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