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