aec4014291be56f5b31f5819507eaaa451cbf705
[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
652                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
653                                            sd->len, &pos, more);
654         }
655
656         return ret;
657 }
658
659 /*
660  * This is a little more tricky than the file -> pipe splicing. There are
661  * basically three cases:
662  *
663  *      - Destination page already exists in the address space and there
664  *        are users of it. For that case we have no other option that
665  *        copying the data. Tough luck.
666  *      - Destination page already exists in the address space, but there
667  *        are no users of it. Make sure it's uptodate, then drop it. Fall
668  *        through to last case.
669  *      - Destination page does not exist, we can add the pipe page to
670  *        the page cache and avoid the copy.
671  *
672  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
673  * sd->flags), we attempt to migrate pages from the pipe to the output
674  * file address space page cache. This is possible if no one else has
675  * the pipe page referenced outside of the pipe and page cache. If
676  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
677  * a new page in the output file page cache and fill/dirty that.
678  */
679 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
680                  struct splice_desc *sd)
681 {
682         struct file *file = sd->u.file;
683         struct address_space *mapping = file->f_mapping;
684         unsigned int offset, this_len;
685         struct page *page;
686         void *fsdata;
687         int ret;
688
689         /*
690          * make sure the data in this buffer is uptodate
691          */
692         ret = buf->ops->confirm(pipe, buf);
693         if (unlikely(ret))
694                 return ret;
695
696         offset = sd->pos & ~PAGE_CACHE_MASK;
697
698         this_len = sd->len;
699         if (this_len + offset > PAGE_CACHE_SIZE)
700                 this_len = PAGE_CACHE_SIZE - offset;
701
702         ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
703                                 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
704         if (unlikely(ret))
705                 goto out;
706
707         if (buf->page != page) {
708                 /*
709                  * Careful, ->map() uses KM_USER0!
710                  */
711                 char *src = buf->ops->map(pipe, buf, 1);
712                 char *dst = kmap_atomic(page, KM_USER1);
713
714                 memcpy(dst + offset, src + buf->offset, this_len);
715                 flush_dcache_page(page);
716                 kunmap_atomic(dst, KM_USER1);
717                 buf->ops->unmap(pipe, buf, src);
718         }
719         ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
720                                 page, fsdata);
721 out:
722         return ret;
723 }
724 EXPORT_SYMBOL(pipe_to_file);
725
726 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
727 {
728         smp_mb();
729         if (waitqueue_active(&pipe->wait))
730                 wake_up_interruptible(&pipe->wait);
731         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
732 }
733
734 /**
735  * splice_from_pipe_feed - feed available data from a pipe to a file
736  * @pipe:       pipe to splice from
737  * @sd:         information to @actor
738  * @actor:      handler that splices the data
739  *
740  * Description:
741  *    This function loops over the pipe and calls @actor to do the
742  *    actual moving of a single struct pipe_buffer to the desired
743  *    destination.  It returns when there's no more buffers left in
744  *    the pipe or if the requested number of bytes (@sd->total_len)
745  *    have been copied.  It returns a positive number (one) if the
746  *    pipe needs to be filled with more data, zero if the required
747  *    number of bytes have been copied and -errno on error.
748  *
749  *    This, together with splice_from_pipe_{begin,end,next}, may be
750  *    used to implement the functionality of __splice_from_pipe() when
751  *    locking is required around copying the pipe buffers to the
752  *    destination.
753  */
754 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
755                           splice_actor *actor)
756 {
757         int ret;
758
759         while (pipe->nrbufs) {
760                 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
761                 const struct pipe_buf_operations *ops = buf->ops;
762
763                 sd->len = buf->len;
764                 if (sd->len > sd->total_len)
765                         sd->len = sd->total_len;
766
767                 ret = actor(pipe, buf, sd);
768                 if (ret <= 0) {
769                         if (ret == -ENODATA)
770                                 ret = 0;
771                         return ret;
772                 }
773                 buf->offset += ret;
774                 buf->len -= ret;
775
776                 sd->num_spliced += ret;
777                 sd->len -= ret;
778                 sd->pos += ret;
779                 sd->total_len -= ret;
780
781                 if (!buf->len) {
782                         buf->ops = NULL;
783                         ops->release(pipe, buf);
784                         pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
785                         pipe->nrbufs--;
786                         if (pipe->inode)
787                                 sd->need_wakeup = true;
788                 }
789
790                 if (!sd->total_len)
791                         return 0;
792         }
793
794         return 1;
795 }
796 EXPORT_SYMBOL(splice_from_pipe_feed);
797
798 /**
799  * splice_from_pipe_next - wait for some data to splice from
800  * @pipe:       pipe to splice from
801  * @sd:         information about the splice operation
802  *
803  * Description:
804  *    This function will wait for some data and return a positive
805  *    value (one) if pipe buffers are available.  It will return zero
806  *    or -errno if no more data needs to be spliced.
807  */
808 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
809 {
810         while (!pipe->nrbufs) {
811                 if (!pipe->writers)
812                         return 0;
813
814                 if (!pipe->waiting_writers && sd->num_spliced)
815                         return 0;
816
817                 if (sd->flags & SPLICE_F_NONBLOCK)
818                         return -EAGAIN;
819
820                 if (signal_pending(current))
821                         return -ERESTARTSYS;
822
823                 if (sd->need_wakeup) {
824                         wakeup_pipe_writers(pipe);
825                         sd->need_wakeup = false;
826                 }
827
828                 pipe_wait(pipe);
829         }
830
831         return 1;
832 }
833 EXPORT_SYMBOL(splice_from_pipe_next);
834
835 /**
836  * splice_from_pipe_begin - start splicing from pipe
837  * @sd:         information about the splice operation
838  *
839  * Description:
840  *    This function should be called before a loop containing
841  *    splice_from_pipe_next() and splice_from_pipe_feed() to
842  *    initialize the necessary fields of @sd.
843  */
844 void splice_from_pipe_begin(struct splice_desc *sd)
845 {
846         sd->num_spliced = 0;
847         sd->need_wakeup = false;
848 }
849 EXPORT_SYMBOL(splice_from_pipe_begin);
850
851 /**
852  * splice_from_pipe_end - finish splicing from pipe
853  * @pipe:       pipe to splice from
854  * @sd:         information about the splice operation
855  *
856  * Description:
857  *    This function will wake up pipe writers if necessary.  It should
858  *    be called after a loop containing splice_from_pipe_next() and
859  *    splice_from_pipe_feed().
860  */
861 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
862 {
863         if (sd->need_wakeup)
864                 wakeup_pipe_writers(pipe);
865 }
866 EXPORT_SYMBOL(splice_from_pipe_end);
867
868 /**
869  * __splice_from_pipe - splice data from a pipe to given actor
870  * @pipe:       pipe to splice from
871  * @sd:         information to @actor
872  * @actor:      handler that splices the data
873  *
874  * Description:
875  *    This function does little more than loop over the pipe and call
876  *    @actor to do the actual moving of a single struct pipe_buffer to
877  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
878  *    pipe_to_user.
879  *
880  */
881 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
882                            splice_actor *actor)
883 {
884         int ret;
885
886         splice_from_pipe_begin(sd);
887         do {
888                 ret = splice_from_pipe_next(pipe, sd);
889                 if (ret > 0)
890                         ret = splice_from_pipe_feed(pipe, sd, actor);
891         } while (ret > 0);
892         splice_from_pipe_end(pipe, sd);
893
894         return sd->num_spliced ? sd->num_spliced : ret;
895 }
896 EXPORT_SYMBOL(__splice_from_pipe);
897
898 /**
899  * splice_from_pipe - splice data from a pipe to a file
900  * @pipe:       pipe to splice from
901  * @out:        file to splice to
902  * @ppos:       position in @out
903  * @len:        how many bytes to splice
904  * @flags:      splice modifier flags
905  * @actor:      handler that splices the data
906  *
907  * Description:
908  *    See __splice_from_pipe. This function locks the pipe inode,
909  *    otherwise it's identical to __splice_from_pipe().
910  *
911  */
912 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
913                          loff_t *ppos, size_t len, unsigned int flags,
914                          splice_actor *actor)
915 {
916         ssize_t ret;
917         struct splice_desc sd = {
918                 .total_len = len,
919                 .flags = flags,
920                 .pos = *ppos,
921                 .u.file = out,
922         };
923
924         pipe_lock(pipe);
925         ret = __splice_from_pipe(pipe, &sd, actor);
926         pipe_unlock(pipe);
927
928         return ret;
929 }
930
931 /**
932  * generic_file_splice_write - splice data from a pipe to a file
933  * @pipe:       pipe info
934  * @out:        file to write to
935  * @ppos:       position in @out
936  * @len:        number of bytes to splice
937  * @flags:      splice modifier flags
938  *
939  * Description:
940  *    Will either move or copy pages (determined by @flags options) from
941  *    the given pipe inode to the given file.
942  *
943  */
944 ssize_t
945 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
946                           loff_t *ppos, size_t len, unsigned int flags)
947 {
948         struct address_space *mapping = out->f_mapping;
949         struct inode *inode = mapping->host;
950         struct splice_desc sd = {
951                 .total_len = len,
952                 .flags = flags,
953                 .pos = *ppos,
954                 .u.file = out,
955         };
956         ssize_t ret;
957
958         pipe_lock(pipe);
959
960         splice_from_pipe_begin(&sd);
961         do {
962                 ret = splice_from_pipe_next(pipe, &sd);
963                 if (ret <= 0)
964                         break;
965
966                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
967                 ret = file_remove_suid(out);
968                 if (!ret) {
969                         file_update_time(out);
970                         ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
971                 }
972                 mutex_unlock(&inode->i_mutex);
973         } while (ret > 0);
974         splice_from_pipe_end(pipe, &sd);
975
976         pipe_unlock(pipe);
977
978         if (sd.num_spliced)
979                 ret = sd.num_spliced;
980
981         if (ret > 0) {
982                 unsigned long nr_pages;
983
984                 *ppos += ret;
985                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
986
987                 /*
988                  * If file or inode is SYNC and we actually wrote some data,
989                  * sync it.
990                  */
991                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
992                         int err;
993
994                         mutex_lock(&inode->i_mutex);
995                         err = generic_osync_inode(inode, mapping,
996                                                   OSYNC_METADATA|OSYNC_DATA);
997                         mutex_unlock(&inode->i_mutex);
998
999                         if (err)
1000                                 ret = err;
1001                 }
1002                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1003         }
1004
1005         return ret;
1006 }
1007
1008 EXPORT_SYMBOL(generic_file_splice_write);
1009
1010 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1011                           struct splice_desc *sd)
1012 {
1013         int ret;
1014         void *data;
1015
1016         ret = buf->ops->confirm(pipe, buf);
1017         if (ret)
1018                 return ret;
1019
1020         data = buf->ops->map(pipe, buf, 0);
1021         ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1022         buf->ops->unmap(pipe, buf, data);
1023
1024         return ret;
1025 }
1026
1027 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1028                                          struct file *out, loff_t *ppos,
1029                                          size_t len, unsigned int flags)
1030 {
1031         ssize_t ret;
1032
1033         ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1034         if (ret > 0)
1035                 *ppos += ret;
1036
1037         return ret;
1038 }
1039
1040 /**
1041  * generic_splice_sendpage - splice data from a pipe to a socket
1042  * @pipe:       pipe to splice from
1043  * @out:        socket to write to
1044  * @ppos:       position in @out
1045  * @len:        number of bytes to splice
1046  * @flags:      splice modifier flags
1047  *
1048  * Description:
1049  *    Will send @len bytes from the pipe to a network socket. No data copying
1050  *    is involved.
1051  *
1052  */
1053 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1054                                 loff_t *ppos, size_t len, unsigned int flags)
1055 {
1056         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1057 }
1058
1059 EXPORT_SYMBOL(generic_splice_sendpage);
1060
1061 /*
1062  * Attempt to initiate a splice from pipe to file.
1063  */
1064 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1065                            loff_t *ppos, size_t len, unsigned int flags)
1066 {
1067         ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1068                                 loff_t *, size_t, unsigned int);
1069         int ret;
1070
1071         if (unlikely(!(out->f_mode & FMODE_WRITE)))
1072                 return -EBADF;
1073
1074         if (unlikely(out->f_flags & O_APPEND))
1075                 return -EINVAL;
1076
1077         ret = rw_verify_area(WRITE, out, ppos, len);
1078         if (unlikely(ret < 0))
1079                 return ret;
1080
1081         splice_write = out->f_op->splice_write;
1082         if (!splice_write)
1083                 splice_write = default_file_splice_write;
1084
1085         return splice_write(pipe, out, ppos, len, flags);
1086 }
1087
1088 /*
1089  * Attempt to initiate a splice from a file to a pipe.
1090  */
1091 static long do_splice_to(struct file *in, loff_t *ppos,
1092                          struct pipe_inode_info *pipe, size_t len,
1093                          unsigned int flags)
1094 {
1095         ssize_t (*splice_read)(struct file *, loff_t *,
1096                                struct pipe_inode_info *, size_t, unsigned int);
1097         int ret;
1098
1099         if (unlikely(!(in->f_mode & FMODE_READ)))
1100                 return -EBADF;
1101
1102         ret = rw_verify_area(READ, in, ppos, len);
1103         if (unlikely(ret < 0))
1104                 return ret;
1105
1106         splice_read = in->f_op->splice_read;
1107         if (!splice_read)
1108                 splice_read = default_file_splice_read;
1109
1110         return splice_read(in, ppos, pipe, len, flags);
1111 }
1112
1113 /**
1114  * splice_direct_to_actor - splices data directly between two non-pipes
1115  * @in:         file to splice from
1116  * @sd:         actor information on where to splice to
1117  * @actor:      handles the data splicing
1118  *
1119  * Description:
1120  *    This is a special case helper to splice directly between two
1121  *    points, without requiring an explicit pipe. Internally an allocated
1122  *    pipe is cached in the process, and reused during the lifetime of
1123  *    that process.
1124  *
1125  */
1126 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1127                                splice_direct_actor *actor)
1128 {
1129         struct pipe_inode_info *pipe;
1130         long ret, bytes;
1131         umode_t i_mode;
1132         size_t len;
1133         int i, flags;
1134
1135         /*
1136          * We require the input being a regular file, as we don't want to
1137          * randomly drop data for eg socket -> socket splicing. Use the
1138          * piped splicing for that!
1139          */
1140         i_mode = in->f_path.dentry->d_inode->i_mode;
1141         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1142                 return -EINVAL;
1143
1144         /*
1145          * neither in nor out is a pipe, setup an internal pipe attached to
1146          * 'out' and transfer the wanted data from 'in' to 'out' through that
1147          */
1148         pipe = current->splice_pipe;
1149         if (unlikely(!pipe)) {
1150                 pipe = alloc_pipe_info(NULL);
1151                 if (!pipe)
1152                         return -ENOMEM;
1153
1154                 /*
1155                  * We don't have an immediate reader, but we'll read the stuff
1156                  * out of the pipe right after the splice_to_pipe(). So set
1157                  * PIPE_READERS appropriately.
1158                  */
1159                 pipe->readers = 1;
1160
1161                 current->splice_pipe = pipe;
1162         }
1163
1164         /*
1165          * Do the splice.
1166          */
1167         ret = 0;
1168         bytes = 0;
1169         len = sd->total_len;
1170         flags = sd->flags;
1171
1172         /*
1173          * Don't block on output, we have to drain the direct pipe.
1174          */
1175         sd->flags &= ~SPLICE_F_NONBLOCK;
1176
1177         while (len) {
1178                 size_t read_len;
1179                 loff_t pos = sd->pos, prev_pos = pos;
1180
1181                 ret = do_splice_to(in, &pos, pipe, len, flags);
1182                 if (unlikely(ret <= 0))
1183                         goto out_release;
1184
1185                 read_len = ret;
1186                 sd->total_len = read_len;
1187
1188                 /*
1189                  * NOTE: nonblocking mode only applies to the input. We
1190                  * must not do the output in nonblocking mode as then we
1191                  * could get stuck data in the internal pipe:
1192                  */
1193                 ret = actor(pipe, sd);
1194                 if (unlikely(ret <= 0)) {
1195                         sd->pos = prev_pos;
1196                         goto out_release;
1197                 }
1198
1199                 bytes += ret;
1200                 len -= ret;
1201                 sd->pos = pos;
1202
1203                 if (ret < read_len) {
1204                         sd->pos = prev_pos + ret;
1205                         goto out_release;
1206                 }
1207         }
1208
1209 done:
1210         pipe->nrbufs = pipe->curbuf = 0;
1211         file_accessed(in);
1212         return bytes;
1213
1214 out_release:
1215         /*
1216          * If we did an incomplete transfer we must release
1217          * the pipe buffers in question:
1218          */
1219         for (i = 0; i < PIPE_BUFFERS; i++) {
1220                 struct pipe_buffer *buf = pipe->bufs + i;
1221
1222                 if (buf->ops) {
1223                         buf->ops->release(pipe, buf);
1224                         buf->ops = NULL;
1225                 }
1226         }
1227
1228         if (!bytes)
1229                 bytes = ret;
1230
1231         goto done;
1232 }
1233 EXPORT_SYMBOL(splice_direct_to_actor);
1234
1235 static int direct_splice_actor(struct pipe_inode_info *pipe,
1236                                struct splice_desc *sd)
1237 {
1238         struct file *file = sd->u.file;
1239
1240         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1241 }
1242
1243 /**
1244  * do_splice_direct - splices data directly between two files
1245  * @in:         file to splice from
1246  * @ppos:       input file offset
1247  * @out:        file to splice to
1248  * @len:        number of bytes to splice
1249  * @flags:      splice modifier flags
1250  *
1251  * Description:
1252  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1253  *    doing it in the application would incur an extra system call
1254  *    (splice in + splice out, as compared to just sendfile()). So this helper
1255  *    can splice directly through a process-private pipe.
1256  *
1257  */
1258 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1259                       size_t len, unsigned int flags)
1260 {
1261         struct splice_desc sd = {
1262                 .len            = len,
1263                 .total_len      = len,
1264                 .flags          = flags,
1265                 .pos            = *ppos,
1266                 .u.file         = out,
1267         };
1268         long ret;
1269
1270         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1271         if (ret > 0)
1272                 *ppos = sd.pos;
1273
1274         return ret;
1275 }
1276
1277 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1278                                struct pipe_inode_info *opipe,
1279                                size_t len, unsigned int flags);
1280 /*
1281  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1282  * location, so checking ->i_pipe is not enough to verify that this is a
1283  * pipe.
1284  */
1285 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1286 {
1287         if (S_ISFIFO(inode->i_mode))
1288                 return inode->i_pipe;
1289
1290         return NULL;
1291 }
1292
1293 /*
1294  * Determine where to splice to/from.
1295  */
1296 static long do_splice(struct file *in, loff_t __user *off_in,
1297                       struct file *out, loff_t __user *off_out,
1298                       size_t len, unsigned int flags)
1299 {
1300         struct pipe_inode_info *ipipe;
1301         struct pipe_inode_info *opipe;
1302         loff_t offset, *off;
1303         long ret;
1304
1305         ipipe = pipe_info(in->f_path.dentry->d_inode);
1306         opipe = pipe_info(out->f_path.dentry->d_inode);
1307
1308         if (ipipe && opipe) {
1309                 if (off_in || off_out)
1310                         return -ESPIPE;
1311
1312                 if (!(in->f_mode & FMODE_READ))
1313                         return -EBADF;
1314
1315                 if (!(out->f_mode & FMODE_WRITE))
1316                         return -EBADF;
1317
1318                 /* Splicing to self would be fun, but... */
1319                 if (ipipe == opipe)
1320                         return -EINVAL;
1321
1322                 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1323         }
1324
1325         if (ipipe) {
1326                 if (off_in)
1327                         return -ESPIPE;
1328                 if (off_out) {
1329                         if (out->f_op->llseek == no_llseek)
1330                                 return -EINVAL;
1331                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1332                                 return -EFAULT;
1333                         off = &offset;
1334                 } else
1335                         off = &out->f_pos;
1336
1337                 ret = do_splice_from(ipipe, out, off, len, flags);
1338
1339                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1340                         ret = -EFAULT;
1341
1342                 return ret;
1343         }
1344
1345         if (opipe) {
1346                 if (off_out)
1347                         return -ESPIPE;
1348                 if (off_in) {
1349                         if (in->f_op->llseek == no_llseek)
1350                                 return -EINVAL;
1351                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1352                                 return -EFAULT;
1353                         off = &offset;
1354                 } else
1355                         off = &in->f_pos;
1356
1357                 ret = do_splice_to(in, off, opipe, len, flags);
1358
1359                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1360                         ret = -EFAULT;
1361
1362                 return ret;
1363         }
1364
1365         return -EINVAL;
1366 }
1367
1368 /*
1369  * Map an iov into an array of pages and offset/length tupples. With the
1370  * partial_page structure, we can map several non-contiguous ranges into
1371  * our ones pages[] map instead of splitting that operation into pieces.
1372  * Could easily be exported as a generic helper for other users, in which
1373  * case one would probably want to add a 'max_nr_pages' parameter as well.
1374  */
1375 static int get_iovec_page_array(const struct iovec __user *iov,
1376                                 unsigned int nr_vecs, struct page **pages,
1377                                 struct partial_page *partial, int aligned)
1378 {
1379         int buffers = 0, error = 0;
1380
1381         while (nr_vecs) {
1382                 unsigned long off, npages;
1383                 struct iovec entry;
1384                 void __user *base;
1385                 size_t len;
1386                 int i;
1387
1388                 error = -EFAULT;
1389                 if (copy_from_user(&entry, iov, sizeof(entry)))
1390                         break;
1391
1392                 base = entry.iov_base;
1393                 len = entry.iov_len;
1394
1395                 /*
1396                  * Sanity check this iovec. 0 read succeeds.
1397                  */
1398                 error = 0;
1399                 if (unlikely(!len))
1400                         break;
1401                 error = -EFAULT;
1402                 if (!access_ok(VERIFY_READ, base, len))
1403                         break;
1404
1405                 /*
1406                  * Get this base offset and number of pages, then map
1407                  * in the user pages.
1408                  */
1409                 off = (unsigned long) base & ~PAGE_MASK;
1410
1411                 /*
1412                  * If asked for alignment, the offset must be zero and the
1413                  * length a multiple of the PAGE_SIZE.
1414                  */
1415                 error = -EINVAL;
1416                 if (aligned && (off || len & ~PAGE_MASK))
1417                         break;
1418
1419                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1420                 if (npages > PIPE_BUFFERS - buffers)
1421                         npages = PIPE_BUFFERS - buffers;
1422
1423                 error = get_user_pages_fast((unsigned long)base, npages,
1424                                         0, &pages[buffers]);
1425
1426                 if (unlikely(error <= 0))
1427                         break;
1428
1429                 /*
1430                  * Fill this contiguous range into the partial page map.
1431                  */
1432                 for (i = 0; i < error; i++) {
1433                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1434
1435                         partial[buffers].offset = off;
1436                         partial[buffers].len = plen;
1437
1438                         off = 0;
1439                         len -= plen;
1440                         buffers++;
1441                 }
1442
1443                 /*
1444                  * We didn't complete this iov, stop here since it probably
1445                  * means we have to move some of this into a pipe to
1446                  * be able to continue.
1447                  */
1448                 if (len)
1449                         break;
1450
1451                 /*
1452                  * Don't continue if we mapped fewer pages than we asked for,
1453                  * or if we mapped the max number of pages that we have
1454                  * room for.
1455                  */
1456                 if (error < npages || buffers == PIPE_BUFFERS)
1457                         break;
1458
1459                 nr_vecs--;
1460                 iov++;
1461         }
1462
1463         if (buffers)
1464                 return buffers;
1465
1466         return error;
1467 }
1468
1469 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1470                         struct splice_desc *sd)
1471 {
1472         char *src;
1473         int ret;
1474
1475         ret = buf->ops->confirm(pipe, buf);
1476         if (unlikely(ret))
1477                 return ret;
1478
1479         /*
1480          * See if we can use the atomic maps, by prefaulting in the
1481          * pages and doing an atomic copy
1482          */
1483         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1484                 src = buf->ops->map(pipe, buf, 1);
1485                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1486                                                         sd->len);
1487                 buf->ops->unmap(pipe, buf, src);
1488                 if (!ret) {
1489                         ret = sd->len;
1490                         goto out;
1491                 }
1492         }
1493
1494         /*
1495          * No dice, use slow non-atomic map and copy
1496          */
1497         src = buf->ops->map(pipe, buf, 0);
1498
1499         ret = sd->len;
1500         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1501                 ret = -EFAULT;
1502
1503         buf->ops->unmap(pipe, buf, src);
1504 out:
1505         if (ret > 0)
1506                 sd->u.userptr += ret;
1507         return ret;
1508 }
1509
1510 /*
1511  * For lack of a better implementation, implement vmsplice() to userspace
1512  * as a simple copy of the pipes pages to the user iov.
1513  */
1514 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1515                              unsigned long nr_segs, unsigned int flags)
1516 {
1517         struct pipe_inode_info *pipe;
1518         struct splice_desc sd;
1519         ssize_t size;
1520         int error;
1521         long ret;
1522
1523         pipe = pipe_info(file->f_path.dentry->d_inode);
1524         if (!pipe)
1525                 return -EBADF;
1526
1527         pipe_lock(pipe);
1528
1529         error = ret = 0;
1530         while (nr_segs) {
1531                 void __user *base;
1532                 size_t len;
1533
1534                 /*
1535                  * Get user address base and length for this iovec.
1536                  */
1537                 error = get_user(base, &iov->iov_base);
1538                 if (unlikely(error))
1539                         break;
1540                 error = get_user(len, &iov->iov_len);
1541                 if (unlikely(error))
1542                         break;
1543
1544                 /*
1545                  * Sanity check this iovec. 0 read succeeds.
1546                  */
1547                 if (unlikely(!len))
1548                         break;
1549                 if (unlikely(!base)) {
1550                         error = -EFAULT;
1551                         break;
1552                 }
1553
1554                 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1555                         error = -EFAULT;
1556                         break;
1557                 }
1558
1559                 sd.len = 0;
1560                 sd.total_len = len;
1561                 sd.flags = flags;
1562                 sd.u.userptr = base;
1563                 sd.pos = 0;
1564
1565                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1566                 if (size < 0) {
1567                         if (!ret)
1568                                 ret = size;
1569
1570                         break;
1571                 }
1572
1573                 ret += size;
1574
1575                 if (size < len)
1576                         break;
1577
1578                 nr_segs--;
1579                 iov++;
1580         }
1581
1582         pipe_unlock(pipe);
1583
1584         if (!ret)
1585                 ret = error;
1586
1587         return ret;
1588 }
1589
1590 /*
1591  * vmsplice splices a user address range into a pipe. It can be thought of
1592  * as splice-from-memory, where the regular splice is splice-from-file (or
1593  * to file). In both cases the output is a pipe, naturally.
1594  */
1595 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1596                              unsigned long nr_segs, unsigned int flags)
1597 {
1598         struct pipe_inode_info *pipe;
1599         struct page *pages[PIPE_BUFFERS];
1600         struct partial_page partial[PIPE_BUFFERS];
1601         struct splice_pipe_desc spd = {
1602                 .pages = pages,
1603                 .partial = partial,
1604                 .flags = flags,
1605                 .ops = &user_page_pipe_buf_ops,
1606                 .spd_release = spd_release_page,
1607         };
1608
1609         pipe = pipe_info(file->f_path.dentry->d_inode);
1610         if (!pipe)
1611                 return -EBADF;
1612
1613         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1614                                             flags & SPLICE_F_GIFT);
1615         if (spd.nr_pages <= 0)
1616                 return spd.nr_pages;
1617
1618         return splice_to_pipe(pipe, &spd);
1619 }
1620
1621 /*
1622  * Note that vmsplice only really supports true splicing _from_ user memory
1623  * to a pipe, not the other way around. Splicing from user memory is a simple
1624  * operation that can be supported without any funky alignment restrictions
1625  * or nasty vm tricks. We simply map in the user memory and fill them into
1626  * a pipe. The reverse isn't quite as easy, though. There are two possible
1627  * solutions for that:
1628  *
1629  *      - memcpy() the data internally, at which point we might as well just
1630  *        do a regular read() on the buffer anyway.
1631  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1632  *        has restriction limitations on both ends of the pipe).
1633  *
1634  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1635  *
1636  */
1637 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1638                 unsigned long, nr_segs, unsigned int, flags)
1639 {
1640         struct file *file;
1641         long error;
1642         int fput;
1643
1644         if (unlikely(nr_segs > UIO_MAXIOV))
1645                 return -EINVAL;
1646         else if (unlikely(!nr_segs))
1647                 return 0;
1648
1649         error = -EBADF;
1650         file = fget_light(fd, &fput);
1651         if (file) {
1652                 if (file->f_mode & FMODE_WRITE)
1653                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1654                 else if (file->f_mode & FMODE_READ)
1655                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1656
1657                 fput_light(file, fput);
1658         }
1659
1660         return error;
1661 }
1662
1663 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1664                 int, fd_out, loff_t __user *, off_out,
1665                 size_t, len, unsigned int, flags)
1666 {
1667         long error;
1668         struct file *in, *out;
1669         int fput_in, fput_out;
1670
1671         if (unlikely(!len))
1672                 return 0;
1673
1674         error = -EBADF;
1675         in = fget_light(fd_in, &fput_in);
1676         if (in) {
1677                 if (in->f_mode & FMODE_READ) {
1678                         out = fget_light(fd_out, &fput_out);
1679                         if (out) {
1680                                 if (out->f_mode & FMODE_WRITE)
1681                                         error = do_splice(in, off_in,
1682                                                           out, off_out,
1683                                                           len, flags);
1684                                 fput_light(out, fput_out);
1685                         }
1686                 }
1687
1688                 fput_light(in, fput_in);
1689         }
1690
1691         return error;
1692 }
1693
1694 /*
1695  * Make sure there's data to read. Wait for input if we can, otherwise
1696  * return an appropriate error.
1697  */
1698 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1699 {
1700         int ret;
1701
1702         /*
1703          * Check ->nrbufs without the inode lock first. This function
1704          * is speculative anyways, so missing one is ok.
1705          */
1706         if (pipe->nrbufs)
1707                 return 0;
1708
1709         ret = 0;
1710         pipe_lock(pipe);
1711
1712         while (!pipe->nrbufs) {
1713                 if (signal_pending(current)) {
1714                         ret = -ERESTARTSYS;
1715                         break;
1716                 }
1717                 if (!pipe->writers)
1718                         break;
1719                 if (!pipe->waiting_writers) {
1720                         if (flags & SPLICE_F_NONBLOCK) {
1721                                 ret = -EAGAIN;
1722                                 break;
1723                         }
1724                 }
1725                 pipe_wait(pipe);
1726         }
1727
1728         pipe_unlock(pipe);
1729         return ret;
1730 }
1731
1732 /*
1733  * Make sure there's writeable room. Wait for room if we can, otherwise
1734  * return an appropriate error.
1735  */
1736 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1737 {
1738         int ret;
1739
1740         /*
1741          * Check ->nrbufs without the inode lock first. This function
1742          * is speculative anyways, so missing one is ok.
1743          */
1744         if (pipe->nrbufs < PIPE_BUFFERS)
1745                 return 0;
1746
1747         ret = 0;
1748         pipe_lock(pipe);
1749
1750         while (pipe->nrbufs >= PIPE_BUFFERS) {
1751                 if (!pipe->readers) {
1752                         send_sig(SIGPIPE, current, 0);
1753                         ret = -EPIPE;
1754                         break;
1755                 }
1756                 if (flags & SPLICE_F_NONBLOCK) {
1757                         ret = -EAGAIN;
1758                         break;
1759                 }
1760                 if (signal_pending(current)) {
1761                         ret = -ERESTARTSYS;
1762                         break;
1763                 }
1764                 pipe->waiting_writers++;
1765                 pipe_wait(pipe);
1766                 pipe->waiting_writers--;
1767         }
1768
1769         pipe_unlock(pipe);
1770         return ret;
1771 }
1772
1773 /*
1774  * Splice contents of ipipe to opipe.
1775  */
1776 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1777                                struct pipe_inode_info *opipe,
1778                                size_t len, unsigned int flags)
1779 {
1780         struct pipe_buffer *ibuf, *obuf;
1781         int ret = 0, nbuf;
1782         bool input_wakeup = false;
1783
1784
1785 retry:
1786         ret = ipipe_prep(ipipe, flags);
1787         if (ret)
1788                 return ret;
1789
1790         ret = opipe_prep(opipe, flags);
1791         if (ret)
1792                 return ret;
1793
1794         /*
1795          * Potential ABBA deadlock, work around it by ordering lock
1796          * grabbing by pipe info address. Otherwise two different processes
1797          * could deadlock (one doing tee from A -> B, the other from B -> A).
1798          */
1799         pipe_double_lock(ipipe, opipe);
1800
1801         do {
1802                 if (!opipe->readers) {
1803                         send_sig(SIGPIPE, current, 0);
1804                         if (!ret)
1805                                 ret = -EPIPE;
1806                         break;
1807                 }
1808
1809                 if (!ipipe->nrbufs && !ipipe->writers)
1810                         break;
1811
1812                 /*
1813                  * Cannot make any progress, because either the input
1814                  * pipe is empty or the output pipe is full.
1815                  */
1816                 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1817                         /* Already processed some buffers, break */
1818                         if (ret)
1819                                 break;
1820
1821                         if (flags & SPLICE_F_NONBLOCK) {
1822                                 ret = -EAGAIN;
1823                                 break;
1824                         }
1825
1826                         /*
1827                          * We raced with another reader/writer and haven't
1828                          * managed to process any buffers.  A zero return
1829                          * value means EOF, so retry instead.
1830                          */
1831                         pipe_unlock(ipipe);
1832                         pipe_unlock(opipe);
1833                         goto retry;
1834                 }
1835
1836                 ibuf = ipipe->bufs + ipipe->curbuf;
1837                 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1838                 obuf = opipe->bufs + nbuf;
1839
1840                 if (len >= ibuf->len) {
1841                         /*
1842                          * Simply move the whole buffer from ipipe to opipe
1843                          */
1844                         *obuf = *ibuf;
1845                         ibuf->ops = NULL;
1846                         opipe->nrbufs++;
1847                         ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1848                         ipipe->nrbufs--;
1849                         input_wakeup = true;
1850                 } else {
1851                         /*
1852                          * Get a reference to this pipe buffer,
1853                          * so we can copy the contents over.
1854                          */
1855                         ibuf->ops->get(ipipe, ibuf);
1856                         *obuf = *ibuf;
1857
1858                         /*
1859                          * Don't inherit the gift flag, we need to
1860                          * prevent multiple steals of this page.
1861                          */
1862                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1863
1864                         obuf->len = len;
1865                         opipe->nrbufs++;
1866                         ibuf->offset += obuf->len;
1867                         ibuf->len -= obuf->len;
1868                 }
1869                 ret += obuf->len;
1870                 len -= obuf->len;
1871         } while (len);
1872
1873         pipe_unlock(ipipe);
1874         pipe_unlock(opipe);
1875
1876         /*
1877          * If we put data in the output pipe, wakeup any potential readers.
1878          */
1879         if (ret > 0) {
1880                 smp_mb();
1881                 if (waitqueue_active(&opipe->wait))
1882                         wake_up_interruptible(&opipe->wait);
1883                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1884         }
1885         if (input_wakeup)
1886                 wakeup_pipe_writers(ipipe);
1887
1888         return ret;
1889 }
1890
1891 /*
1892  * Link contents of ipipe to opipe.
1893  */
1894 static int link_pipe(struct pipe_inode_info *ipipe,
1895                      struct pipe_inode_info *opipe,
1896                      size_t len, unsigned int flags)
1897 {
1898         struct pipe_buffer *ibuf, *obuf;
1899         int ret = 0, i = 0, nbuf;
1900
1901         /*
1902          * Potential ABBA deadlock, work around it by ordering lock
1903          * grabbing by pipe info address. Otherwise two different processes
1904          * could deadlock (one doing tee from A -> B, the other from B -> A).
1905          */
1906         pipe_double_lock(ipipe, opipe);
1907
1908         do {
1909                 if (!opipe->readers) {
1910                         send_sig(SIGPIPE, current, 0);
1911                         if (!ret)
1912                                 ret = -EPIPE;
1913                         break;
1914                 }
1915
1916                 /*
1917                  * If we have iterated all input buffers or ran out of
1918                  * output room, break.
1919                  */
1920                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1921                         break;
1922
1923                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1924                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1925
1926                 /*
1927                  * Get a reference to this pipe buffer,
1928                  * so we can copy the contents over.
1929                  */
1930                 ibuf->ops->get(ipipe, ibuf);
1931
1932                 obuf = opipe->bufs + nbuf;
1933                 *obuf = *ibuf;
1934
1935                 /*
1936                  * Don't inherit the gift flag, we need to
1937                  * prevent multiple steals of this page.
1938                  */
1939                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1940
1941                 if (obuf->len > len)
1942                         obuf->len = len;
1943
1944                 opipe->nrbufs++;
1945                 ret += obuf->len;
1946                 len -= obuf->len;
1947                 i++;
1948         } while (len);
1949
1950         /*
1951          * return EAGAIN if we have the potential of some data in the
1952          * future, otherwise just return 0
1953          */
1954         if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1955                 ret = -EAGAIN;
1956
1957         pipe_unlock(ipipe);
1958         pipe_unlock(opipe);
1959
1960         /*
1961          * If we put data in the output pipe, wakeup any potential readers.
1962          */
1963         if (ret > 0) {
1964                 smp_mb();
1965                 if (waitqueue_active(&opipe->wait))
1966                         wake_up_interruptible(&opipe->wait);
1967                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1968         }
1969
1970         return ret;
1971 }
1972
1973 /*
1974  * This is a tee(1) implementation that works on pipes. It doesn't copy
1975  * any data, it simply references the 'in' pages on the 'out' pipe.
1976  * The 'flags' used are the SPLICE_F_* variants, currently the only
1977  * applicable one is SPLICE_F_NONBLOCK.
1978  */
1979 static long do_tee(struct file *in, struct file *out, size_t len,
1980                    unsigned int flags)
1981 {
1982         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1983         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1984         int ret = -EINVAL;
1985
1986         /*
1987          * Duplicate the contents of ipipe to opipe without actually
1988          * copying the data.
1989          */
1990         if (ipipe && opipe && ipipe != opipe) {
1991                 /*
1992                  * Keep going, unless we encounter an error. The ipipe/opipe
1993                  * ordering doesn't really matter.
1994                  */
1995                 ret = ipipe_prep(ipipe, flags);
1996                 if (!ret) {
1997                         ret = opipe_prep(opipe, flags);
1998                         if (!ret)
1999                                 ret = link_pipe(ipipe, opipe, len, flags);
2000                 }
2001         }
2002
2003         return ret;
2004 }
2005
2006 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2007 {
2008         struct file *in;
2009         int error, fput_in;
2010
2011         if (unlikely(!len))
2012                 return 0;
2013
2014         error = -EBADF;
2015         in = fget_light(fdin, &fput_in);
2016         if (in) {
2017                 if (in->f_mode & FMODE_READ) {
2018                         int fput_out;
2019                         struct file *out = fget_light(fdout, &fput_out);
2020
2021                         if (out) {
2022                                 if (out->f_mode & FMODE_WRITE)
2023                                         error = do_tee(in, out, len, flags);
2024                                 fput_light(out, fput_out);
2025                         }
2026                 }
2027                 fput_light(in, fput_in);
2028         }
2029
2030         return error;
2031 }