_cache_release(page);
329                                 if (error == -EEXIST)
330                                         continue;
331                                 break;
332                         }
333                         /*
334                          * add_to_page_cache() locks the page, unlock it
335                          * to avoid convoluting the logic below even more.
336                          */
337                         unlock_page(page);
338                 }
339
340                 pages[spd.nr_pages++] = page;
341                 index++;
342         }
343
344         /*
345          * Now loop over the map and see if we need to start IO on any
346          * pages, fill in the partial map, etc.
347          */
348         index = *ppos >> PAGE_CACHE_SHIFT;
349         nr_pages = spd.nr_pages;
350         spd.nr_pages = 0;
351         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
352                 unsigned int this_len;
353
354                 if (!len)
355                         break;
356
357                 /*
358                  * this_len is the max we'll use from this page
359                  */
360                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
361                 page = pages[page_nr];
362
363                 if (PageReadahead(page))
364                         page_cache_async_readahead(mapping, &in->f_ra, in,
365                                         page, index, req_pages - page_nr);
366
367                 /*
368                  * If the page isn't uptodate, we may need to start io on it
369                  */
370                 if (!PageUptodate(page)) {
371                         /*
372                          * If in nonblock mode then dont block on waiting
373                          * for an in-flight io page
374                          */
375                         if (flags & SPLICE_F_NONBLOCK) {
376                                 if (!trylock_page(page)) {
377                                         error = -EAGAIN;
378                                         break;
379                                 }
380                         } else
381                                 lock_page(page);
382
383                         /*
384                          * Page was truncated, or invalidated by the
385                          * filesystem.  Redo the find/create, but this time the
386                          * page is kept locked, so there's no chance of another
387                          * race with truncate/invalidate.
388                          */
389                         if (!page->mapping) {
390                                 unlock_page(page);
391                                 page = find_or_create_page(mapping, index,
392                                                 mapping_gfp_mask(mapping));
393
394                                 if (!page) {
395                                         error = -ENOMEM;
396                                         break;
397                                 }
398                                 page_cache_release(pages[page_nr]);
399                                 pages[page_nr] = page;
400                         }
401                         /*
402                          * page was already under io and is now done, great
403                          */
404                         if (PageUptodate(page)) {
405                                 unlock_page(page);
406                                 goto fill_it;
407                         }
408
409                         /*
410                          * need to read in the page
411                          */
412                         error = mapping->a_ops->readpage(in, page);
413                         if (unlikely(error)) {
414                                 /*
415                                  * We really should re-lookup the page here,
416                                  * but it complicates things a lot. Instead
417                                  * lets just do what we already stored, and
418                                  * we'll get it the next time we are called.
419                                  */
420                                 if (error == AOP_TRUNCATED_PAGE)
421                                         error = 0;
422
423                                 break;
424                         }
425                 }
426 fill_it:
427                 /*
428                  * i_size must be checked after PageUptodate.
429                  */
430                 isize = i_size_read(mapping->host);
431                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
432                 if (unlikely(!isize || index > end_index))
433                         break;
434
435                 /*
436                  * if this is the last page, see if we need to shrink
437                  * the length and stop
438                  */
439                 if (end_index == index) {
440                         unsigned int plen;
441
442                         /*
443                          * max good bytes in this page
444                          */
445                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
446                         if (plen <= loff)
447                                 break;
448
449                         /*
450                          * force quit after adding this page
451                          */
452                         this_len = min(this_len, plen - loff);
453                         len = this_len;
454                 }
455
456                 partial[page_nr].offset = loff;
457                 partial[page_nr].len = this_len;
458                 len -= this_len;
459                 loff = 0;
460                 spd.nr_pages++;
461                 index++;
462         }
463
464         /*
465          * Release any pages at the end, if we quit early. 'page_nr' is how far
466          * we got, 'nr_pages' is how many pages are in the map.
467          */
468         while (page_nr < nr_pages)
469                 page_cache_release(pages[page_nr++]);
470         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
471
472         if (spd.nr_pages)
473                 return splice_to_pipe(pipe, &spd);
474
475         return error;
476 }
477
478 /**
479  * generic_file_splice_read - splice data from file to a pipe
480  * @in:         file to splice from
481  * @ppos:       position in @in
482  * @pipe:       pipe to splice to
483  * @len:        number of bytes to splice
484  * @flags:      splice modifier flags
485  *
486  * Description:
487  *    Will read pages from given file and fill them into a pipe. Can be
488  *    used as long as the address_space operations for the source implements
489  *    a readpage() hook.
490  *
491  */
492 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
493                                  struct pipe_inode_info *pipe, size_t len,
494                                  unsigned int flags)
495 {
496         loff_t isize, left;
497         int ret;
498
499         isize = i_size_read(in->f_mapping->host);
500         if (unlikely(*ppos >= isize))
501                 return 0;
502
503         left = isize - *ppos;
504         if (unlikely(left < len))
505                 len = left;
506
507         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
508         if (ret > 0)
509                 *ppos += ret;
510
511         return ret;
512 }
513
514 EXPORT_SYMBOL(generic_file_splice_read);
515
516 /*
517  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
518  * using sendpage(). Return the number of bytes sent.
519  */
520 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
521                             struct pipe_buffer *buf, struct splice_desc *sd)
522 {
523         struct file *file = sd->u.file;
524         loff_t pos = sd->pos;
525         int ret, more;
526
527         ret = buf->ops->confirm(pipe, buf);
528         if (!ret) {
529                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
530
531                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
532                                            sd->len, &pos, more);
533         }
534
535         return ret;
536 }
537
538 /*
539  * This is a little more tricky than the file -> pipe splicing. There are
540  * basically three cases:
541  *
542  *      - Destination page already exists in the address space and there
543  *        are users of it. For that case we have no other option that
544  *        copying the data. Tough luck.
545  *      - Destination page already exists in the address space, but there
546  *        are no users of it. Make sure it's uptodate, then drop it. Fall
547  *        through to last case.
548  *      - Destination page does not exist, we can add the pipe page to
549  *        the page cache and avoid the copy.
550  *
551  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
552  * sd->flags), we attempt to migrate pages from the pipe to the output
553  * file address space page cache. This is possible if no one else has
554  * the pipe page referenced outside of the pipe and page cache. If
555  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
556  * a new page in the output file page cache and fill/dirty that.
557  */
558 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
559                         struct splice_desc *sd)
560 {
561         struct file *file = sd->u.file;
562         struct address_space *mapping = file->f_mapping;
563         unsigned int offset, this_len;
564         struct page *page;
565         void *fsdata;
566         int ret;
567
568         /*
569          * make sure the data in this buffer is uptodate
570          */
571         ret = buf->ops->confirm(pipe, buf);
572         if (unlikely(ret))
573                 return ret;
574
575         offset = sd->pos & ~PAGE_CACHE_MASK;
576
577         this_len = sd->len;
578         if (this_len + offset > PAGE_CACHE_SIZE)
579                 this_len = PAGE_CACHE_SIZE - offset;
580
581         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
582                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
583         if (unlikely(ret))
584                 goto out;
585
586         if (buf->page != page) {
587                 /*
588                  * Careful, ->map() uses KM_USER0!
589                  */
590                 char *src = buf->ops->map(pipe, buf, 1);
591                 char *dst = kmap_atomic(page, KM_USER1);
592
593                 memcpy(dst + offset, src + buf->offset, this_len);
594                 flush_dcache_page(page);
595                 kunmap_atomic(dst, KM_USER1);
596                 buf->ops->unmap(pipe, buf, src);
597         }
598         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
599                                 page, fsdata);
600 out:
601         return ret;
602 }
603
604 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
605 {
606         smp_mb();
607         if (waitqueue_active(&pipe->wait))
608                 wake_up_interruptible(&pipe->wait);
609         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
610 }
611
612 /**
613  * splice_from_pipe_feed - feed available data from a pipe to a file
614  * @pipe:       pipe to splice from
615  * @sd:         information to @actor
616  * @actor:      handler that splices the data
617  *
618  * Description:
619
620  *    This function loops over the pipe and calls @actor to do the
621  *    actual moving of a single struct pipe_buffer to the desired
622  *    destination.  It returns when there's no more buffers left in
623  *    the pipe or if the requested number of bytes (@sd->total_len)
624  *    have been copied.  It returns a positive number (one) if the
625  *    pipe needs to be filled with more data, zero if the required
626  *    number of bytes have been copied and -errno on error.
627  *
628  *    This, together with splice_from_pipe_{begin,end,next}, may be
629  *    used to implement the functionality of __splice_from_pipe() when
630  *    locking is required around copying the pipe buffers to the
631  *    destination.
632  */
633 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
634                           splice_actor *actor)
635 {
636         int ret;
637
638         while (pipe->nrbufs) {
639                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
640                 const struct pipe_buf_operations *ops = buf->ops;
641
642                 sd->len = buf->len;
643                 if (sd->len > sd->total_len)
644                         sd->len = sd->total_len;
645
646                 ret = actor(pipe, buf, sd);
647                 if (ret <= 0) {
648                         if (ret == -ENODATA)
649                                 ret = 0;
650                         return ret;
651                 }
652                 buf->offset += ret;
653                 buf->len -= ret;
654
655                 sd->num_spliced += ret;
656                 sd->len -= ret;
657                 sd->pos += ret;
658                 sd->total_len -= ret;
659
660                 if (!buf->len) {
661                         buf->ops = NULL;
662                         ops->release(pipe, buf);
663                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
664                         pipe->nrbufs--;
665                         if (pipe->inode)
666                                 sd->need_wakeup = true;
667                 }
668
669                 if (!sd->total_len)
670                         return 0;
671         }
672
673         return 1;
674 }
675 EXPORT_SYMBOL(splice_from_pipe_feed);
676
677 /**
678  * splice_from_pipe_next - wait for some data to splice from
679  * @pipe:       pipe to splice from
680  * @sd:         information about the splice operation
681  *
682  * Description:
683  *    This function will wait for some data and return a positive
684  *    value (one) if pipe buffers are available.  It will return zero
685  *    or -errno if no more data needs to be spliced.
686  */
687 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
688 {
689         while (!pipe->nrbufs) {
690                 if (!pipe->writers)
691                         return 0;
692
693                 if (!pipe->waiting_writers && sd->num_spliced)
694                         return 0;
695
696                 if (sd->flags & SPLICE_F_NONBLOCK)
697                         return -EAGAIN;
698
699                 if (signal_pending(current))
700                         return -ERESTARTSYS;
701
702                 if (sd->need_wakeup) {
703                         wakeup_pipe_writers(pipe);
704                         sd->need_wakeup = false;
705                 }
706
707                 pipe_wait(pipe);
708         }
709
710         return 1;
711 }
712 EXPORT_SYMBOL(splice_from_pipe_next);
713
714 /**
715  * splice_from_pipe_begin - start splicing from pipe
716  * @pipe:       pipe to splice from
717  *
718  * Description:
719  *    This function should be called before a loop containing
720  *    splice_from_pipe_next() and splice_from_pipe_feed() to
721  *    initialize the necessary fields of @sd.
722  */
723 void splice_from_pipe_begin(struct splice_desc *sd)
724 {
725         sd->num_spliced = 0;
726         sd->need_wakeup = false;
727 }
728 EXPORT_SYMBOL(splice_from_pipe_begin);
729
730 /**
731  * splice_from_pipe_end - finish splicing from pipe
732  * @pipe:       pipe to splice from
733  * @sd:         information about the splice operation
734  *
735  * Description:
736  *    This function will wake up pipe writers if necessary.  It should
737  *    be called after a loop containing splice_from_pipe_next() and
738  *    splice_from_pipe_feed().
739  */
740 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
741 {
742         if (sd->need_wakeup)
743                 wakeup_pipe_writers(pipe);
744 }
745 EXPORT_SYMBOL(splice_from_pipe_end);
746
747 /**
748  * __splice_from_pipe - splice data from a pipe to given actor
749  * @pipe:       pipe to splice from
750  * @sd:         information to @actor
751  * @actor:      handler that splices the data
752  *
753  * Description:
754  *    This function does little more than loop over the pipe and call
755  *    @actor to do the actual moving of a single struct pipe_buffer to
756  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
757  *    pipe_to_user.
758  *
759  */
760 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
761                            splice_actor *actor)
762 {
763         int ret;
764
765         splice_from_pipe_begin(sd);
766         do {
767                 ret = splice_from_pipe_next(pipe, sd);
768                 if (ret > 0)
769                         ret = splice_from_pipe_feed(pipe, sd, actor);
770         } while (ret > 0);
771         splice_from_pipe_end(pipe, sd);
772
773         return sd->num_spliced ? sd->num_spliced : ret;
774 }
775 EXPORT_SYMBOL(__splice_from_pipe);
776
777 /**
778  * splice_from_pipe - splice data from a pipe to a file
779  * @pipe:       pipe to splice from
780  * @out:        file to splice to
781  * @ppos:       position in @out
782  * @len:        how many bytes to splice
783  * @flags:      splice modifier flags
784  * @actor:      handler that splices the data
785  *
786  * Description:
787  *    See __splice_from_pipe. This function locks the pipe inode,
788  *    otherwise it's identical to __splice_from_pipe().
789  *
790  */
791 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
792                          loff_t *ppos, size_t len, unsigned int flags,
793                          splice_actor *actor)
794 {
795         ssize_t ret;
796         struct splice_desc sd = {
797                 .total_len = len,
798                 .flags = flags,
799                 .pos = *ppos,
800                 .u.file = out,
801         };
802
803         if (pipe->inode)
804                 mutex_lock(&pipe->inode->i_mutex);
805         ret = __splice_from_pipe(pipe, &sd, actor);
806         if (pipe->inode)
807                 mutex_unlock(&pipe->inode->i_mutex);
808
809         return ret;
810 }
811
812 /**
813  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
814  * @pipe:       pipe info
815  * @out:        file to write to
816  * @ppos:       position in @out
817  * @len:        number of bytes to splice
818  * @flags:      splice modifier flags
819  *
820  * Description:
821  *    Will either move or copy pages (determined by @flags options) from
822  *    the given pipe inode to the given file. The caller is responsible
823  *    for acquiring i_mutex on both inodes.
824  *
825  */
826 ssize_t
827 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
828                                  loff_t *ppos, size_t len, unsigned int flags)
829 {
830         struct address_space *mapping = out->f_mapping;
831         struct inode *inode = mapping->host;
832         struct splice_desc sd = {
833                 .total_len = len,
834                 .flags = flags,
835                 .pos = *ppos,
836                 .u.file = out,
837         };
838         ssize_t ret;
839         int err;
840
841         err = file_remove_suid(out);
842         if (unlikely(err))
843                 return err;
844
845         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
846         if (ret > 0) {
847                 unsigned long nr_pages;
848
849                 *ppos += ret;
850                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
851
852                 /*
853                  * If file or inode is SYNC and we actually wrote some data,
854                  * sync it.
855                  */
856                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
857                         err = generic_osync_inode(inode, mapping,
858                                                   OSYNC_METADATA|OSYNC_DATA);
859
860                         if (err)
861                                 ret = err;
862                 }
863                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
864         }
865
866         return ret;
867 }
868
869 EXPORT_SYMBOL(generic_file_splice_write_nolock);
870
871 /**
872  * generic_file_splice_write - splice data from a pipe to a file
873  * @pipe:       pipe info
874  * @out:        file to write to
875  * @ppos:       position in @out
876  * @len:        number of bytes to splice
877  * @flags:      splice modifier flags
878  *
879  * Description:
880  *    Will either move or copy pages (determined by @flags options) from
881  *    the given pipe inode to the given file.
882  *
883  */
884 ssize_t
885 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
886                           loff_t *ppos, size_t len, unsigned int flags)
887 {
888         struct address_space *mapping = out->f_mapping;
889         struct inode *inode = mapping->host;
890         struct splice_desc sd = {
891                 .total_len = len,
892                 .flags = flags,
893                 .pos = *ppos,
894                 .u.file = out,
895         };
896         ssize_t ret;
897
898         WARN_ON(S_ISFIFO(inode->i_mode));
899         mutex_lock_nested(&inode->i_mutex, I_MUTEX_PARENT);
900         ret = file_remove_suid(out);
901         if (likely(!ret)) {
902                 if (pipe->inode)
903                         mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_CHILD);
904                 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
905                 if (pipe->inode)
906                         mutex_unlock(&pipe->inode->i_mutex);
907         }
908         mutex_unlock(&inode->i_mutex);
909         if (ret > 0) {
910                 unsigned long nr_pages;
911
912                 *ppos += ret;
913                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
914
915                 /*
916                  * If file or inode is SYNC and we actually wrote some data,
917                  * sync it.
918                  */
919                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
920                         int err;
921
922                         mutex_lock(&inode->i_mutex);
923                         err = generic_osync_inode(inode, mapping,
924                                                   OSYNC_METADATA|OSYNC_DATA);
925                         mutex_unlock(&inode->i_mutex);
926
927                         if (err)
928                                 ret = err;
929                 }
930                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
931         }
932
933         return ret;
934 }
935
936 EXPORT_SYMBOL(generic_file_splice_write);
937
938 /**
939  * generic_splice_sendpage - splice data from a pipe to a socket
940  * @pipe:       pipe to splice from
941  * @out:        socket to write to
942  * @ppos:       position in @out
943  * @len:        number of bytes to splice
944  * @flags:      splice modifier flags
945  *
946  * Description:
947  *    Will send @len bytes from the pipe to a network socket. No data copying
948  *    is involved.
949  *
950  */
951 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
952                                 loff_t *ppos, size_t len, unsigned int flags)
953 {
954         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
955 }
956
957 EXPORT_SYMBOL(generic_splice_sendpage);
958
959 /*
960  * Attempt to initiate a splice from pipe to file.
961  */
962 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
963                            loff_t *ppos, size_t len, unsigned int flags)
964 {
965         int ret;
966
967         if (unlikely(!out->f_op || !out->f_op->splice_write))
968                 return -EINVAL;
969
970         if (unlikely(!(out->f_mode & FMODE_WRITE)))
971                 return -EBADF;
972
973         if (unlikely(out->f_flags & O_APPEND))
974                 return -EINVAL;
975
976         ret = rw_verify_area(WRITE, out, ppos, len);
977         if (unlikely(ret < 0))
978                 return ret;
979
980         return out->f_op->splice_write(pipe, out, ppos, len, flags);
981 }
982
983 /*
984  * Attempt to initiate a splice from a file to a pipe.
985  */
986 static long do_splice_to(struct file *in, loff_t *ppos,
987                          struct pipe_inode_info *pipe, size_t len,
988                          unsigned int flags)
989 {
990         int ret;
991
992         if (unlikely(!in->f_op || !in->f_op->splice_read))
993                 return -EINVAL;
994
995         if (unlikely(!(in->f_mode & FMODE_READ)))
996                 return -EBADF;
997
998         ret = rw_verify_area(READ, in, ppos, len);
999         if (unlikely(ret < 0))
1000                 return ret;
1001
1002         return in->f_op->splice_read(in, ppos, pipe, len, flags);
1003 }
1004
1005 /**
1006  * splice_direct_to_actor - splices data directly between two non-pipes
1007  * @in:         file to splice from
1008  * @sd:         actor information on where to splice to
1009  * @actor:      handles the data splicing
1010  *
1011  * Description:
1012  *    This is a special case helper to splice directly between two
1013  *    points, without requiring an explicit pipe. Internally an allocated
1014  *    pipe is cached in the process, and reused during the lifetime of
1015  *    that process.
1016  *
1017  */
1018 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1019                                splice_direct_actor *actor)
1020 {
1021         struct pipe_inode_info *pipe;
1022         long ret, bytes;
1023         umode_t i_mode;
1024         size_t len;
1025         int i, flags;
1026
1027         /*
1028          * We require the input being a regular file, as we don't want to
1029          * randomly drop data for eg socket -> socket splicing. Use the
1030          * piped splicing for that!
1031          */
1032         i_mode = in->f_path.dentry->d_inode->i_mode;
1033         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1034                 return -EINVAL;
1035
1036         /*
1037          * neither in nor out is a pipe, setup an internal pipe attached to
1038          * 'out' and transfer the wanted data from 'in' to 'out' through that
1039          */
1040         pipe = current->splice_pipe;
1041         if (unlikely(!pipe)) {
1042                 pipe = alloc_pipe_info(NULL);
1043                 if (!pipe)
1044                         return -ENOMEM;
1045
1046                 /*
1047                  * We don't have an immediate reader, but we'll read the stuff
1048                  * out of the pipe right after the splice_to_pipe(). So set
1049                  * PIPE_READERS appropriately.
1050                  */
1051                 pipe->readers = 1;
1052
1053                 current->splice_pipe = pipe;
1054         }
1055
1056         /*
1057          * Do the splice.
1058          */
1059         ret = 0;
1060         bytes = 0;
1061         len = sd->total_len;
1062         flags = sd->flags;
1063
1064         /*
1065          * Don't block on output, we have to drain the direct pipe.
1066          */
1067         sd->flags &= ~SPLICE_F_NONBLOCK;
1068
1069         while (len) {
1070                 size_t read_len;
1071                 loff_t pos = sd->pos, prev_pos = pos;
1072
1073                 ret = do_splice_to(in, &pos, pipe, len, flags);
1074                 if (unlikely(ret <= 0))
1075                         goto out_release;
1076
1077                 read_len = ret;
1078                 sd->total_len = read_len;
1079
1080                 /*
1081                  * NOTE: nonblocking mode only applies to the input. We
1082                  * must not do the output in nonblocking mode as then we
1083                  * could get stuck data in the internal pipe:
1084                  */
1085                 ret = actor(pipe, sd);
1086                 if (unlikely(ret <= 0)) {
1087                         sd->pos = prev_pos;
1088                         goto out_release;
1089                 }
1090
1091                 bytes += ret;
1092                 len -= ret;
1093                 sd->pos = pos;
1094
1095                 if (ret < read_len) {
1096                         sd->pos = prev_pos + ret;
1097                         goto out_release;
1098                 }
1099         }
1100
1101 done:
1102         pipe->nrbufs = pipe->curbuf = 0;
1103         file_accessed(in);
1104         return bytes;
1105
1106 out_release:
1107         /*
1108          * If we did an incomplete transfer we must release
1109          * the pipe buffers in question:
1110          */
1111         for (i = 0; i < PIPE_BUFFERS; i++) {
1112                 struct pipe_buffer *buf = pipe->bufs + i;
1113
1114                 if (buf->ops) {
1115                         buf->ops->release(pipe, buf);
1116                         buf->ops = NULL;
1117                 }
1118         }
1119
1120         if (!bytes)
1121                 bytes = ret;
1122
1123         goto done;
1124 }
1125 EXPORT_SYMBOL(splice_direct_to_actor);
1126
1127 static int direct_splice_actor(struct pipe_inode_info *pipe,
1128                                struct splice_desc *sd)
1129 {
1130         struct file *file = sd->u.file;
1131
1132         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1133 }
1134
1135 /**
1136  * do_splice_direct - splices data directly between two files
1137  * @in:         file to splice from
1138  * @ppos:       input file offset
1139  * @out:        file to splice to
1140  * @len:        number of bytes to splice
1141  * @flags:      splice modifier flags
1142  *
1143  * Description:
1144  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1145  *    doing it in the application would incur an extra system call
1146  *    (splice in + splice out, as compared to just sendfile()). So this helper
1147  *    can splice directly through a process-private pipe.
1148  *
1149  */
1150 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1151                       size_t len, unsigned int flags)
1152 {
1153         struct splice_desc sd = {
1154                 .len            = len,
1155                 .total_len      = len,
1156                 .flags          = flags,
1157                 .pos            = *ppos,
1158                 .u.file         = out,
1159         };
1160         long ret;
1161
1162         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1163         if (ret > 0)
1164                 *ppos = sd.pos;
1165
1166         return ret;
1167 }
1168
1169 /*
1170  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1171  * location, so checking ->i_pipe is not enough to verify that this is a
1172  * pipe.
1173  */
1174 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1175 {
1176         if (S_ISFIFO(inode->i_mode))
1177                 return inode->i_pipe;
1178
1179         return NULL;
1180 }
1181
1182 /*
1183  * Determine where to splice to/from.
1184  */
1185 static long do_splice(struct file *in, loff_t __user *off_in,
1186                       struct file *out, loff_t __user *off_out,
1187                       size_t len, unsigned int flags)
1188 {
1189         struct pipe_inode_info *pipe;
1190         loff_t offset, *off;
1191         long ret;
1192
1193         pipe = pipe_info(in->f_path.dentry->d_inode);
1194         if (pipe) {
1195                 if (off_in)
1196                         return -ESPIPE;
1197                 if (off_out) {
1198                         if (out->f_op->llseek == no_llseek)
1199                                 return -EINVAL;
1200                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1201                                 return -EFAULT;
1202                         off = &offset;
1203                 } else
1204                         off = &out->f_pos;
1205
1206                 ret = do_splice_from(pipe, out, off, len, flags);
1207
1208                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1209                         ret = -EFAULT;
1210
1211                 return ret;
1212         }
1213
1214         pipe = pipe_info(out->f_path.dentry->d_inode);
1215         if (pipe) {
1216                 if (off_out)
1217                         return -ESPIPE;
1218                 if (off_in) {
1219                         if (in->f_op->llseek == no_llseek)
1220                                 return -EINVAL;
1221                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1222                                 return -EFAULT;
1223                         off = &offset;
1224                 } else
1225                         off = &in->f_pos;
1226
1227                 ret = do_splice_to(in, off, pipe, len, flags);
1228
1229                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1230                         ret = -EFAULT;
1231
1232                 return ret;
1233         }
1234
1235         return -EINVAL;
1236 }
1237
1238 /*
1239  * Map an iov into an array of pages and offset/length tupples. With the
1240  * partial_page structure, we can map several non-contiguous ranges into
1241  * our ones pages[] map instead of splitting that operation into pieces.
1242  * Could easily be exported as a generic helper for other users, in which
1243  * case one would probably want to add a 'max_nr_pages' parameter as well.
1244  */
1245 static int get_iovec_page_array(const struct iovec __user *iov,
1246                                 unsigned int nr_vecs, struct page **pages,
1247                                 struct partial_page *partial, int aligned)
1248 {
1249         int buffers = 0, error = 0;
1250
1251         while (nr_vecs) {
1252                 unsigned long off, npages;
1253                 struct iovec entry;
1254                 void __user *base;
1255                 size_t len;
1256                 int i;
1257
1258                 error = -EFAULT;
1259                 if (copy_from_user(&entry, iov, sizeof(entry)))
1260                         break;
1261
1262                 base = entry.iov_base;
1263                 len = entry.iov_len;
1264
1265                 /*
1266                  * Sanity check this iovec. 0 read succeeds.
1267                  */
1268                 error = 0;
1269                 if (unlikely(!len))
1270                         break;
1271                 error = -EFAULT;
1272                 if (!access_ok(VERIFY_READ, base, len))
1273                         break;
1274
1275                 /*
1276                  * Get this base offset and number of pages, then map
1277                  * in the user pages.
1278                  */
1279                 off = (unsigned long) base & ~PAGE_MASK;
1280
1281                 /*
1282                  * If asked for alignment, the offset must be zero and the
1283                  * length a multiple of the PAGE_SIZE.
1284                  */
1285                 error = -EINVAL;
1286                 if (aligned && (off || len & ~PAGE_MASK))
1287                         break;
1288
1289                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1290                 if (npages > PIPE_BUFFERS - buffers)
1291                         npages = PIPE_BUFFERS - buffers;
1292
1293                 error = get_user_pages_fast((unsigned long)base, npages,
1294                                         0, &pages[buffers]);
1295
1296                 if (unlikely(error <= 0))
1297                         break;
1298
1299                 /*
1300                  * Fill this contiguous range into the partial page map.
1301                  */
1302                 for (i = 0; i < error; i++) {
1303                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1304
1305                         partial[buffers].offset = off;
1306                         partial[buffers].len = plen;
1307
1308                         off = 0;
1309                         len -= plen;
1310                         buffers++;
1311                 }
1312
1313                 /*
1314                  * We didn't complete this iov, stop here since it probably
1315                  * means we have to move some of this into a pipe to
1316                  * be able to continue.
1317                  */
1318                 if (len)
1319                         break;
1320
1321                 /*
1322                  * Don't continue if we mapped fewer pages than we asked for,
1323                  * or if we mapped the max number of pages that we have
1324                  * room for.
1325                  */
1326                 if (error < npages || buffers == PIPE_BUFFERS)
1327                         break;
1328
1329                 nr_vecs--;
1330                 iov++;
1331         }
1332
1333         if (buffers)
1334                 return buffers;
1335
1336         return error;
1337 }
1338
1339 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1340                         struct splice_desc *sd)
1341 {
1342         char *src;
1343         int ret;
1344
1345         ret = buf->ops->confirm(pipe, buf);
1346         if (unlikely(ret))
1347                 return ret;
1348
1349         /*
1350          * See if we can use the atomic maps, by prefaulting in the
1351          * pages and doing an atomic copy
1352          */
1353         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1354                 src = buf->ops->map(pipe, buf, 1);
1355                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1356                                                         sd->len);
1357                 buf->ops->unmap(pipe, buf, src);
1358                 if (!ret) {
1359                         ret = sd->len;
1360                         goto out;
1361                 }
1362         }
1363
1364         /*
1365          * No dice, use slow non-atomic map and copy
1366          */
1367         src = buf->ops->map(pipe, buf, 0);
1368
1369         ret = sd->len;
1370         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1371                 ret = -EFAULT;
1372
1373         buf->ops->unmap(pipe, buf, src);
1374 out:
1375         if (ret > 0)
1376                 sd->u.userptr += ret;
1377         return ret;
1378 }
1379
1380 /*
1381  * For lack of a better implementation, implement vmsplice() to userspace
1382  * as a simple copy of the pipes pages to the user iov.
1383  */
1384 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1385                              unsigned long nr_segs, unsigned int flags)
1386 {
1387         struct pipe_inode_info *pipe;
1388         struct splice_desc sd;
1389         ssize_t size;
1390         int error;
1391         long ret;
1392
1393         pipe = pipe_info(file->f_path.dentry->d_inode);
1394         if (!pipe)
1395                 return -EBADF;
1396
1397         if (pipe->inode)
1398                 mutex_lock(&pipe->inode->i_mutex);
1399
1400         error = ret = 0;
1401         while (nr_segs) {
1402                 void __user *base;
1403                 size_t len;
1404
1405                 /*
1406                  * Get user address base and length for this iovec.
1407                  */
1408                 error = get_user(base, &iov->iov_base);
1409                 if (unlikely(error))
1410                         break;
1411                 error = get_user(len, &iov->iov_len);
1412                 if (unlikely(error))
1413                         break;
1414
1415                 /*
1416                  * Sanity check this iovec. 0 read succeeds.
1417                  */
1418                 if (unlikely(!len))
1419                         break;
1420                 if (unlikely(!base)) {
1421                         error = -EFAULT;
1422                         break;
1423                 }
1424
1425                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1426                         error = -EFAULT;
1427                         break;
1428                 }
1429
1430                 sd.len = 0;
1431                 sd.total_len = len;
1432                 sd.flags = flags;
1433                 sd.u.userptr = base;
1434                 sd.pos = 0;
1435
1436                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1437                 if (size < 0) {
1438                         if (!ret)
1439                                 ret = size;
1440
1441                         break;
1442                 }
1443
1444                 ret += size;
1445
1446                 if (size < len)
1447                         break;
1448
1449                 nr_segs--;
1450                 iov++;
1451         }
1452
1453         if (pipe->inode)
1454                 mutex_unlock(&pipe->inode->i_mutex);
1455
1456         if (!ret)
1457                 ret = error;
1458
1459         return ret;
1460 }
1461
1462 /*
1463  * vmsplice splices a user address range into a pipe. It can be thought of
1464  * as splice-from-memory, where the regular splice is splice-from-file (or
1465  * to file). In both cases the output is a pipe, naturally.
1466  */
1467 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1468                              unsigned long nr_segs, unsigned int flags)
1469 {
1470         struct pipe_inode_info *pipe;
1471         struct page *pages[PIPE_BUFFERS];
1472         struct partial_page partial[PIPE_BUFFERS];
1473         struct splice_pipe_desc spd = {
1474                 .pages = pages,
1475                 .partial = partial,
1476                 .flags = flags,
1477                 .ops = &user_page_pipe_buf_ops,
1478                 .spd_release = spd_release_page,
1479         };
1480
1481         pipe = pipe_info(file->f_path.dentry->d_inode);
1482         if (!pipe)
1483                 return -EBADF;
1484
1485         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1486                                             flags & SPLICE_F_GIFT);
1487         if (spd.nr_pages <= 0)
1488                 return spd.nr_pages;
1489
1490         return splice_to_pipe(pipe, &spd);
1491 }
1492
1493 /*
1494  * Note that vmsplice only really supports true splicing _from_ user memory
1495  * to a pipe, not the other way around. Splicing from user memory is a simple
1496  * operation that can be supported without any funky alignment restrictions
1497  * or nasty vm tricks. We simply map in the user memory and fill them into
1498  * a pipe. The reverse isn't quite as easy, though. There are two possible
1499  * solutions for that:
1500  *
1501  *      - memcpy() the data internally, at which point we might as well just
1502  *        do a regular read() on the buffer anyway.
1503  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1504  *        has restriction limitations on both ends of the pipe).
1505  *
1506  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1507  *
1508  */
1509 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1510                 unsigned long, nr_segs, unsigned int, flags)
1511 {
1512         struct file *file;
1513         long error;
1514         int fput;
1515
1516         if (unlikely(nr_segs > UIO_MAXIOV))
1517                 return -EINVAL;
1518         else if (unlikely(!nr_segs))
1519                 return 0;
1520
1521         error = -EBADF;
1522         file = fget_light(fd, &fput);
1523         if (file) {
1524                 if (file->f_mode & FMODE_WRITE)
1525                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1526                 else if (file->f_mode & FMODE_READ)
1527                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1528
1529                 fput_light(file, fput);
1530         }
1531
1532         return error;
1533 }
1534
1535 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1536                 int, fd_out, loff_t __user *, off_out,
1537                 size_t, len, unsigned int, flags)
1538 {
1539         long error;
1540         struct file *in, *out;
1541         int fput_in, fput_out;
1542
1543         if (unlikely(!len))
1544                 return 0;
1545
1546         error = -EBADF;
1547         in = fget_light(fd_in, &fput_in);
1548         if (in) {
1549                 if (in->f_mode & FMODE_READ) {
1550                         out = fget_light(fd_out, &fput_out);
1551                         if (out) {
1552                                 if (out->f_mode & FMODE_WRITE)
1553                                         error = do_splice(in, off_in,
1554                                                           out, off_out,
1555                                                           len, flags);
1556                                 fput_light(out, fput_out);
1557                         }
1558                 }
1559
1560                 fput_light(in, fput_in);
1561         }
1562
1563         return error;
1564 }
1565
1566 /*
1567  * Make sure there's data to read. Wait for input if we can, otherwise
1568  * return an appropriate error.
1569  */
1570 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1571 {
1572         int ret;
1573
1574         /*
1575          * Check ->nrbufs without the inode lock first. This function
1576          * is speculative anyways, so missing one is ok.
1577          */
1578         if (pipe->nrbufs)
1579                 return 0;
1580
1581         ret = 0;
1582         mutex_lock(&pipe->inode->i_mutex);
1583
1584         while (!pipe->nrbufs) {
1585                 if (signal_pending(current)) {
1586                         ret = -ERESTARTSYS;
1587                         break;
1588                 }
1589                 if (!pipe->writers)
1590                         break;
1591                 if (!pipe->waiting_writers) {
1592                         if (flags & SPLICE_F_NONBLOCK) {
1593                                 ret = -EAGAIN;
1594                                 break;
1595                         }
1596                 }
1597                 pipe_wait(pipe);
1598         }
1599
1600         mutex_unlock(&pipe->inode->i_mutex);
1601         return ret;
1602 }
1603
1604 /*
1605  * Make sure there's writeable room. Wait for room if we can, otherwise
1606  * return an appropriate error.
1607  */
1608 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1609 {
1610         int ret;
1611
1612         /*
1613          * Check ->nrbufs without the inode lock first. This function
1614          * is speculative anyways, so missing one is ok.
1615          */
1616         if (pipe->nrbufs < PIPE_BUFFERS)
1617                 return 0;
1618
1619         ret = 0;
1620         mutex_lock(&pipe->inode->i_mutex);
1621
1622         while (pipe->nrbufs >= PIPE_BUFFERS) {
1623                 if (!pipe->readers) {
1624                         send_sig(SIGPIPE, current, 0);
1625                         ret = -EPIPE;
1626                         break;
1627                 }
1628                 if (flags & SPLICE_F_NONBLOCK) {
1629                         ret = -EAGAIN;
1630                         break;
1631                 }
1632                 if (signal_pending(current)) {
1633                         ret = -ERESTARTSYS;
1634                         break;
1635                 }
1636                 pipe->waiting_writers++;
1637                 pipe_wait(pipe);
1638                 pipe->waiting_writers--;
1639         }
1640
1641         mutex_unlock(&pipe->inode->i_mutex);
1642         return ret;
1643 }
1644
1645 /*
1646  * Link contents of ipipe to opipe.
1647  */
1648 static int link_pipe(struct pipe_inode_info *ipipe,
1649                      struct pipe_inode_info *opipe,
1650                      size_t len, unsigned int flags)
1651 {
1652         struct pipe_buffer *ibuf, *obuf;
1653         int ret = 0, i = 0, nbuf;
1654
1655         /*
1656          * Potential ABBA deadlock, work around it by ordering lock
1657          * grabbing by inode address. Otherwise two different processes
1658          * could deadlock (one doing tee from A -> B, the other from B -> A).
1659          */
1660         inode_double_lock(ipipe->inode, opipe->inode);
1661
1662         do {
1663                 if (!opipe->readers) {
1664                         send_sig(SIGPIPE, current, 0);
1665                         if (!ret)
1666                                 ret = -EPIPE;
1667                         break;
1668                 }
1669
1670                 /*
1671                  * If we have iterated all input buffers or ran out of
1672                  * output room, break.
1673                  */
1674                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1675                         break;
1676
1677                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1678                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1679
1680                 /*
1681                  * Get a reference to this pipe buffer,
1682                  * so we can copy the contents over.
1683                  */
1684                 ibuf->ops->get(ipipe, ibuf);
1685
1686                 obuf = opipe->bufs + nbuf;
1687                 *obuf = *ibuf;
1688
1689                 /*
1690                  * Don't inherit the gift flag, we need to
1691                  * prevent multiple steals of this page.
1692                  */
1693                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1694
1695                 if (obuf->len > len)
1696                         obuf->len = len;
1697
1698                 opipe->nrbufs++;
1699                 ret += obuf->len;
1700                 len -= obuf->len;
1701                 i++;
1702         } while (len);
1703
1704         /*
1705          * return EAGAIN if we have the potential of some data in the
1706          * future, otherwise just return 0
1707          */
1708         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1709                 ret = -EAGAIN;
1710
1711         inode_double_unlock(ipipe->inode, opipe->inode);
1712
1713         /*
1714          * If we put data in the output pipe, wakeup any potential readers.
1715          */
1716         if (ret > 0) {
1717                 smp_mb();
1718                 if (waitqueue_active(&opipe->wait))
1719                         wake_up_interruptible(&opipe->wait);
1720                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1721         }
1722
1723         return ret;
1724 }
1725
1726 /*
1727  * This is a tee(1) implementation that works on pipes. It doesn't copy
1728  * any data, it simply references the 'in' pages on the 'out' pipe.
1729  * The 'flags' used are the SPLICE_F_* variants, currently the only
1730  * applicable one is SPLICE_F_NONBLOCK.
1731  */
1732 static long do_tee(struct file *in, struct file *out, size_t len,
1733                    unsigned int flags)
1734 {
1735         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1736         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1737         int ret = -EINVAL;
1738
1739         /*
1740          * Duplicate the contents of ipipe to opipe without actually
1741          * copying the data.
1742          */
1743         if (ipipe && opipe && ipipe != opipe) {
1744                 /*
1745                  * Keep going, unless we encounter an error. The ipipe/opipe
1746                  * ordering doesn't really matter.
1747                  */
1748                 ret = link_ipipe_prep(ipipe, flags);
1749                 if (!ret) {
1750                         ret = link_opipe_prep(opipe, flags);
1751                         if (!ret)
1752                                 ret = link_pipe(ipipe, opipe, len, flags);
1753                 }
1754         }
1755
1756         return ret;
1757 }
1758
1759 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1760 {
1761         struct file *in;
1762         int error, fput_in;
1763
1764         if (unlikely(!len))
1765                 return 0;
1766
1767         error = -EBADF;
1768         in = fget_light(fdin, &fput_in);
1769         if (in) {
1770                 if (in->f_mode & FMODE_READ) {
1771                         int fput_out;
1772                         struct file *out = fget_light(fdout, &fput_out);
1773
1774                         if (out) {
1775                                 if (out->f_mode & FMODE_WRITE)
1776                                         error = do_tee(in, out, len, flags);
1777                                 fput_light(out, fput_out);
1778                         }
1779                 }
1780                 fput_light(in, fput_in);
1781         }
1782
1783         return error;
1784 }