splice: fix generic_file_splice_read() race with page invalidation
[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/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32
33 /*
34  * Attempt to steal a page from a pipe buffer. This should perhaps go into
35  * a vm helper function, it's already simplified quite a bit by the
36  * addition of remove_mapping(). If success is returned, the caller may
37  * attempt to reuse this page for another destination.
38  */
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
40                                      struct pipe_buffer *buf)
41 {
42         struct page *page = buf->page;
43         struct address_space *mapping;
44
45         lock_page(page);
46
47         mapping = page_mapping(page);
48         if (mapping) {
49                 WARN_ON(!PageUptodate(page));
50
51                 /*
52                  * At least for ext2 with nobh option, we need to wait on
53                  * writeback completing on this page, since we'll remove it
54                  * from the pagecache.  Otherwise truncate wont wait on the
55                  * page, allowing the disk blocks to be reused by someone else
56                  * before we actually wrote our data to them. fs corruption
57                  * ensues.
58                  */
59                 wait_on_page_writeback(page);
60
61                 if (PagePrivate(page) && !try_to_release_page(page, GFP_KERNEL))
62                         goto out_unlock;
63
64                 /*
65                  * If we succeeded in removing the mapping, set LRU flag
66                  * and return good.
67                  */
68                 if (remove_mapping(mapping, page)) {
69                         buf->flags |= PIPE_BUF_FLAG_LRU;
70                         return 0;
71                 }
72         }
73
74         /*
75          * Raced with truncate or failed to remove page from current
76          * address space, unlock and return failure.
77          */
78 out_unlock:
79         unlock_page(page);
80         return 1;
81 }
82
83 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
84                                         struct pipe_buffer *buf)
85 {
86         page_cache_release(buf->page);
87         buf->flags &= ~PIPE_BUF_FLAG_LRU;
88 }
89
90 /*
91  * Check whether the contents of buf is OK to access. Since the content
92  * is a page cache page, IO may be in flight.
93  */
94 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
95                                        struct pipe_buffer *buf)
96 {
97         struct page *page = buf->page;
98         int err;
99
100         if (!PageUptodate(page)) {
101                 lock_page(page);
102
103                 /*
104                  * Page got truncated/unhashed. This will cause a 0-byte
105                  * splice, if this is the first page.
106                  */
107                 if (!page->mapping) {
108                         err = -ENODATA;
109                         goto error;
110                 }
111
112                 /*
113                  * Uh oh, read-error from disk.
114                  */
115                 if (!PageUptodate(page)) {
116                         err = -EIO;
117                         goto error;
118                 }
119
120                 /*
121                  * Page is ok afterall, we are done.
122                  */
123                 unlock_page(page);
124         }
125
126         return 0;
127 error:
128         unlock_page(page);
129         return err;
130 }
131
132 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
133         .can_merge = 0,
134         .map = generic_pipe_buf_map,
135         .unmap = generic_pipe_buf_unmap,
136         .confirm = page_cache_pipe_buf_confirm,
137         .release = page_cache_pipe_buf_release,
138         .steal = page_cache_pipe_buf_steal,
139         .get = generic_pipe_buf_get,
140 };
141
142 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
143                                     struct pipe_buffer *buf)
144 {
145         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
146                 return 1;
147
148         buf->flags |= PIPE_BUF_FLAG_LRU;
149         return generic_pipe_buf_steal(pipe, buf);
150 }
151
152 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
153         .can_merge = 0,
154         .map = generic_pipe_buf_map,
155         .unmap = generic_pipe_buf_unmap,
156         .confirm = generic_pipe_buf_confirm,
157         .release = page_cache_pipe_buf_release,
158         .steal = user_page_pipe_buf_steal,
159         .get = generic_pipe_buf_get,
160 };
161
162 /**
163  * splice_to_pipe - fill passed data into a pipe
164  * @pipe:       pipe to fill
165  * @spd:        data to fill
166  *
167  * Description:
168  *    @spd contains a map of pages and len/offset tuples, along with
169  *    the struct pipe_buf_operations associated with these pages. This
170  *    function will link that data to the pipe.
171  *
172  */
173 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
174                        struct splice_pipe_desc *spd)
175 {
176         unsigned int spd_pages = spd->nr_pages;
177         int ret, do_wakeup, page_nr;
178
179         ret = 0;
180         do_wakeup = 0;
181         page_nr = 0;
182
183         if (pipe->inode)
184                 mutex_lock(&pipe->inode->i_mutex);
185
186         for (;;) {
187                 if (!pipe->readers) {
188                         send_sig(SIGPIPE, current, 0);
189                         if (!ret)
190                                 ret = -EPIPE;
191                         break;
192                 }
193
194                 if (pipe->nrbufs < PIPE_BUFFERS) {
195                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
196                         struct pipe_buffer *buf = pipe->bufs + newbuf;
197
198                         buf->page = spd->pages[page_nr];
199                         buf->offset = spd->partial[page_nr].offset;
200                         buf->len = spd->partial[page_nr].len;
201                         buf->private = spd->partial[page_nr].private;
202                         buf->ops = spd->ops;
203                         if (spd->flags & SPLICE_F_GIFT)
204                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
205
206                         pipe->nrbufs++;
207                         page_nr++;
208                         ret += buf->len;
209
210                         if (pipe->inode)
211                                 do_wakeup = 1;
212
213                         if (!--spd->nr_pages)
214                                 break;
215                         if (pipe->nrbufs < PIPE_BUFFERS)
216                                 continue;
217
218                         break;
219                 }
220
221                 if (spd->flags & SPLICE_F_NONBLOCK) {
222                         if (!ret)
223                                 ret = -EAGAIN;
224                         break;
225                 }
226
227                 if (signal_pending(current)) {
228                         if (!ret)
229                                 ret = -ERESTARTSYS;
230                         break;
231                 }
232
233                 if (do_wakeup) {
234                         smp_mb();
235                         if (waitqueue_active(&pipe->wait))
236                                 wake_up_interruptible_sync(&pipe->wait);
237                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
238                         do_wakeup = 0;
239                 }
240
241                 pipe->waiting_writers++;
242                 pipe_wait(pipe);
243                 pipe->waiting_writers--;
244         }
245
246         if (pipe->inode) {
247                 mutex_unlock(&pipe->inode->i_mutex);
248
249                 if (do_wakeup) {
250                         smp_mb();
251                         if (waitqueue_active(&pipe->wait))
252                                 wake_up_interruptible(&pipe->wait);
253                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
254                 }
255         }
256
257         while (page_nr < spd_pages)
258                 spd->spd_release(spd, page_nr++);
259
260         return ret;
261 }
262
263 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
264 {
265         page_cache_release(spd->pages[i]);
266 }
267
268 static int
269 __generic_file_splice_read(struct file *in, loff_t *ppos,
270                            struct pipe_inode_info *pipe, size_t len,
271                            unsigned int flags)
272 {
273         struct address_space *mapping = in->f_mapping;
274         unsigned int loff, nr_pages, req_pages;
275         struct page *pages[PIPE_BUFFERS];
276         struct partial_page partial[PIPE_BUFFERS];
277         struct page *page;
278         pgoff_t index, end_index;
279         loff_t isize;
280         int error, page_nr;
281         struct splice_pipe_desc spd = {
282                 .pages = pages,
283                 .partial = partial,
284                 .flags = flags,
285                 .ops = &page_cache_pipe_buf_ops,
286                 .spd_release = spd_release_page,
287         };
288
289         index = *ppos >> PAGE_CACHE_SHIFT;
290         loff = *ppos & ~PAGE_CACHE_MASK;
291         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
292         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
293
294         /*
295          * Lookup the (hopefully) full range of pages we need.
296          */
297         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
298         index += spd.nr_pages;
299
300         /*
301          * If find_get_pages_contig() returned fewer pages than we needed,
302          * readahead/allocate the rest and fill in the holes.
303          */
304         if (spd.nr_pages < nr_pages)
305                 page_cache_sync_readahead(mapping, &in->f_ra, in,
306                                 index, req_pages - spd.nr_pages);
307
308         error = 0;
309         while (spd.nr_pages < nr_pages) {
310                 /*
311                  * Page could be there, find_get_pages_contig() breaks on
312                  * the first hole.
313                  */
314                 page = find_get_page(mapping, index);
315                 if (!page) {
316                         /*
317                          * page didn't exist, allocate one.
318                          */
319                         page = page_cache_alloc_cold(mapping);
320                         if (!page)
321                                 break;
322
323                         error = add_to_page_cache_lru(page, mapping, index,
324                                                 mapping_gfp_mask(mapping));
325                         if (unlikely(error)) {
326                                 page_cache_release(page);
327                                 if (error == -EEXIST)
328                                         continue;
329                                 break;
330                         }
331                         /*
332                          * add_to_page_cache() locks the page, unlock it
333                          * to avoid convoluting the logic below even more.
334                          */
335                         unlock_page(page);
336                 }
337
338                 pages[spd.nr_pages++] = page;
339                 index++;
340         }
341
342         /*
343          * Now loop over the map and see if we need to start IO on any
344          * pages, fill in the partial map, etc.
345          */
346         index = *ppos >> PAGE_CACHE_SHIFT;
347         nr_pages = spd.nr_pages;
348         spd.nr_pages = 0;
349         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
350                 unsigned int this_len;
351
352                 if (!len)
353                         break;
354
355                 /*
356                  * this_len is the max we'll use from this page
357                  */
358                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
359                 page = pages[page_nr];
360
361                 if (PageReadahead(page))
362                         page_cache_async_readahead(mapping, &in->f_ra, in,
363                                         page, index, req_pages - page_nr);
364
365                 /*
366                  * If the page isn't uptodate, we may need to start io on it
367                  */
368                 if (!PageUptodate(page)) {
369                         /*
370                          * If in nonblock mode then dont block on waiting
371                          * for an in-flight io page
372                          */
373                         if (flags & SPLICE_F_NONBLOCK) {
374                                 if (TestSetPageLocked(page)) {
375                                         error = -EAGAIN;
376                                         break;
377                                 }
378                         } else
379                                 lock_page(page);
380
381                         /*
382                          * Page was truncated, or invalidated by the
383                          * filesystem.  Redo the find/create, but this time the
384                          * page is kept locked, so there's no chance of another
385                          * race with truncate/invalidate.
386                          */
387                         if (!page->mapping) {
388                                 unlock_page(page);
389                                 page = find_or_create_page(mapping, index,
390                                                 mapping_gfp_mask(mapping));
391
392                                 if (!page) {
393                                         error = -ENOMEM;
394                                         break;
395                                 }
396                                 page_cache_release(pages[page_nr]);
397                                 pages[page_nr] = page;
398                         }
399                         /*
400                          * page was already under io and is now done, great
401                          */
402                         if (PageUptodate(page)) {
403                                 unlock_page(page);
404                                 goto fill_it;
405                         }
406
407                         /*
408                          * need to read in the page
409                          */
410                         error = mapping->a_ops->readpage(in, page);
411                         if (unlikely(error)) {
412                                 /*
413                                  * We really should re-lookup the page here,
414                                  * but it complicates things a lot. Instead
415                                  * lets just do what we already stored, and
416                                  * we'll get it the next time we are called.
417                                  */
418                                 if (error == AOP_TRUNCATED_PAGE)
419                                         error = 0;
420
421                                 break;
422                         }
423                 }
424 fill_it:
425                 /*
426                  * i_size must be checked after PageUptodate.
427                  */
428                 isize = i_size_read(mapping->host);
429                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
430                 if (unlikely(!isize || index > end_index))
431                         break;
432
433                 /*
434                  * if this is the last page, see if we need to shrink
435                  * the length and stop
436                  */
437                 if (end_index == index) {
438                         unsigned int plen;
439
440                         /*
441                          * max good bytes in this page
442                          */
443                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
444                         if (plen <= loff)
445                                 break;
446
447                         /*
448                          * force quit after adding this page
449                          */
450                         this_len = min(this_len, plen - loff);
451                         len = this_len;
452                 }
453
454                 partial[page_nr].offset = loff;
455                 partial[page_nr].len = this_len;
456                 len -= this_len;
457                 loff = 0;
458                 spd.nr_pages++;
459                 index++;
460         }
461
462         /*
463          * Release any pages at the end, if we quit early. 'page_nr' is how far
464          * we got, 'nr_pages' is how many pages are in the map.
465          */
466         while (page_nr < nr_pages)
467                 page_cache_release(pages[page_nr++]);
468         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
469
470         if (spd.nr_pages)
471                 return splice_to_pipe(pipe, &spd);
472
473         return error;
474 }
475
476 /**
477  * generic_file_splice_read - splice data from file to a pipe
478  * @in:         file to splice from
479  * @ppos:       position in @in
480  * @pipe:       pipe to splice to
481  * @len:        number of bytes to splice
482  * @flags:      splice modifier flags
483  *
484  * Description:
485  *    Will read pages from given file and fill them into a pipe. Can be
486  *    used as long as the address_space operations for the source implements
487  *    a readpage() hook.
488  *
489  */
490 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
491                                  struct pipe_inode_info *pipe, size_t len,
492                                  unsigned int flags)
493 {
494         loff_t isize, left;
495         int ret;
496
497         isize = i_size_read(in->f_mapping->host);
498         if (unlikely(*ppos >= isize))
499                 return 0;
500
501         left = isize - *ppos;
502         if (unlikely(left < len))
503                 len = left;
504
505         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
506         if (ret > 0)
507                 *ppos += ret;
508
509         return ret;
510 }
511
512 EXPORT_SYMBOL(generic_file_splice_read);
513
514 /*
515  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
516  * using sendpage(). Return the number of bytes sent.
517  */
518 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
519                             struct pipe_buffer *buf, struct splice_desc *sd)
520 {
521         struct file *file = sd->u.file;
522         loff_t pos = sd->pos;
523         int ret, more;
524
525         ret = buf->ops->confirm(pipe, buf);
526         if (!ret) {
527                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
528
529                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
530                                            sd->len, &pos, more);
531         }
532
533         return ret;
534 }
535
536 /*
537  * This is a little more tricky than the file -> pipe splicing. There are
538  * basically three cases:
539  *
540  *      - Destination page already exists in the address space and there
541  *        are users of it. For that case we have no other option that
542  *        copying the data. Tough luck.
543  *      - Destination page already exists in the address space, but there
544  *        are no users of it. Make sure it's uptodate, then drop it. Fall
545  *        through to last case.
546  *      - Destination page does not exist, we can add the pipe page to
547  *        the page cache and avoid the copy.
548  *
549  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
550  * sd->flags), we attempt to migrate pages from the pipe to the output
551  * file address space page cache. This is possible if no one else has
552  * the pipe page referenced outside of the pipe and page cache. If
553  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
554  * a new page in the output file page cache and fill/dirty that.
555  */
556 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
557                         struct splice_desc *sd)
558 {
559         struct file *file = sd->u.file;
560         struct address_space *mapping = file->f_mapping;
561         unsigned int offset, this_len;
562         struct page *page;
563         void *fsdata;
564         int ret;
565
566         /*
567          * make sure the data in this buffer is uptodate
568          */
569         ret = buf->ops->confirm(pipe, buf);
570         if (unlikely(ret))
571                 return ret;
572
573         offset = sd->pos & ~PAGE_CACHE_MASK;
574
575         this_len = sd->len;
576         if (this_len + offset > PAGE_CACHE_SIZE)
577                 this_len = PAGE_CACHE_SIZE - offset;
578
579         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
580                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
581         if (unlikely(ret))
582                 goto out;
583
584         if (buf->page != page) {
585                 /*
586                  * Careful, ->map() uses KM_USER0!
587                  */
588                 char *src = buf->ops->map(pipe, buf, 1);
589                 char *dst = kmap_atomic(page, KM_USER1);
590
591                 memcpy(dst + offset, src + buf->offset, this_len);
592                 flush_dcache_page(page);
593                 kunmap_atomic(dst, KM_USER1);
594                 buf->ops->unmap(pipe, buf, src);
595         }
596         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
597                                 page, fsdata);
598 out:
599         return ret;
600 }
601
602 /**
603  * __splice_from_pipe - splice data from a pipe to given actor
604  * @pipe:       pipe to splice from
605  * @sd:         information to @actor
606  * @actor:      handler that splices the data
607  *
608  * Description:
609  *    This function does little more than loop over the pipe and call
610  *    @actor to do the actual moving of a single struct pipe_buffer to
611  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
612  *    pipe_to_user.
613  *
614  */
615 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
616                            splice_actor *actor)
617 {
618         int ret, do_wakeup, err;
619
620         ret = 0;
621         do_wakeup = 0;
622
623         for (;;) {
624                 if (pipe->nrbufs) {
625                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
626                         const struct pipe_buf_operations *ops = buf->ops;
627
628                         sd->len = buf->len;
629                         if (sd->len > sd->total_len)
630                                 sd->len = sd->total_len;
631
632                         err = actor(pipe, buf, sd);
633                         if (err <= 0) {
634                                 if (!ret && err != -ENODATA)
635                                         ret = err;
636
637                                 break;
638                         }
639
640                         ret += err;
641                         buf->offset += err;
642                         buf->len -= err;
643
644                         sd->len -= err;
645                         sd->pos += err;
646                         sd->total_len -= err;
647                         if (sd->len)
648                                 continue;
649
650                         if (!buf->len) {
651                                 buf->ops = NULL;
652                                 ops->release(pipe, buf);
653                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
654                                 pipe->nrbufs--;
655                                 if (pipe->inode)
656                                         do_wakeup = 1;
657                         }
658
659                         if (!sd->total_len)
660                                 break;
661                 }
662
663                 if (pipe->nrbufs)
664                         continue;
665                 if (!pipe->writers)
666                         break;
667                 if (!pipe->waiting_writers) {
668                         if (ret)
669                                 break;
670                 }
671
672                 if (sd->flags & SPLICE_F_NONBLOCK) {
673                         if (!ret)
674                                 ret = -EAGAIN;
675                         break;
676                 }
677
678                 if (signal_pending(current)) {
679                         if (!ret)
680                                 ret = -ERESTARTSYS;
681                         break;
682                 }
683
684                 if (do_wakeup) {
685                         smp_mb();
686                         if (waitqueue_active(&pipe->wait))
687                                 wake_up_interruptible_sync(&pipe->wait);
688                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
689                         do_wakeup = 0;
690                 }
691
692                 pipe_wait(pipe);
693         }
694
695         if (do_wakeup) {
696                 smp_mb();
697                 if (waitqueue_active(&pipe->wait))
698                         wake_up_interruptible(&pipe->wait);
699                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
700         }
701
702         return ret;
703 }
704 EXPORT_SYMBOL(__splice_from_pipe);
705
706 /**
707  * splice_from_pipe - splice data from a pipe to a file
708  * @pipe:       pipe to splice from
709  * @out:        file to splice to
710  * @ppos:       position in @out
711  * @len:        how many bytes to splice
712  * @flags:      splice modifier flags
713  * @actor:      handler that splices the data
714  *
715  * Description:
716  *    See __splice_from_pipe. This function locks the input and output inodes,
717  *    otherwise it's identical to __splice_from_pipe().
718  *
719  */
720 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
721                          loff_t *ppos, size_t len, unsigned int flags,
722                          splice_actor *actor)
723 {
724         ssize_t ret;
725         struct inode *inode = out->f_mapping->host;
726         struct splice_desc sd = {
727                 .total_len = len,
728                 .flags = flags,
729                 .pos = *ppos,
730                 .u.file = out,
731         };
732
733         /*
734          * The actor worker might be calling ->prepare_write and
735          * ->commit_write. Most of the time, these expect i_mutex to
736          * be held. Since this may result in an ABBA deadlock with
737          * pipe->inode, we have to order lock acquiry here.
738          */
739         inode_double_lock(inode, pipe->inode);
740         ret = __splice_from_pipe(pipe, &sd, actor);
741         inode_double_unlock(inode, pipe->inode);
742
743         return ret;
744 }
745
746 /**
747  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
748  * @pipe:       pipe info
749  * @out:        file to write to
750  * @ppos:       position in @out
751  * @len:        number of bytes to splice
752  * @flags:      splice modifier flags
753  *
754  * Description:
755  *    Will either move or copy pages (determined by @flags options) from
756  *    the given pipe inode to the given file. The caller is responsible
757  *    for acquiring i_mutex on both inodes.
758  *
759  */
760 ssize_t
761 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
762                                  loff_t *ppos, size_t len, unsigned int flags)
763 {
764         struct address_space *mapping = out->f_mapping;
765         struct inode *inode = mapping->host;
766         struct splice_desc sd = {
767                 .total_len = len,
768                 .flags = flags,
769                 .pos = *ppos,
770                 .u.file = out,
771         };
772         ssize_t ret;
773         int err;
774
775         err = remove_suid(out->f_path.dentry);
776         if (unlikely(err))
777                 return err;
778
779         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
780         if (ret > 0) {
781                 unsigned long nr_pages;
782
783                 *ppos += ret;
784                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
785
786                 /*
787                  * If file or inode is SYNC and we actually wrote some data,
788                  * sync it.
789                  */
790                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
791                         err = generic_osync_inode(inode, mapping,
792                                                   OSYNC_METADATA|OSYNC_DATA);
793
794                         if (err)
795                                 ret = err;
796                 }
797                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
798         }
799
800         return ret;
801 }
802
803 EXPORT_SYMBOL(generic_file_splice_write_nolock);
804
805 /**
806  * generic_file_splice_write - splice data from a pipe to a file
807  * @pipe:       pipe info
808  * @out:        file to write to
809  * @ppos:       position in @out
810  * @len:        number of bytes to splice
811  * @flags:      splice modifier flags
812  *
813  * Description:
814  *    Will either move or copy pages (determined by @flags options) from
815  *    the given pipe inode to the given file.
816  *
817  */
818 ssize_t
819 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
820                           loff_t *ppos, size_t len, unsigned int flags)
821 {
822         struct address_space *mapping = out->f_mapping;
823         struct inode *inode = mapping->host;
824         struct splice_desc sd = {
825                 .total_len = len,
826                 .flags = flags,
827                 .pos = *ppos,
828                 .u.file = out,
829         };
830         ssize_t ret;
831
832         inode_double_lock(inode, pipe->inode);
833         ret = remove_suid(out->f_path.dentry);
834         if (likely(!ret))
835                 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
836         inode_double_unlock(inode, pipe->inode);
837         if (ret > 0) {
838                 unsigned long nr_pages;
839
840                 *ppos += ret;
841                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
842
843                 /*
844                  * If file or inode is SYNC and we actually wrote some data,
845                  * sync it.
846                  */
847                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
848                         int err;
849
850                         mutex_lock(&inode->i_mutex);
851                         err = generic_osync_inode(inode, mapping,
852                                                   OSYNC_METADATA|OSYNC_DATA);
853                         mutex_unlock(&inode->i_mutex);
854
855                         if (err)
856                                 ret = err;
857                 }
858                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
859         }
860
861         return ret;
862 }
863
864 EXPORT_SYMBOL(generic_file_splice_write);
865
866 /**
867  * generic_splice_sendpage - splice data from a pipe to a socket
868  * @pipe:       pipe to splice from
869  * @out:        socket to write to
870  * @ppos:       position in @out
871  * @len:        number of bytes to splice
872  * @flags:      splice modifier flags
873  *
874  * Description:
875  *    Will send @len bytes from the pipe to a network socket. No data copying
876  *    is involved.
877  *
878  */
879 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
880                                 loff_t *ppos, size_t len, unsigned int flags)
881 {
882         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
883 }
884
885 EXPORT_SYMBOL(generic_splice_sendpage);
886
887 /*
888  * Attempt to initiate a splice from pipe to file.
889  */
890 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
891                            loff_t *ppos, size_t len, unsigned int flags)
892 {
893         int ret;
894
895         if (unlikely(!out->f_op || !out->f_op->splice_write))
896                 return -EINVAL;
897
898         if (unlikely(!(out->f_mode & FMODE_WRITE)))
899                 return -EBADF;
900
901         ret = rw_verify_area(WRITE, out, ppos, len);
902         if (unlikely(ret < 0))
903                 return ret;
904
905         return out->f_op->splice_write(pipe, out, ppos, len, flags);
906 }
907
908 /*
909  * Attempt to initiate a splice from a file to a pipe.
910  */
911 static long do_splice_to(struct file *in, loff_t *ppos,
912                          struct pipe_inode_info *pipe, size_t len,
913                          unsigned int flags)
914 {
915         int ret;
916
917         if (unlikely(!in->f_op || !in->f_op->splice_read))
918                 return -EINVAL;
919
920         if (unlikely(!(in->f_mode & FMODE_READ)))
921                 return -EBADF;
922
923         ret = rw_verify_area(READ, in, ppos, len);
924         if (unlikely(ret < 0))
925                 return ret;
926
927         return in->f_op->splice_read(in, ppos, pipe, len, flags);
928 }
929
930 /**
931  * splice_direct_to_actor - splices data directly between two non-pipes
932  * @in:         file to splice from
933  * @sd:         actor information on where to splice to
934  * @actor:      handles the data splicing
935  *
936  * Description:
937  *    This is a special case helper to splice directly between two
938  *    points, without requiring an explicit pipe. Internally an allocated
939  *    pipe is cached in the process, and reused during the lifetime of
940  *    that process.
941  *
942  */
943 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
944                                splice_direct_actor *actor)
945 {
946         struct pipe_inode_info *pipe;
947         long ret, bytes;
948         umode_t i_mode;
949         size_t len;
950         int i, flags;
951
952         /*
953          * We require the input being a regular file, as we don't want to
954          * randomly drop data for eg socket -> socket splicing. Use the
955          * piped splicing for that!
956          */
957         i_mode = in->f_path.dentry->d_inode->i_mode;
958         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
959                 return -EINVAL;
960
961         /*
962          * neither in nor out is a pipe, setup an internal pipe attached to
963          * 'out' and transfer the wanted data from 'in' to 'out' through that
964          */
965         pipe = current->splice_pipe;
966         if (unlikely(!pipe)) {
967                 pipe = alloc_pipe_info(NULL);
968                 if (!pipe)
969                         return -ENOMEM;
970
971                 /*
972                  * We don't have an immediate reader, but we'll read the stuff
973                  * out of the pipe right after the splice_to_pipe(). So set
974                  * PIPE_READERS appropriately.
975                  */
976                 pipe->readers = 1;
977
978                 current->splice_pipe = pipe;
979         }
980
981         /*
982          * Do the splice.
983          */
984         ret = 0;
985         bytes = 0;
986         len = sd->total_len;
987         flags = sd->flags;
988
989         /*
990          * Don't block on output, we have to drain the direct pipe.
991          */
992         sd->flags &= ~SPLICE_F_NONBLOCK;
993
994         while (len) {
995                 size_t read_len;
996                 loff_t pos = sd->pos, prev_pos = pos;
997
998                 ret = do_splice_to(in, &pos, pipe, len, flags);
999                 if (unlikely(ret <= 0))
1000                         goto out_release;
1001
1002                 read_len = ret;
1003                 sd->total_len = read_len;
1004
1005                 /*
1006                  * NOTE: nonblocking mode only applies to the input. We
1007                  * must not do the output in nonblocking mode as then we
1008                  * could get stuck data in the internal pipe:
1009                  */
1010                 ret = actor(pipe, sd);
1011                 if (unlikely(ret <= 0)) {
1012                         sd->pos = prev_pos;
1013                         goto out_release;
1014                 }
1015
1016                 bytes += ret;
1017                 len -= ret;
1018                 sd->pos = pos;
1019
1020                 if (ret < read_len) {
1021                         sd->pos = prev_pos + ret;
1022                         goto out_release;
1023                 }
1024         }
1025
1026 done:
1027         pipe->nrbufs = pipe->curbuf = 0;
1028         file_accessed(in);
1029         return bytes;
1030
1031 out_release:
1032         /*
1033          * If we did an incomplete transfer we must release
1034          * the pipe buffers in question:
1035          */
1036         for (i = 0; i < PIPE_BUFFERS; i++) {
1037                 struct pipe_buffer *buf = pipe->bufs + i;
1038
1039                 if (buf->ops) {
1040                         buf->ops->release(pipe, buf);
1041                         buf->ops = NULL;
1042                 }
1043         }
1044
1045         if (!bytes)
1046                 bytes = ret;
1047
1048         goto done;
1049 }
1050 EXPORT_SYMBOL(splice_direct_to_actor);
1051
1052 static int direct_splice_actor(struct pipe_inode_info *pipe,
1053                                struct splice_desc *sd)
1054 {
1055         struct file *file = sd->u.file;
1056
1057         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1058 }
1059
1060 /**
1061  * do_splice_direct - splices data directly between two files
1062  * @in:         file to splice from
1063  * @ppos:       input file offset
1064  * @out:        file to splice to
1065  * @len:        number of bytes to splice
1066  * @flags:      splice modifier flags
1067  *
1068  * Description:
1069  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1070  *    doing it in the application would incur an extra system call
1071  *    (splice in + splice out, as compared to just sendfile()). So this helper
1072  *    can splice directly through a process-private pipe.
1073  *
1074  */
1075 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1076                       size_t len, unsigned int flags)
1077 {
1078         struct splice_desc sd = {
1079                 .len            = len,
1080                 .total_len      = len,
1081                 .flags          = flags,
1082                 .pos            = *ppos,
1083                 .u.file         = out,
1084         };
1085         long ret;
1086
1087         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1088         if (ret > 0)
1089                 *ppos = sd.pos;
1090
1091         return ret;
1092 }
1093
1094 /*
1095  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1096  * location, so checking ->i_pipe is not enough to verify that this is a
1097  * pipe.
1098  */
1099 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1100 {
1101         if (S_ISFIFO(inode->i_mode))
1102                 return inode->i_pipe;
1103
1104         return NULL;
1105 }
1106
1107 /*
1108  * Determine where to splice to/from.
1109  */
1110 static long do_splice(struct file *in, loff_t __user *off_in,
1111                       struct file *out, loff_t __user *off_out,
1112                       size_t len, unsigned int flags)
1113 {
1114         struct pipe_inode_info *pipe;
1115         loff_t offset, *off;
1116         long ret;
1117
1118         pipe = pipe_info(in->f_path.dentry->d_inode);
1119         if (pipe) {
1120                 if (off_in)
1121                         return -ESPIPE;
1122                 if (off_out) {
1123                         if (out->f_op->llseek == no_llseek)
1124                                 return -EINVAL;
1125                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1126                                 return -EFAULT;
1127                         off = &offset;
1128                 } else
1129                         off = &out->f_pos;
1130
1131                 ret = do_splice_from(pipe, out, off, len, flags);
1132
1133                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1134                         ret = -EFAULT;
1135
1136                 return ret;
1137         }
1138
1139         pipe = pipe_info(out->f_path.dentry->d_inode);
1140         if (pipe) {
1141                 if (off_out)
1142                         return -ESPIPE;
1143                 if (off_in) {
1144                         if (in->f_op->llseek == no_llseek)
1145                                 return -EINVAL;
1146                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1147                                 return -EFAULT;
1148                         off = &offset;
1149                 } else
1150                         off = &in->f_pos;
1151
1152                 ret = do_splice_to(in, off, pipe, len, flags);
1153
1154                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1155                         ret = -EFAULT;
1156
1157                 return ret;
1158         }
1159
1160         return -EINVAL;
1161 }
1162
1163 /*
1164  * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1165  * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1166  * for writing) and page faulting on the user memory pointed to by src.
1167  * This assumes that we will very rarely hit the partial != 0 path, or this
1168  * will not be a win.
1169  */
1170 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1171 {
1172         int partial;
1173
1174         if (!access_ok(VERIFY_READ, src, n))
1175                 return -EFAULT;
1176
1177         pagefault_disable();
1178         partial = __copy_from_user_inatomic(dst, src, n);
1179         pagefault_enable();
1180
1181         /*
1182          * Didn't copy everything, drop the mmap_sem and do a faulting copy
1183          */
1184         if (unlikely(partial)) {
1185                 up_read(&current->mm->mmap_sem);
1186                 partial = copy_from_user(dst, src, n);
1187                 down_read(&current->mm->mmap_sem);
1188         }
1189
1190         return partial;
1191 }
1192
1193 /*
1194  * Map an iov into an array of pages and offset/length tupples. With the
1195  * partial_page structure, we can map several non-contiguous ranges into
1196  * our ones pages[] map instead of splitting that operation into pieces.
1197  * Could easily be exported as a generic helper for other users, in which
1198  * case one would probably want to add a 'max_nr_pages' parameter as well.
1199  */
1200 static int get_iovec_page_array(const struct iovec __user *iov,
1201                                 unsigned int nr_vecs, struct page **pages,
1202                                 struct partial_page *partial, int aligned)
1203 {
1204         int buffers = 0, error = 0;
1205
1206         down_read(&current->mm->mmap_sem);
1207
1208         while (nr_vecs) {
1209                 unsigned long off, npages;
1210                 struct iovec entry;
1211                 void __user *base;
1212                 size_t len;
1213                 int i;
1214
1215                 error = -EFAULT;
1216                 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1217                         break;
1218
1219                 base = entry.iov_base;
1220                 len = entry.iov_len;
1221
1222                 /*
1223                  * Sanity check this iovec. 0 read succeeds.
1224                  */
1225                 error = 0;
1226                 if (unlikely(!len))
1227                         break;
1228                 error = -EFAULT;
1229                 if (!access_ok(VERIFY_READ, base, len))
1230                         break;
1231
1232                 /*
1233                  * Get this base offset and number of pages, then map
1234                  * in the user pages.
1235                  */
1236                 off = (unsigned long) base & ~PAGE_MASK;
1237
1238                 /*
1239                  * If asked for alignment, the offset must be zero and the
1240                  * length a multiple of the PAGE_SIZE.
1241                  */
1242                 error = -EINVAL;
1243                 if (aligned && (off || len & ~PAGE_MASK))
1244                         break;
1245
1246                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1247                 if (npages > PIPE_BUFFERS - buffers)
1248                         npages = PIPE_BUFFERS - buffers;
1249
1250                 error = get_user_pages(current, current->mm,
1251                                        (unsigned long) base, npages, 0, 0,
1252                                        &pages[buffers], NULL);
1253
1254                 if (unlikely(error <= 0))
1255                         break;
1256
1257                 /*
1258                  * Fill this contiguous range into the partial page map.
1259                  */
1260                 for (i = 0; i < error; i++) {
1261                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1262
1263                         partial[buffers].offset = off;
1264                         partial[buffers].len = plen;
1265
1266                         off = 0;
1267                         len -= plen;
1268                         buffers++;
1269                 }
1270
1271                 /*
1272                  * We didn't complete this iov, stop here since it probably
1273                  * means we have to move some of this into a pipe to
1274                  * be able to continue.
1275                  */
1276                 if (len)
1277                         break;
1278
1279                 /*
1280                  * Don't continue if we mapped fewer pages than we asked for,
1281                  * or if we mapped the max number of pages that we have
1282                  * room for.
1283                  */
1284                 if (error < npages || buffers == PIPE_BUFFERS)
1285                         break;
1286
1287                 nr_vecs--;
1288                 iov++;
1289         }
1290
1291         up_read(&current->mm->mmap_sem);
1292
1293         if (buffers)
1294                 return buffers;
1295
1296         return error;
1297 }
1298
1299 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1300                         struct splice_desc *sd)
1301 {
1302         char *src;
1303         int ret;
1304
1305         ret = buf->ops->confirm(pipe, buf);
1306         if (unlikely(ret))
1307                 return ret;
1308
1309         /*
1310          * See if we can use the atomic maps, by prefaulting in the
1311          * pages and doing an atomic copy
1312          */
1313         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1314                 src = buf->ops->map(pipe, buf, 1);
1315                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1316                                                         sd->len);
1317                 buf->ops->unmap(pipe, buf, src);
1318                 if (!ret) {
1319                         ret = sd->len;
1320                         goto out;
1321                 }
1322         }
1323
1324         /*
1325          * No dice, use slow non-atomic map and copy
1326          */
1327         src = buf->ops->map(pipe, buf, 0);
1328
1329         ret = sd->len;
1330         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1331                 ret = -EFAULT;
1332
1333         buf->ops->unmap(pipe, buf, src);
1334 out:
1335         if (ret > 0)
1336                 sd->u.userptr += ret;
1337         return ret;
1338 }
1339
1340 /*
1341  * For lack of a better implementation, implement vmsplice() to userspace
1342  * as a simple copy of the pipes pages to the user iov.
1343  */
1344 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1345                              unsigned long nr_segs, unsigned int flags)
1346 {
1347         struct pipe_inode_info *pipe;
1348         struct splice_desc sd;
1349         ssize_t size;
1350         int error;
1351         long ret;
1352
1353         pipe = pipe_info(file->f_path.dentry->d_inode);
1354         if (!pipe)
1355                 return -EBADF;
1356
1357         if (pipe->inode)
1358                 mutex_lock(&pipe->inode->i_mutex);
1359
1360         error = ret = 0;
1361         while (nr_segs) {
1362                 void __user *base;
1363                 size_t len;
1364
1365                 /*
1366                  * Get user address base and length for this iovec.
1367                  */
1368                 error = get_user(base, &iov->iov_base);
1369                 if (unlikely(error))
1370                         break;
1371                 error = get_user(len, &iov->iov_len);
1372                 if (unlikely(error))
1373                         break;
1374
1375                 /*
1376                  * Sanity check this iovec. 0 read succeeds.
1377                  */
1378                 if (unlikely(!len))
1379                         break;
1380                 if (unlikely(!base)) {
1381                         error = -EFAULT;
1382                         break;
1383                 }
1384
1385                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1386                         error = -EFAULT;
1387                         break;
1388                 }
1389
1390                 sd.len = 0;
1391                 sd.total_len = len;
1392                 sd.flags = flags;
1393                 sd.u.userptr = base;
1394                 sd.pos = 0;
1395
1396                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1397                 if (size < 0) {
1398                         if (!ret)
1399                                 ret = size;
1400
1401                         break;
1402                 }
1403
1404                 ret += size;
1405
1406                 if (size < len)
1407                         break;
1408
1409                 nr_segs--;
1410                 iov++;
1411         }
1412
1413         if (pipe->inode)
1414                 mutex_unlock(&pipe->inode->i_mutex);
1415
1416         if (!ret)
1417                 ret = error;
1418
1419         return ret;
1420 }
1421
1422 /*
1423  * vmsplice splices a user address range into a pipe. It can be thought of
1424  * as splice-from-memory, where the regular splice is splice-from-file (or
1425  * to file). In both cases the output is a pipe, naturally.
1426  */
1427 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1428                              unsigned long nr_segs, unsigned int flags)
1429 {
1430         struct pipe_inode_info *pipe;
1431         struct page *pages[PIPE_BUFFERS];
1432         struct partial_page partial[PIPE_BUFFERS];
1433         struct splice_pipe_desc spd = {
1434                 .pages = pages,
1435                 .partial = partial,
1436                 .flags = flags,
1437                 .ops = &user_page_pipe_buf_ops,
1438                 .spd_release = spd_release_page,
1439         };
1440
1441         pipe = pipe_info(file->f_path.dentry->d_inode);
1442         if (!pipe)
1443                 return -EBADF;
1444
1445         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1446                                             flags & SPLICE_F_GIFT);
1447         if (spd.nr_pages <= 0)
1448                 return spd.nr_pages;
1449
1450         return splice_to_pipe(pipe, &spd);
1451 }
1452
1453 /*
1454  * Note that vmsplice only really supports true splicing _from_ user memory
1455  * to a pipe, not the other way around. Splicing from user memory is a simple
1456  * operation that can be supported without any funky alignment restrictions
1457  * or nasty vm tricks. We simply map in the user memory and fill them into
1458  * a pipe. The reverse isn't quite as easy, though. There are two possible
1459  * solutions for that:
1460  *
1461  *      - memcpy() the data internally, at which point we might as well just
1462  *        do a regular read() on the buffer anyway.
1463  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1464  *        has restriction limitations on both ends of the pipe).
1465  *
1466  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1467  *
1468  */
1469 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1470                              unsigned long nr_segs, unsigned int flags)
1471 {
1472         struct file *file;
1473         long error;
1474         int fput;
1475
1476         if (unlikely(nr_segs > UIO_MAXIOV))
1477                 return -EINVAL;
1478         else if (unlikely(!nr_segs))
1479                 return 0;
1480
1481         error = -EBADF;
1482         file = fget_light(fd, &fput);
1483         if (file) {
1484                 if (file->f_mode & FMODE_WRITE)
1485                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1486                 else if (file->f_mode & FMODE_READ)
1487                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1488
1489                 fput_light(file, fput);
1490         }
1491
1492         return error;
1493 }
1494
1495 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1496                            int fd_out, loff_t __user *off_out,
1497                            size_t len, unsigned int flags)
1498 {
1499         long error;
1500         struct file *in, *out;
1501         int fput_in, fput_out;
1502
1503         if (unlikely(!len))
1504                 return 0;
1505
1506         error = -EBADF;
1507         in = fget_light(fd_in, &fput_in);
1508         if (in) {
1509                 if (in->f_mode & FMODE_READ) {
1510                         out = fget_light(fd_out, &fput_out);
1511                         if (out) {
1512                                 if (out->f_mode & FMODE_WRITE)
1513                                         error = do_splice(in, off_in,
1514                                                           out, off_out,
1515                                                           len, flags);
1516                                 fput_light(out, fput_out);
1517                         }
1518                 }
1519
1520                 fput_light(in, fput_in);
1521         }
1522
1523         return error;
1524 }
1525
1526 /*
1527  * Make sure there's data to read. Wait for input if we can, otherwise
1528  * return an appropriate error.
1529  */
1530 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1531 {
1532         int ret;
1533
1534         /*
1535          * Check ->nrbufs without the inode lock first. This function
1536          * is speculative anyways, so missing one is ok.
1537          */
1538         if (pipe->nrbufs)
1539                 return 0;
1540
1541         ret = 0;
1542         mutex_lock(&pipe->inode->i_mutex);
1543
1544         while (!pipe->nrbufs) {
1545                 if (signal_pending(current)) {
1546                         ret = -ERESTARTSYS;
1547                         break;
1548                 }
1549                 if (!pipe->writers)
1550                         break;
1551                 if (!pipe->waiting_writers) {
1552                         if (flags & SPLICE_F_NONBLOCK) {
1553                                 ret = -EAGAIN;
1554                                 break;
1555                         }
1556                 }
1557                 pipe_wait(pipe);
1558         }
1559
1560         mutex_unlock(&pipe->inode->i_mutex);
1561         return ret;
1562 }
1563
1564 /*
1565  * Make sure there's writeable room. Wait for room if we can, otherwise
1566  * return an appropriate error.
1567  */
1568 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1569 {
1570         int ret;
1571
1572         /*
1573          * Check ->nrbufs without the inode lock first. This function
1574          * is speculative anyways, so missing one is ok.
1575          */
1576         if (pipe->nrbufs < PIPE_BUFFERS)
1577                 return 0;
1578
1579         ret = 0;
1580         mutex_lock(&pipe->inode->i_mutex);
1581
1582         while (pipe->nrbufs >= PIPE_BUFFERS) {
1583                 if (!pipe->readers) {
1584                         send_sig(SIGPIPE, current, 0);
1585                         ret = -EPIPE;
1586                         break;
1587                 }
1588                 if (flags & SPLICE_F_NONBLOCK) {
1589                         ret = -EAGAIN;
1590                         break;
1591                 }
1592                 if (signal_pending(current)) {
1593                         ret = -ERESTARTSYS;
1594                         break;
1595                 }
1596                 pipe->waiting_writers++;
1597                 pipe_wait(pipe);
1598                 pipe->waiting_writers--;
1599         }
1600
1601         mutex_unlock(&pipe->inode->i_mutex);
1602         return ret;
1603 }
1604
1605 /*
1606  * Link contents of ipipe to opipe.
1607  */
1608 static int link_pipe(struct pipe_inode_info *ipipe,
1609                      struct pipe_inode_info *opipe,
1610                      size_t len, unsigned int flags)
1611 {
1612         struct pipe_buffer *ibuf, *obuf;
1613         int ret = 0, i = 0, nbuf;
1614
1615         /*
1616          * Potential ABBA deadlock, work around it by ordering lock
1617          * grabbing by inode address. Otherwise two different processes
1618          * could deadlock (one doing tee from A -> B, the other from B -> A).
1619          */
1620         inode_double_lock(ipipe->inode, opipe->inode);
1621
1622         do {
1623                 if (!opipe->readers) {
1624                         send_sig(SIGPIPE, current, 0);
1625                         if (!ret)
1626                                 ret = -EPIPE;
1627                         break;
1628                 }
1629
1630                 /*
1631                  * If we have iterated all input buffers or ran out of
1632                  * output room, break.
1633                  */
1634                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1635                         break;
1636
1637                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1638                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1639
1640                 /*
1641                  * Get a reference to this pipe buffer,
1642                  * so we can copy the contents over.
1643                  */
1644                 ibuf->ops->get(ipipe, ibuf);
1645
1646                 obuf = opipe->bufs + nbuf;
1647                 *obuf = *ibuf;
1648
1649                 /*
1650                  * Don't inherit the gift flag, we need to
1651                  * prevent multiple steals of this page.
1652                  */
1653                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1654
1655                 if (obuf->len > len)
1656                         obuf->len = len;
1657
1658                 opipe->nrbufs++;
1659                 ret += obuf->len;
1660                 len -= obuf->len;
1661                 i++;
1662         } while (len);
1663
1664         /*
1665          * return EAGAIN if we have the potential of some data in the
1666          * future, otherwise just return 0
1667          */
1668         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1669                 ret = -EAGAIN;
1670
1671         inode_double_unlock(ipipe->inode, opipe->inode);
1672
1673         /*
1674          * If we put data in the output pipe, wakeup any potential readers.
1675          */
1676         if (ret > 0) {
1677                 smp_mb();
1678                 if (waitqueue_active(&opipe->wait))
1679                         wake_up_interruptible(&opipe->wait);
1680                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1681         }
1682
1683         return ret;
1684 }
1685
1686 /*
1687  * This is a tee(1) implementation that works on pipes. It doesn't copy
1688  * any data, it simply references the 'in' pages on the 'out' pipe.
1689  * The 'flags' used are the SPLICE_F_* variants, currently the only
1690  * applicable one is SPLICE_F_NONBLOCK.
1691  */
1692 static long do_tee(struct file *in, struct file *out, size_t len,
1693                    unsigned int flags)
1694 {
1695         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1696         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1697         int ret = -EINVAL;
1698
1699         /*
1700          * Duplicate the contents of ipipe to opipe without actually
1701          * copying the data.
1702          */
1703         if (ipipe && opipe && ipipe != opipe) {
1704                 /*
1705                  * Keep going, unless we encounter an error. The ipipe/opipe
1706                  * ordering doesn't really matter.
1707                  */
1708                 ret = link_ipipe_prep(ipipe, flags);
1709                 if (!ret) {
1710                         ret = link_opipe_prep(opipe, flags);
1711                         if (!ret)
1712                                 ret = link_pipe(ipipe, opipe, len, flags);
1713                 }
1714         }
1715
1716         return ret;
1717 }
1718
1719 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1720 {
1721         struct file *in;
1722         int error, fput_in;
1723
1724         if (unlikely(!len))
1725                 return 0;
1726
1727         error = -EBADF;
1728         in = fget_light(fdin, &fput_in);
1729         if (in) {
1730                 if (in->f_mode & FMODE_READ) {
1731                         int fput_out;
1732                         struct file *out = fget_light(fdout, &fput_out);
1733
1734                         if (out) {
1735                                 if (out->f_mode & FMODE_WRITE)
1736                                         error = do_tee(in, out, len, flags);
1737                                 fput_light(out, fput_out);
1738                         }
1739                 }
1740                 fput_light(in, fput_in);
1741         }
1742
1743         return error;
1744 }