2 * "splice": joining two ropes together by interweaving their strands.
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.
8 * The traditional unix read/write is extended with a "splice()" operation
9 * that transfers data buffers to or from a pipe buffer.
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.
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>
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/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32 #include <linux/gfp.h>
33 #include <linux/socket.h>
36 * Attempt to steal a page from a pipe buffer. This should perhaps go into
37 * a vm helper function, it's already simplified quite a bit by the
38 * addition of remove_mapping(). If success is returned, the caller may
39 * attempt to reuse this page for another destination.
41 static int page_cache_pipe_buf_steal(struct pipe_inode_info
*pipe
,
42 struct pipe_buffer
*buf
)
44 struct page
*page
= buf
->page
;
45 struct address_space
*mapping
;
49 mapping
= page_mapping(page
);
51 WARN_ON(!PageUptodate(page
));
54 * At least for ext2 with nobh option, we need to wait on
55 * writeback completing on this page, since we'll remove it
56 * from the pagecache. Otherwise truncate wont wait on the
57 * page, allowing the disk blocks to be reused by someone else
58 * before we actually wrote our data to them. fs corruption
61 wait_on_page_writeback(page
);
63 if (page_has_private(page
) &&
64 !try_to_release_page(page
, GFP_KERNEL
))
68 * If we succeeded in removing the mapping, set LRU flag
71 if (remove_mapping(mapping
, page
)) {
72 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
78 * Raced with truncate or failed to remove page from current
79 * address space, unlock and return failure.
86 static void page_cache_pipe_buf_release(struct pipe_inode_info
*pipe
,
87 struct pipe_buffer
*buf
)
89 page_cache_release(buf
->page
);
90 buf
->flags
&= ~PIPE_BUF_FLAG_LRU
;
94 * Check whether the contents of buf is OK to access. Since the content
95 * is a page cache page, IO may be in flight.
97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info
*pipe
,
98 struct pipe_buffer
*buf
)
100 struct page
*page
= buf
->page
;
103 if (!PageUptodate(page
)) {
107 * Page got truncated/unhashed. This will cause a 0-byte
108 * splice, if this is the first page.
110 if (!page
->mapping
) {
116 * Uh oh, read-error from disk.
118 if (!PageUptodate(page
)) {
124 * Page is ok afterall, we are done.
135 const struct pipe_buf_operations page_cache_pipe_buf_ops
= {
137 .map
= generic_pipe_buf_map
,
138 .unmap
= generic_pipe_buf_unmap
,
139 .confirm
= page_cache_pipe_buf_confirm
,
140 .release
= page_cache_pipe_buf_release
,
141 .steal
= page_cache_pipe_buf_steal
,
142 .get
= generic_pipe_buf_get
,
145 static int user_page_pipe_buf_steal(struct pipe_inode_info
*pipe
,
146 struct pipe_buffer
*buf
)
148 if (!(buf
->flags
& PIPE_BUF_FLAG_GIFT
))
151 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
152 return generic_pipe_buf_steal(pipe
, buf
);
155 static const struct pipe_buf_operations user_page_pipe_buf_ops
= {
157 .map
= generic_pipe_buf_map
,
158 .unmap
= generic_pipe_buf_unmap
,
159 .confirm
= generic_pipe_buf_confirm
,
160 .release
= page_cache_pipe_buf_release
,
161 .steal
= user_page_pipe_buf_steal
,
162 .get
= generic_pipe_buf_get
,
165 static void wakeup_pipe_readers(struct pipe_inode_info
*pipe
)
168 if (waitqueue_active(&pipe
->wait
))
169 wake_up_interruptible(&pipe
->wait
);
170 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
174 * splice_to_pipe - fill passed data into a pipe
175 * @pipe: pipe to fill
179 * @spd contains a map of pages and len/offset tuples, along with
180 * the struct pipe_buf_operations associated with these pages. This
181 * function will link that data to the pipe.
184 ssize_t
splice_to_pipe(struct pipe_inode_info
*pipe
,
185 struct splice_pipe_desc
*spd
)
187 unsigned int spd_pages
= spd
->nr_pages
;
188 int ret
, do_wakeup
, page_nr
;
197 if (!pipe
->readers
) {
198 send_sig(SIGPIPE
, current
, 0);
204 if (pipe
->nrbufs
< pipe
->buffers
) {
205 int newbuf
= (pipe
->curbuf
+ pipe
->nrbufs
) & (pipe
->buffers
- 1);
206 struct pipe_buffer
*buf
= pipe
->bufs
+ newbuf
;
208 buf
->page
= spd
->pages
[page_nr
];
209 buf
->offset
= spd
->partial
[page_nr
].offset
;
210 buf
->len
= spd
->partial
[page_nr
].len
;
211 buf
->private = spd
->partial
[page_nr
].private;
213 if (spd
->flags
& SPLICE_F_GIFT
)
214 buf
->flags
|= PIPE_BUF_FLAG_GIFT
;
223 if (!--spd
->nr_pages
)
225 if (pipe
->nrbufs
< pipe
->buffers
)
231 if (spd
->flags
& SPLICE_F_NONBLOCK
) {
237 if (signal_pending(current
)) {
245 if (waitqueue_active(&pipe
->wait
))
246 wake_up_interruptible_sync(&pipe
->wait
);
247 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
251 pipe
->waiting_writers
++;
253 pipe
->waiting_writers
--;
259 wakeup_pipe_readers(pipe
);
261 while (page_nr
< spd_pages
)
262 spd
->spd_release(spd
, page_nr
++);
267 void spd_release_page(struct splice_pipe_desc
*spd
, unsigned int i
)
269 page_cache_release(spd
->pages
[i
]);
273 * Check if we need to grow the arrays holding pages and partial page
276 int splice_grow_spd(struct pipe_inode_info
*pipe
, struct splice_pipe_desc
*spd
)
278 if (pipe
->buffers
<= PIPE_DEF_BUFFERS
)
281 spd
->pages
= kmalloc(pipe
->buffers
* sizeof(struct page
*), GFP_KERNEL
);
282 spd
->partial
= kmalloc(pipe
->buffers
* sizeof(struct partial_page
), GFP_KERNEL
);
284 if (spd
->pages
&& spd
->partial
)
292 void splice_shrink_spd(struct pipe_inode_info
*pipe
,
293 struct splice_pipe_desc
*spd
)
295 if (pipe
->buffers
<= PIPE_DEF_BUFFERS
)
303 __generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
304 struct pipe_inode_info
*pipe
, size_t len
,
307 struct address_space
*mapping
= in
->f_mapping
;
308 unsigned int loff
, nr_pages
, req_pages
;
309 struct page
*pages
[PIPE_DEF_BUFFERS
];
310 struct partial_page partial
[PIPE_DEF_BUFFERS
];
312 pgoff_t index
, end_index
;
315 struct splice_pipe_desc spd
= {
319 .ops
= &page_cache_pipe_buf_ops
,
320 .spd_release
= spd_release_page
,
323 if (splice_grow_spd(pipe
, &spd
))
326 index
= *ppos
>> PAGE_CACHE_SHIFT
;
327 loff
= *ppos
& ~PAGE_CACHE_MASK
;
328 req_pages
= (len
+ loff
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
329 nr_pages
= min(req_pages
, pipe
->buffers
);
332 * Lookup the (hopefully) full range of pages we need.
334 spd
.nr_pages
= find_get_pages_contig(mapping
, index
, nr_pages
, spd
.pages
);
335 index
+= spd
.nr_pages
;
338 * If find_get_pages_contig() returned fewer pages than we needed,
339 * readahead/allocate the rest and fill in the holes.
341 if (spd
.nr_pages
< nr_pages
)
342 page_cache_sync_readahead(mapping
, &in
->f_ra
, in
,
343 index
, req_pages
- spd
.nr_pages
);
346 while (spd
.nr_pages
< nr_pages
) {
348 * Page could be there, find_get_pages_contig() breaks on
351 page
= find_get_page(mapping
, index
);
354 * page didn't exist, allocate one.
356 page
= page_cache_alloc_cold(mapping
);
360 error
= add_to_page_cache_lru(page
, mapping
, index
,
362 if (unlikely(error
)) {
363 page_cache_release(page
);
364 if (error
== -EEXIST
)
369 * add_to_page_cache() locks the page, unlock it
370 * to avoid convoluting the logic below even more.
375 spd
.pages
[spd
.nr_pages
++] = page
;
380 * Now loop over the map and see if we need to start IO on any
381 * pages, fill in the partial map, etc.
383 index
= *ppos
>> PAGE_CACHE_SHIFT
;
384 nr_pages
= spd
.nr_pages
;
386 for (page_nr
= 0; page_nr
< nr_pages
; page_nr
++) {
387 unsigned int this_len
;
393 * this_len is the max we'll use from this page
395 this_len
= min_t(unsigned long, len
, PAGE_CACHE_SIZE
- loff
);
396 page
= spd
.pages
[page_nr
];
398 if (PageReadahead(page
))
399 page_cache_async_readahead(mapping
, &in
->f_ra
, in
,
400 page
, index
, req_pages
- page_nr
);
403 * If the page isn't uptodate, we may need to start io on it
405 if (!PageUptodate(page
)) {
409 * Page was truncated, or invalidated by the
410 * filesystem. Redo the find/create, but this time the
411 * page is kept locked, so there's no chance of another
412 * race with truncate/invalidate.
414 if (!page
->mapping
) {
416 page
= find_or_create_page(mapping
, index
,
417 mapping_gfp_mask(mapping
));
423 page_cache_release(spd
.pages
[page_nr
]);
424 spd
.pages
[page_nr
] = page
;
427 * page was already under io and is now done, great
429 if (PageUptodate(page
)) {
435 * need to read in the page
437 error
= mapping
->a_ops
->readpage(in
, page
);
438 if (unlikely(error
)) {
440 * We really should re-lookup the page here,
441 * but it complicates things a lot. Instead
442 * lets just do what we already stored, and
443 * we'll get it the next time we are called.
445 if (error
== AOP_TRUNCATED_PAGE
)
453 * i_size must be checked after PageUptodate.
455 isize
= i_size_read(mapping
->host
);
456 end_index
= (isize
- 1) >> PAGE_CACHE_SHIFT
;
457 if (unlikely(!isize
|| index
> end_index
))
461 * if this is the last page, see if we need to shrink
462 * the length and stop
464 if (end_index
== index
) {
468 * max good bytes in this page
470 plen
= ((isize
- 1) & ~PAGE_CACHE_MASK
) + 1;
475 * force quit after adding this page
477 this_len
= min(this_len
, plen
- loff
);
481 spd
.partial
[page_nr
].offset
= loff
;
482 spd
.partial
[page_nr
].len
= this_len
;
490 * Release any pages at the end, if we quit early. 'page_nr' is how far
491 * we got, 'nr_pages' is how many pages are in the map.
493 while (page_nr
< nr_pages
)
494 page_cache_release(spd
.pages
[page_nr
++]);
495 in
->f_ra
.prev_pos
= (loff_t
)index
<< PAGE_CACHE_SHIFT
;
498 error
= splice_to_pipe(pipe
, &spd
);
500 splice_shrink_spd(pipe
, &spd
);
505 * generic_file_splice_read - splice data from file to a pipe
506 * @in: file to splice from
507 * @ppos: position in @in
508 * @pipe: pipe to splice to
509 * @len: number of bytes to splice
510 * @flags: splice modifier flags
513 * Will read pages from given file and fill them into a pipe. Can be
514 * used as long as the address_space operations for the source implements
518 ssize_t
generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
519 struct pipe_inode_info
*pipe
, size_t len
,
525 isize
= i_size_read(in
->f_mapping
->host
);
526 if (unlikely(*ppos
>= isize
))
529 left
= isize
- *ppos
;
530 if (unlikely(left
< len
))
533 ret
= __generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
541 EXPORT_SYMBOL(generic_file_splice_read
);
543 static const struct pipe_buf_operations default_pipe_buf_ops
= {
545 .map
= generic_pipe_buf_map
,
546 .unmap
= generic_pipe_buf_unmap
,
547 .confirm
= generic_pipe_buf_confirm
,
548 .release
= generic_pipe_buf_release
,
549 .steal
= generic_pipe_buf_steal
,
550 .get
= generic_pipe_buf_get
,
553 static ssize_t
kernel_readv(struct file
*file
, const struct iovec
*vec
,
554 unsigned long vlen
, loff_t offset
)
562 /* The cast to a user pointer is valid due to the set_fs() */
563 res
= vfs_readv(file
, (const struct iovec __user
*)vec
, vlen
, &pos
);
569 static ssize_t
kernel_write(struct file
*file
, const char *buf
, size_t count
,
577 /* The cast to a user pointer is valid due to the set_fs() */
578 res
= vfs_write(file
, (const char __user
*)buf
, count
, &pos
);
584 ssize_t
default_file_splice_read(struct file
*in
, loff_t
*ppos
,
585 struct pipe_inode_info
*pipe
, size_t len
,
588 unsigned int nr_pages
;
589 unsigned int nr_freed
;
591 struct page
*pages
[PIPE_DEF_BUFFERS
];
592 struct partial_page partial
[PIPE_DEF_BUFFERS
];
593 struct iovec
*vec
, __vec
[PIPE_DEF_BUFFERS
];
598 struct splice_pipe_desc spd
= {
602 .ops
= &default_pipe_buf_ops
,
603 .spd_release
= spd_release_page
,
606 if (splice_grow_spd(pipe
, &spd
))
611 if (pipe
->buffers
> PIPE_DEF_BUFFERS
) {
612 vec
= kmalloc(pipe
->buffers
* sizeof(struct iovec
), GFP_KERNEL
);
617 offset
= *ppos
& ~PAGE_CACHE_MASK
;
618 nr_pages
= (len
+ offset
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
620 for (i
= 0; i
< nr_pages
&& i
< pipe
->buffers
&& len
; i
++) {
623 page
= alloc_page(GFP_USER
);
628 this_len
= min_t(size_t, len
, PAGE_CACHE_SIZE
- offset
);
629 vec
[i
].iov_base
= (void __user
*) page_address(page
);
630 vec
[i
].iov_len
= this_len
;
637 res
= kernel_readv(in
, vec
, spd
.nr_pages
, *ppos
);
648 for (i
= 0; i
< spd
.nr_pages
; i
++) {
649 this_len
= min_t(size_t, vec
[i
].iov_len
, res
);
650 spd
.partial
[i
].offset
= 0;
651 spd
.partial
[i
].len
= this_len
;
653 __free_page(spd
.pages
[i
]);
659 spd
.nr_pages
-= nr_freed
;
661 res
= splice_to_pipe(pipe
, &spd
);
668 splice_shrink_spd(pipe
, &spd
);
672 for (i
= 0; i
< spd
.nr_pages
; i
++)
673 __free_page(spd
.pages
[i
]);
678 EXPORT_SYMBOL(default_file_splice_read
);
681 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
682 * using sendpage(). Return the number of bytes sent.
684 static int pipe_to_sendpage(struct pipe_inode_info
*pipe
,
685 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
687 struct file
*file
= sd
->u
.file
;
688 loff_t pos
= sd
->pos
;
691 if (!likely(file
->f_op
&& file
->f_op
->sendpage
))
694 more
= (sd
->flags
& SPLICE_F_MORE
) ? MSG_MORE
: 0;
695 if (sd
->len
< sd
->total_len
)
696 more
|= MSG_SENDPAGE_NOTLAST
;
697 return file
->f_op
->sendpage(file
, buf
->page
, buf
->offset
,
698 sd
->len
, &pos
, more
);
702 * This is a little more tricky than the file -> pipe splicing. There are
703 * basically three cases:
705 * - Destination page already exists in the address space and there
706 * are users of it. For that case we have no other option that
707 * copying the data. Tough luck.
708 * - Destination page already exists in the address space, but there
709 * are no users of it. Make sure it's uptodate, then drop it. Fall
710 * through to last case.
711 * - Destination page does not exist, we can add the pipe page to
712 * the page cache and avoid the copy.
714 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
715 * sd->flags), we attempt to migrate pages from the pipe to the output
716 * file address space page cache. This is possible if no one else has
717 * the pipe page referenced outside of the pipe and page cache. If
718 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
719 * a new page in the output file page cache and fill/dirty that.
721 int pipe_to_file(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
722 struct splice_desc
*sd
)
724 struct file
*file
= sd
->u
.file
;
725 struct address_space
*mapping
= file
->f_mapping
;
726 unsigned int offset
, this_len
;
731 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
734 if (this_len
+ offset
> PAGE_CACHE_SIZE
)
735 this_len
= PAGE_CACHE_SIZE
- offset
;
737 ret
= pagecache_write_begin(file
, mapping
, sd
->pos
, this_len
,
738 AOP_FLAG_UNINTERRUPTIBLE
, &page
, &fsdata
);
742 if (buf
->page
!= page
) {
744 * Careful, ->map() uses KM_USER0!
746 char *src
= buf
->ops
->map(pipe
, buf
, 1);
747 char *dst
= kmap_atomic(page
, KM_USER1
);
749 memcpy(dst
+ offset
, src
+ buf
->offset
, this_len
);
750 flush_dcache_page(page
);
751 kunmap_atomic(dst
, KM_USER1
);
752 buf
->ops
->unmap(pipe
, buf
, src
);
754 ret
= pagecache_write_end(file
, mapping
, sd
->pos
, this_len
, this_len
,
759 EXPORT_SYMBOL(pipe_to_file
);
761 static void wakeup_pipe_writers(struct pipe_inode_info
*pipe
)
764 if (waitqueue_active(&pipe
->wait
))
765 wake_up_interruptible(&pipe
->wait
);
766 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
770 * splice_from_pipe_feed - feed available data from a pipe to a file
771 * @pipe: pipe to splice from
772 * @sd: information to @actor
773 * @actor: handler that splices the data
776 * This function loops over the pipe and calls @actor to do the
777 * actual moving of a single struct pipe_buffer to the desired
778 * destination. It returns when there's no more buffers left in
779 * the pipe or if the requested number of bytes (@sd->total_len)
780 * have been copied. It returns a positive number (one) if the
781 * pipe needs to be filled with more data, zero if the required
782 * number of bytes have been copied and -errno on error.
784 * This, together with splice_from_pipe_{begin,end,next}, may be
785 * used to implement the functionality of __splice_from_pipe() when
786 * locking is required around copying the pipe buffers to the
789 int splice_from_pipe_feed(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
794 while (pipe
->nrbufs
) {
795 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
796 const struct pipe_buf_operations
*ops
= buf
->ops
;
799 if (sd
->len
> sd
->total_len
)
800 sd
->len
= sd
->total_len
;
802 ret
= buf
->ops
->confirm(pipe
, buf
);
809 ret
= actor(pipe
, buf
, sd
);
816 sd
->num_spliced
+= ret
;
819 sd
->total_len
-= ret
;
823 ops
->release(pipe
, buf
);
824 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (pipe
->buffers
- 1);
827 sd
->need_wakeup
= true;
836 EXPORT_SYMBOL(splice_from_pipe_feed
);
839 * splice_from_pipe_next - wait for some data to splice from
840 * @pipe: pipe to splice from
841 * @sd: information about the splice operation
844 * This function will wait for some data and return a positive
845 * value (one) if pipe buffers are available. It will return zero
846 * or -errno if no more data needs to be spliced.
848 int splice_from_pipe_next(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
850 while (!pipe
->nrbufs
) {
854 if (!pipe
->waiting_writers
&& sd
->num_spliced
)
857 if (sd
->flags
& SPLICE_F_NONBLOCK
)
860 if (signal_pending(current
))
863 if (sd
->need_wakeup
) {
864 wakeup_pipe_writers(pipe
);
865 sd
->need_wakeup
= false;
873 EXPORT_SYMBOL(splice_from_pipe_next
);
876 * splice_from_pipe_begin - start splicing from pipe
877 * @sd: information about the splice operation
880 * This function should be called before a loop containing
881 * splice_from_pipe_next() and splice_from_pipe_feed() to
882 * initialize the necessary fields of @sd.
884 void splice_from_pipe_begin(struct splice_desc
*sd
)
887 sd
->need_wakeup
= false;
889 EXPORT_SYMBOL(splice_from_pipe_begin
);
892 * splice_from_pipe_end - finish splicing from pipe
893 * @pipe: pipe to splice from
894 * @sd: information about the splice operation
897 * This function will wake up pipe writers if necessary. It should
898 * be called after a loop containing splice_from_pipe_next() and
899 * splice_from_pipe_feed().
901 void splice_from_pipe_end(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
904 wakeup_pipe_writers(pipe
);
906 EXPORT_SYMBOL(splice_from_pipe_end
);
909 * __splice_from_pipe - splice data from a pipe to given actor
910 * @pipe: pipe to splice from
911 * @sd: information to @actor
912 * @actor: handler that splices the data
915 * This function does little more than loop over the pipe and call
916 * @actor to do the actual moving of a single struct pipe_buffer to
917 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
921 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
926 splice_from_pipe_begin(sd
);
928 ret
= splice_from_pipe_next(pipe
, sd
);
930 ret
= splice_from_pipe_feed(pipe
, sd
, actor
);
932 splice_from_pipe_end(pipe
, sd
);
934 return sd
->num_spliced
? sd
->num_spliced
: ret
;
936 EXPORT_SYMBOL(__splice_from_pipe
);
939 * splice_from_pipe - splice data from a pipe to a file
940 * @pipe: pipe to splice from
941 * @out: file to splice to
942 * @ppos: position in @out
943 * @len: how many bytes to splice
944 * @flags: splice modifier flags
945 * @actor: handler that splices the data
948 * See __splice_from_pipe. This function locks the pipe inode,
949 * otherwise it's identical to __splice_from_pipe().
952 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
953 loff_t
*ppos
, size_t len
, unsigned int flags
,
957 struct splice_desc sd
= {
965 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
972 * generic_file_splice_write - splice data from a pipe to a file
974 * @out: file to write to
975 * @ppos: position in @out
976 * @len: number of bytes to splice
977 * @flags: splice modifier flags
980 * Will either move or copy pages (determined by @flags options) from
981 * the given pipe inode to the given file.
985 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
986 loff_t
*ppos
, size_t len
, unsigned int flags
)
988 struct address_space
*mapping
= out
->f_mapping
;
989 struct inode
*inode
= mapping
->host
;
990 struct splice_desc sd
= {
1000 splice_from_pipe_begin(&sd
);
1002 ret
= splice_from_pipe_next(pipe
, &sd
);
1006 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_CHILD
);
1007 ret
= file_remove_suid(out
);
1009 file_update_time(out
);
1010 ret
= splice_from_pipe_feed(pipe
, &sd
, pipe_to_file
);
1012 mutex_unlock(&inode
->i_mutex
);
1014 splice_from_pipe_end(pipe
, &sd
);
1019 ret
= sd
.num_spliced
;
1022 unsigned long nr_pages
;
1025 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
1027 err
= generic_write_sync(out
, *ppos
, ret
);
1032 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
1038 EXPORT_SYMBOL(generic_file_splice_write
);
1040 static int write_pipe_buf(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1041 struct splice_desc
*sd
)
1046 data
= buf
->ops
->map(pipe
, buf
, 0);
1047 ret
= kernel_write(sd
->u
.file
, data
+ buf
->offset
, sd
->len
, sd
->pos
);
1048 buf
->ops
->unmap(pipe
, buf
, data
);
1053 static ssize_t
default_file_splice_write(struct pipe_inode_info
*pipe
,
1054 struct file
*out
, loff_t
*ppos
,
1055 size_t len
, unsigned int flags
)
1059 ret
= splice_from_pipe(pipe
, out
, ppos
, len
, flags
, write_pipe_buf
);
1067 * generic_splice_sendpage - splice data from a pipe to a socket
1068 * @pipe: pipe to splice from
1069 * @out: socket to write to
1070 * @ppos: position in @out
1071 * @len: number of bytes to splice
1072 * @flags: splice modifier flags
1075 * Will send @len bytes from the pipe to a network socket. No data copying
1079 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
1080 loff_t
*ppos
, size_t len
, unsigned int flags
)
1082 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
1085 EXPORT_SYMBOL(generic_splice_sendpage
);
1088 * Attempt to initiate a splice from pipe to file.
1090 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
1091 loff_t
*ppos
, size_t len
, unsigned int flags
)
1093 ssize_t (*splice_write
)(struct pipe_inode_info
*, struct file
*,
1094 loff_t
*, size_t, unsigned int);
1097 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
1100 if (unlikely(out
->f_flags
& O_APPEND
))
1103 ret
= rw_verify_area(WRITE
, out
, ppos
, len
);
1104 if (unlikely(ret
< 0))
1107 if (out
->f_op
&& out
->f_op
->splice_write
)
1108 splice_write
= out
->f_op
->splice_write
;
1110 splice_write
= default_file_splice_write
;
1112 return splice_write(pipe
, out
, ppos
, len
, flags
);
1116 * Attempt to initiate a splice from a file to a pipe.
1118 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
1119 struct pipe_inode_info
*pipe
, size_t len
,
1122 ssize_t (*splice_read
)(struct file
*, loff_t
*,
1123 struct pipe_inode_info
*, size_t, unsigned int);
1126 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
1129 ret
= rw_verify_area(READ
, in
, ppos
, len
);
1130 if (unlikely(ret
< 0))
1133 if (in
->f_op
&& in
->f_op
->splice_read
)
1134 splice_read
= in
->f_op
->splice_read
;
1136 splice_read
= default_file_splice_read
;
1138 return splice_read(in
, ppos
, pipe
, len
, flags
);
1142 * splice_direct_to_actor - splices data directly between two non-pipes
1143 * @in: file to splice from
1144 * @sd: actor information on where to splice to
1145 * @actor: handles the data splicing
1148 * This is a special case helper to splice directly between two
1149 * points, without requiring an explicit pipe. Internally an allocated
1150 * pipe is cached in the process, and reused during the lifetime of
1154 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
1155 splice_direct_actor
*actor
)
1157 struct pipe_inode_info
*pipe
;
1164 * We require the input being a regular file, as we don't want to
1165 * randomly drop data for eg socket -> socket splicing. Use the
1166 * piped splicing for that!
1168 i_mode
= in
->f_path
.dentry
->d_inode
->i_mode
;
1169 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
1173 * neither in nor out is a pipe, setup an internal pipe attached to
1174 * 'out' and transfer the wanted data from 'in' to 'out' through that
1176 pipe
= current
->splice_pipe
;
1177 if (unlikely(!pipe
)) {
1178 pipe
= alloc_pipe_info(NULL
);
1183 * We don't have an immediate reader, but we'll read the stuff
1184 * out of the pipe right after the splice_to_pipe(). So set
1185 * PIPE_READERS appropriately.
1189 current
->splice_pipe
= pipe
;
1197 len
= sd
->total_len
;
1201 * Don't block on output, we have to drain the direct pipe.
1203 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1207 loff_t pos
= sd
->pos
, prev_pos
= pos
;
1209 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1210 if (unlikely(ret
<= 0))
1214 sd
->total_len
= read_len
;
1217 * NOTE: nonblocking mode only applies to the input. We
1218 * must not do the output in nonblocking mode as then we
1219 * could get stuck data in the internal pipe:
1221 ret
= actor(pipe
, sd
);
1222 if (unlikely(ret
<= 0)) {
1231 if (ret
< read_len
) {
1232 sd
->pos
= prev_pos
+ ret
;
1238 pipe
->nrbufs
= pipe
->curbuf
= 0;
1244 * If we did an incomplete transfer we must release
1245 * the pipe buffers in question:
1247 for (i
= 0; i
< pipe
->buffers
; i
++) {
1248 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1251 buf
->ops
->release(pipe
, buf
);
1261 EXPORT_SYMBOL(splice_direct_to_actor
);
1263 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1264 struct splice_desc
*sd
)
1266 struct file
*file
= sd
->u
.file
;
1268 return do_splice_from(pipe
, file
, &file
->f_pos
, sd
->total_len
,
1273 * do_splice_direct - splices data directly between two files
1274 * @in: file to splice from
1275 * @ppos: input file offset
1276 * @out: file to splice to
1277 * @len: number of bytes to splice
1278 * @flags: splice modifier flags
1281 * For use by do_sendfile(). splice can easily emulate sendfile, but
1282 * doing it in the application would incur an extra system call
1283 * (splice in + splice out, as compared to just sendfile()). So this helper
1284 * can splice directly through a process-private pipe.
1287 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1288 size_t len
, unsigned int flags
)
1290 struct splice_desc sd
= {
1299 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1306 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1307 struct pipe_inode_info
*opipe
,
1308 size_t len
, unsigned int flags
);
1311 * Determine where to splice to/from.
1313 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1314 struct file
*out
, loff_t __user
*off_out
,
1315 size_t len
, unsigned int flags
)
1317 struct pipe_inode_info
*ipipe
;
1318 struct pipe_inode_info
*opipe
;
1319 loff_t offset
, *off
;
1322 ipipe
= get_pipe_info(in
);
1323 opipe
= get_pipe_info(out
);
1325 if (ipipe
&& opipe
) {
1326 if (off_in
|| off_out
)
1329 if (!(in
->f_mode
& FMODE_READ
))
1332 if (!(out
->f_mode
& FMODE_WRITE
))
1335 /* Splicing to self would be fun, but... */
1339 return splice_pipe_to_pipe(ipipe
, opipe
, len
, flags
);
1346 if (!(out
->f_mode
& FMODE_PWRITE
))
1348 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1354 ret
= do_splice_from(ipipe
, out
, off
, len
, flags
);
1356 if (off_out
&& copy_to_user(off_out
, off
, sizeof(loff_t
)))
1366 if (!(in
->f_mode
& FMODE_PREAD
))
1368 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1374 ret
= do_splice_to(in
, off
, opipe
, len
, flags
);
1376 if (off_in
&& copy_to_user(off_in
, off
, sizeof(loff_t
)))
1386 * Map an iov into an array of pages and offset/length tupples. With the
1387 * partial_page structure, we can map several non-contiguous ranges into
1388 * our ones pages[] map instead of splitting that operation into pieces.
1389 * Could easily be exported as a generic helper for other users, in which
1390 * case one would probably want to add a 'max_nr_pages' parameter as well.
1392 static int get_iovec_page_array(const struct iovec __user
*iov
,
1393 unsigned int nr_vecs
, struct page
**pages
,
1394 struct partial_page
*partial
, int aligned
,
1395 unsigned int pipe_buffers
)
1397 int buffers
= 0, error
= 0;
1400 unsigned long off
, npages
;
1407 if (copy_from_user(&entry
, iov
, sizeof(entry
)))
1410 base
= entry
.iov_base
;
1411 len
= entry
.iov_len
;
1414 * Sanity check this iovec. 0 read succeeds.
1420 if (!access_ok(VERIFY_READ
, base
, len
))
1424 * Get this base offset and number of pages, then map
1425 * in the user pages.
1427 off
= (unsigned long) base
& ~PAGE_MASK
;
1430 * If asked for alignment, the offset must be zero and the
1431 * length a multiple of the PAGE_SIZE.
1434 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1437 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1438 if (npages
> pipe_buffers
- buffers
)
1439 npages
= pipe_buffers
- buffers
;
1441 error
= get_user_pages_fast((unsigned long)base
, npages
,
1442 0, &pages
[buffers
]);
1444 if (unlikely(error
<= 0))
1448 * Fill this contiguous range into the partial page map.
1450 for (i
= 0; i
< error
; i
++) {
1451 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1453 partial
[buffers
].offset
= off
;
1454 partial
[buffers
].len
= plen
;
1462 * We didn't complete this iov, stop here since it probably
1463 * means we have to move some of this into a pipe to
1464 * be able to continue.
1470 * Don't continue if we mapped fewer pages than we asked for,
1471 * or if we mapped the max number of pages that we have
1474 if (error
< npages
|| buffers
== pipe_buffers
)
1487 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1488 struct splice_desc
*sd
)
1494 * See if we can use the atomic maps, by prefaulting in the
1495 * pages and doing an atomic copy
1497 if (!fault_in_pages_writeable(sd
->u
.userptr
, sd
->len
)) {
1498 src
= buf
->ops
->map(pipe
, buf
, 1);
1499 ret
= __copy_to_user_inatomic(sd
->u
.userptr
, src
+ buf
->offset
,
1501 buf
->ops
->unmap(pipe
, buf
, src
);
1509 * No dice, use slow non-atomic map and copy
1511 src
= buf
->ops
->map(pipe
, buf
, 0);
1514 if (copy_to_user(sd
->u
.userptr
, src
+ buf
->offset
, sd
->len
))
1517 buf
->ops
->unmap(pipe
, buf
, src
);
1520 sd
->u
.userptr
+= ret
;
1525 * For lack of a better implementation, implement vmsplice() to userspace
1526 * as a simple copy of the pipes pages to the user iov.
1528 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*iov
,
1529 unsigned long nr_segs
, unsigned int flags
)
1531 struct pipe_inode_info
*pipe
;
1532 struct splice_desc sd
;
1537 pipe
= get_pipe_info(file
);
1549 * Get user address base and length for this iovec.
1551 error
= get_user(base
, &iov
->iov_base
);
1552 if (unlikely(error
))
1554 error
= get_user(len
, &iov
->iov_len
);
1555 if (unlikely(error
))
1559 * Sanity check this iovec. 0 read succeeds.
1563 if (unlikely(!base
)) {
1568 if (unlikely(!access_ok(VERIFY_WRITE
, base
, len
))) {
1576 sd
.u
.userptr
= base
;
1579 size
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1605 * vmsplice splices a user address range into a pipe. It can be thought of
1606 * as splice-from-memory, where the regular splice is splice-from-file (or
1607 * to file). In both cases the output is a pipe, naturally.
1609 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1610 unsigned long nr_segs
, unsigned int flags
)
1612 struct pipe_inode_info
*pipe
;
1613 struct page
*pages
[PIPE_DEF_BUFFERS
];
1614 struct partial_page partial
[PIPE_DEF_BUFFERS
];
1615 struct splice_pipe_desc spd
= {
1619 .ops
= &user_page_pipe_buf_ops
,
1620 .spd_release
= spd_release_page
,
1624 pipe
= get_pipe_info(file
);
1628 if (splice_grow_spd(pipe
, &spd
))
1631 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, spd
.pages
,
1632 spd
.partial
, flags
& SPLICE_F_GIFT
,
1634 if (spd
.nr_pages
<= 0)
1637 ret
= splice_to_pipe(pipe
, &spd
);
1639 splice_shrink_spd(pipe
, &spd
);
1644 * Note that vmsplice only really supports true splicing _from_ user memory
1645 * to a pipe, not the other way around. Splicing from user memory is a simple
1646 * operation that can be supported without any funky alignment restrictions
1647 * or nasty vm tricks. We simply map in the user memory and fill them into
1648 * a pipe. The reverse isn't quite as easy, though. There are two possible
1649 * solutions for that:
1651 * - memcpy() the data internally, at which point we might as well just
1652 * do a regular read() on the buffer anyway.
1653 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1654 * has restriction limitations on both ends of the pipe).
1656 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1659 SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct iovec __user
*, iov
,
1660 unsigned long, nr_segs
, unsigned int, flags
)
1666 if (unlikely(nr_segs
> UIO_MAXIOV
))
1668 else if (unlikely(!nr_segs
))
1672 file
= fget_light(fd
, &fput
);
1674 if (file
->f_mode
& FMODE_WRITE
)
1675 error
= vmsplice_to_pipe(file
, iov
, nr_segs
, flags
);
1676 else if (file
->f_mode
& FMODE_READ
)
1677 error
= vmsplice_to_user(file
, iov
, nr_segs
, flags
);
1679 fput_light(file
, fput
);
1685 SYSCALL_DEFINE6(splice
, int, fd_in
, loff_t __user
*, off_in
,
1686 int, fd_out
, loff_t __user
*, off_out
,
1687 size_t, len
, unsigned int, flags
)
1690 struct file
*in
, *out
;
1691 int fput_in
, fput_out
;
1697 in
= fget_light(fd_in
, &fput_in
);
1699 if (in
->f_mode
& FMODE_READ
) {
1700 out
= fget_light(fd_out
, &fput_out
);
1702 if (out
->f_mode
& FMODE_WRITE
)
1703 error
= do_splice(in
, off_in
,
1706 fput_light(out
, fput_out
);
1710 fput_light(in
, fput_in
);
1717 * Make sure there's data to read. Wait for input if we can, otherwise
1718 * return an appropriate error.
1720 static int ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1725 * Check ->nrbufs without the inode lock first. This function
1726 * is speculative anyways, so missing one is ok.
1734 while (!pipe
->nrbufs
) {
1735 if (signal_pending(current
)) {
1741 if (!pipe
->waiting_writers
) {
1742 if (flags
& SPLICE_F_NONBLOCK
) {
1755 * Make sure there's writeable room. Wait for room if we can, otherwise
1756 * return an appropriate error.
1758 static int opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1763 * Check ->nrbufs without the inode lock first. This function
1764 * is speculative anyways, so missing one is ok.
1766 if (pipe
->nrbufs
< pipe
->buffers
)
1772 while (pipe
->nrbufs
>= pipe
->buffers
) {
1773 if (!pipe
->readers
) {
1774 send_sig(SIGPIPE
, current
, 0);
1778 if (flags
& SPLICE_F_NONBLOCK
) {
1782 if (signal_pending(current
)) {
1786 pipe
->waiting_writers
++;
1788 pipe
->waiting_writers
--;
1796 * Splice contents of ipipe to opipe.
1798 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1799 struct pipe_inode_info
*opipe
,
1800 size_t len
, unsigned int flags
)
1802 struct pipe_buffer
*ibuf
, *obuf
;
1804 bool input_wakeup
= false;
1808 ret
= ipipe_prep(ipipe
, flags
);
1812 ret
= opipe_prep(opipe
, flags
);
1817 * Potential ABBA deadlock, work around it by ordering lock
1818 * grabbing by pipe info address. Otherwise two different processes
1819 * could deadlock (one doing tee from A -> B, the other from B -> A).
1821 pipe_double_lock(ipipe
, opipe
);
1824 if (!opipe
->readers
) {
1825 send_sig(SIGPIPE
, current
, 0);
1831 if (!ipipe
->nrbufs
&& !ipipe
->writers
)
1835 * Cannot make any progress, because either the input
1836 * pipe is empty or the output pipe is full.
1838 if (!ipipe
->nrbufs
|| opipe
->nrbufs
>= opipe
->buffers
) {
1839 /* Already processed some buffers, break */
1843 if (flags
& SPLICE_F_NONBLOCK
) {
1849 * We raced with another reader/writer and haven't
1850 * managed to process any buffers. A zero return
1851 * value means EOF, so retry instead.
1858 ibuf
= ipipe
->bufs
+ ipipe
->curbuf
;
1859 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (opipe
->buffers
- 1);
1860 obuf
= opipe
->bufs
+ nbuf
;
1862 if (len
>= ibuf
->len
) {
1864 * Simply move the whole buffer from ipipe to opipe
1869 ipipe
->curbuf
= (ipipe
->curbuf
+ 1) & (ipipe
->buffers
- 1);
1871 input_wakeup
= true;
1874 * Get a reference to this pipe buffer,
1875 * so we can copy the contents over.
1877 ibuf
->ops
->get(ipipe
, ibuf
);
1881 * Don't inherit the gift flag, we need to
1882 * prevent multiple steals of this page.
1884 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1888 ibuf
->offset
+= obuf
->len
;
1889 ibuf
->len
-= obuf
->len
;
1899 * If we put data in the output pipe, wakeup any potential readers.
1902 wakeup_pipe_readers(opipe
);
1905 wakeup_pipe_writers(ipipe
);
1911 * Link contents of ipipe to opipe.
1913 static int link_pipe(struct pipe_inode_info
*ipipe
,
1914 struct pipe_inode_info
*opipe
,
1915 size_t len
, unsigned int flags
)
1917 struct pipe_buffer
*ibuf
, *obuf
;
1918 int ret
= 0, i
= 0, nbuf
;
1921 * Potential ABBA deadlock, work around it by ordering lock
1922 * grabbing by pipe info address. Otherwise two different processes
1923 * could deadlock (one doing tee from A -> B, the other from B -> A).
1925 pipe_double_lock(ipipe
, opipe
);
1928 if (!opipe
->readers
) {
1929 send_sig(SIGPIPE
, current
, 0);
1936 * If we have iterated all input buffers or ran out of
1937 * output room, break.
1939 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= opipe
->buffers
)
1942 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (ipipe
->buffers
-1));
1943 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (opipe
->buffers
- 1);
1946 * Get a reference to this pipe buffer,
1947 * so we can copy the contents over.
1949 ibuf
->ops
->get(ipipe
, ibuf
);
1951 obuf
= opipe
->bufs
+ nbuf
;
1955 * Don't inherit the gift flag, we need to
1956 * prevent multiple steals of this page.
1958 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1960 if (obuf
->len
> len
)
1970 * return EAGAIN if we have the potential of some data in the
1971 * future, otherwise just return 0
1973 if (!ret
&& ipipe
->waiting_writers
&& (flags
& SPLICE_F_NONBLOCK
))
1980 * If we put data in the output pipe, wakeup any potential readers.
1983 wakeup_pipe_readers(opipe
);
1989 * This is a tee(1) implementation that works on pipes. It doesn't copy
1990 * any data, it simply references the 'in' pages on the 'out' pipe.
1991 * The 'flags' used are the SPLICE_F_* variants, currently the only
1992 * applicable one is SPLICE_F_NONBLOCK.
1994 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1997 struct pipe_inode_info
*ipipe
= get_pipe_info(in
);
1998 struct pipe_inode_info
*opipe
= get_pipe_info(out
);
2002 * Duplicate the contents of ipipe to opipe without actually
2005 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
2007 * Keep going, unless we encounter an error. The ipipe/opipe
2008 * ordering doesn't really matter.
2010 ret
= ipipe_prep(ipipe
, flags
);
2012 ret
= opipe_prep(opipe
, flags
);
2014 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
2021 SYSCALL_DEFINE4(tee
, int, fdin
, int, fdout
, size_t, len
, unsigned int, flags
)
2030 in
= fget_light(fdin
, &fput_in
);
2032 if (in
->f_mode
& FMODE_READ
) {
2034 struct file
*out
= fget_light(fdout
, &fput_out
);
2037 if (out
->f_mode
& FMODE_WRITE
)
2038 error
= do_tee(in
, out
, len
, flags
);
2039 fput_light(out
, fput_out
);
2042 fput_light(in
, fput_in
);