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/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
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.
40 static int page_cache_pipe_buf_steal(struct pipe_inode_info
*pipe
,
41 struct pipe_buffer
*buf
)
43 struct page
*page
= buf
->page
;
44 struct address_space
*mapping
;
48 mapping
= page_mapping(page
);
50 WARN_ON(!PageUptodate(page
));
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
60 wait_on_page_writeback(page
);
62 if (page_has_private(page
) &&
63 !try_to_release_page(page
, GFP_KERNEL
))
67 * If we succeeded in removing the mapping, set LRU flag
70 if (remove_mapping(mapping
, page
)) {
71 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
77 * Raced with truncate or failed to remove page from current
78 * address space, unlock and return failure.
85 static void page_cache_pipe_buf_release(struct pipe_inode_info
*pipe
,
86 struct pipe_buffer
*buf
)
88 page_cache_release(buf
->page
);
89 buf
->flags
&= ~PIPE_BUF_FLAG_LRU
;
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.
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info
*pipe
,
97 struct pipe_buffer
*buf
)
99 struct page
*page
= buf
->page
;
102 if (!PageUptodate(page
)) {
106 * Page got truncated/unhashed. This will cause a 0-byte
107 * splice, if this is the first page.
109 if (!page
->mapping
) {
115 * Uh oh, read-error from disk.
117 if (!PageUptodate(page
)) {
123 * Page is ok afterall, we are done.
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops
= {
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
,
144 static int user_page_pipe_buf_steal(struct pipe_inode_info
*pipe
,
145 struct pipe_buffer
*buf
)
147 if (!(buf
->flags
& PIPE_BUF_FLAG_GIFT
))
150 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
151 return generic_pipe_buf_steal(pipe
, buf
);
154 static const struct pipe_buf_operations user_page_pipe_buf_ops
= {
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
,
165 * splice_to_pipe - fill passed data into a pipe
166 * @pipe: pipe to fill
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.
175 ssize_t
splice_to_pipe(struct pipe_inode_info
*pipe
,
176 struct splice_pipe_desc
*spd
)
178 unsigned int spd_pages
= spd
->nr_pages
;
179 int ret
, do_wakeup
, page_nr
;
188 if (!pipe
->readers
) {
189 send_sig(SIGPIPE
, current
, 0);
195 if (pipe
->nrbufs
< PIPE_BUFFERS
) {
196 int newbuf
= (pipe
->curbuf
+ pipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
197 struct pipe_buffer
*buf
= pipe
->bufs
+ newbuf
;
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;
204 if (spd
->flags
& SPLICE_F_GIFT
)
205 buf
->flags
|= PIPE_BUF_FLAG_GIFT
;
214 if (!--spd
->nr_pages
)
216 if (pipe
->nrbufs
< PIPE_BUFFERS
)
222 if (spd
->flags
& SPLICE_F_NONBLOCK
) {
228 if (signal_pending(current
)) {
236 if (waitqueue_active(&pipe
->wait
))
237 wake_up_interruptible_sync(&pipe
->wait
);
238 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
242 pipe
->waiting_writers
++;
244 pipe
->waiting_writers
--;
251 if (waitqueue_active(&pipe
->wait
))
252 wake_up_interruptible(&pipe
->wait
);
253 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
256 while (page_nr
< spd_pages
)
257 spd
->spd_release(spd
, page_nr
++);
262 static void spd_release_page(struct splice_pipe_desc
*spd
, unsigned int i
)
264 page_cache_release(spd
->pages
[i
]);
268 __generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
269 struct pipe_inode_info
*pipe
, size_t len
,
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
];
277 pgoff_t index
, end_index
;
280 struct splice_pipe_desc spd
= {
284 .ops
= &page_cache_pipe_buf_ops
,
285 .spd_release
= spd_release_page
,
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
);
294 * Lookup the (hopefully) full range of pages we need.
296 spd
.nr_pages
= find_get_pages_contig(mapping
, index
, nr_pages
, pages
);
297 index
+= spd
.nr_pages
;
300 * If find_get_pages_contig() returned fewer pages than we needed,
301 * readahead/allocate the rest and fill in the holes.
303 if (spd
.nr_pages
< nr_pages
)
304 page_cache_sync_readahead(mapping
, &in
->f_ra
, in
,
305 index
, req_pages
- spd
.nr_pages
);
308 while (spd
.nr_pages
< nr_pages
) {
310 * Page could be there, find_get_pages_contig() breaks on
313 page
= find_get_page(mapping
, index
);
316 * page didn't exist, allocate one.
318 page
= page_cache_alloc_cold(mapping
);
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
)
331 * add_to_page_cache() locks the page, unlock it
332 * to avoid convoluting the logic below even more.
337 pages
[spd
.nr_pages
++] = page
;
342 * Now loop over the map and see if we need to start IO on any
343 * pages, fill in the partial map, etc.
345 index
= *ppos
>> PAGE_CACHE_SHIFT
;
346 nr_pages
= spd
.nr_pages
;
348 for (page_nr
= 0; page_nr
< nr_pages
; page_nr
++) {
349 unsigned int this_len
;
355 * this_len is the max we'll use from this page
357 this_len
= min_t(unsigned long, len
, PAGE_CACHE_SIZE
- loff
);
358 page
= pages
[page_nr
];
360 if (PageReadahead(page
))
361 page_cache_async_readahead(mapping
, &in
->f_ra
, in
,
362 page
, index
, req_pages
- page_nr
);
365 * If the page isn't uptodate, we may need to start io on it
367 if (!PageUptodate(page
)) {
369 * If in nonblock mode then dont block on waiting
370 * for an in-flight io page
372 if (flags
& SPLICE_F_NONBLOCK
) {
373 if (!trylock_page(page
)) {
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.
386 if (!page
->mapping
) {
388 page
= find_or_create_page(mapping
, index
,
389 mapping_gfp_mask(mapping
));
395 page_cache_release(pages
[page_nr
]);
396 pages
[page_nr
] = page
;
399 * page was already under io and is now done, great
401 if (PageUptodate(page
)) {
407 * need to read in the page
409 error
= mapping
->a_ops
->readpage(in
, page
);
410 if (unlikely(error
)) {
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.
417 if (error
== AOP_TRUNCATED_PAGE
)
425 * i_size must be checked after PageUptodate.
427 isize
= i_size_read(mapping
->host
);
428 end_index
= (isize
- 1) >> PAGE_CACHE_SHIFT
;
429 if (unlikely(!isize
|| index
> end_index
))
433 * if this is the last page, see if we need to shrink
434 * the length and stop
436 if (end_index
== index
) {
440 * max good bytes in this page
442 plen
= ((isize
- 1) & ~PAGE_CACHE_MASK
) + 1;
447 * force quit after adding this page
449 this_len
= min(this_len
, plen
- loff
);
453 partial
[page_nr
].offset
= loff
;
454 partial
[page_nr
].len
= this_len
;
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.
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
;
470 return splice_to_pipe(pipe
, &spd
);
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
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
489 ssize_t
generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
490 struct pipe_inode_info
*pipe
, size_t len
,
496 isize
= i_size_read(in
->f_mapping
->host
);
497 if (unlikely(*ppos
>= isize
))
500 left
= isize
- *ppos
;
501 if (unlikely(left
< len
))
504 ret
= __generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
510 EXPORT_SYMBOL(generic_file_splice_read
);
512 static const struct pipe_buf_operations default_pipe_buf_ops
= {
514 .map
= generic_pipe_buf_map
,
515 .unmap
= generic_pipe_buf_unmap
,
516 .confirm
= generic_pipe_buf_confirm
,
517 .release
= generic_pipe_buf_release
,
518 .steal
= generic_pipe_buf_steal
,
519 .get
= generic_pipe_buf_get
,
522 static ssize_t
kernel_readv(struct file
*file
, const struct iovec
*vec
,
523 unsigned long vlen
, loff_t offset
)
531 /* The cast to a user pointer is valid due to the set_fs() */
532 res
= vfs_readv(file
, (const struct iovec __user
*)vec
, vlen
, &pos
);
538 ssize_t
default_file_splice_read(struct file
*in
, loff_t
*ppos
,
539 struct pipe_inode_info
*pipe
, size_t len
,
542 unsigned int nr_pages
;
543 unsigned int nr_freed
;
545 struct page
*pages
[PIPE_BUFFERS
];
546 struct partial_page partial
[PIPE_BUFFERS
];
547 struct iovec vec
[PIPE_BUFFERS
];
553 struct splice_pipe_desc spd
= {
557 .ops
= &default_pipe_buf_ops
,
558 .spd_release
= spd_release_page
,
561 index
= *ppos
>> PAGE_CACHE_SHIFT
;
562 offset
= *ppos
& ~PAGE_CACHE_MASK
;
563 nr_pages
= (len
+ offset
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
565 for (i
= 0; i
< nr_pages
&& i
< PIPE_BUFFERS
&& len
; i
++) {
568 page
= alloc_page(GFP_HIGHUSER
);
573 this_len
= min_t(size_t, len
, PAGE_CACHE_SIZE
- offset
);
574 vec
[i
].iov_base
= (void __user
*) kmap(page
);
575 vec
[i
].iov_len
= this_len
;
582 res
= kernel_readv(in
, vec
, spd
.nr_pages
, *ppos
);
591 for (i
= 0; i
< spd
.nr_pages
; i
++) {
593 this_len
= min_t(size_t, vec
[i
].iov_len
, res
);
594 partial
[i
].offset
= 0;
595 partial
[i
].len
= this_len
;
597 __free_page(pages
[i
]);
603 spd
.nr_pages
-= nr_freed
;
605 res
= splice_to_pipe(pipe
, &spd
);
612 for (i
= 0; i
< spd
.nr_pages
; i
++) {
614 __free_page(pages
[i
]);
618 EXPORT_SYMBOL(default_file_splice_read
);
621 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
622 * using sendpage(). Return the number of bytes sent.
624 static int pipe_to_sendpage(struct pipe_inode_info
*pipe
,
625 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
627 struct file
*file
= sd
->u
.file
;
628 loff_t pos
= sd
->pos
;
631 ret
= buf
->ops
->confirm(pipe
, buf
);
633 more
= (sd
->flags
& SPLICE_F_MORE
) || sd
->len
< sd
->total_len
;
635 ret
= file
->f_op
->sendpage(file
, buf
->page
, buf
->offset
,
636 sd
->len
, &pos
, more
);
643 * This is a little more tricky than the file -> pipe splicing. There are
644 * basically three cases:
646 * - Destination page already exists in the address space and there
647 * are users of it. For that case we have no other option that
648 * copying the data. Tough luck.
649 * - Destination page already exists in the address space, but there
650 * are no users of it. Make sure it's uptodate, then drop it. Fall
651 * through to last case.
652 * - Destination page does not exist, we can add the pipe page to
653 * the page cache and avoid the copy.
655 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
656 * sd->flags), we attempt to migrate pages from the pipe to the output
657 * file address space page cache. This is possible if no one else has
658 * the pipe page referenced outside of the pipe and page cache. If
659 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
660 * a new page in the output file page cache and fill/dirty that.
662 int pipe_to_file(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
663 struct splice_desc
*sd
)
665 struct file
*file
= sd
->u
.file
;
666 struct address_space
*mapping
= file
->f_mapping
;
667 unsigned int offset
, this_len
;
673 * make sure the data in this buffer is uptodate
675 ret
= buf
->ops
->confirm(pipe
, buf
);
679 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
682 if (this_len
+ offset
> PAGE_CACHE_SIZE
)
683 this_len
= PAGE_CACHE_SIZE
- offset
;
685 ret
= pagecache_write_begin(file
, mapping
, sd
->pos
, this_len
,
686 AOP_FLAG_UNINTERRUPTIBLE
, &page
, &fsdata
);
690 if (buf
->page
!= page
) {
692 * Careful, ->map() uses KM_USER0!
694 char *src
= buf
->ops
->map(pipe
, buf
, 1);
695 char *dst
= kmap_atomic(page
, KM_USER1
);
697 memcpy(dst
+ offset
, src
+ buf
->offset
, this_len
);
698 flush_dcache_page(page
);
699 kunmap_atomic(dst
, KM_USER1
);
700 buf
->ops
->unmap(pipe
, buf
, src
);
702 ret
= pagecache_write_end(file
, mapping
, sd
->pos
, this_len
, this_len
,
707 EXPORT_SYMBOL(pipe_to_file
);
709 static void wakeup_pipe_writers(struct pipe_inode_info
*pipe
)
712 if (waitqueue_active(&pipe
->wait
))
713 wake_up_interruptible(&pipe
->wait
);
714 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
718 * splice_from_pipe_feed - feed available data from a pipe to a file
719 * @pipe: pipe to splice from
720 * @sd: information to @actor
721 * @actor: handler that splices the data
724 * This function loops over the pipe and calls @actor to do the
725 * actual moving of a single struct pipe_buffer to the desired
726 * destination. It returns when there's no more buffers left in
727 * the pipe or if the requested number of bytes (@sd->total_len)
728 * have been copied. It returns a positive number (one) if the
729 * pipe needs to be filled with more data, zero if the required
730 * number of bytes have been copied and -errno on error.
732 * This, together with splice_from_pipe_{begin,end,next}, may be
733 * used to implement the functionality of __splice_from_pipe() when
734 * locking is required around copying the pipe buffers to the
737 int splice_from_pipe_feed(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
742 while (pipe
->nrbufs
) {
743 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
744 const struct pipe_buf_operations
*ops
= buf
->ops
;
747 if (sd
->len
> sd
->total_len
)
748 sd
->len
= sd
->total_len
;
750 ret
= actor(pipe
, buf
, sd
);
759 sd
->num_spliced
+= ret
;
762 sd
->total_len
-= ret
;
766 ops
->release(pipe
, buf
);
767 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (PIPE_BUFFERS
- 1);
770 sd
->need_wakeup
= true;
779 EXPORT_SYMBOL(splice_from_pipe_feed
);
782 * splice_from_pipe_next - wait for some data to splice from
783 * @pipe: pipe to splice from
784 * @sd: information about the splice operation
787 * This function will wait for some data and return a positive
788 * value (one) if pipe buffers are available. It will return zero
789 * or -errno if no more data needs to be spliced.
791 int splice_from_pipe_next(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
793 while (!pipe
->nrbufs
) {
797 if (!pipe
->waiting_writers
&& sd
->num_spliced
)
800 if (sd
->flags
& SPLICE_F_NONBLOCK
)
803 if (signal_pending(current
))
806 if (sd
->need_wakeup
) {
807 wakeup_pipe_writers(pipe
);
808 sd
->need_wakeup
= false;
816 EXPORT_SYMBOL(splice_from_pipe_next
);
819 * splice_from_pipe_begin - start splicing from pipe
820 * @sd: information about the splice operation
823 * This function should be called before a loop containing
824 * splice_from_pipe_next() and splice_from_pipe_feed() to
825 * initialize the necessary fields of @sd.
827 void splice_from_pipe_begin(struct splice_desc
*sd
)
830 sd
->need_wakeup
= false;
832 EXPORT_SYMBOL(splice_from_pipe_begin
);
835 * splice_from_pipe_end - finish splicing from pipe
836 * @pipe: pipe to splice from
837 * @sd: information about the splice operation
840 * This function will wake up pipe writers if necessary. It should
841 * be called after a loop containing splice_from_pipe_next() and
842 * splice_from_pipe_feed().
844 void splice_from_pipe_end(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
847 wakeup_pipe_writers(pipe
);
849 EXPORT_SYMBOL(splice_from_pipe_end
);
852 * __splice_from_pipe - splice data from a pipe to given actor
853 * @pipe: pipe to splice from
854 * @sd: information to @actor
855 * @actor: handler that splices the data
858 * This function does little more than loop over the pipe and call
859 * @actor to do the actual moving of a single struct pipe_buffer to
860 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
864 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
869 splice_from_pipe_begin(sd
);
871 ret
= splice_from_pipe_next(pipe
, sd
);
873 ret
= splice_from_pipe_feed(pipe
, sd
, actor
);
875 splice_from_pipe_end(pipe
, sd
);
877 return sd
->num_spliced
? sd
->num_spliced
: ret
;
879 EXPORT_SYMBOL(__splice_from_pipe
);
882 * splice_from_pipe - splice data from a pipe to a file
883 * @pipe: pipe to splice from
884 * @out: file to splice to
885 * @ppos: position in @out
886 * @len: how many bytes to splice
887 * @flags: splice modifier flags
888 * @actor: handler that splices the data
891 * See __splice_from_pipe. This function locks the pipe inode,
892 * otherwise it's identical to __splice_from_pipe().
895 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
896 loff_t
*ppos
, size_t len
, unsigned int flags
,
900 struct splice_desc sd
= {
908 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
915 * generic_file_splice_write - splice data from a pipe to a file
917 * @out: file to write to
918 * @ppos: position in @out
919 * @len: number of bytes to splice
920 * @flags: splice modifier flags
923 * Will either move or copy pages (determined by @flags options) from
924 * the given pipe inode to the given file.
928 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
929 loff_t
*ppos
, size_t len
, unsigned int flags
)
931 struct address_space
*mapping
= out
->f_mapping
;
932 struct inode
*inode
= mapping
->host
;
933 struct splice_desc sd
= {
943 splice_from_pipe_begin(&sd
);
945 ret
= splice_from_pipe_next(pipe
, &sd
);
949 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_CHILD
);
950 ret
= file_remove_suid(out
);
952 ret
= splice_from_pipe_feed(pipe
, &sd
, pipe_to_file
);
953 mutex_unlock(&inode
->i_mutex
);
955 splice_from_pipe_end(pipe
, &sd
);
960 ret
= sd
.num_spliced
;
963 unsigned long nr_pages
;
966 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
969 * If file or inode is SYNC and we actually wrote some data,
972 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
975 mutex_lock(&inode
->i_mutex
);
976 err
= generic_osync_inode(inode
, mapping
,
977 OSYNC_METADATA
|OSYNC_DATA
);
978 mutex_unlock(&inode
->i_mutex
);
983 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
989 EXPORT_SYMBOL(generic_file_splice_write
);
992 * generic_splice_sendpage - splice data from a pipe to a socket
993 * @pipe: pipe to splice from
994 * @out: socket to write to
995 * @ppos: position in @out
996 * @len: number of bytes to splice
997 * @flags: splice modifier flags
1000 * Will send @len bytes from the pipe to a network socket. No data copying
1004 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
1005 loff_t
*ppos
, size_t len
, unsigned int flags
)
1007 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
1010 EXPORT_SYMBOL(generic_splice_sendpage
);
1013 * Attempt to initiate a splice from pipe to file.
1015 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
1016 loff_t
*ppos
, size_t len
, unsigned int flags
)
1020 if (unlikely(!out
->f_op
|| !out
->f_op
->splice_write
))
1023 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
1026 if (unlikely(out
->f_flags
& O_APPEND
))
1029 ret
= rw_verify_area(WRITE
, out
, ppos
, len
);
1030 if (unlikely(ret
< 0))
1033 return out
->f_op
->splice_write(pipe
, out
, ppos
, len
, flags
);
1037 * Attempt to initiate a splice from a file to a pipe.
1039 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
1040 struct pipe_inode_info
*pipe
, size_t len
,
1043 ssize_t (*splice_read
)(struct file
*, loff_t
*,
1044 struct pipe_inode_info
*, size_t, unsigned int);
1047 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
1050 ret
= rw_verify_area(READ
, in
, ppos
, len
);
1051 if (unlikely(ret
< 0))
1054 splice_read
= in
->f_op
->splice_read
;
1056 splice_read
= default_file_splice_read
;
1058 return splice_read(in
, ppos
, pipe
, len
, flags
);
1062 * splice_direct_to_actor - splices data directly between two non-pipes
1063 * @in: file to splice from
1064 * @sd: actor information on where to splice to
1065 * @actor: handles the data splicing
1068 * This is a special case helper to splice directly between two
1069 * points, without requiring an explicit pipe. Internally an allocated
1070 * pipe is cached in the process, and reused during the lifetime of
1074 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
1075 splice_direct_actor
*actor
)
1077 struct pipe_inode_info
*pipe
;
1084 * We require the input being a regular file, as we don't want to
1085 * randomly drop data for eg socket -> socket splicing. Use the
1086 * piped splicing for that!
1088 i_mode
= in
->f_path
.dentry
->d_inode
->i_mode
;
1089 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
1093 * neither in nor out is a pipe, setup an internal pipe attached to
1094 * 'out' and transfer the wanted data from 'in' to 'out' through that
1096 pipe
= current
->splice_pipe
;
1097 if (unlikely(!pipe
)) {
1098 pipe
= alloc_pipe_info(NULL
);
1103 * We don't have an immediate reader, but we'll read the stuff
1104 * out of the pipe right after the splice_to_pipe(). So set
1105 * PIPE_READERS appropriately.
1109 current
->splice_pipe
= pipe
;
1117 len
= sd
->total_len
;
1121 * Don't block on output, we have to drain the direct pipe.
1123 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1127 loff_t pos
= sd
->pos
, prev_pos
= pos
;
1129 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1130 if (unlikely(ret
<= 0))
1134 sd
->total_len
= read_len
;
1137 * NOTE: nonblocking mode only applies to the input. We
1138 * must not do the output in nonblocking mode as then we
1139 * could get stuck data in the internal pipe:
1141 ret
= actor(pipe
, sd
);
1142 if (unlikely(ret
<= 0)) {
1151 if (ret
< read_len
) {
1152 sd
->pos
= prev_pos
+ ret
;
1158 pipe
->nrbufs
= pipe
->curbuf
= 0;
1164 * If we did an incomplete transfer we must release
1165 * the pipe buffers in question:
1167 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
1168 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1171 buf
->ops
->release(pipe
, buf
);
1181 EXPORT_SYMBOL(splice_direct_to_actor
);
1183 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1184 struct splice_desc
*sd
)
1186 struct file
*file
= sd
->u
.file
;
1188 return do_splice_from(pipe
, file
, &sd
->pos
, sd
->total_len
, sd
->flags
);
1192 * do_splice_direct - splices data directly between two files
1193 * @in: file to splice from
1194 * @ppos: input file offset
1195 * @out: file to splice to
1196 * @len: number of bytes to splice
1197 * @flags: splice modifier flags
1200 * For use by do_sendfile(). splice can easily emulate sendfile, but
1201 * doing it in the application would incur an extra system call
1202 * (splice in + splice out, as compared to just sendfile()). So this helper
1203 * can splice directly through a process-private pipe.
1206 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1207 size_t len
, unsigned int flags
)
1209 struct splice_desc sd
= {
1218 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1225 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1226 struct pipe_inode_info
*opipe
,
1227 size_t len
, unsigned int flags
);
1229 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1230 * location, so checking ->i_pipe is not enough to verify that this is a
1233 static inline struct pipe_inode_info
*pipe_info(struct inode
*inode
)
1235 if (S_ISFIFO(inode
->i_mode
))
1236 return inode
->i_pipe
;
1242 * Determine where to splice to/from.
1244 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1245 struct file
*out
, loff_t __user
*off_out
,
1246 size_t len
, unsigned int flags
)
1248 struct pipe_inode_info
*ipipe
;
1249 struct pipe_inode_info
*opipe
;
1250 loff_t offset
, *off
;
1253 ipipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1254 opipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1256 if (ipipe
&& opipe
) {
1257 if (off_in
|| off_out
)
1260 if (!(in
->f_mode
& FMODE_READ
))
1263 if (!(out
->f_mode
& FMODE_WRITE
))
1266 /* Splicing to self would be fun, but... */
1270 return splice_pipe_to_pipe(ipipe
, opipe
, len
, flags
);
1277 if (out
->f_op
->llseek
== no_llseek
)
1279 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1285 ret
= do_splice_from(ipipe
, out
, off
, len
, flags
);
1287 if (off_out
&& copy_to_user(off_out
, off
, sizeof(loff_t
)))
1297 if (in
->f_op
->llseek
== no_llseek
)
1299 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1305 ret
= do_splice_to(in
, off
, opipe
, len
, flags
);
1307 if (off_in
&& copy_to_user(off_in
, off
, sizeof(loff_t
)))
1317 * Map an iov into an array of pages and offset/length tupples. With the
1318 * partial_page structure, we can map several non-contiguous ranges into
1319 * our ones pages[] map instead of splitting that operation into pieces.
1320 * Could easily be exported as a generic helper for other users, in which
1321 * case one would probably want to add a 'max_nr_pages' parameter as well.
1323 static int get_iovec_page_array(const struct iovec __user
*iov
,
1324 unsigned int nr_vecs
, struct page
**pages
,
1325 struct partial_page
*partial
, int aligned
)
1327 int buffers
= 0, error
= 0;
1330 unsigned long off
, npages
;
1337 if (copy_from_user(&entry
, iov
, sizeof(entry
)))
1340 base
= entry
.iov_base
;
1341 len
= entry
.iov_len
;
1344 * Sanity check this iovec. 0 read succeeds.
1350 if (!access_ok(VERIFY_READ
, base
, len
))
1354 * Get this base offset and number of pages, then map
1355 * in the user pages.
1357 off
= (unsigned long) base
& ~PAGE_MASK
;
1360 * If asked for alignment, the offset must be zero and the
1361 * length a multiple of the PAGE_SIZE.
1364 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1367 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1368 if (npages
> PIPE_BUFFERS
- buffers
)
1369 npages
= PIPE_BUFFERS
- buffers
;
1371 error
= get_user_pages_fast((unsigned long)base
, npages
,
1372 0, &pages
[buffers
]);
1374 if (unlikely(error
<= 0))
1378 * Fill this contiguous range into the partial page map.
1380 for (i
= 0; i
< error
; i
++) {
1381 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1383 partial
[buffers
].offset
= off
;
1384 partial
[buffers
].len
= plen
;
1392 * We didn't complete this iov, stop here since it probably
1393 * means we have to move some of this into a pipe to
1394 * be able to continue.
1400 * Don't continue if we mapped fewer pages than we asked for,
1401 * or if we mapped the max number of pages that we have
1404 if (error
< npages
|| buffers
== PIPE_BUFFERS
)
1417 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1418 struct splice_desc
*sd
)
1423 ret
= buf
->ops
->confirm(pipe
, buf
);
1428 * See if we can use the atomic maps, by prefaulting in the
1429 * pages and doing an atomic copy
1431 if (!fault_in_pages_writeable(sd
->u
.userptr
, sd
->len
)) {
1432 src
= buf
->ops
->map(pipe
, buf
, 1);
1433 ret
= __copy_to_user_inatomic(sd
->u
.userptr
, src
+ buf
->offset
,
1435 buf
->ops
->unmap(pipe
, buf
, src
);
1443 * No dice, use slow non-atomic map and copy
1445 src
= buf
->ops
->map(pipe
, buf
, 0);
1448 if (copy_to_user(sd
->u
.userptr
, src
+ buf
->offset
, sd
->len
))
1451 buf
->ops
->unmap(pipe
, buf
, src
);
1454 sd
->u
.userptr
+= ret
;
1459 * For lack of a better implementation, implement vmsplice() to userspace
1460 * as a simple copy of the pipes pages to the user iov.
1462 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*iov
,
1463 unsigned long nr_segs
, unsigned int flags
)
1465 struct pipe_inode_info
*pipe
;
1466 struct splice_desc sd
;
1471 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1483 * Get user address base and length for this iovec.
1485 error
= get_user(base
, &iov
->iov_base
);
1486 if (unlikely(error
))
1488 error
= get_user(len
, &iov
->iov_len
);
1489 if (unlikely(error
))
1493 * Sanity check this iovec. 0 read succeeds.
1497 if (unlikely(!base
)) {
1502 if (unlikely(!access_ok(VERIFY_WRITE
, base
, len
))) {
1510 sd
.u
.userptr
= base
;
1513 size
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1539 * vmsplice splices a user address range into a pipe. It can be thought of
1540 * as splice-from-memory, where the regular splice is splice-from-file (or
1541 * to file). In both cases the output is a pipe, naturally.
1543 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1544 unsigned long nr_segs
, unsigned int flags
)
1546 struct pipe_inode_info
*pipe
;
1547 struct page
*pages
[PIPE_BUFFERS
];
1548 struct partial_page partial
[PIPE_BUFFERS
];
1549 struct splice_pipe_desc spd
= {
1553 .ops
= &user_page_pipe_buf_ops
,
1554 .spd_release
= spd_release_page
,
1557 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1561 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, pages
, partial
,
1562 flags
& SPLICE_F_GIFT
);
1563 if (spd
.nr_pages
<= 0)
1564 return spd
.nr_pages
;
1566 return splice_to_pipe(pipe
, &spd
);
1570 * Note that vmsplice only really supports true splicing _from_ user memory
1571 * to a pipe, not the other way around. Splicing from user memory is a simple
1572 * operation that can be supported without any funky alignment restrictions
1573 * or nasty vm tricks. We simply map in the user memory and fill them into
1574 * a pipe. The reverse isn't quite as easy, though. There are two possible
1575 * solutions for that:
1577 * - memcpy() the data internally, at which point we might as well just
1578 * do a regular read() on the buffer anyway.
1579 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1580 * has restriction limitations on both ends of the pipe).
1582 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1585 SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct iovec __user
*, iov
,
1586 unsigned long, nr_segs
, unsigned int, flags
)
1592 if (unlikely(nr_segs
> UIO_MAXIOV
))
1594 else if (unlikely(!nr_segs
))
1598 file
= fget_light(fd
, &fput
);
1600 if (file
->f_mode
& FMODE_WRITE
)
1601 error
= vmsplice_to_pipe(file
, iov
, nr_segs
, flags
);
1602 else if (file
->f_mode
& FMODE_READ
)
1603 error
= vmsplice_to_user(file
, iov
, nr_segs
, flags
);
1605 fput_light(file
, fput
);
1611 SYSCALL_DEFINE6(splice
, int, fd_in
, loff_t __user
*, off_in
,
1612 int, fd_out
, loff_t __user
*, off_out
,
1613 size_t, len
, unsigned int, flags
)
1616 struct file
*in
, *out
;
1617 int fput_in
, fput_out
;
1623 in
= fget_light(fd_in
, &fput_in
);
1625 if (in
->f_mode
& FMODE_READ
) {
1626 out
= fget_light(fd_out
, &fput_out
);
1628 if (out
->f_mode
& FMODE_WRITE
)
1629 error
= do_splice(in
, off_in
,
1632 fput_light(out
, fput_out
);
1636 fput_light(in
, fput_in
);
1643 * Make sure there's data to read. Wait for input if we can, otherwise
1644 * return an appropriate error.
1646 static int ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1651 * Check ->nrbufs without the inode lock first. This function
1652 * is speculative anyways, so missing one is ok.
1660 while (!pipe
->nrbufs
) {
1661 if (signal_pending(current
)) {
1667 if (!pipe
->waiting_writers
) {
1668 if (flags
& SPLICE_F_NONBLOCK
) {
1681 * Make sure there's writeable room. Wait for room if we can, otherwise
1682 * return an appropriate error.
1684 static int opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1689 * Check ->nrbufs without the inode lock first. This function
1690 * is speculative anyways, so missing one is ok.
1692 if (pipe
->nrbufs
< PIPE_BUFFERS
)
1698 while (pipe
->nrbufs
>= PIPE_BUFFERS
) {
1699 if (!pipe
->readers
) {
1700 send_sig(SIGPIPE
, current
, 0);
1704 if (flags
& SPLICE_F_NONBLOCK
) {
1708 if (signal_pending(current
)) {
1712 pipe
->waiting_writers
++;
1714 pipe
->waiting_writers
--;
1722 * Splice contents of ipipe to opipe.
1724 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1725 struct pipe_inode_info
*opipe
,
1726 size_t len
, unsigned int flags
)
1728 struct pipe_buffer
*ibuf
, *obuf
;
1730 bool input_wakeup
= false;
1734 ret
= ipipe_prep(ipipe
, flags
);
1738 ret
= opipe_prep(opipe
, flags
);
1743 * Potential ABBA deadlock, work around it by ordering lock
1744 * grabbing by pipe info address. Otherwise two different processes
1745 * could deadlock (one doing tee from A -> B, the other from B -> A).
1747 pipe_double_lock(ipipe
, opipe
);
1750 if (!opipe
->readers
) {
1751 send_sig(SIGPIPE
, current
, 0);
1757 if (!ipipe
->nrbufs
&& !ipipe
->writers
)
1761 * Cannot make any progress, because either the input
1762 * pipe is empty or the output pipe is full.
1764 if (!ipipe
->nrbufs
|| opipe
->nrbufs
>= PIPE_BUFFERS
) {
1765 /* Already processed some buffers, break */
1769 if (flags
& SPLICE_F_NONBLOCK
) {
1775 * We raced with another reader/writer and haven't
1776 * managed to process any buffers. A zero return
1777 * value means EOF, so retry instead.
1784 ibuf
= ipipe
->bufs
+ ipipe
->curbuf
;
1785 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) % PIPE_BUFFERS
;
1786 obuf
= opipe
->bufs
+ nbuf
;
1788 if (len
>= ibuf
->len
) {
1790 * Simply move the whole buffer from ipipe to opipe
1795 ipipe
->curbuf
= (ipipe
->curbuf
+ 1) % PIPE_BUFFERS
;
1797 input_wakeup
= true;
1800 * Get a reference to this pipe buffer,
1801 * so we can copy the contents over.
1803 ibuf
->ops
->get(ipipe
, ibuf
);
1807 * Don't inherit the gift flag, we need to
1808 * prevent multiple steals of this page.
1810 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1814 ibuf
->offset
+= obuf
->len
;
1815 ibuf
->len
-= obuf
->len
;
1825 * If we put data in the output pipe, wakeup any potential readers.
1829 if (waitqueue_active(&opipe
->wait
))
1830 wake_up_interruptible(&opipe
->wait
);
1831 kill_fasync(&opipe
->fasync_readers
, SIGIO
, POLL_IN
);
1834 wakeup_pipe_writers(ipipe
);
1840 * Link contents of ipipe to opipe.
1842 static int link_pipe(struct pipe_inode_info
*ipipe
,
1843 struct pipe_inode_info
*opipe
,
1844 size_t len
, unsigned int flags
)
1846 struct pipe_buffer
*ibuf
, *obuf
;
1847 int ret
= 0, i
= 0, nbuf
;
1850 * Potential ABBA deadlock, work around it by ordering lock
1851 * grabbing by pipe info address. Otherwise two different processes
1852 * could deadlock (one doing tee from A -> B, the other from B -> A).
1854 pipe_double_lock(ipipe
, opipe
);
1857 if (!opipe
->readers
) {
1858 send_sig(SIGPIPE
, current
, 0);
1865 * If we have iterated all input buffers or ran out of
1866 * output room, break.
1868 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= PIPE_BUFFERS
)
1871 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (PIPE_BUFFERS
- 1));
1872 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
1875 * Get a reference to this pipe buffer,
1876 * so we can copy the contents over.
1878 ibuf
->ops
->get(ipipe
, ibuf
);
1880 obuf
= opipe
->bufs
+ nbuf
;
1884 * Don't inherit the gift flag, we need to
1885 * prevent multiple steals of this page.
1887 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1889 if (obuf
->len
> len
)
1899 * return EAGAIN if we have the potential of some data in the
1900 * future, otherwise just return 0
1902 if (!ret
&& ipipe
->waiting_writers
&& (flags
& SPLICE_F_NONBLOCK
))
1909 * If we put data in the output pipe, wakeup any potential readers.
1913 if (waitqueue_active(&opipe
->wait
))
1914 wake_up_interruptible(&opipe
->wait
);
1915 kill_fasync(&opipe
->fasync_readers
, SIGIO
, POLL_IN
);
1922 * This is a tee(1) implementation that works on pipes. It doesn't copy
1923 * any data, it simply references the 'in' pages on the 'out' pipe.
1924 * The 'flags' used are the SPLICE_F_* variants, currently the only
1925 * applicable one is SPLICE_F_NONBLOCK.
1927 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1930 struct pipe_inode_info
*ipipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1931 struct pipe_inode_info
*opipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1935 * Duplicate the contents of ipipe to opipe without actually
1938 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
1940 * Keep going, unless we encounter an error. The ipipe/opipe
1941 * ordering doesn't really matter.
1943 ret
= ipipe_prep(ipipe
, flags
);
1945 ret
= opipe_prep(opipe
, flags
);
1947 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
1954 SYSCALL_DEFINE4(tee
, int, fdin
, int, fdout
, size_t, len
, unsigned int, flags
)
1963 in
= fget_light(fdin
, &fput_in
);
1965 if (in
->f_mode
& FMODE_READ
) {
1967 struct file
*out
= fget_light(fdout
, &fput_out
);
1970 if (out
->f_mode
& FMODE_WRITE
)
1971 error
= do_tee(in
, out
, len
, flags
);
1972 fput_light(out
, fput_out
);
1975 fput_light(in
, fput_in
);