memcg: synchronized LRU
[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 (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
63                         goto out_unlock;
64
65                 /*
66                  * If we succeeded in removing the mapping, set LRU flag
67                  * and return good.
68                  */
69                 if (remove_mapping(mapping, page)) {
70                         buf->flags |= PIPE_BUF_FLAG_LRU;
71                         return 0;
72                 }
73         }
74
75         /*
76          * Raced with truncate or failed to remove page from current
77          * address space, unlock and return failure.
78          */
79 out_unlock:
80         unlock_page(page);
81         return 1;
82 }
83
84 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
85                                         struct pipe_buffer *buf)
86 {
87         page_cache_release(buf->page);
88         buf->flags &= ~PIPE_BUF_FLAG_LRU;
89 }
90
91 /*
92  * Check whether the contents of buf is OK to access. Since the content
93  * is a page cache page, IO may be in flight.
94  */
95 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
96                                        struct pipe_buffer *buf)
97 {
98         struct page *page = buf->page;
99         int err;
100
101         if (!PageUptodate(page)) {
102                 lock_page(page);
103
104                 /*
105                  * Page got truncated/unhashed. This will cause a 0-byte
106                  * splice, if this is the first page.
107                  */
108                 if (!page->mapping) {
109                         err = -ENODATA;
110                         goto error;
111                 }
112
113                 /*
114                  * Uh oh, read-error from disk.
115                  */
116                 if (!PageUptodate(page)) {
117                         err = -EIO;
118                         goto error;
119                 }
120
121                 /*
122                  * Page is ok afterall, we are done.
123                  */
124                 unlock_page(page);
125         }
126
127         return 0;
128 error:
129         unlock_page(page);
130         return err;
131 }
132
133 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
134         .can_merge = 0,
135         .map = generic_pipe_buf_map,
136         .unmap = generic_pipe_buf_unmap,
137         .confirm = page_cache_pipe_buf_confirm,
138         .release = page_cache_pipe_buf_release,
139         .steal = page_cache_pipe_buf_steal,
140         .get = generic_pipe_buf_get,
141 };
142
143 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
144                                     struct pipe_buffer *buf)
145 {
146         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
147                 return 1;
148
149         buf->flags |= PIPE_BUF_FLAG_LRU;
150         return generic_pipe_buf_steal(pipe, buf);
151 }
152
153 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
154         .can_merge = 0,
155         .map = generic_pipe_buf_map,
156         .unmap = generic_pipe_buf_unmap,
157         .confirm = generic_pipe_buf_confirm,
158         .release = page_cache_pipe_buf_release,
159         .steal = user_page_pipe_buf_steal,
160         .get = generic_pipe_buf_get,
161 };
162
163 /**
164  * splice_to_pipe - fill passed data into a pipe
165  * @pipe:       pipe to fill
166  * @spd:        data to fill
167  *
168  * Description:
169  *    @spd contains a map of pages and len/offset tuples, along with
170  *    the struct pipe_buf_operations associated with these pages. This
171  *    function will link that data to the pipe.
172  *
173  */
174 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
175                        struct splice_pipe_desc *spd)
176 {
177         unsigned int spd_pages = spd->nr_pages;
178         int ret, do_wakeup, page_nr;
179
180         ret = 0;
181         do_wakeup = 0;
182         page_nr = 0;
183
184         if (pipe->inode)
185                 mutex_lock(&pipe->inode->i_mutex);
186
187         for (;;) {
188                 if (!pipe->readers) {
189                         send_sig(SIGPIPE, current, 0);
190                         if (!ret)
191                                 ret = -EPIPE;
192                         break;
193                 }
194
195                 if (pipe->nrbufs < PIPE_BUFFERS) {
196                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
197                         struct pipe_buffer *buf = pipe->bufs + newbuf;
198
199                         buf->page = spd->pages[page_nr];
200                         buf->offset = spd->partial[page_nr].offset;
201                         buf->len = spd->partial[page_nr].len;
202                         buf->private = spd->partial[page_nr].private;
203                         buf->ops = spd->ops;
204                         if (spd->flags & SPLICE_F_GIFT)
205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
206
207                         pipe->nrbufs++;
208                         page_nr++;
209                         ret += buf->len;
210
211                         if (pipe->inode)
212                                 do_wakeup = 1;
213
214                         if (!--spd->nr_pages)
215                                 break;
216                         if (pipe->nrbufs < PIPE_BUFFERS)
217                                 continue;
218
219                         break;
220                 }
221
222                 if (spd->flags & SPLICE_F_NONBLOCK) {
223                         if (!ret)
224                                 ret = -EAGAIN;
225                         break;
226                 }
227
228                 if (signal_pending(current)) {
229                         if (!ret)
230                                 ret = -ERESTARTSYS;
231                         break;
232                 }
233
234                 if (do_wakeup) {
235                         smp_mb();
236                         if (waitqueue_active(&pipe->wait))
237                                 wake_up_interruptible_sync(&pipe->wait);
238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
239                         do_wakeup = 0;
240                 }
241
242                 pipe->waiting_writers++;
243                 pipe_wait(pipe);
244                 pipe->waiting_writers--;
245         }
246
247         if (pipe->inode) {
248                 mutex_unlock(&pipe->inode->i_mutex);
249
250                 if (do_wakeup) {
251                         smp_mb();
252                         if (waitqueue_active(&pipe->wait))
253                                 wake_up_interruptible(&pipe->wait);
254                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
255                 }
256         }
257
258         while (page_nr < spd_pages)
259                 spd->spd_release(spd, page_nr++);
260
261         return ret;
262 }
263
264 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
265 {
266         page_cache_release(spd->pages[i]);
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, req_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         int error, page_nr;
282         struct splice_pipe_desc spd = {
283                 .pages = pages,
284                 .partial = partial,
285                 .flags = flags,
286                 .ops = &page_cache_pipe_buf_ops,
287                 .spd_release = spd_release_page,
288         };
289
290         index = *ppos >> PAGE_CACHE_SHIFT;
291         loff = *ppos & ~PAGE_CACHE_MASK;
292         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
293         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
294
295         /*
296          * Lookup the (hopefully) full range of pages we need.
297          */
298         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
299         index += spd.nr_pages;
300
301         /*
302          * If find_get_pages_contig() returned fewer pages than we needed,
303          * readahead/allocate the rest and fill in the holes.
304          */
305         if (spd.nr_pages < nr_pages)
306                 page_cache_sync_readahead(mapping, &in->f_ra, in,
307                                 index, req_pages - spd.nr_pages);
308
309         error = 0;
310         while (spd.nr_pages < nr_pages) {
311                 /*
312                  * Page could be there, find_get_pages_contig() breaks on
313                  * the first hole.
314                  */
315                 page = find_get_page(mapping, index);
316                 if (!page) {
317                         /*
318                          * page didn't exist, allocate one.
319                          */
320                         page = page_cache_alloc_cold(mapping);
321                         if (!page)
322                                 break;
323
324                         error = add_to_page_cache_lru(page, mapping, index,
325                                                 mapping_gfp_mask(mapping));
326                         if (unlikely(error)) {
327                                 page_cache_release(page);
328                                 if (error == -EEXIST)
329                                         continue;
330                                 break;
331                         }
332                         /*
333                          * add_to_page_cache() locks the page, unlock it
334                          * to avoid convoluting the logic below even more.
335                          */
336                         unlock_page(page);
337                 }
338
339                 pages[spd.nr_pages++] = page;
340                 index++;
341         }
342
343         /*
344          * Now loop over the map and see if we need to start IO on any
345          * pages, fill in the partial map, etc.
346          */
347         index = *ppos >> PAGE_CACHE_SHIFT;
348         nr_pages = spd.nr_pages;
349         spd.nr_pages = 0;
350         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
351                 unsigned int this_len;
352
353                 if (!len)
354                         break;
355
356                 /*
357                  * this_len is the max we'll use from this page
358                  */
359                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
360                 page = pages[page_nr];
361
362                 if (PageReadahead(page))
363                         page_cache_async_readahead(mapping, &in->f_ra, in,
364                                         page, index, req_pages - page_nr);
365
366                 /*
367                  * If the page isn't uptodate, we may need to start io on it
368                  */
369                 if (!PageUptodate(page)) {
370                         /*
371                          * If in nonblock mode then dont block on waiting
372                          * for an in-flight io page
373                          */
374                         if (flags & SPLICE_F_NONBLOCK) {
375                                 if (!trylock_page(page)) {
376                                         error = -EAGAIN;
377                                         break;
378                                 }
379                         } else
380                                 lock_page(page);
381
382                         /*
383                          * Page was truncated, or invalidated by the
384                          * filesystem.  Redo the find/create, but this time the
385                          * page is kept locked, so there's no chance of another
386                          * race with truncate/invalidate.
387                          */
388                         if (!page->mapping) {
389                                 unlock_page(page);
390                                 page = find_or_create_page(mapping, index,
391                                                 mapping_gfp_mask(mapping));
392
393                                 if (!page) {
394                                         error = -ENOMEM;
395                                         break;
396                                 }
397                                 page_cache_release(pages[page_nr]);
398                                 pages[page_nr] = page;
399                         }
400                         /*
401                          * page was already under io and is now done, great
402                          */
403                         if (PageUptodate(page)) {
404                                 unlock_page(page);
405                                 goto fill_it;
406                         }
407
408                         /*
409                          * need to read in the page
410                          */
411                         error = mapping->a_ops->readpage(in, page);
412                         if (unlikely(error)) {
413                                 /*
414                                  * We really should re-lookup the page here,
415                                  * but it complicates things a lot. Instead
416                                  * lets just do what we already stored, and
417                                  * we'll get it the next time we are called.
418                                  */
419                                 if (error == AOP_TRUNCATED_PAGE)
420                                         error = 0;
421
422                                 break;
423                         }
424                 }
425 fill_it:
426                 /*
427                  * i_size must be checked after PageUptodate.
428                  */
429                 isize = i_size_read(mapping->host);
430                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
431                 if (unlikely(!isize || index > end_index))
432                         break;
433
434                 /*
435                  * if this is the last page, see if we need to shrink
436                  * the length and stop
437                  */
438                 if (end_index == index) {
439                         unsigned int plen;
440
441                         /*
442                          * max good bytes in this page
443                          */
444                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
445                         if (plen <= loff)
446                                 break;
447
448                         /*
449                          * force quit after adding this page
450                          */
451                         this_len = min(this_len, plen - loff);
452                         len = this_len;
453                 }
454
455                 partial[page_nr].offset = loff;
456                 partial[page_nr].len = this_len;
457                 len -= this_len;
458                 loff = 0;
459                 spd.nr_pages++;
460                 index++;
461         }
462
463         /*
464          * Release any pages at the end, if we quit early. 'page_nr' is how far
465          * we got, 'nr_pages' is how many pages are in the map.
466          */
467         while (page_nr < nr_pages)
468                 page_cache_release(pages[page_nr++]);
469         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
470
471         if (spd.nr_pages)
472                 return splice_to_pipe(pipe, &spd);
473
474         return error;
475 }
476
477 /**
478  * generic_file_splice_read - splice data from file to a pipe
479  * @in:         file to splice from
480  * @ppos:       position in @in
481  * @pipe:       pipe to splice to
482  * @len:        number of bytes to splice
483  * @flags:      splice modifier flags
484  *
485  * Description:
486  *    Will read pages from given file and fill them into a pipe. Can be
487  *    used as long as the address_space operations for the source implements
488  *    a readpage() hook.
489  *
490  */
491 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
492                                  struct pipe_inode_info *pipe, size_t len,
493                                  unsigned int flags)
494 {
495         loff_t isize, left;
496         int ret;
497
498         isize = i_size_read(in->f_mapping->host);
499         if (unlikely(*ppos >= isize))
500                 return 0;
501
502         left = isize - *ppos;
503         if (unlikely(left < len))
504                 len = left;
505
506         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
507         if (ret > 0)
508                 *ppos += ret;
509
510         return ret;
511 }
512
513 EXPORT_SYMBOL(generic_file_splice_read);
514
515 /*
516  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
517  * using sendpage(). Return the number of bytes sent.
518  */
519 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
520                             struct pipe_buffer *buf, struct splice_desc *sd)
521 {
522         struct file *file = sd->u.file;
523         loff_t pos = sd->pos;
524         int ret, more;
525
526         ret = buf->ops->confirm(pipe, buf);
527         if (!ret) {
528                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
529
530                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
531                                            sd->len, &pos, more);
532         }
533
534         return ret;
535 }
536
537 /*
538  * This is a little more tricky than the file -> pipe splicing. There are
539  * basically three cases:
540  *
541  *      - Destination page already exists in the address space and there
542  *        are users of it. For that case we have no other option that
543  *        copying the data. Tough luck.
544  *      - Destination page already exists in the address space, but there
545  *        are no users of it. Make sure it's uptodate, then drop it. Fall
546  *        through to last case.
547  *      - Destination page does not exist, we can add the pipe page to
548  *        the page cache and avoid the copy.
549  *
550  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
551  * sd->flags), we attempt to migrate pages from the pipe to the output
552  * file address space page cache. This is possible if no one else has
553  * the pipe page referenced outside of the pipe and page cache. If
554  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
555  * a new page in the output file page cache and fill/dirty that.
556  */
557 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
558                         struct splice_desc *sd)
559 {
560         struct file *file = sd->u.file;
561         struct address_space *mapping = file->f_mapping;
562         unsigned int offset, this_len;
563         struct page *page;
564         void *fsdata;
565         int ret;
566
567         /*
568          * make sure the data in this buffer is uptodate
569          */
570         ret = buf->ops->confirm(pipe, buf);
571         if (unlikely(ret))
572                 return ret;
573
574         offset = sd->pos & ~PAGE_CACHE_MASK;
575
576         this_len = sd->len;
577         if (this_len + offset > PAGE_CACHE_SIZE)
578                 this_len = PAGE_CACHE_SIZE - offset;
579
580         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
581                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
582         if (unlikely(ret))
583                 goto out;
584
585         if (buf->page != page) {
586                 /*
587                  * Careful, ->map() uses KM_USER0!
588                  */
589                 char *src = buf->ops->map(pipe, buf, 1);
590                 char *dst = kmap_atomic(page, KM_USER1);
591
592                 memcpy(dst + offset, src + buf->offset, this_len);
593                 flush_dcache_page(page);
594                 kunmap_atomic(dst, KM_USER1);
595                 buf->ops->unmap(pipe, buf, src);
596         }
597         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
598                                 page, fsdata);
599 out:
600         return ret;
601 }
602
603 /**
604  * __splice_from_pipe - splice data from a pipe to given actor
605  * @pipe:       pipe to splice from
606  * @sd:         information to @actor
607  * @actor:      handler that splices the data
608  *
609  * Description:
610  *    This function does little more than loop over the pipe and call
611  *    @actor to do the actual moving of a single struct pipe_buffer to
612  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
613  *    pipe_to_user.
614  *
615  */
616 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
617                            splice_actor *actor)
618 {
619         int ret, do_wakeup, err;
620
621         ret = 0;
622         do_wakeup = 0;
623
624         for (;;) {
625                 if (pipe->nrbufs) {
626                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
627                         const struct pipe_buf_operations *ops = buf->ops;
628
629                         sd->len = buf->len;
630                         if (sd->len > sd->total_len)
631                                 sd->len = sd->total_len;
632
633                         err = actor(pipe, buf, sd);
634                         if (err <= 0) {
635                                 if (!ret && err != -ENODATA)
636                                         ret = err;
637
638                                 break;
639                         }
640
641                         ret += err;
642                         buf->offset += err;
643                         buf->len -= err;
644
645                         sd->len -= err;
646                         sd->pos += err;
647                         sd->total_len -= err;
648                         if (sd->len)
649                                 continue;
650
651                         if (!buf->len) {
652                                 buf->ops = NULL;
653                                 ops->release(pipe, buf);
654                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
655                                 pipe->nrbufs--;
656                                 if (pipe->inode)
657                                         do_wakeup = 1;
658                         }
659
660                         if (!sd->total_len)
661                                 break;
662                 }
663
664                 if (pipe->nrbufs)
665                         continue;
666                 if (!pipe->writers)
667                         break;
668                 if (!pipe->waiting_writers) {
669                         if (ret)
670                                 break;
671                 }
672
673                 if (sd->flags & SPLICE_F_NONBLOCK) {
674                         if (!ret)
675                                 ret = -EAGAIN;
676                         break;
677                 }
678
679                 if (signal_pending(current)) {
680                         if (!ret)
681                                 ret = -ERESTARTSYS;
682                         break;
683                 }
684
685                 if (do_wakeup) {
686                         smp_mb();
687                         if (waitqueue_active(&pipe->wait))
688                                 wake_up_interruptible_sync(&pipe->wait);
689                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
690                         do_wakeup = 0;
691                 }
692
693                 pipe_wait(pipe);
694         }
695
696         if (do_wakeup) {
697                 smp_mb();
698                 if (waitqueue_active(&pipe->wait))
699                         wake_up_interruptible(&pipe->wait);
700                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
701         }
702
703         return ret;
704 }
705 EXPORT_SYMBOL(__splice_from_pipe);
706
707 /**
708  * splice_from_pipe - splice data from a pipe to a file
709  * @pipe:       pipe to splice from
710  * @out:        file to splice to
711  * @ppos:       position in @out
712  * @len:        how many bytes to splice
713  * @flags:      splice modifier flags
714  * @actor:      handler that splices the data
715  *
716  * Description:
717  *    See __splice_from_pipe. This function locks the input and output inodes,
718  *    otherwise it's identical to __splice_from_pipe().
719  *
720  */
721 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
722                          loff_t *ppos, size_t len, unsigned int flags,
723                          splice_actor *actor)
724 {
725         ssize_t ret;
726         struct inode *inode = out->f_mapping->host;
727         struct splice_desc sd = {
728                 .total_len = len,
729                 .flags = flags,
730                 .pos = *ppos,
731                 .u.file = out,
732         };
733
734         /*
735          * The actor worker might be calling ->write_begin and
736          * ->write_end. Most of the time, these expect i_mutex to
737          * be held. Since this may result in an ABBA deadlock with
738          * pipe->inode, we have to order lock acquiry here.
739          */
740         inode_double_lock(inode, pipe->inode);
741         ret = __splice_from_pipe(pipe, &sd, actor);
742         inode_double_unlock(inode, pipe->inode);
743
744         return ret;
745 }
746
747 /**
748  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
749  * @pipe:       pipe info
750  * @out:        file to write to
751  * @ppos:       position in @out
752  * @len:        number of bytes to splice
753  * @flags:      splice modifier flags
754  *
755  * Description:
756  *    Will either move or copy pages (determined by @flags options) from
757  *    the given pipe inode to the given file. The caller is responsible
758  *    for acquiring i_mutex on both inodes.
759  *
760  */
761 ssize_t
762 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
763                                  loff_t *ppos, size_t len, unsigned int flags)
764 {
765         struct address_space *mapping = out->f_mapping;
766         struct inode *inode = mapping->host;
767         struct splice_desc sd = {
768                 .total_len = len,
769                 .flags = flags,
770                 .pos = *ppos,
771                 .u.file = out,
772         };
773         ssize_t ret;
774         int err;
775
776         err = file_remove_suid(out);
777         if (unlikely(err))
778                 return err;
779
780         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
781         if (ret > 0) {
782                 unsigned long nr_pages;
783
784                 *ppos += ret;
785                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
786
787                 /*
788                  * If file or inode is SYNC and we actually wrote some data,
789                  * sync it.
790                  */
791                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
792                         err = generic_osync_inode(inode, mapping,
793                                                   OSYNC_METADATA|OSYNC_DATA);
794
795                         if (err)
796                                 ret = err;
797                 }
798                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
799         }
800
801         return ret;
802 }
803
804 EXPORT_SYMBOL(generic_file_splice_write_nolock);
805
806 /**
807  * generic_file_splice_write - splice data from a pipe to a file
808  * @pipe:       pipe info
809  * @out:        file to write to
810  * @ppos:       position in @out
811  * @len:        number of bytes to splice
812  * @flags:      splice modifier flags
813  *
814  * Description:
815  *    Will either move or copy pages (determined by @flags options) from
816  *    the given pipe inode to the given file.
817  *
818  */
819 ssize_t
820 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
821                           loff_t *ppos, size_t len, unsigned int flags)
822 {
823         struct address_space *mapping = out->f_mapping;
824         struct inode *inode = mapping->host;
825         struct splice_desc sd = {
826                 .total_len = len,
827                 .flags = flags,
828                 .pos = *ppos,
829                 .u.file = out,
830         };
831         ssize_t ret;
832
833         inode_double_lock(inode, pipe->inode);
834         ret = file_remove_suid(out);
835         if (likely(!ret))
836                 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
837         inode_double_unlock(inode, pipe->inode);
838         if (ret > 0) {
839                 unsigned long nr_pages;
840
841                 *ppos += ret;
842                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
843
844                 /*
845                  * If file or inode is SYNC and we actually wrote some data,
846                  * sync it.
847                  */
848                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
849                         int err;
850
851                         mutex_lock(&inode->i_mutex);
852                         err = generic_osync_inode(inode, mapping,
853                                                   OSYNC_METADATA|OSYNC_DATA);
854                         mutex_unlock(&inode->i_mutex);
855
856                         if (err)
857                                 ret = err;
858                 }
859                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
860         }
861
862         return ret;
863 }
864
865 EXPORT_SYMBOL(generic_file_splice_write);
866
867 /**
868  * generic_splice_sendpage - splice data from a pipe to a socket
869  * @pipe:       pipe to splice from
870  * @out:        socket to write to
871  * @ppos:       position in @out
872  * @len:        number of bytes to splice
873  * @flags:      splice modifier flags
874  *
875  * Description:
876  *    Will send @len bytes from the pipe to a network socket. No data copying
877  *    is involved.
878  *
879  */
880 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
881                                 loff_t *ppos, size_t len, unsigned int flags)
882 {
883         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
884 }
885
886 EXPORT_SYMBOL(generic_splice_sendpage);
887
888 /*
889  * Attempt to initiate a splice from pipe to file.
890  */
891 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
892                            loff_t *ppos, size_t len, unsigned int flags)
893 {
894         int ret;
895
896         if (unlikely(!out->f_op || !out->f_op->splice_write))
897                 return -EINVAL;
898
899         if (unlikely(!(out->f_mode & FMODE_WRITE)))
900                 return -EBADF;
901
902         if (unlikely(out->f_flags & O_APPEND))
903                 return -EINVAL;
904
905         ret = rw_verify_area(WRITE, out, ppos, len);
906         if (unlikely(ret < 0))
907                 return ret;
908
909         return out->f_op->splice_write(pipe, out, ppos, len, flags);
910 }
911
912 /*
913  * Attempt to initiate a splice from a file to a pipe.
914  */
915 static long do_splice_to(struct file *in, loff_t *ppos,
916                          struct pipe_inode_info *pipe, size_t len,
917                          unsigned int flags)
918 {
919         int ret;
920
921         if (unlikely(!in->f_op || !in->f_op->splice_read))
922                 return -EINVAL;
923
924         if (unlikely(!(in->f_mode & FMODE_READ)))
925                 return -EBADF;
926
927         ret = rw_verify_area(READ, in, ppos, len);
928         if (unlikely(ret < 0))
929                 return ret;
930
931         return in->f_op->splice_read(in, ppos, pipe, len, flags);
932 }
933
934 /**
935  * splice_direct_to_actor - splices data directly between two non-pipes
936  * @in:         file to splice from
937  * @sd:         actor information on where to splice to
938  * @actor:      handles the data splicing
939  *
940  * Description:
941  *    This is a special case helper to splice directly between two
942  *    points, without requiring an explicit pipe. Internally an allocated
943  *    pipe is cached in the process, and reused during the lifetime of
944  *    that process.
945  *
946  */
947 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
948                                splice_direct_actor *actor)
949 {
950         struct pipe_inode_info *pipe;
951         long ret, bytes;
952         umode_t i_mode;
953         size_t len;
954         int i, flags;
955
956         /*
957          * We require the input being a regular file, as we don't want to
958          * randomly drop data for eg socket -> socket splicing. Use the
959          * piped splicing for that!
960          */
961         i_mode = in->f_path.dentry->d_inode->i_mode;
962         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
963                 return -EINVAL;
964
965         /*
966          * neither in nor out is a pipe, setup an internal pipe attached to
967          * 'out' and transfer the wanted data from 'in' to 'out' through that
968          */
969         pipe = current->splice_pipe;
970         if (unlikely(!pipe)) {
971                 pipe = alloc_pipe_info(NULL);
972                 if (!pipe)
973                         return -ENOMEM;
974
975                 /*
976                  * We don't have an immediate reader, but we'll read the stuff
977                  * out of the pipe right after the splice_to_pipe(). So set
978                  * PIPE_READERS appropriately.
979                  */
980                 pipe->readers = 1;
981
982                 current->splice_pipe = pipe;
983         }
984
985         /*
986          * Do the splice.
987          */
988         ret = 0;
989         bytes = 0;
990         len = sd->total_len;
991         flags = sd->flags;
992
993         /*
994          * Don't block on output, we have to drain the direct pipe.
995          */
996         sd->flags &= ~SPLICE_F_NONBLOCK;
997
998         while (len) {
999                 size_t read_len;
1000                 loff_t pos = sd->pos, prev_pos = pos;
1001
1002                 ret = do_splice_to(in, &pos, pipe, len, flags);
1003                 if (unlikely(ret <= 0))
1004                         goto out_release;
1005
1006                 read_len = ret;
1007                 sd->total_len = read_len;
1008
1009                 /*
1010                  * NOTE: nonblocking mode only applies to the input. We
1011                  * must not do the output in nonblocking mode as then we
1012                  * could get stuck data in the internal pipe:
1013                  */
1014                 ret = actor(pipe, sd);
1015                 if (unlikely(ret <= 0)) {
1016                         sd->pos = prev_pos;
1017                         goto out_release;
1018                 }
1019
1020                 bytes += ret;
1021                 len -= ret;
1022                 sd->pos = pos;
1023
1024                 if (ret < read_len) {
1025                         sd->pos = prev_pos + ret;
1026                         goto out_release;
1027                 }
1028         }
1029
1030 done:
1031         pipe->nrbufs = pipe->curbuf = 0;
1032         file_accessed(in);
1033         return bytes;
1034
1035 out_release:
1036         /*
1037          * If we did an incomplete transfer we must release
1038          * the pipe buffers in question:
1039          */
1040         for (i = 0; i < PIPE_BUFFERS; i++) {
1041                 struct pipe_buffer *buf = pipe->bufs + i;
1042
1043                 if (buf->ops) {
1044                         buf->ops->release(pipe, buf);
1045                         buf->ops = NULL;
1046                 }
1047         }
1048
1049         if (!bytes)
1050                 bytes = ret;
1051
1052         goto done;
1053 }
1054 EXPORT_SYMBOL(splice_direct_to_actor);
1055
1056 static int direct_splice_actor(struct pipe_inode_info *pipe,
1057                                struct splice_desc *sd)
1058 {
1059         struct file *file = sd->u.file;
1060
1061         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1062 }
1063
1064 /**
1065  * do_splice_direct - splices data directly between two files
1066  * @in:         file to splice from
1067  * @ppos:       input file offset
1068  * @out:        file to splice to
1069  * @len:        number of bytes to splice
1070  * @flags:      splice modifier flags
1071  *
1072  * Description:
1073  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1074  *    doing it in the application would incur an extra system call
1075  *    (splice in + splice out, as compared to just sendfile()). So this helper
1076  *    can splice directly through a process-private pipe.
1077  *
1078  */
1079 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1080                       size_t len, unsigned int flags)
1081 {
1082         struct splice_desc sd = {
1083                 .len            = len,
1084                 .total_len      = len,
1085                 .flags          = flags,
1086                 .pos            = *ppos,
1087                 .u.file         = out,
1088         };
1089         long ret;
1090
1091         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1092         if (ret > 0)
1093                 *ppos = sd.pos;
1094
1095         return ret;
1096 }
1097
1098 /*
1099  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1100  * location, so checking ->i_pipe is not enough to verify that this is a
1101  * pipe.
1102  */
1103 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1104 {
1105         if (S_ISFIFO(inode->i_mode))
1106                 return inode->i_pipe;
1107
1108         return NULL;
1109 }
1110
1111 /*
1112  * Determine where to splice to/from.
1113  */
1114 static long do_splice(struct file *in, loff_t __user *off_in,
1115                       struct file *out, loff_t __user *off_out,
1116                       size_t len, unsigned int flags)
1117 {
1118         struct pipe_inode_info *pipe;
1119         loff_t offset, *off;
1120         long ret;
1121
1122         pipe = pipe_info(in->f_path.dentry->d_inode);
1123         if (pipe) {
1124                 if (off_in)
1125                         return -ESPIPE;
1126                 if (off_out) {
1127                         if (out->f_op->llseek == no_llseek)
1128                                 return -EINVAL;
1129                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1130                                 return -EFAULT;
1131                         off = &offset;
1132                 } else
1133                         off = &out->f_pos;
1134
1135                 ret = do_splice_from(pipe, out, off, len, flags);
1136
1137                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1138                         ret = -EFAULT;
1139
1140                 return ret;
1141         }
1142
1143         pipe = pipe_info(out->f_path.dentry->d_inode);
1144         if (pipe) {
1145                 if (off_out)
1146                         return -ESPIPE;
1147                 if (off_in) {
1148                         if (in->f_op->llseek == no_llseek)
1149                                 return -EINVAL;
1150                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1151                                 return -EFAULT;
1152                         off = &offset;
1153                 } else
1154                         off = &in->f_pos;
1155
1156                 ret = do_splice_to(in, off, pipe, len, flags);
1157
1158                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1159                         ret = -EFAULT;
1160
1161                 return ret;
1162         }
1163
1164         return -EINVAL;
1165 }
1166
1167 /*
1168  * Map an iov into an array of pages and offset/length tupples. With the
1169  * partial_page structure, we can map several non-contiguous ranges into
1170  * our ones pages[] map instead of splitting that operation into pieces.
1171  * Could easily be exported as a generic helper for other users, in which
1172  * case one would probably want to add a 'max_nr_pages' parameter as well.
1173  */
1174 static int get_iovec_page_array(const struct iovec __user *iov,
1175                                 unsigned int nr_vecs, struct page **pages,
1176                                 struct partial_page *partial, int aligned)
1177 {
1178         int buffers = 0, error = 0;
1179
1180         while (nr_vecs) {
1181                 unsigned long off, npages;
1182                 struct iovec entry;
1183                 void __user *base;
1184                 size_t len;
1185                 int i;
1186
1187                 error = -EFAULT;
1188                 if (copy_from_user(&entry, iov, sizeof(entry)))
1189                         break;
1190
1191                 base = entry.iov_base;
1192                 len = entry.iov_len;
1193
1194                 /*
1195                  * Sanity check this iovec. 0 read succeeds.
1196                  */
1197                 error = 0;
1198                 if (unlikely(!len))
1199                         break;
1200                 error = -EFAULT;
1201                 if (!access_ok(VERIFY_READ, base, len))
1202                         break;
1203
1204                 /*
1205                  * Get this base offset and number of pages, then map
1206                  * in the user pages.
1207                  */
1208                 off = (unsigned long) base & ~PAGE_MASK;
1209
1210                 /*
1211                  * If asked for alignment, the offset must be zero and the
1212                  * length a multiple of the PAGE_SIZE.
1213                  */
1214                 error = -EINVAL;
1215                 if (aligned && (off || len & ~PAGE_MASK))
1216                         break;
1217
1218                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1219                 if (npages > PIPE_BUFFERS - buffers)
1220                         npages = PIPE_BUFFERS - buffers;
1221
1222                 error = get_user_pages_fast((unsigned long)base, npages,
1223                                         0, &pages[buffers]);
1224
1225                 if (unlikely(error <= 0))
1226                         break;
1227
1228                 /*
1229                  * Fill this contiguous range into the partial page map.
1230                  */
1231                 for (i = 0; i < error; i++) {
1232                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1233
1234                         partial[buffers].offset = off;
1235                         partial[buffers].len = plen;
1236
1237                         off = 0;
1238                         len -= plen;
1239                         buffers++;
1240                 }
1241
1242                 /*
1243                  * We didn't complete this iov, stop here since it probably
1244                  * means we have to move some of this into a pipe to
1245                  * be able to continue.
1246                  */
1247                 if (len)
1248                         break;
1249
1250                 /*
1251                  * Don't continue if we mapped fewer pages than we asked for,
1252                  * or if we mapped the max number of pages that we have
1253                  * room for.
1254                  */
1255                 if (error < npages || buffers == PIPE_BUFFERS)
1256                         break;
1257
1258                 nr_vecs--;
1259                 iov++;
1260         }
1261
1262         if (buffers)
1263                 return buffers;
1264
1265         return error;
1266 }
1267
1268 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1269                         struct splice_desc *sd)
1270 {
1271         char *src;
1272         int ret;
1273
1274         ret = buf->ops->confirm(pipe, buf);
1275         if (unlikely(ret))
1276                 return ret;
1277
1278         /*
1279          * See if we can use the atomic maps, by prefaulting in the
1280          * pages and doing an atomic copy
1281          */
1282         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1283                 src = buf->ops->map(pipe, buf, 1);
1284                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1285                                                         sd->len);
1286                 buf->ops->unmap(pipe, buf, src);
1287                 if (!ret) {
1288                         ret = sd->len;
1289                         goto out;
1290                 }
1291         }
1292
1293         /*
1294          * No dice, use slow non-atomic map and copy
1295          */
1296         src = buf->ops->map(pipe, buf, 0);
1297
1298         ret = sd->len;
1299         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1300                 ret = -EFAULT;
1301
1302         buf->ops->unmap(pipe, buf, src);
1303 out:
1304         if (ret > 0)
1305                 sd->u.userptr += ret;
1306         return ret;
1307 }
1308
1309 /*
1310  * For lack of a better implementation, implement vmsplice() to userspace
1311  * as a simple copy of the pipes pages to the user iov.
1312  */
1313 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1314                              unsigned long nr_segs, unsigned int flags)
1315 {
1316         struct pipe_inode_info *pipe;
1317         struct splice_desc sd;
1318         ssize_t size;
1319         int error;
1320         long ret;
1321
1322         pipe = pipe_info(file->f_path.dentry->d_inode);
1323         if (!pipe)
1324                 return -EBADF;
1325
1326         if (pipe->inode)
1327                 mutex_lock(&pipe->inode->i_mutex);
1328
1329         error = ret = 0;
1330         while (nr_segs) {
1331                 void __user *base;
1332                 size_t len;
1333
1334                 /*
1335                  * Get user address base and length for this iovec.
1336                  */
1337                 error = get_user(base, &iov->iov_base);
1338                 if (unlikely(error))
1339                         break;
1340                 error = get_user(len, &iov->iov_len);
1341                 if (unlikely(error))
1342                         break;
1343
1344                 /*
1345                  * Sanity check this iovec. 0 read succeeds.
1346                  */
1347                 if (unlikely(!len))
1348                         break;
1349                 if (unlikely(!base)) {
1350                         error = -EFAULT;
1351                         break;
1352                 }
1353
1354                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1355                         error = -EFAULT;
1356                         break;
1357                 }
1358
1359                 sd.len = 0;
1360                 sd.total_len = len;
1361                 sd.flags = flags;
1362                 sd.u.userptr = base;
1363                 sd.pos = 0;
1364
1365                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1366                 if (size < 0) {
1367                         if (!ret)
1368                                 ret = size;
1369
1370                         break;
1371                 }
1372
1373                 ret += size;
1374
1375                 if (size < len)
1376                         break;
1377
1378                 nr_segs--;
1379                 iov++;
1380         }
1381
1382         if (pipe->inode)
1383                 mutex_unlock(&pipe->inode->i_mutex);
1384
1385         if (!ret)
1386                 ret = error;
1387
1388         return ret;
1389 }
1390
1391 /*
1392  * vmsplice splices a user address range into a pipe. It can be thought of
1393  * as splice-from-memory, where the regular splice is splice-from-file (or
1394  * to file). In both cases the output is a pipe, naturally.
1395  */
1396 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1397                              unsigned long nr_segs, unsigned int flags)
1398 {
1399         struct pipe_inode_info *pipe;
1400         struct page *pages[PIPE_BUFFERS];
1401         struct partial_page partial[PIPE_BUFFERS];
1402         struct splice_pipe_desc spd = {
1403                 .pages = pages,
1404                 .partial = partial,
1405                 .flags = flags,
1406                 .ops = &user_page_pipe_buf_ops,
1407                 .spd_release = spd_release_page,
1408         };
1409
1410         pipe = pipe_info(file->f_path.dentry->d_inode);
1411         if (!pipe)
1412                 return -EBADF;
1413
1414         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1415                                             flags & SPLICE_F_GIFT);
1416         if (spd.nr_pages <= 0)
1417                 return spd.nr_pages;
1418
1419         return splice_to_pipe(pipe, &spd);
1420 }
1421
1422 /*
1423  * Note that vmsplice only really supports true splicing _from_ user memory
1424  * to a pipe, not the other way around. Splicing from user memory is a simple
1425  * operation that can be supported without any funky alignment restrictions
1426  * or nasty vm tricks. We simply map in the user memory and fill them into
1427  * a pipe. The reverse isn't quite as easy, though. There are two possible
1428  * solutions for that:
1429  *
1430  *      - memcpy() the data internally, at which point we might as well just
1431  *        do a regular read() on the buffer anyway.
1432  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1433  *        has restriction limitations on both ends of the pipe).
1434  *
1435  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1436  *
1437  */
1438 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1439                              unsigned long nr_segs, unsigned int flags)
1440 {
1441         struct file *file;
1442         long error;
1443         int fput;
1444
1445         if (unlikely(nr_segs > UIO_MAXIOV))
1446                 return -EINVAL;
1447         else if (unlikely(!nr_segs))
1448                 return 0;
1449
1450         error = -EBADF;
1451         file = fget_light(fd, &fput);
1452         if (file) {
1453                 if (file->f_mode & FMODE_WRITE)
1454                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1455                 else if (file->f_mode & FMODE_READ)
1456                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1457
1458                 fput_light(file, fput);
1459         }
1460
1461         return error;
1462 }
1463
1464 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1465                            int fd_out, loff_t __user *off_out,
1466                            size_t len, unsigned int flags)
1467 {
1468         long error;
1469         struct file *in, *out;
1470         int fput_in, fput_out;
1471
1472         if (unlikely(!len))
1473                 return 0;
1474
1475         error = -EBADF;
1476         in = fget_light(fd_in, &fput_in);
1477         if (in) {
1478                 if (in->f_mode & FMODE_READ) {
1479                         out = fget_light(fd_out, &fput_out);
1480                         if (out) {
1481                                 if (out->f_mode & FMODE_WRITE)
1482                                         error = do_splice(in, off_in,
1483                                                           out, off_out,
1484                                                           len, flags);
1485                                 fput_light(out, fput_out);
1486                         }
1487                 }
1488
1489                 fput_light(in, fput_in);
1490         }
1491
1492         return error;
1493 }
1494
1495 /*
1496  * Make sure there's data to read. Wait for input if we can, otherwise
1497  * return an appropriate error.
1498  */
1499 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1500 {
1501         int ret;
1502
1503         /*
1504          * Check ->nrbufs without the inode lock first. This function
1505          * is speculative anyways, so missing one is ok.
1506          */
1507         if (pipe->nrbufs)
1508                 return 0;
1509
1510         ret = 0;
1511         mutex_lock(&pipe->inode->i_mutex);
1512
1513         while (!pipe->nrbufs) {
1514                 if (signal_pending(current)) {
1515                         ret = -ERESTARTSYS;
1516                         break;
1517                 }
1518                 if (!pipe->writers)
1519                         break;
1520                 if (!pipe->waiting_writers) {
1521                         if (flags & SPLICE_F_NONBLOCK) {
1522                                 ret = -EAGAIN;
1523                                 break;
1524                         }
1525                 }
1526                 pipe_wait(pipe);
1527         }
1528
1529         mutex_unlock(&pipe->inode->i_mutex);
1530         return ret;
1531 }
1532
1533 /*
1534  * Make sure there's writeable room. Wait for room if we can, otherwise
1535  * return an appropriate error.
1536  */
1537 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1538 {
1539         int ret;
1540
1541         /*
1542          * Check ->nrbufs without the inode lock first. This function
1543          * is speculative anyways, so missing one is ok.
1544          */
1545         if (pipe->nrbufs < PIPE_BUFFERS)
1546                 return 0;
1547
1548         ret = 0;
1549         mutex_lock(&pipe->inode->i_mutex);
1550
1551         while (pipe->nrbufs >= PIPE_BUFFERS) {
1552                 if (!pipe->readers) {
1553                         send_sig(SIGPIPE, current, 0);
1554                         ret = -EPIPE;
1555                         break;
1556                 }
1557                 if (flags & SPLICE_F_NONBLOCK) {
1558                         ret = -EAGAIN;
1559                         break;
1560                 }
1561                 if (signal_pending(current)) {
1562                         ret = -ERESTARTSYS;
1563                         break;
1564                 }
1565                 pipe->waiting_writers++;
1566                 pipe_wait(pipe);
1567                 pipe->waiting_writers--;
1568         }
1569
1570         mutex_unlock(&pipe->inode->i_mutex);
1571         return ret;
1572 }
1573
1574 /*
1575  * Link contents of ipipe to opipe.
1576  */
1577 static int link_pipe(struct pipe_inode_info *ipipe,
1578                      struct pipe_inode_info *opipe,
1579                      size_t len, unsigned int flags)
1580 {
1581         struct pipe_buffer *ibuf, *obuf;
1582         int ret = 0, i = 0, nbuf;
1583
1584         /*
1585          * Potential ABBA deadlock, work around it by ordering lock
1586          * grabbing by inode address. Otherwise two different processes
1587          * could deadlock (one doing tee from A -> B, the other from B -> A).
1588          */
1589         inode_double_lock(ipipe->inode, opipe->inode);
1590
1591         do {
1592                 if (!opipe->readers) {
1593                         send_sig(SIGPIPE, current, 0);
1594                         if (!ret)
1595                                 ret = -EPIPE;
1596                         break;
1597                 }
1598
1599                 /*
1600                  * If we have iterated all input buffers or ran out of
1601                  * output room, break.
1602                  */
1603                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1604                         break;
1605
1606                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1607                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1608
1609                 /*
1610                  * Get a reference to this pipe buffer,
1611                  * so we can copy the contents over.
1612                  */
1613                 ibuf->ops->get(ipipe, ibuf);
1614
1615                 obuf = opipe->bufs + nbuf;
1616                 *obuf = *ibuf;
1617
1618                 /*
1619                  * Don't inherit the gift flag, we need to
1620                  * prevent multiple steals of this page.
1621                  */
1622                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1623
1624                 if (obuf->len > len)
1625                         obuf->len = len;
1626
1627                 opipe->nrbufs++;
1628                 ret += obuf->len;
1629                 len -= obuf->len;
1630                 i++;
1631         } while (len);
1632
1633         /*
1634          * return EAGAIN if we have the potential of some data in the
1635          * future, otherwise just return 0
1636          */
1637         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1638                 ret = -EAGAIN;
1639
1640         inode_double_unlock(ipipe->inode, opipe->inode);
1641
1642         /*
1643          * If we put data in the output pipe, wakeup any potential readers.
1644          */
1645         if (ret > 0) {
1646                 smp_mb();
1647                 if (waitqueue_active(&opipe->wait))
1648                         wake_up_interruptible(&opipe->wait);
1649                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1650         }
1651
1652         return ret;
1653 }
1654
1655 /*
1656  * This is a tee(1) implementation that works on pipes. It doesn't copy
1657  * any data, it simply references the 'in' pages on the 'out' pipe.
1658  * The 'flags' used are the SPLICE_F_* variants, currently the only
1659  * applicable one is SPLICE_F_NONBLOCK.
1660  */
1661 static long do_tee(struct file *in, struct file *out, size_t len,
1662                    unsigned int flags)
1663 {
1664         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1665         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1666         int ret = -EINVAL;
1667
1668         /*
1669          * Duplicate the contents of ipipe to opipe without actually
1670          * copying the data.
1671          */
1672         if (ipipe && opipe && ipipe != opipe) {
1673                 /*
1674                  * Keep going, unless we encounter an error. The ipipe/opipe
1675                  * ordering doesn't really matter.
1676                  */
1677                 ret = link_ipipe_prep(ipipe, flags);
1678                 if (!ret) {
1679                         ret = link_opipe_prep(opipe, flags);
1680                         if (!ret)
1681                                 ret = link_pipe(ipipe, opipe, len, flags);
1682                 }
1683         }
1684
1685         return ret;
1686 }
1687
1688 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1689 {
1690         struct file *in;
1691         int error, fput_in;
1692
1693         if (unlikely(!len))
1694                 return 0;
1695
1696         error = -EBADF;
1697         in = fget_light(fdin, &fput_in);
1698         if (in) {
1699                 if (in->f_mode & FMODE_READ) {
1700                         int fput_out;
1701                         struct file *out = fget_light(fdout, &fput_out);
1702
1703                         if (out) {
1704                                 if (out->f_mode & FMODE_WRITE)
1705                                         error = do_tee(in, out, len, flags);
1706                                 fput_light(out, fput_out);
1707                         }
1708                 }
1709                 fput_light(in, fput_in);
1710         }
1711
1712         return error;
1713 }