ocfs2: fix i_mutex locking in ocfs2_splice_to_file()
[linux-2.6] / 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 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 EXPORT_SYMBOL(pipe_to_file);
604
605 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
606 {
607         smp_mb();
608         if (waitqueue_active(&pipe->wait))
609                 wake_up_interruptible(&pipe->wait);
610         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
611 }
612
613 /**
614  * splice_from_pipe_feed - feed available data from a pipe to a file
615  * @pipe:       pipe to splice from
616  * @sd:         information to @actor
617  * @actor:      handler that splices the data
618  *
619  * Description:
620
621  *    This function loops over the pipe and calls @actor to do the
622  *    actual moving of a single struct pipe_buffer to the desired
623  *    destination.  It returns when there's no more buffers left in
624  *    the pipe or if the requested number of bytes (@sd->total_len)
625  *    have been copied.  It returns a positive number (one) if the
626  *    pipe needs to be filled with more data, zero if the required
627  *    number of bytes have been copied and -errno on error.
628  *
629  *    This, together with splice_from_pipe_{begin,end,next}, may be
630  *    used to implement the functionality of __splice_from_pipe() when
631  *    locking is required around copying the pipe buffers to the
632  *    destination.
633  */
634 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
635                           splice_actor *actor)
636 {
637         int ret;
638
639         while (pipe->nrbufs) {
640                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
641                 const struct pipe_buf_operations *ops = buf->ops;
642
643                 sd->len = buf->len;
644                 if (sd->len > sd->total_len)
645                         sd->len = sd->total_len;
646
647                 ret = actor(pipe, buf, sd);
648                 if (ret <= 0) {
649                         if (ret == -ENODATA)
650                                 ret = 0;
651                         return ret;
652                 }
653                 buf->offset += ret;
654                 buf->len -= ret;
655
656                 sd->num_spliced += ret;
657                 sd->len -= ret;
658                 sd->pos += ret;
659                 sd->total_len -= ret;
660
661                 if (!buf->len) {
662                         buf->ops = NULL;
663                         ops->release(pipe, buf);
664                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
665                         pipe->nrbufs--;
666                         if (pipe->inode)
667                                 sd->need_wakeup = true;
668                 }
669
670                 if (!sd->total_len)
671                         return 0;
672         }
673
674         return 1;
675 }
676 EXPORT_SYMBOL(splice_from_pipe_feed);
677
678 /**
679  * splice_from_pipe_next - wait for some data to splice from
680  * @pipe:       pipe to splice from
681  * @sd:         information about the splice operation
682  *
683  * Description:
684  *    This function will wait for some data and return a positive
685  *    value (one) if pipe buffers are available.  It will return zero
686  *    or -errno if no more data needs to be spliced.
687  */
688 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
689 {
690         while (!pipe->nrbufs) {
691                 if (!pipe->writers)
692                         return 0;
693
694                 if (!pipe->waiting_writers && sd->num_spliced)
695                         return 0;
696
697                 if (sd->flags & SPLICE_F_NONBLOCK)
698                         return -EAGAIN;
699
700                 if (signal_pending(current))
701                         return -ERESTARTSYS;
702
703                 if (sd->need_wakeup) {
704                         wakeup_pipe_writers(pipe);
705                         sd->need_wakeup = false;
706                 }
707
708                 pipe_wait(pipe);
709         }
710
711         return 1;
712 }
713 EXPORT_SYMBOL(splice_from_pipe_next);
714
715 /**
716  * splice_from_pipe_begin - start splicing from pipe
717  * @pipe:       pipe to splice from
718  *
719  * Description:
720  *    This function should be called before a loop containing
721  *    splice_from_pipe_next() and splice_from_pipe_feed() to
722  *    initialize the necessary fields of @sd.
723  */
724 void splice_from_pipe_begin(struct splice_desc *sd)
725 {
726         sd->num_spliced = 0;
727         sd->need_wakeup = false;
728 }
729 EXPORT_SYMBOL(splice_from_pipe_begin);
730
731 /**
732  * splice_from_pipe_end - finish splicing from pipe
733  * @pipe:       pipe to splice from
734  * @sd:         information about the splice operation
735  *
736  * Description:
737  *    This function will wake up pipe writers if necessary.  It should
738  *    be called after a loop containing splice_from_pipe_next() and
739  *    splice_from_pipe_feed().
740  */
741 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
742 {
743         if (sd->need_wakeup)
744                 wakeup_pipe_writers(pipe);
745 }
746 EXPORT_SYMBOL(splice_from_pipe_end);
747
748 /**
749  * __splice_from_pipe - splice data from a pipe to given actor
750  * @pipe:       pipe to splice from
751  * @sd:         information to @actor
752  * @actor:      handler that splices the data
753  *
754  * Description:
755  *    This function does little more than loop over the pipe and call
756  *    @actor to do the actual moving of a single struct pipe_buffer to
757  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
758  *    pipe_to_user.
759  *
760  */
761 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
762                            splice_actor *actor)
763 {
764         int ret;
765
766         splice_from_pipe_begin(sd);
767         do {
768                 ret = splice_from_pipe_next(pipe, sd);
769                 if (ret > 0)
770                         ret = splice_from_pipe_feed(pipe, sd, actor);
771         } while (ret > 0);
772         splice_from_pipe_end(pipe, sd);
773
774         return sd->num_spliced ? sd->num_spliced : ret;
775 }
776 EXPORT_SYMBOL(__splice_from_pipe);
777
778 /**
779  * splice_from_pipe - splice data from a pipe to a file
780  * @pipe:       pipe to splice from
781  * @out:        file to splice to
782  * @ppos:       position in @out
783  * @len:        how many bytes to splice
784  * @flags:      splice modifier flags
785  * @actor:      handler that splices the data
786  *
787  * Description:
788  *    See __splice_from_pipe. This function locks the pipe inode,
789  *    otherwise it's identical to __splice_from_pipe().
790  *
791  */
792 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
793                          loff_t *ppos, size_t len, unsigned int flags,
794                          splice_actor *actor)
795 {
796         ssize_t ret;
797         struct splice_desc sd = {
798                 .total_len = len,
799                 .flags = flags,
800                 .pos = *ppos,
801                 .u.file = out,
802         };
803
804         if (pipe->inode)
805                 mutex_lock(&pipe->inode->i_mutex);
806         ret = __splice_from_pipe(pipe, &sd, actor);
807         if (pipe->inode)
808                 mutex_unlock(&pipe->inode->i_mutex);
809
810         return ret;
811 }
812
813 /**
814  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
815  * @pipe:       pipe info
816  * @out:        file to write to
817  * @ppos:       position in @out
818  * @len:        number of bytes to splice
819  * @flags:      splice modifier flags
820  *
821  * Description:
822  *    Will either move or copy pages (determined by @flags options) from
823  *    the given pipe inode to the given file. The caller is responsible
824  *    for acquiring i_mutex on both inodes.
825  *
826  */
827 ssize_t
828 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
829                                  loff_t *ppos, size_t len, unsigned int flags)
830 {
831         struct address_space *mapping = out->f_mapping;
832         struct inode *inode = mapping->host;
833         struct splice_desc sd = {
834                 .total_len = len,
835                 .flags = flags,
836                 .pos = *ppos,
837                 .u.file = out,
838         };
839         ssize_t ret;
840         int err;
841
842         err = file_remove_suid(out);
843         if (unlikely(err))
844                 return err;
845
846         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
847         if (ret > 0) {
848                 unsigned long nr_pages;
849
850                 *ppos += ret;
851                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
852
853                 /*
854                  * If file or inode is SYNC and we actually wrote some data,
855                  * sync it.
856                  */
857                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
858                         err = generic_osync_inode(inode, mapping,
859                                                   OSYNC_METADATA|OSYNC_DATA);
860
861                         if (err)
862                                 ret = err;
863                 }
864                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
865         }
866
867         return ret;
868 }
869
870 EXPORT_SYMBOL(generic_file_splice_write_nolock);
871
872 /**
873  * generic_file_splice_write - splice data from a pipe to a file
874  * @pipe:       pipe info
875  * @out:        file to write to
876  * @ppos:       position in @out
877  * @len:        number of bytes to splice
878  * @flags:      splice modifier flags
879  *
880  * Description:
881  *    Will either move or copy pages (determined by @flags options) from
882  *    the given pipe inode to the given file.
883  *
884  */
885 ssize_t
886 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
887                           loff_t *ppos, size_t len, unsigned int flags)
888 {
889         struct address_space *mapping = out->f_mapping;
890         struct inode *inode = mapping->host;
891         struct splice_desc sd = {
892                 .total_len = len,
893                 .flags = flags,
894                 .pos = *ppos,
895                 .u.file = out,
896         };
897         ssize_t ret;
898
899         if (pipe->inode)
900                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
901
902         splice_from_pipe_begin(&sd);
903         do {
904                 ret = splice_from_pipe_next(pipe, &sd);
905                 if (ret <= 0)
906                         break;
907
908                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
909                 ret = file_remove_suid(out);
910                 if (!ret)
911                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
912                 mutex_unlock(&inode->i_mutex);
913         } while (ret > 0);
914         splice_from_pipe_end(pipe, &sd);
915
916         if (pipe->inode)
917                 mutex_unlock(&pipe->inode->i_mutex);
918
919         if (sd.num_spliced)
920                 ret = sd.num_spliced;
921
922         if (ret > 0) {
923                 unsigned long nr_pages;
924
925                 *ppos += ret;
926                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
927
928                 /*
929                  * If file or inode is SYNC and we actually wrote some data,
930                  * sync it.
931                  */
932                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
933                         int err;
934
935                         mutex_lock(&inode->i_mutex);
936                         err = generic_osync_inode(inode, mapping,
937                                                   OSYNC_METADATA|OSYNC_DATA);
938                         mutex_unlock(&inode->i_mutex);
939
940                         if (err)
941                                 ret = err;
942                 }
943                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
944         }
945
946         return ret;
947 }
948
949 EXPORT_SYMBOL(generic_file_splice_write);
950
951 /**
952  * generic_splice_sendpage - splice data from a pipe to a socket
953  * @pipe:       pipe to splice from
954  * @out:        socket to write to
955  * @ppos:       position in @out
956  * @len:        number of bytes to splice
957  * @flags:      splice modifier flags
958  *
959  * Description:
960  *    Will send @len bytes from the pipe to a network socket. No data copying
961  *    is involved.
962  *
963  */
964 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
965                                 loff_t *ppos, size_t len, unsigned int flags)
966 {
967         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
968 }
969
970 EXPORT_SYMBOL(generic_splice_sendpage);
971
972 /*
973  * Attempt to initiate a splice from pipe to file.
974  */
975 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
976                            loff_t *ppos, size_t len, unsigned int flags)
977 {
978         int ret;
979
980         if (unlikely(!out->f_op || !out->f_op->splice_write))
981                 return -EINVAL;
982
983         if (unlikely(!(out->f_mode & FMODE_WRITE)))
984                 return -EBADF;
985
986         if (unlikely(out->f_flags & O_APPEND))
987                 return -EINVAL;
988
989         ret = rw_verify_area(WRITE, out, ppos, len);
990         if (unlikely(ret < 0))
991                 return ret;
992
993         return out->f_op->splice_write(pipe, out, ppos, len, flags);
994 }
995
996 /*
997  * Attempt to initiate a splice from a file to a pipe.
998  */
999 static long do_splice_to(struct file *in, loff_t *ppos,
1000                          struct pipe_inode_info *pipe, size_t len,
1001                          unsigned int flags)
1002 {
1003         int ret;
1004
1005         if (unlikely(!in->f_op || !in->f_op->splice_read))
1006                 return -EINVAL;
1007
1008         if (unlikely(!(in->f_mode & FMODE_READ)))
1009                 return -EBADF;
1010
1011         ret = rw_verify_area(READ, in, ppos, len);
1012         if (unlikely(ret < 0))
1013                 return ret;
1014
1015         return in->f_op->splice_read(in, ppos, pipe, len, flags);
1016 }
1017
1018 /**
1019  * splice_direct_to_actor - splices data directly between two non-pipes
1020  * @in:         file to splice from
1021  * @sd:         actor information on where to splice to
1022  * @actor:      handles the data splicing
1023  *
1024  * Description:
1025  *    This is a special case helper to splice directly between two
1026  *    points, without requiring an explicit pipe. Internally an allocated
1027  *    pipe is cached in the process, and reused during the lifetime of
1028  *    that process.
1029  *
1030  */
1031 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1032                                splice_direct_actor *actor)
1033 {
1034         struct pipe_inode_info *pipe;
1035         long ret, bytes;
1036         umode_t i_mode;
1037         size_t len;
1038         int i, flags;
1039
1040         /*
1041          * We require the input being a regular file, as we don't want to
1042          * randomly drop data for eg socket -> socket splicing. Use the
1043          * piped splicing for that!
1044          */
1045         i_mode = in->f_path.dentry->d_inode->i_mode;
1046         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1047                 return -EINVAL;
1048
1049         /*
1050          * neither in nor out is a pipe, setup an internal pipe attached to
1051          * 'out' and transfer the wanted data from 'in' to 'out' through that
1052          */
1053         pipe = current->splice_pipe;
1054         if (unlikely(!pipe)) {
1055                 pipe = alloc_pipe_info(NULL);
1056                 if (!pipe)
1057                         return -ENOMEM;
1058
1059                 /*
1060                  * We don't have an immediate reader, but we'll read the stuff
1061                  * out of the pipe right after the splice_to_pipe(). So set
1062                  * PIPE_READERS appropriately.
1063                  */
1064                 pipe->readers = 1;
1065
1066                 current->splice_pipe = pipe;
1067         }
1068
1069         /*
1070          * Do the splice.
1071          */
1072         ret = 0;
1073         bytes = 0;
1074         len = sd->total_len;
1075         flags = sd->flags;
1076
1077         /*
1078          * Don't block on output, we have to drain the direct pipe.
1079          */
1080         sd->flags &= ~SPLICE_F_NONBLOCK;
1081
1082         while (len) {
1083                 size_t read_len;
1084                 loff_t pos = sd->pos, prev_pos = pos;
1085
1086                 ret = do_splice_to(in, &pos, pipe, len, flags);
1087                 if (unlikely(ret <= 0))
1088                         goto out_release;
1089
1090                 read_len = ret;
1091                 sd->total_len = read_len;
1092
1093                 /*
1094                  * NOTE: nonblocking mode only applies to the input. We
1095                  * must not do the output in nonblocking mode as then we
1096                  * could get stuck data in the internal pipe:
1097                  */
1098                 ret = actor(pipe, sd);
1099                 if (unlikely(ret <= 0)) {
1100                         sd->pos = prev_pos;
1101                         goto out_release;
1102                 }
1103
1104                 bytes += ret;
1105                 len -= ret;
1106                 sd->pos = pos;
1107
1108                 if (ret < read_len) {
1109                         sd->pos = prev_pos + ret;
1110                         goto out_release;
1111                 }
1112         }
1113
1114 done:
1115         pipe->nrbufs = pipe->curbuf = 0;
1116         file_accessed(in);
1117         return bytes;
1118
1119 out_release:
1120         /*
1121          * If we did an incomplete transfer we must release
1122          * the pipe buffers in question:
1123          */
1124         for (i = 0; i < PIPE_BUFFERS; i++) {
1125                 struct pipe_buffer *buf = pipe->bufs + i;
1126
1127                 if (buf->ops) {
1128                         buf->ops->release(pipe, buf);
1129                         buf->ops = NULL;
1130                 }
1131         }
1132
1133         if (!bytes)
1134                 bytes = ret;
1135
1136         goto done;
1137 }
1138 EXPORT_SYMBOL(splice_direct_to_actor);
1139
1140 static int direct_splice_actor(struct pipe_inode_info *pipe,
1141                                struct splice_desc *sd)
1142 {
1143         struct file *file = sd->u.file;
1144
1145         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1146 }
1147
1148 /**
1149  * do_splice_direct - splices data directly between two files
1150  * @in:         file to splice from
1151  * @ppos:       input file offset
1152  * @out:        file to splice to
1153  * @len:        number of bytes to splice
1154  * @flags:      splice modifier flags
1155  *
1156  * Description:
1157  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1158  *    doing it in the application would incur an extra system call
1159  *    (splice in + splice out, as compared to just sendfile()). So this helper
1160  *    can splice directly through a process-private pipe.
1161  *
1162  */
1163 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1164                       size_t len, unsigned int flags)
1165 {
1166         struct splice_desc sd = {
1167                 .len            = len,
1168                 .total_len      = len,
1169                 .flags          = flags,
1170                 .pos            = *ppos,
1171                 .u.file         = out,
1172         };
1173         long ret;
1174
1175         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1176         if (ret > 0)
1177                 *ppos = sd.pos;
1178
1179         return ret;
1180 }
1181
1182 /*
1183  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1184  * location, so checking ->i_pipe is not enough to verify that this is a
1185  * pipe.
1186  */
1187 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1188 {
1189         if (S_ISFIFO(inode->i_mode))
1190                 return inode->i_pipe;
1191
1192         return NULL;
1193 }
1194
1195 /*
1196  * Determine where to splice to/from.
1197  */
1198 static long do_splice(struct file *in, loff_t __user *off_in,
1199                       struct file *out, loff_t __user *off_out,
1200                       size_t len, unsigned int flags)
1201 {
1202         struct pipe_inode_info *pipe;
1203         loff_t offset, *off;
1204         long ret;
1205
1206         pipe = pipe_info(in->f_path.dentry->d_inode);
1207         if (pipe) {
1208                 if (off_in)
1209                         return -ESPIPE;
1210                 if (off_out) {
1211                         if (out->f_op->llseek == no_llseek)
1212                                 return -EINVAL;
1213                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1214                                 return -EFAULT;
1215                         off = &offset;
1216                 } else
1217                         off = &out->f_pos;
1218
1219                 ret = do_splice_from(pipe, out, off, len, flags);
1220
1221                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1222                         ret = -EFAULT;
1223
1224                 return ret;
1225         }
1226
1227         pipe = pipe_info(out->f_path.dentry->d_inode);
1228         if (pipe) {
1229                 if (off_out)
1230                         return -ESPIPE;
1231                 if (off_in) {
1232                         if (in->f_op->llseek == no_llseek)
1233                                 return -EINVAL;
1234                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1235                                 return -EFAULT;
1236                         off = &offset;
1237                 } else
1238                         off = &in->f_pos;
1239
1240                 ret = do_splice_to(in, off, pipe, len, flags);
1241
1242                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1243                         ret = -EFAULT;
1244
1245                 return ret;
1246         }
1247
1248         return -EINVAL;
1249 }
1250
1251 /*
1252  * Map an iov into an array of pages and offset/length tupples. With the
1253  * partial_page structure, we can map several non-contiguous ranges into
1254  * our ones pages[] map instead of splitting that operation into pieces.
1255  * Could easily be exported as a generic helper for other users, in which
1256  * case one would probably want to add a 'max_nr_pages' parameter as well.
1257  */
1258 static int get_iovec_page_array(const struct iovec __user *iov,
1259                                 unsigned int nr_vecs, struct page **pages,
1260                                 struct partial_page *partial, int aligned)
1261 {
1262         int buffers = 0, error = 0;
1263
1264         while (nr_vecs) {
1265                 unsigned long off, npages;
1266                 struct iovec entry;
1267                 void __user *base;
1268                 size_t len;
1269                 int i;
1270
1271                 error = -EFAULT;
1272                 if (copy_from_user(&entry, iov, sizeof(entry)))
1273                         break;
1274
1275                 base = entry.iov_base;
1276                 len = entry.iov_len;
1277
1278                 /*
1279                  * Sanity check this iovec. 0 read succeeds.
1280                  */
1281                 error = 0;
1282                 if (unlikely(!len))
1283                         break;
1284                 error = -EFAULT;
1285                 if (!access_ok(VERIFY_READ, base, len))
1286                         break;
1287
1288                 /*
1289                  * Get this base offset and number of pages, then map
1290                  * in the user pages.
1291                  */
1292                 off = (unsigned long) base & ~PAGE_MASK;
1293
1294                 /*
1295                  * If asked for alignment, the offset must be zero and the
1296                  * length a multiple of the PAGE_SIZE.
1297                  */
1298                 error = -EINVAL;
1299                 if (aligned && (off || len & ~PAGE_MASK))
1300                         break;
1301
1302                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1303                 if (npages > PIPE_BUFFERS - buffers)
1304                         npages = PIPE_BUFFERS - buffers;
1305
1306                 error = get_user_pages_fast((unsigned long)base, npages,
1307                                         0, &pages[buffers]);
1308
1309                 if (unlikely(error <= 0))
1310                         break;
1311
1312                 /*
1313                  * Fill this contiguous range into the partial page map.
1314                  */
1315                 for (i = 0; i < error; i++) {
1316                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1317
1318                         partial[buffers].offset = off;
1319                         partial[buffers].len = plen;
1320
1321                         off = 0;
1322                         len -= plen;
1323                         buffers++;
1324                 }
1325
1326                 /*
1327                  * We didn't complete this iov, stop here since it probably
1328                  * means we have to move some of this into a pipe to
1329                  * be able to continue.
1330                  */
1331                 if (len)
1332                         break;
1333
1334                 /*
1335                  * Don't continue if we mapped fewer pages than we asked for,
1336                  * or if we mapped the max number of pages that we have
1337                  * room for.
1338                  */
1339                 if (error < npages || buffers == PIPE_BUFFERS)
1340                         break;
1341
1342                 nr_vecs--;
1343                 iov++;
1344         }
1345
1346         if (buffers)
1347                 return buffers;
1348
1349         return error;
1350 }
1351
1352 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1353                         struct splice_desc *sd)
1354 {
1355         char *src;
1356         int ret;
1357
1358         ret = buf->ops->confirm(pipe, buf);
1359         if (unlikely(ret))
1360                 return ret;
1361
1362         /*
1363          * See if we can use the atomic maps, by prefaulting in the
1364          * pages and doing an atomic copy
1365          */
1366         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1367                 src = buf->ops->map(pipe, buf, 1);
1368                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1369                                                         sd->len);
1370                 buf->ops->unmap(pipe, buf, src);
1371                 if (!ret) {
1372                         ret = sd->len;
1373                         goto out;
1374                 }
1375         }
1376
1377         /*
1378          * No dice, use slow non-atomic map and copy
1379          */
1380         src = buf->ops->map(pipe, buf, 0);
1381
1382         ret = sd->len;
1383         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1384                 ret = -EFAULT;
1385
1386         buf->ops->unmap(pipe, buf, src);
1387 out:
1388         if (ret > 0)
1389                 sd->u.userptr += ret;
1390         return ret;
1391 }
1392
1393 /*
1394  * For lack of a better implementation, implement vmsplice() to userspace
1395  * as a simple copy of the pipes pages to the user iov.
1396  */
1397 static long vmsplice_to_user(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 splice_desc sd;
1402         ssize_t size;
1403         int error;
1404         long ret;
1405
1406         pipe = pipe_info(file->f_path.dentry->d_inode);
1407         if (!pipe)
1408                 return -EBADF;
1409
1410         if (pipe->inode)
1411                 mutex_lock(&pipe->inode->i_mutex);
1412
1413         error = ret = 0;
1414         while (nr_segs) {
1415                 void __user *base;
1416                 size_t len;
1417
1418                 /*
1419                  * Get user address base and length for this iovec.
1420                  */
1421                 error = get_user(base, &iov->iov_base);
1422                 if (unlikely(error))
1423                         break;
1424                 error = get_user(len, &iov->iov_len);
1425                 if (unlikely(error))
1426                         break;
1427
1428                 /*
1429                  * Sanity check this iovec. 0 read succeeds.
1430                  */
1431                 if (unlikely(!len))
1432                         break;
1433                 if (unlikely(!base)) {
1434                         error = -EFAULT;
1435                         break;
1436                 }
1437
1438                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1439                         error = -EFAULT;
1440                         break;
1441                 }
1442
1443                 sd.len = 0;
1444                 sd.total_len = len;
1445                 sd.flags = flags;
1446                 sd.u.userptr = base;
1447                 sd.pos = 0;
1448
1449                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1450                 if (size < 0) {
1451                         if (!ret)
1452                                 ret = size;
1453
1454                         break;
1455                 }
1456
1457                 ret += size;
1458
1459                 if (size < len)
1460                         break;
1461
1462                 nr_segs--;
1463                 iov++;
1464         }
1465
1466         if (pipe->inode)
1467                 mutex_unlock(&pipe->inode->i_mutex);
1468
1469         if (!ret)
1470                 ret = error;
1471
1472         return ret;
1473 }
1474
1475 /*
1476  * vmsplice splices a user address range into a pipe. It can be thought of
1477  * as splice-from-memory, where the regular splice is splice-from-file (or
1478  * to file). In both cases the output is a pipe, naturally.
1479  */
1480 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1481                              unsigned long nr_segs, unsigned int flags)
1482 {
1483         struct pipe_inode_info *pipe;
1484         struct page *pages[PIPE_BUFFERS];
1485         struct partial_page partial[PIPE_BUFFERS];
1486         struct splice_pipe_desc spd = {
1487                 .pages = pages,
1488                 .partial = partial,
1489                 .flags = flags,
1490                 .ops = &user_page_pipe_buf_ops,
1491                 .spd_release = spd_release_page,
1492         };
1493
1494         pipe = pipe_info(file->f_path.dentry->d_inode);
1495         if (!pipe)
1496                 return -EBADF;
1497
1498         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1499                                             flags & SPLICE_F_GIFT);
1500         if (spd.nr_pages <= 0)
1501                 return spd.nr_pages;
1502
1503         return splice_to_pipe(pipe, &spd);
1504 }
1505
1506 /*
1507  * Note that vmsplice only really supports true splicing _from_ user memory
1508  * to a pipe, not the other way around. Splicing from user memory is a simple
1509  * operation that can be supported without any funky alignment restrictions
1510  * or nasty vm tricks. We simply map in the user memory and fill them into
1511  * a pipe. The reverse isn't quite as easy, though. There are two possible
1512  * solutions for that:
1513  *
1514  *      - memcpy() the data internally, at which point we might as well just
1515  *        do a regular read() on the buffer anyway.
1516  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1517  *        has restriction limitations on both ends of the pipe).
1518  *
1519  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1520  *
1521  */
1522 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1523                 unsigned long, nr_segs, unsigned int, flags)
1524 {
1525         struct file *file;
1526         long error;
1527         int fput;
1528
1529         if (unlikely(nr_segs > UIO_MAXIOV))
1530                 return -EINVAL;
1531         else if (unlikely(!nr_segs))
1532                 return 0;
1533
1534         error = -EBADF;
1535         file = fget_light(fd, &fput);
1536         if (file) {
1537                 if (file->f_mode & FMODE_WRITE)
1538                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1539                 else if (file->f_mode & FMODE_READ)
1540                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1541
1542                 fput_light(file, fput);
1543         }
1544
1545         return error;
1546 }
1547
1548 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1549                 int, fd_out, loff_t __user *, off_out,
1550                 size_t, len, unsigned int, flags)
1551 {
1552         long error;
1553         struct file *in, *out;
1554         int fput_in, fput_out;
1555
1556         if (unlikely(!len))
1557                 return 0;
1558
1559         error = -EBADF;
1560         in = fget_light(fd_in, &fput_in);
1561         if (in) {
1562                 if (in->f_mode & FMODE_READ) {
1563                         out = fget_light(fd_out, &fput_out);
1564                         if (out) {
1565                                 if (out->f_mode & FMODE_WRITE)
1566                                         error = do_splice(in, off_in,
1567                                                           out, off_out,
1568                                                           len, flags);
1569                                 fput_light(out, fput_out);
1570                         }
1571                 }
1572
1573                 fput_light(in, fput_in);
1574         }
1575
1576         return error;
1577 }
1578
1579 /*
1580  * Make sure there's data to read. Wait for input if we can, otherwise
1581  * return an appropriate error.
1582  */
1583 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1584 {
1585         int ret;
1586
1587         /*
1588          * Check ->nrbufs without the inode lock first. This function
1589          * is speculative anyways, so missing one is ok.
1590          */
1591         if (pipe->nrbufs)
1592                 return 0;
1593
1594         ret = 0;
1595         mutex_lock(&pipe->inode->i_mutex);
1596
1597         while (!pipe->nrbufs) {
1598                 if (signal_pending(current)) {
1599                         ret = -ERESTARTSYS;
1600                         break;
1601                 }
1602                 if (!pipe->writers)
1603                         break;
1604                 if (!pipe->waiting_writers) {
1605                         if (flags & SPLICE_F_NONBLOCK) {
1606                                 ret = -EAGAIN;
1607                                 break;
1608                         }
1609                 }
1610                 pipe_wait(pipe);
1611         }
1612
1613         mutex_unlock(&pipe->inode->i_mutex);
1614         return ret;
1615 }
1616
1617 /*
1618  * Make sure there's writeable room. Wait for room if we can, otherwise
1619  * return an appropriate error.
1620  */
1621 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1622 {
1623         int ret;
1624
1625         /*
1626          * Check ->nrbufs without the inode lock first. This function
1627          * is speculative anyways, so missing one is ok.
1628          */
1629         if (pipe->nrbufs < PIPE_BUFFERS)
1630                 return 0;
1631
1632         ret = 0;
1633         mutex_lock(&pipe->inode->i_mutex);
1634
1635         while (pipe->nrbufs >= PIPE_BUFFERS) {
1636                 if (!pipe->readers) {
1637                         send_sig(SIGPIPE, current, 0);
1638                         ret = -EPIPE;
1639                         break;
1640                 }
1641                 if (flags & SPLICE_F_NONBLOCK) {
1642                         ret = -EAGAIN;
1643                         break;
1644                 }
1645                 if (signal_pending(current)) {
1646                         ret = -ERESTARTSYS;
1647                         break;
1648                 }
1649                 pipe->waiting_writers++;
1650                 pipe_wait(pipe);
1651                 pipe->waiting_writers--;
1652         }
1653
1654         mutex_unlock(&pipe->inode->i_mutex);
1655         return ret;
1656 }
1657
1658 /*
1659  * Link contents of ipipe to opipe.
1660  */
1661 static int link_pipe(struct pipe_inode_info *ipipe,
1662                      struct pipe_inode_info *opipe,
1663                      size_t len, unsigned int flags)
1664 {
1665         struct pipe_buffer *ibuf, *obuf;
1666         int ret = 0, i = 0, nbuf;
1667
1668         /*
1669          * Potential ABBA deadlock, work around it by ordering lock
1670          * grabbing by inode address. Otherwise two different processes
1671          * could deadlock (one doing tee from A -> B, the other from B -> A).
1672          */
1673         inode_double_lock(ipipe->inode, opipe->inode);
1674
1675         do {
1676                 if (!opipe->readers) {
1677                         send_sig(SIGPIPE, current, 0);
1678                         if (!ret)
1679                                 ret = -EPIPE;
1680                         break;
1681                 }
1682
1683                 /*
1684                  * If we have iterated all input buffers or ran out of
1685                  * output room, break.
1686                  */
1687                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1688                         break;
1689
1690                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1691                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1692
1693                 /*
1694                  * Get a reference to this pipe buffer,
1695                  * so we can copy the contents over.
1696                  */
1697                 ibuf->ops->get(ipipe, ibuf);
1698
1699                 obuf = opipe->bufs + nbuf;
1700                 *obuf = *ibuf;
1701
1702                 /*
1703                  * Don't inherit the gift flag, we need to
1704                  * prevent multiple steals of this page.
1705                  */
1706                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1707
1708                 if (obuf->len > len)
1709                         obuf->len = len;
1710
1711                 opipe->nrbufs++;
1712                 ret += obuf->len;
1713                 len -= obuf->len;
1714                 i++;
1715         } while (len);
1716
1717         /*
1718          * return EAGAIN if we have the potential of some data in the
1719          * future, otherwise just return 0
1720          */
1721         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1722                 ret = -EAGAIN;
1723
1724         inode_double_unlock(ipipe->inode, opipe->inode);
1725
1726         /*
1727          * If we put data in the output pipe, wakeup any potential readers.
1728          */
1729         if (ret > 0) {
1730                 smp_mb();
1731                 if (waitqueue_active(&opipe->wait))
1732                         wake_up_interruptible(&opipe->wait);
1733                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1734         }
1735
1736         return ret;
1737 }
1738
1739 /*
1740  * This is a tee(1) implementation that works on pipes. It doesn't copy
1741  * any data, it simply references the 'in' pages on the 'out' pipe.
1742  * The 'flags' used are the SPLICE_F_* variants, currently the only
1743  * applicable one is SPLICE_F_NONBLOCK.
1744  */
1745 static long do_tee(struct file *in, struct file *out, size_t len,
1746                    unsigned int flags)
1747 {
1748         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1749         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1750         int ret = -EINVAL;
1751
1752         /*
1753          * Duplicate the contents of ipipe to opipe without actually
1754          * copying the data.
1755          */
1756         if (ipipe && opipe && ipipe != opipe) {
1757                 /*
1758                  * Keep going, unless we encounter an error. The ipipe/opipe
1759                  * ordering doesn't really matter.
1760                  */
1761                 ret = link_ipipe_prep(ipipe, flags);
1762                 if (!ret) {
1763                         ret = link_opipe_prep(opipe, flags);
1764                         if (!ret)
1765                                 ret = link_pipe(ipipe, opipe, len, flags);
1766                 }
1767         }
1768
1769         return ret;
1770 }
1771
1772 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1773 {
1774         struct file *in;
1775         int error, fput_in;
1776
1777         if (unlikely(!len))
1778                 return 0;
1779
1780         error = -EBADF;
1781         in = fget_light(fdin, &fput_in);
1782         if (in) {
1783                 if (in->f_mode & FMODE_READ) {
1784                         int fput_out;
1785                         struct file *out = fget_light(fdout, &fput_out);
1786
1787                         if (out) {
1788                                 if (out->f_mode & FMODE_WRITE)
1789                                         error = do_tee(in, out, len, flags);
1790                                 fput_light(out, fput_out);
1791                         }
1792                 }
1793                 fput_light(in, fput_in);
1794         }
1795
1796         return error;
1797 }