Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/sparc-2.6
[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         pipe_lock(pipe);
186
187         for (;;) {
188                 if (!pipe->readers) {
189                         send_sig(SIGPIPE, current, 0);
190                         if (!ret)
191                                 ret = -EPIPE;
192                         break;
193                 }
194
195                 if (pipe->nrbufs < PIPE_BUFFERS) {
196                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
197                         struct pipe_buffer *buf = pipe->bufs + newbuf;
198
199                         buf->page = spd->pages[page_nr];
200                         buf->offset = spd->partial[page_nr].offset;
201                         buf->len = spd->partial[page_nr].len;
202                         buf->private = spd->partial[page_nr].private;
203                         buf->ops = spd->ops;
204                         if (spd->flags & SPLICE_F_GIFT)
205                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
206
207                         pipe->nrbufs++;
208                         page_nr++;
209                         ret += buf->len;
210
211                         if (pipe->inode)
212                                 do_wakeup = 1;
213
214                         if (!--spd->nr_pages)
215                                 break;
216                         if (pipe->nrbufs < PIPE_BUFFERS)
217                                 continue;
218
219                         break;
220                 }
221
222                 if (spd->flags & SPLICE_F_NONBLOCK) {
223                         if (!ret)
224                                 ret = -EAGAIN;
225                         break;
226                 }
227
228                 if (signal_pending(current)) {
229                         if (!ret)
230                                 ret = -ERESTARTSYS;
231                         break;
232                 }
233
234                 if (do_wakeup) {
235                         smp_mb();
236                         if (waitqueue_active(&pipe->wait))
237                                 wake_up_interruptible_sync(&pipe->wait);
238                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
239                         do_wakeup = 0;
240                 }
241
242                 pipe->waiting_writers++;
243                 pipe_wait(pipe);
244                 pipe->waiting_writers--;
245         }
246
247         pipe_unlock(pipe);
248
249         if (do_wakeup) {
250                 smp_mb();
251                 if (waitqueue_active(&pipe->wait))
252                         wake_up_interruptible(&pipe->wait);
253                 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
254         }
255
256         while (page_nr < spd_pages)
257                 spd->spd_release(spd, page_nr++);
258
259         return ret;
260 }
261
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
263 {
264         page_cache_release(spd->pages[i]);
265 }
266
267 static int
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269                            struct pipe_inode_info *pipe, size_t len,
270                            unsigned int flags)
271 {
272         struct address_space *mapping = in->f_mapping;
273         unsigned int loff, nr_pages, req_pages;
274         struct page *pages[PIPE_BUFFERS];
275         struct partial_page partial[PIPE_BUFFERS];
276         struct page *page;
277         pgoff_t index, end_index;
278         loff_t isize;
279         int error, page_nr;
280         struct splice_pipe_desc spd = {
281                 .pages = pages,
282                 .partial = partial,
283                 .flags = flags,
284                 .ops = &page_cache_pipe_buf_ops,
285                 .spd_release = spd_release_page,
286         };
287
288         index = *ppos >> PAGE_CACHE_SHIFT;
289         loff = *ppos & ~PAGE_CACHE_MASK;
290         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
292
293         /*
294          * Lookup the (hopefully) full range of pages we need.
295          */
296         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297         index += spd.nr_pages;
298
299         /*
300          * If find_get_pages_contig() returned fewer pages than we needed,
301          * readahead/allocate the rest and fill in the holes.
302          */
303         if (spd.nr_pages < nr_pages)
304                 page_cache_sync_readahead(mapping, &in->f_ra, in,
305                                 index, req_pages - spd.nr_pages);
306
307         error = 0;
308         while (spd.nr_pages < nr_pages) {
309                 /*
310                  * Page could be there, find_get_pages_contig() breaks on
311                  * the first hole.
312                  */
313                 page = find_get_page(mapping, index);
314                 if (!page) {
315                         /*
316                          * page didn't exist, allocate one.
317                          */
318                         page = page_cache_alloc_cold(mapping);
319                         if (!page)
320                                 break;
321
322                         error = add_to_page_cache_lru(page, mapping, index,
323                                                 mapping_gfp_mask(mapping));
324                         if (unlikely(error)) {
325                                 page_cache_release(page);
326                                 if (error == -EEXIST)
327                                         continue;
328                                 break;
329                         }
330                         /*
331                          * add_to_page_cache() locks the page, unlock it
332                          * to avoid convoluting the logic below even more.
333                          */
334                         unlock_page(page);
335                 }
336
337                 pages[spd.nr_pages++] = page;
338                 index++;
339         }
340
341         /*
342          * Now loop over the map and see if we need to start IO on any
343          * pages, fill in the partial map, etc.
344          */
345         index = *ppos >> PAGE_CACHE_SHIFT;
346         nr_pages = spd.nr_pages;
347         spd.nr_pages = 0;
348         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349                 unsigned int this_len;
350
351                 if (!len)
352                         break;
353
354                 /*
355                  * this_len is the max we'll use from this page
356                  */
357                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358                 page = pages[page_nr];
359
360                 if (PageReadahead(page))
361                         page_cache_async_readahead(mapping, &in->f_ra, in,
362                                         page, index, req_pages - page_nr);
363
364                 /*
365                  * If the page isn't uptodate, we may need to start io on it
366                  */
367                 if (!PageUptodate(page)) {
368                         /*
369                          * If in nonblock mode then dont block on waiting
370                          * for an in-flight io page
371                          */
372                         if (flags & SPLICE_F_NONBLOCK) {
373                                 if (!trylock_page(page)) {
374                                         error = -EAGAIN;
375                                         break;
376                                 }
377                         } else
378                                 lock_page(page);
379
380                         /*
381                          * Page was truncated, or invalidated by the
382                          * filesystem.  Redo the find/create, but this time the
383                          * page is kept locked, so there's no chance of another
384                          * race with truncate/invalidate.
385                          */
386                         if (!page->mapping) {
387                                 unlock_page(page);
388                                 page = find_or_create_page(mapping, index,
389                                                 mapping_gfp_mask(mapping));
390
391                                 if (!page) {
392                                         error = -ENOMEM;
393                                         break;
394                                 }
395                                 page_cache_release(pages[page_nr]);
396                                 pages[page_nr] = page;
397                         }
398                         /*
399                          * page was already under io and is now done, great
400                          */
401                         if (PageUptodate(page)) {
402                                 unlock_page(page);
403                                 goto fill_it;
404                         }
405
406                         /*
407                          * need to read in the page
408                          */
409                         error = mapping->a_ops->readpage(in, page);
410                         if (unlikely(error)) {
411                                 /*
412                                  * We really should re-lookup the page here,
413                                  * but it complicates things a lot. Instead
414                                  * lets just do what we already stored, and
415                                  * we'll get it the next time we are called.
416                                  */
417                                 if (error == AOP_TRUNCATED_PAGE)
418                                         error = 0;
419
420                                 break;
421                         }
422                 }
423 fill_it:
424                 /*
425                  * i_size must be checked after PageUptodate.
426                  */
427                 isize = i_size_read(mapping->host);
428                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
429                 if (unlikely(!isize || index > end_index))
430                         break;
431
432                 /*
433                  * if this is the last page, see if we need to shrink
434                  * the length and stop
435                  */
436                 if (end_index == index) {
437                         unsigned int plen;
438
439                         /*
440                          * max good bytes in this page
441                          */
442                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
443                         if (plen <= loff)
444                                 break;
445
446                         /*
447                          * force quit after adding this page
448                          */
449                         this_len = min(this_len, plen - loff);
450                         len = this_len;
451                 }
452
453                 partial[page_nr].offset = loff;
454                 partial[page_nr].len = this_len;
455                 len -= this_len;
456                 loff = 0;
457                 spd.nr_pages++;
458                 index++;
459         }
460
461         /*
462          * Release any pages at the end, if we quit early. 'page_nr' is how far
463          * we got, 'nr_pages' is how many pages are in the map.
464          */
465         while (page_nr < nr_pages)
466                 page_cache_release(pages[page_nr++]);
467         in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
468
469         if (spd.nr_pages)
470                 return splice_to_pipe(pipe, &spd);
471
472         return error;
473 }
474
475 /**
476  * generic_file_splice_read - splice data from file to a pipe
477  * @in:         file to splice from
478  * @ppos:       position in @in
479  * @pipe:       pipe to splice to
480  * @len:        number of bytes to splice
481  * @flags:      splice modifier flags
482  *
483  * Description:
484  *    Will read pages from given file and fill them into a pipe. Can be
485  *    used as long as the address_space operations for the source implements
486  *    a readpage() hook.
487  *
488  */
489 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
490                                  struct pipe_inode_info *pipe, size_t len,
491                                  unsigned int flags)
492 {
493         loff_t isize, left;
494         int ret;
495
496         isize = i_size_read(in->f_mapping->host);
497         if (unlikely(*ppos >= isize))
498                 return 0;
499
500         left = isize - *ppos;
501         if (unlikely(left < len))
502                 len = left;
503
504         ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
505         if (ret > 0) {
506                 *ppos += ret;
507                 file_accessed(in);
508         }
509
510         return ret;
511 }
512 EXPORT_SYMBOL(generic_file_splice_read);
513
514 static const struct pipe_buf_operations default_pipe_buf_ops = {
515         .can_merge = 0,
516         .map = generic_pipe_buf_map,
517         .unmap = generic_pipe_buf_unmap,
518         .confirm = generic_pipe_buf_confirm,
519         .release = generic_pipe_buf_release,
520         .steal = generic_pipe_buf_steal,
521         .get = generic_pipe_buf_get,
522 };
523
524 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
525                             unsigned long vlen, loff_t offset)
526 {
527         mm_segment_t old_fs;
528         loff_t pos = offset;
529         ssize_t res;
530
531         old_fs = get_fs();
532         set_fs(get_ds());
533         /* The cast to a user pointer is valid due to the set_fs() */
534         res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
535         set_fs(old_fs);
536
537         return res;
538 }
539
540 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
541                             loff_t pos)
542 {
543         mm_segment_t old_fs;
544         ssize_t res;
545
546         old_fs = get_fs();
547         set_fs(get_ds());
548         /* The cast to a user pointer is valid due to the set_fs() */
549         res = vfs_write(file, (const char __user *)buf, count, &pos);
550         set_fs(old_fs);
551
552         return res;
553 }
554
555 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
556                                  struct pipe_inode_info *pipe, size_t len,
557                                  unsigned int flags)
558 {
559         unsigned int nr_pages;
560         unsigned int nr_freed;
561         size_t offset;
562         struct page *pages[PIPE_BUFFERS];
563         struct partial_page partial[PIPE_BUFFERS];
564         struct iovec vec[PIPE_BUFFERS];
565         pgoff_t index;
566         ssize_t res;
567         size_t this_len;
568         int error;
569         int i;
570         struct splice_pipe_desc spd = {
571                 .pages = pages,
572                 .partial = partial,
573                 .flags = flags,
574                 .ops = &default_pipe_buf_ops,
575                 .spd_release = spd_release_page,
576         };
577
578         index = *ppos >> PAGE_CACHE_SHIFT;
579         offset = *ppos & ~PAGE_CACHE_MASK;
580         nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
581
582         for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
583                 struct page *page;
584
585                 page = alloc_page(GFP_USER);
586                 error = -ENOMEM;
587                 if (!page)
588                         goto err;
589
590                 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
591                 vec[i].iov_base = (void __user *) page_address(page);
592                 vec[i].iov_len = this_len;
593                 pages[i] = page;
594                 spd.nr_pages++;
595                 len -= this_len;
596                 offset = 0;
597         }
598
599         res = kernel_readv(in, vec, spd.nr_pages, *ppos);
600         if (res < 0) {
601                 error = res;
602                 goto err;
603         }
604
605         error = 0;
606         if (!res)
607                 goto err;
608
609         nr_freed = 0;
610         for (i = 0; i < spd.nr_pages; i++) {
611                 this_len = min_t(size_t, vec[i].iov_len, res);
612                 partial[i].offset = 0;
613                 partial[i].len = this_len;
614                 if (!this_len) {
615                         __free_page(pages[i]);
616                         pages[i] = NULL;
617                         nr_freed++;
618                 }
619                 res -= this_len;
620         }
621         spd.nr_pages -= nr_freed;
622
623         res = splice_to_pipe(pipe, &spd);
624         if (res > 0)
625                 *ppos += res;
626
627         return res;
628
629 err:
630         for (i = 0; i < spd.nr_pages; i++)
631                 __free_page(pages[i]);
632
633         return error;
634 }
635 EXPORT_SYMBOL(default_file_splice_read);
636
637 /*
638  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
639  * using sendpage(). Return the number of bytes sent.
640  */
641 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
642                             struct pipe_buffer *buf, struct splice_desc *sd)
643 {
644         struct file *file = sd->u.file;
645         loff_t pos = sd->pos;
646         int ret, more;
647
648         ret = buf->ops->confirm(pipe, buf);
649         if (!ret) {
650                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
651                 if (file->f_op && file->f_op->sendpage)
652                         ret = file->f_op->sendpage(file, buf->page, buf->offset,
653                                                    sd->len, &pos, more);
654                 else
655                         ret = -EINVAL;
656         }
657
658         return ret;
659 }
660
661 /*
662  * This is a little more tricky than the file -> pipe splicing. There are
663  * basically three cases:
664  *
665  *      - Destination page already exists in the address space and there
666  *        are users of it. For that case we have no other option that
667  *        copying the data. Tough luck.
668  *      - Destination page already exists in the address space, but there
669  *        are no users of it. Make sure it's uptodate, then drop it. Fall
670  *        through to last case.
671  *      - Destination page does not exist, we can add the pipe page to
672  *        the page cache and avoid the copy.
673  *
674  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
675  * sd->flags), we attempt to migrate pages from the pipe to the output
676  * file address space page cache. This is possible if no one else has
677  * the pipe page referenced outside of the pipe and page cache. If
678  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
679  * a new page in the output file page cache and fill/dirty that.
680  */
681 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
682                  struct splice_desc *sd)
683 {
684         struct file *file = sd->u.file;
685         struct address_space *mapping = file->f_mapping;
686         unsigned int offset, this_len;
687         struct page *page;
688         void *fsdata;
689         int ret;
690
691         /*
692          * make sure the data in this buffer is uptodate
693          */
694         ret = buf->ops->confirm(pipe, buf);
695         if (unlikely(ret))
696                 return ret;
697
698         offset = sd->pos & ~PAGE_CACHE_MASK;
699
700         this_len = sd->len;
701         if (this_len + offset > PAGE_CACHE_SIZE)
702                 this_len = PAGE_CACHE_SIZE - offset;
703
704         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
705                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
706         if (unlikely(ret))
707                 goto out;
708
709         if (buf->page != page) {
710                 /*
711                  * Careful, ->map() uses KM_USER0!
712                  */
713                 char *src = buf->ops->map(pipe, buf, 1);
714                 char *dst = kmap_atomic(page, KM_USER1);
715
716                 memcpy(dst + offset, src + buf->offset, this_len);
717                 flush_dcache_page(page);
718                 kunmap_atomic(dst, KM_USER1);
719                 buf->ops->unmap(pipe, buf, src);
720         }
721         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
722                                 page, fsdata);
723 out:
724         return ret;
725 }
726 EXPORT_SYMBOL(pipe_to_file);
727
728 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
729 {
730         smp_mb();
731         if (waitqueue_active(&pipe->wait))
732                 wake_up_interruptible(&pipe->wait);
733         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
734 }
735
736 /**
737  * splice_from_pipe_feed - feed available data from a pipe to a file
738  * @pipe:       pipe to splice from
739  * @sd:         information to @actor
740  * @actor:      handler that splices the data
741  *
742  * Description:
743  *    This function loops over the pipe and calls @actor to do the
744  *    actual moving of a single struct pipe_buffer to the desired
745  *    destination.  It returns when there's no more buffers left in
746  *    the pipe or if the requested number of bytes (@sd->total_len)
747  *    have been copied.  It returns a positive number (one) if the
748  *    pipe needs to be filled with more data, zero if the required
749  *    number of bytes have been copied and -errno on error.
750  *
751  *    This, together with splice_from_pipe_{begin,end,next}, may be
752  *    used to implement the functionality of __splice_from_pipe() when
753  *    locking is required around copying the pipe buffers to the
754  *    destination.
755  */
756 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
757                           splice_actor *actor)
758 {
759         int ret;
760
761         while (pipe->nrbufs) {
762                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
763                 const struct pipe_buf_operations *ops = buf->ops;
764
765                 sd->len = buf->len;
766                 if (sd->len > sd->total_len)
767                         sd->len = sd->total_len;
768
769                 ret = actor(pipe, buf, sd);
770                 if (ret <= 0) {
771                         if (ret == -ENODATA)
772                                 ret = 0;
773                         return ret;
774                 }
775                 buf->offset += ret;
776                 buf->len -= ret;
777
778                 sd->num_spliced += ret;
779                 sd->len -= ret;
780                 sd->pos += ret;
781                 sd->total_len -= ret;
782
783                 if (!buf->len) {
784                         buf->ops = NULL;
785                         ops->release(pipe, buf);
786                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
787                         pipe->nrbufs--;
788                         if (pipe->inode)
789                                 sd->need_wakeup = true;
790                 }
791
792                 if (!sd->total_len)
793                         return 0;
794         }
795
796         return 1;
797 }
798 EXPORT_SYMBOL(splice_from_pipe_feed);
799
800 /**
801  * splice_from_pipe_next - wait for some data to splice from
802  * @pipe:       pipe to splice from
803  * @sd:         information about the splice operation
804  *
805  * Description:
806  *    This function will wait for some data and return a positive
807  *    value (one) if pipe buffers are available.  It will return zero
808  *    or -errno if no more data needs to be spliced.
809  */
810 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
811 {
812         while (!pipe->nrbufs) {
813                 if (!pipe->writers)
814                         return 0;
815
816                 if (!pipe->waiting_writers && sd->num_spliced)
817                         return 0;
818
819                 if (sd->flags & SPLICE_F_NONBLOCK)
820                         return -EAGAIN;
821
822                 if (signal_pending(current))
823                         return -ERESTARTSYS;
824
825                 if (sd->need_wakeup) {
826                         wakeup_pipe_writers(pipe);
827                         sd->need_wakeup = false;
828                 }
829
830                 pipe_wait(pipe);
831         }
832
833         return 1;
834 }
835 EXPORT_SYMBOL(splice_from_pipe_next);
836
837 /**
838  * splice_from_pipe_begin - start splicing from pipe
839  * @sd:         information about the splice operation
840  *
841  * Description:
842  *    This function should be called before a loop containing
843  *    splice_from_pipe_next() and splice_from_pipe_feed() to
844  *    initialize the necessary fields of @sd.
845  */
846 void splice_from_pipe_begin(struct splice_desc *sd)
847 {
848         sd->num_spliced = 0;
849         sd->need_wakeup = false;
850 }
851 EXPORT_SYMBOL(splice_from_pipe_begin);
852
853 /**
854  * splice_from_pipe_end - finish splicing from pipe
855  * @pipe:       pipe to splice from
856  * @sd:         information about the splice operation
857  *
858  * Description:
859  *    This function will wake up pipe writers if necessary.  It should
860  *    be called after a loop containing splice_from_pipe_next() and
861  *    splice_from_pipe_feed().
862  */
863 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
864 {
865         if (sd->need_wakeup)
866                 wakeup_pipe_writers(pipe);
867 }
868 EXPORT_SYMBOL(splice_from_pipe_end);
869
870 /**
871  * __splice_from_pipe - splice data from a pipe to given actor
872  * @pipe:       pipe to splice from
873  * @sd:         information to @actor
874  * @actor:      handler that splices the data
875  *
876  * Description:
877  *    This function does little more than loop over the pipe and call
878  *    @actor to do the actual moving of a single struct pipe_buffer to
879  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
880  *    pipe_to_user.
881  *
882  */
883 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
884                            splice_actor *actor)
885 {
886         int ret;
887
888         splice_from_pipe_begin(sd);
889         do {
890                 ret = splice_from_pipe_next(pipe, sd);
891                 if (ret > 0)
892                         ret = splice_from_pipe_feed(pipe, sd, actor);
893         } while (ret > 0);
894         splice_from_pipe_end(pipe, sd);
895
896         return sd->num_spliced ? sd->num_spliced : ret;
897 }
898 EXPORT_SYMBOL(__splice_from_pipe);
899
900 /**
901  * splice_from_pipe - splice data from a pipe to a file
902  * @pipe:       pipe to splice from
903  * @out:        file to splice to
904  * @ppos:       position in @out
905  * @len:        how many bytes to splice
906  * @flags:      splice modifier flags
907  * @actor:      handler that splices the data
908  *
909  * Description:
910  *    See __splice_from_pipe. This function locks the pipe inode,
911  *    otherwise it's identical to __splice_from_pipe().
912  *
913  */
914 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
915                          loff_t *ppos, size_t len, unsigned int flags,
916                          splice_actor *actor)
917 {
918         ssize_t ret;
919         struct splice_desc sd = {
920                 .total_len = len,
921                 .flags = flags,
922                 .pos = *ppos,
923                 .u.file = out,
924         };
925
926         pipe_lock(pipe);
927         ret = __splice_from_pipe(pipe, &sd, actor);
928         pipe_unlock(pipe);
929
930         return ret;
931 }
932
933 /**
934  * generic_file_splice_write - splice data from a pipe to a file
935  * @pipe:       pipe info
936  * @out:        file to write to
937  * @ppos:       position in @out
938  * @len:        number of bytes to splice
939  * @flags:      splice modifier flags
940  *
941  * Description:
942  *    Will either move or copy pages (determined by @flags options) from
943  *    the given pipe inode to the given file.
944  *
945  */
946 ssize_t
947 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
948                           loff_t *ppos, size_t len, unsigned int flags)
949 {
950         struct address_space *mapping = out->f_mapping;
951         struct inode *inode = mapping->host;
952         struct splice_desc sd = {
953                 .total_len = len,
954                 .flags = flags,
955                 .pos = *ppos,
956                 .u.file = out,
957         };
958         ssize_t ret;
959
960         pipe_lock(pipe);
961
962         splice_from_pipe_begin(&sd);
963         do {
964                 ret = splice_from_pipe_next(pipe, &sd);
965                 if (ret <= 0)
966                         break;
967
968                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
969                 ret = file_remove_suid(out);
970                 if (!ret) {
971                         file_update_time(out);
972                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
973                 }
974                 mutex_unlock(&inode->i_mutex);
975         } while (ret > 0);
976         splice_from_pipe_end(pipe, &sd);
977
978         pipe_unlock(pipe);
979
980         if (sd.num_spliced)
981                 ret = sd.num_spliced;
982
983         if (ret > 0) {
984                 unsigned long nr_pages;
985                 int err;
986
987                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
988
989                 err = generic_write_sync(out, *ppos, ret);
990                 if (err)
991                         ret = err;
992                 else
993                         *ppos += ret;
994                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
995         }
996
997         return ret;
998 }
999
1000 EXPORT_SYMBOL(generic_file_splice_write);
1001
1002 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1003                           struct splice_desc *sd)
1004 {
1005         int ret;
1006         void *data;
1007
1008         ret = buf->ops->confirm(pipe, buf);
1009         if (ret)
1010                 return ret;
1011
1012         data = buf->ops->map(pipe, buf, 0);
1013         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1014         buf->ops->unmap(pipe, buf, data);
1015
1016         return ret;
1017 }
1018
1019 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1020                                          struct file *out, loff_t *ppos,
1021                                          size_t len, unsigned int flags)
1022 {
1023         ssize_t ret;
1024
1025         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1026         if (ret > 0)
1027                 *ppos += ret;
1028
1029         return ret;
1030 }
1031
1032 /**
1033  * generic_splice_sendpage - splice data from a pipe to a socket
1034  * @pipe:       pipe to splice from
1035  * @out:        socket to write to
1036  * @ppos:       position in @out
1037  * @len:        number of bytes to splice
1038  * @flags:      splice modifier flags
1039  *
1040  * Description:
1041  *    Will send @len bytes from the pipe to a network socket. No data copying
1042  *    is involved.
1043  *
1044  */
1045 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1046                                 loff_t *ppos, size_t len, unsigned int flags)
1047 {
1048         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1049 }
1050
1051 EXPORT_SYMBOL(generic_splice_sendpage);
1052
1053 /*
1054  * Attempt to initiate a splice from pipe to file.
1055  */
1056 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1057                            loff_t *ppos, size_t len, unsigned int flags)
1058 {
1059         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1060                                 loff_t *, size_t, unsigned int);
1061         int ret;
1062
1063         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1064                 return -EBADF;
1065
1066         if (unlikely(out->f_flags & O_APPEND))
1067                 return -EINVAL;
1068
1069         ret = rw_verify_area(WRITE, out, ppos, len);
1070         if (unlikely(ret < 0))
1071                 return ret;
1072
1073         if (out->f_op && out->f_op->splice_write)
1074                 splice_write = out->f_op->splice_write;
1075         else
1076                 splice_write = default_file_splice_write;
1077
1078         return splice_write(pipe, out, ppos, len, flags);
1079 }
1080
1081 /*
1082  * Attempt to initiate a splice from a file to a pipe.
1083  */
1084 static long do_splice_to(struct file *in, loff_t *ppos,
1085                          struct pipe_inode_info *pipe, size_t len,
1086                          unsigned int flags)
1087 {
1088         ssize_t (*splice_read)(struct file *, loff_t *,
1089                                struct pipe_inode_info *, size_t, unsigned int);
1090         int ret;
1091
1092         if (unlikely(!(in->f_mode & FMODE_READ)))
1093                 return -EBADF;
1094
1095         ret = rw_verify_area(READ, in, ppos, len);
1096         if (unlikely(ret < 0))
1097                 return ret;
1098
1099         if (in->f_op && in->f_op->splice_read)
1100                 splice_read = in->f_op->splice_read;
1101         else
1102                 splice_read = default_file_splice_read;
1103
1104         return splice_read(in, ppos, pipe, len, flags);
1105 }
1106
1107 /**
1108  * splice_direct_to_actor - splices data directly between two non-pipes
1109  * @in:         file to splice from
1110  * @sd:         actor information on where to splice to
1111  * @actor:      handles the data splicing
1112  *
1113  * Description:
1114  *    This is a special case helper to splice directly between two
1115  *    points, without requiring an explicit pipe. Internally an allocated
1116  *    pipe is cached in the process, and reused during the lifetime of
1117  *    that process.
1118  *
1119  */
1120 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1121                                splice_direct_actor *actor)
1122 {
1123         struct pipe_inode_info *pipe;
1124         long ret, bytes;
1125         umode_t i_mode;
1126         size_t len;
1127         int i, flags;
1128
1129         /*
1130          * We require the input being a regular file, as we don't want to
1131          * randomly drop data for eg socket -> socket splicing. Use the
1132          * piped splicing for that!
1133          */
1134         i_mode = in->f_path.dentry->d_inode->i_mode;
1135         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1136                 return -EINVAL;
1137
1138         /*
1139          * neither in nor out is a pipe, setup an internal pipe attached to
1140          * 'out' and transfer the wanted data from 'in' to 'out' through that
1141          */
1142         pipe = current->splice_pipe;
1143         if (unlikely(!pipe)) {
1144                 pipe = alloc_pipe_info(NULL);
1145                 if (!pipe)
1146                         return -ENOMEM;
1147
1148                 /*
1149                  * We don't have an immediate reader, but we'll read the stuff
1150                  * out of the pipe right after the splice_to_pipe(). So set
1151                  * PIPE_READERS appropriately.
1152                  */
1153                 pipe->readers = 1;
1154
1155                 current->splice_pipe = pipe;
1156         }
1157
1158         /*
1159          * Do the splice.
1160          */
1161         ret = 0;
1162         bytes = 0;
1163         len = sd->total_len;
1164         flags = sd->flags;
1165
1166         /*
1167          * Don't block on output, we have to drain the direct pipe.
1168          */
1169         sd->flags &= ~SPLICE_F_NONBLOCK;
1170
1171         while (len) {
1172                 size_t read_len;
1173                 loff_t pos = sd->pos, prev_pos = pos;
1174
1175                 ret = do_splice_to(in, &pos, pipe, len, flags);
1176                 if (unlikely(ret <= 0))
1177                         goto out_release;
1178
1179                 read_len = ret;
1180                 sd->total_len = read_len;
1181
1182                 /*
1183                  * NOTE: nonblocking mode only applies to the input. We
1184                  * must not do the output in nonblocking mode as then we
1185                  * could get stuck data in the internal pipe:
1186                  */
1187                 ret = actor(pipe, sd);
1188                 if (unlikely(ret <= 0)) {
1189                         sd->pos = prev_pos;
1190                         goto out_release;
1191                 }
1192
1193                 bytes += ret;
1194                 len -= ret;
1195                 sd->pos = pos;
1196
1197                 if (ret < read_len) {
1198                         sd->pos = prev_pos + ret;
1199                         goto out_release;
1200                 }
1201         }
1202
1203 done:
1204         pipe->nrbufs = pipe->curbuf = 0;
1205         file_accessed(in);
1206         return bytes;
1207
1208 out_release:
1209         /*
1210          * If we did an incomplete transfer we must release
1211          * the pipe buffers in question:
1212          */
1213         for (i = 0; i < PIPE_BUFFERS; i++) {
1214                 struct pipe_buffer *buf = pipe->bufs + i;
1215
1216                 if (buf->ops) {
1217                         buf->ops->release(pipe, buf);
1218                         buf->ops = NULL;
1219                 }
1220         }
1221
1222         if (!bytes)
1223                 bytes = ret;
1224
1225         goto done;
1226 }
1227 EXPORT_SYMBOL(splice_direct_to_actor);
1228
1229 static int direct_splice_actor(struct pipe_inode_info *pipe,
1230                                struct splice_desc *sd)
1231 {
1232         struct file *file = sd->u.file;
1233
1234         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1235 }
1236
1237 /**
1238  * do_splice_direct - splices data directly between two files
1239  * @in:         file to splice from
1240  * @ppos:       input file offset
1241  * @out:        file to splice to
1242  * @len:        number of bytes to splice
1243  * @flags:      splice modifier flags
1244  *
1245  * Description:
1246  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1247  *    doing it in the application would incur an extra system call
1248  *    (splice in + splice out, as compared to just sendfile()). So this helper
1249  *    can splice directly through a process-private pipe.
1250  *
1251  */
1252 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1253                       size_t len, unsigned int flags)
1254 {
1255         struct splice_desc sd = {
1256                 .len            = len,
1257                 .total_len      = len,
1258                 .flags          = flags,
1259                 .pos            = *ppos,
1260                 .u.file         = out,
1261         };
1262         long ret;
1263
1264         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1265         if (ret > 0)
1266                 *ppos = sd.pos;
1267
1268         return ret;
1269 }
1270
1271 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1272                                struct pipe_inode_info *opipe,
1273                                size_t len, unsigned int flags);
1274 /*
1275  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1276  * location, so checking ->i_pipe is not enough to verify that this is a
1277  * pipe.
1278  */
1279 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1280 {
1281         if (S_ISFIFO(inode->i_mode))
1282                 return inode->i_pipe;
1283
1284         return NULL;
1285 }
1286
1287 /*
1288  * Determine where to splice to/from.
1289  */
1290 static long do_splice(struct file *in, loff_t __user *off_in,
1291                       struct file *out, loff_t __user *off_out,
1292                       size_t len, unsigned int flags)
1293 {
1294         struct pipe_inode_info *ipipe;
1295         struct pipe_inode_info *opipe;
1296         loff_t offset, *off;
1297         long ret;
1298
1299         ipipe = pipe_info(in->f_path.dentry->d_inode);
1300         opipe = pipe_info(out->f_path.dentry->d_inode);
1301
1302         if (ipipe && opipe) {
1303                 if (off_in || off_out)
1304                         return -ESPIPE;
1305
1306                 if (!(in->f_mode & FMODE_READ))
1307                         return -EBADF;
1308
1309                 if (!(out->f_mode & FMODE_WRITE))
1310                         return -EBADF;
1311
1312                 /* Splicing to self would be fun, but... */
1313                 if (ipipe == opipe)
1314                         return -EINVAL;
1315
1316                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1317         }
1318
1319         if (ipipe) {
1320                 if (off_in)
1321                         return -ESPIPE;
1322                 if (off_out) {
1323                         if (!out->f_op || !out->f_op->llseek ||
1324                             out->f_op->llseek == no_llseek)
1325                                 return -EINVAL;
1326                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1327                                 return -EFAULT;
1328                         off = &offset;
1329                 } else
1330                         off = &out->f_pos;
1331
1332                 ret = do_splice_from(ipipe, out, off, len, flags);
1333
1334                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1335                         ret = -EFAULT;
1336
1337                 return ret;
1338         }
1339
1340         if (opipe) {
1341                 if (off_out)
1342                         return -ESPIPE;
1343                 if (off_in) {
1344                         if (!in->f_op || !in->f_op->llseek ||
1345                             in->f_op->llseek == no_llseek)
1346                                 return -EINVAL;
1347                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1348                                 return -EFAULT;
1349                         off = &offset;
1350                 } else
1351                         off = &in->f_pos;
1352
1353                 ret = do_splice_to(in, off, opipe, len, flags);
1354
1355                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1356                         ret = -EFAULT;
1357
1358                 return ret;
1359         }
1360
1361         return -EINVAL;
1362 }
1363
1364 /*
1365  * Map an iov into an array of pages and offset/length tupples. With the
1366  * partial_page structure, we can map several non-contiguous ranges into
1367  * our ones pages[] map instead of splitting that operation into pieces.
1368  * Could easily be exported as a generic helper for other users, in which
1369  * case one would probably want to add a 'max_nr_pages' parameter as well.
1370  */
1371 static int get_iovec_page_array(const struct iovec __user *iov,
1372                                 unsigned int nr_vecs, struct page **pages,
1373                                 struct partial_page *partial, int aligned)
1374 {
1375         int buffers = 0, error = 0;
1376
1377         while (nr_vecs) {
1378                 unsigned long off, npages;
1379                 struct iovec entry;
1380                 void __user *base;
1381                 size_t len;
1382                 int i;
1383
1384                 error = -EFAULT;
1385                 if (copy_from_user(&entry, iov, sizeof(entry)))
1386                         break;
1387
1388                 base = entry.iov_base;
1389                 len = entry.iov_len;
1390
1391                 /*
1392                  * Sanity check this iovec. 0 read succeeds.
1393                  */
1394                 error = 0;
1395                 if (unlikely(!len))
1396                         break;
1397                 error = -EFAULT;
1398                 if (!access_ok(VERIFY_READ, base, len))
1399                         break;
1400
1401                 /*
1402                  * Get this base offset and number of pages, then map
1403                  * in the user pages.
1404                  */
1405                 off = (unsigned long) base & ~PAGE_MASK;
1406
1407                 /*
1408                  * If asked for alignment, the offset must be zero and the
1409                  * length a multiple of the PAGE_SIZE.
1410                  */
1411                 error = -EINVAL;
1412                 if (aligned && (off || len & ~PAGE_MASK))
1413                         break;
1414
1415                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1416                 if (npages > PIPE_BUFFERS - buffers)
1417                         npages = PIPE_BUFFERS - buffers;
1418
1419                 error = get_user_pages_fast((unsigned long)base, npages,
1420                                         0, &pages[buffers]);
1421
1422                 if (unlikely(error <= 0))
1423                         break;
1424
1425                 /*
1426                  * Fill this contiguous range into the partial page map.
1427                  */
1428                 for (i = 0; i < error; i++) {
1429                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1430
1431                         partial[buffers].offset = off;
1432                         partial[buffers].len = plen;
1433
1434                         off = 0;
1435                         len -= plen;
1436                         buffers++;
1437                 }
1438
1439                 /*
1440                  * We didn't complete this iov, stop here since it probably
1441                  * means we have to move some of this into a pipe to
1442                  * be able to continue.
1443                  */
1444                 if (len)
1445                         break;
1446
1447                 /*
1448                  * Don't continue if we mapped fewer pages than we asked for,
1449                  * or if we mapped the max number of pages that we have
1450                  * room for.
1451                  */
1452                 if (error < npages || buffers == PIPE_BUFFERS)
1453                         break;
1454
1455                 nr_vecs--;
1456                 iov++;
1457         }
1458
1459         if (buffers)
1460                 return buffers;
1461
1462         return error;
1463 }
1464
1465 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1466                         struct splice_desc *sd)
1467 {
1468         char *src;
1469         int ret;
1470
1471         ret = buf->ops->confirm(pipe, buf);
1472         if (unlikely(ret))
1473                 return ret;
1474
1475         /*
1476          * See if we can use the atomic maps, by prefaulting in the
1477          * pages and doing an atomic copy
1478          */
1479         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1480                 src = buf->ops->map(pipe, buf, 1);
1481                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1482                                                         sd->len);
1483                 buf->ops->unmap(pipe, buf, src);
1484                 if (!ret) {
1485                         ret = sd->len;
1486                         goto out;
1487                 }
1488         }
1489
1490         /*
1491          * No dice, use slow non-atomic map and copy
1492          */
1493         src = buf->ops->map(pipe, buf, 0);
1494
1495         ret = sd->len;
1496         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1497                 ret = -EFAULT;
1498
1499         buf->ops->unmap(pipe, buf, src);
1500 out:
1501         if (ret > 0)
1502                 sd->u.userptr += ret;
1503         return ret;
1504 }
1505
1506 /*
1507  * For lack of a better implementation, implement vmsplice() to userspace
1508  * as a simple copy of the pipes pages to the user iov.
1509  */
1510 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1511                              unsigned long nr_segs, unsigned int flags)
1512 {
1513         struct pipe_inode_info *pipe;
1514         struct splice_desc sd;
1515         ssize_t size;
1516         int error;
1517         long ret;
1518
1519         pipe = pipe_info(file->f_path.dentry->d_inode);
1520         if (!pipe)
1521                 return -EBADF;
1522
1523         pipe_lock(pipe);
1524
1525         error = ret = 0;
1526         while (nr_segs) {
1527                 void __user *base;
1528                 size_t len;
1529
1530                 /*
1531                  * Get user address base and length for this iovec.
1532                  */
1533                 error = get_user(base, &iov->iov_base);
1534                 if (unlikely(error))
1535                         break;
1536                 error = get_user(len, &iov->iov_len);
1537                 if (unlikely(error))
1538                         break;
1539
1540                 /*
1541                  * Sanity check this iovec. 0 read succeeds.
1542                  */
1543                 if (unlikely(!len))
1544                         break;
1545                 if (unlikely(!base)) {
1546                         error = -EFAULT;
1547                         break;
1548                 }
1549
1550                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1551                         error = -EFAULT;
1552                         break;
1553                 }
1554
1555                 sd.len = 0;
1556                 sd.total_len = len;
1557                 sd.flags = flags;
1558                 sd.u.userptr = base;
1559                 sd.pos = 0;
1560
1561                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1562                 if (size < 0) {
1563                         if (!ret)
1564                                 ret = size;
1565
1566                         break;
1567                 }
1568
1569                 ret += size;
1570
1571                 if (size < len)
1572                         break;
1573
1574                 nr_segs--;
1575                 iov++;
1576         }
1577
1578         pipe_unlock(pipe);
1579
1580         if (!ret)
1581                 ret = error;
1582
1583         return ret;
1584 }
1585
1586 /*
1587  * vmsplice splices a user address range into a pipe. It can be thought of
1588  * as splice-from-memory, where the regular splice is splice-from-file (or
1589  * to file). In both cases the output is a pipe, naturally.
1590  */
1591 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1592                              unsigned long nr_segs, unsigned int flags)
1593 {
1594         struct pipe_inode_info *pipe;
1595         struct page *pages[PIPE_BUFFERS];
1596         struct partial_page partial[PIPE_BUFFERS];
1597         struct splice_pipe_desc spd = {
1598                 .pages = pages,
1599                 .partial = partial,
1600                 .flags = flags,
1601                 .ops = &user_page_pipe_buf_ops,
1602                 .spd_release = spd_release_page,
1603         };
1604
1605         pipe = pipe_info(file->f_path.dentry->d_inode);
1606         if (!pipe)
1607                 return -EBADF;
1608
1609         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1610                                             flags & SPLICE_F_GIFT);
1611         if (spd.nr_pages <= 0)
1612                 return spd.nr_pages;
1613
1614         return splice_to_pipe(pipe, &spd);
1615 }
1616
1617 /*
1618  * Note that vmsplice only really supports true splicing _from_ user memory
1619  * to a pipe, not the other way around. Splicing from user memory is a simple
1620  * operation that can be supported without any funky alignment restrictions
1621  * or nasty vm tricks. We simply map in the user memory and fill them into
1622  * a pipe. The reverse isn't quite as easy, though. There are two possible
1623  * solutions for that:
1624  *
1625  *      - memcpy() the data internally, at which point we might as well just
1626  *        do a regular read() on the buffer anyway.
1627  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1628  *        has restriction limitations on both ends of the pipe).
1629  *
1630  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1631  *
1632  */
1633 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1634                 unsigned long, nr_segs, unsigned int, flags)
1635 {
1636         struct file *file;
1637         long error;
1638         int fput;
1639
1640         if (unlikely(nr_segs > UIO_MAXIOV))
1641                 return -EINVAL;
1642         else if (unlikely(!nr_segs))
1643                 return 0;
1644
1645         error = -EBADF;
1646         file = fget_light(fd, &fput);
1647         if (file) {
1648                 if (file->f_mode & FMODE_WRITE)
1649                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1650                 else if (file->f_mode & FMODE_READ)
1651                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1652
1653                 fput_light(file, fput);
1654         }
1655
1656         return error;
1657 }
1658
1659 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1660                 int, fd_out, loff_t __user *, off_out,
1661                 size_t, len, unsigned int, flags)
1662 {
1663         long error;
1664         struct file *in, *out;
1665         int fput_in, fput_out;
1666
1667         if (unlikely(!len))
1668                 return 0;
1669
1670         error = -EBADF;
1671         in = fget_light(fd_in, &fput_in);
1672         if (in) {
1673                 if (in->f_mode & FMODE_READ) {
1674                         out = fget_light(fd_out, &fput_out);
1675                         if (out) {
1676                                 if (out->f_mode & FMODE_WRITE)
1677                                         error = do_splice(in, off_in,
1678                                                           out, off_out,
1679                                                           len, flags);
1680                                 fput_light(out, fput_out);
1681                         }
1682                 }
1683
1684                 fput_light(in, fput_in);
1685         }
1686
1687         return error;
1688 }
1689
1690 /*
1691  * Make sure there's data to read. Wait for input if we can, otherwise
1692  * return an appropriate error.
1693  */
1694 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1695 {
1696         int ret;
1697
1698         /*
1699          * Check ->nrbufs without the inode lock first. This function
1700          * is speculative anyways, so missing one is ok.
1701          */
1702         if (pipe->nrbufs)
1703                 return 0;
1704
1705         ret = 0;
1706         pipe_lock(pipe);
1707
1708         while (!pipe->nrbufs) {
1709                 if (signal_pending(current)) {
1710                         ret = -ERESTARTSYS;
1711                         break;
1712                 }
1713                 if (!pipe->writers)
1714                         break;
1715                 if (!pipe->waiting_writers) {
1716                         if (flags & SPLICE_F_NONBLOCK) {
1717                                 ret = -EAGAIN;
1718                                 break;
1719                         }
1720                 }
1721                 pipe_wait(pipe);
1722         }
1723
1724         pipe_unlock(pipe);
1725         return ret;
1726 }
1727
1728 /*
1729  * Make sure there's writeable room. Wait for room if we can, otherwise
1730  * return an appropriate error.
1731  */
1732 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1733 {
1734         int ret;
1735
1736         /*
1737          * Check ->nrbufs without the inode lock first. This function
1738          * is speculative anyways, so missing one is ok.
1739          */
1740         if (pipe->nrbufs < PIPE_BUFFERS)
1741                 return 0;
1742
1743         ret = 0;
1744         pipe_lock(pipe);
1745
1746         while (pipe->nrbufs >= PIPE_BUFFERS) {
1747                 if (!pipe->readers) {
1748                         send_sig(SIGPIPE, current, 0);
1749                         ret = -EPIPE;
1750                         break;
1751                 }
1752                 if (flags & SPLICE_F_NONBLOCK) {
1753                         ret = -EAGAIN;
1754                         break;
1755                 }
1756                 if (signal_pending(current)) {
1757                         ret = -ERESTARTSYS;
1758                         break;
1759                 }
1760                 pipe->waiting_writers++;
1761                 pipe_wait(pipe);
1762                 pipe->waiting_writers--;
1763         }
1764
1765         pipe_unlock(pipe);
1766         return ret;
1767 }
1768
1769 /*
1770  * Splice contents of ipipe to opipe.
1771  */
1772 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1773                                struct pipe_inode_info *opipe,
1774                                size_t len, unsigned int flags)
1775 {
1776         struct pipe_buffer *ibuf, *obuf;
1777         int ret = 0, nbuf;
1778         bool input_wakeup = false;
1779
1780
1781 retry:
1782         ret = ipipe_prep(ipipe, flags);
1783         if (ret)
1784                 return ret;
1785
1786         ret = opipe_prep(opipe, flags);
1787         if (ret)
1788                 return ret;
1789
1790         /*
1791          * Potential ABBA deadlock, work around it by ordering lock
1792          * grabbing by pipe info address. Otherwise two different processes
1793          * could deadlock (one doing tee from A -> B, the other from B -> A).
1794          */
1795         pipe_double_lock(ipipe, opipe);
1796
1797         do {
1798                 if (!opipe->readers) {
1799                         send_sig(SIGPIPE, current, 0);
1800                         if (!ret)
1801                                 ret = -EPIPE;
1802                         break;
1803                 }
1804
1805                 if (!ipipe->nrbufs && !ipipe->writers)
1806                         break;
1807
1808                 /*
1809                  * Cannot make any progress, because either the input
1810                  * pipe is empty or the output pipe is full.
1811                  */
1812                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1813                         /* Already processed some buffers, break */
1814                         if (ret)
1815                                 break;
1816
1817                         if (flags & SPLICE_F_NONBLOCK) {
1818                                 ret = -EAGAIN;
1819                                 break;
1820                         }
1821
1822                         /*
1823                          * We raced with another reader/writer and haven't
1824                          * managed to process any buffers.  A zero return
1825                          * value means EOF, so retry instead.
1826                          */
1827                         pipe_unlock(ipipe);
1828                         pipe_unlock(opipe);
1829                         goto retry;
1830                 }
1831
1832                 ibuf = ipipe->bufs + ipipe->curbuf;
1833                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1834                 obuf = opipe->bufs + nbuf;
1835
1836                 if (len >= ibuf->len) {
1837                         /*
1838                          * Simply move the whole buffer from ipipe to opipe
1839                          */
1840                         *obuf = *ibuf;
1841                         ibuf->ops = NULL;
1842                         opipe->nrbufs++;
1843                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1844                         ipipe->nrbufs--;
1845                         input_wakeup = true;
1846                 } else {
1847                         /*
1848                          * Get a reference to this pipe buffer,
1849                          * so we can copy the contents over.
1850                          */
1851                         ibuf->ops->get(ipipe, ibuf);
1852                         *obuf = *ibuf;
1853
1854                         /*
1855                          * Don't inherit the gift flag, we need to
1856                          * prevent multiple steals of this page.
1857                          */
1858                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1859
1860                         obuf->len = len;
1861                         opipe->nrbufs++;
1862                         ibuf->offset += obuf->len;
1863                         ibuf->len -= obuf->len;
1864                 }
1865                 ret += obuf->len;
1866                 len -= obuf->len;
1867         } while (len);
1868
1869         pipe_unlock(ipipe);
1870         pipe_unlock(opipe);
1871
1872         /*
1873          * If we put data in the output pipe, wakeup any potential readers.
1874          */
1875         if (ret > 0) {
1876                 smp_mb();
1877                 if (waitqueue_active(&opipe->wait))
1878                         wake_up_interruptible(&opipe->wait);
1879                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1880         }
1881         if (input_wakeup)
1882                 wakeup_pipe_writers(ipipe);
1883
1884         return ret;
1885 }
1886
1887 /*
1888  * Link contents of ipipe to opipe.
1889  */
1890 static int link_pipe(struct pipe_inode_info *ipipe,
1891                      struct pipe_inode_info *opipe,
1892                      size_t len, unsigned int flags)
1893 {
1894         struct pipe_buffer *ibuf, *obuf;
1895         int ret = 0, i = 0, nbuf;
1896
1897         /*
1898          * Potential ABBA deadlock, work around it by ordering lock
1899          * grabbing by pipe info address. Otherwise two different processes
1900          * could deadlock (one doing tee from A -> B, the other from B -> A).
1901          */
1902         pipe_double_lock(ipipe, opipe);
1903
1904         do {
1905                 if (!opipe->readers) {
1906                         send_sig(SIGPIPE, current, 0);
1907                         if (!ret)
1908                                 ret = -EPIPE;
1909                         break;
1910                 }
1911
1912                 /*
1913                  * If we have iterated all input buffers or ran out of
1914                  * output room, break.
1915                  */
1916                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1917                         break;
1918
1919                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1920                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1921
1922                 /*
1923                  * Get a reference to this pipe buffer,
1924                  * so we can copy the contents over.
1925                  */
1926                 ibuf->ops->get(ipipe, ibuf);
1927
1928                 obuf = opipe->bufs + nbuf;
1929                 *obuf = *ibuf;
1930
1931                 /*
1932                  * Don't inherit the gift flag, we need to
1933                  * prevent multiple steals of this page.
1934                  */
1935                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1936
1937                 if (obuf->len > len)
1938                         obuf->len = len;
1939
1940                 opipe->nrbufs++;
1941                 ret += obuf->len;
1942                 len -= obuf->len;
1943                 i++;
1944         } while (len);
1945
1946         /*
1947          * return EAGAIN if we have the potential of some data in the
1948          * future, otherwise just return 0
1949          */
1950         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1951                 ret = -EAGAIN;
1952
1953         pipe_unlock(ipipe);
1954         pipe_unlock(opipe);
1955
1956         /*
1957          * If we put data in the output pipe, wakeup any potential readers.
1958          */
1959         if (ret > 0) {
1960                 smp_mb();
1961                 if (waitqueue_active(&opipe->wait))
1962                         wake_up_interruptible(&opipe->wait);
1963                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1964         }
1965
1966         return ret;
1967 }
1968
1969 /*
1970  * This is a tee(1) implementation that works on pipes. It doesn't copy
1971  * any data, it simply references the 'in' pages on the 'out' pipe.
1972  * The 'flags' used are the SPLICE_F_* variants, currently the only
1973  * applicable one is SPLICE_F_NONBLOCK.
1974  */
1975 static long do_tee(struct file *in, struct file *out, size_t len,
1976                    unsigned int flags)
1977 {
1978         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1979         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1980         int ret = -EINVAL;
1981
1982         /*
1983          * Duplicate the contents of ipipe to opipe without actually
1984          * copying the data.
1985          */
1986         if (ipipe && opipe && ipipe != opipe) {
1987                 /*
1988                  * Keep going, unless we encounter an error. The ipipe/opipe
1989                  * ordering doesn't really matter.
1990                  */
1991                 ret = ipipe_prep(ipipe, flags);
1992                 if (!ret) {
1993                         ret = opipe_prep(opipe, flags);
1994                         if (!ret)
1995                                 ret = link_pipe(ipipe, opipe, len, flags);
1996                 }
1997         }
1998
1999         return ret;
2000 }
2001
2002 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2003 {
2004         struct file *in;
2005         int error, fput_in;
2006
2007         if (unlikely(!len))
2008                 return 0;
2009
2010         error = -EBADF;
2011         in = fget_light(fdin, &fput_in);
2012         if (in) {
2013                 if (in->f_mode & FMODE_READ) {
2014                         int fput_out;
2015                         struct file *out = fget_light(fdout, &fput_out);
2016
2017                         if (out) {
2018                                 if (out->f_mode & FMODE_WRITE)
2019                                         error = do_tee(in, out, len, flags);
2020                                 fput_light(out, fput_out);
2021                         }
2022                 }
2023                 fput_light(in, fput_in);
2024         }
2025
2026         return error;
2027 }