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/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
34 * Attempt to steal a page from a pipe buffer. This should perhaps go into
35 * a vm helper function, it's already simplified quite a bit by the
36 * addition of remove_mapping(). If success is returned, the caller may
37 * attempt to reuse this page for another destination.
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info
*pipe
,
40 struct pipe_buffer
*buf
)
42 struct page
*page
= buf
->page
;
43 struct address_space
*mapping
;
47 mapping
= page_mapping(page
);
49 WARN_ON(!PageUptodate(page
));
52 * At least for ext2 with nobh option, we need to wait on
53 * writeback completing on this page, since we'll remove it
54 * from the pagecache. Otherwise truncate wont wait on the
55 * page, allowing the disk blocks to be reused by someone else
56 * before we actually wrote our data to them. fs corruption
59 wait_on_page_writeback(page
);
61 if (PagePrivate(page
))
62 try_to_release_page(page
, GFP_KERNEL
);
65 * If we succeeded in removing the mapping, set LRU flag
68 if (remove_mapping(mapping
, page
)) {
69 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
75 * Raced with truncate or failed to remove page from current
76 * address space, unlock and return failure.
82 static void page_cache_pipe_buf_release(struct pipe_inode_info
*pipe
,
83 struct pipe_buffer
*buf
)
85 page_cache_release(buf
->page
);
86 buf
->flags
&= ~PIPE_BUF_FLAG_LRU
;
90 * Check whether the contents of buf is OK to access. Since the content
91 * is a page cache page, IO may be in flight.
93 static int page_cache_pipe_buf_confirm(struct pipe_inode_info
*pipe
,
94 struct pipe_buffer
*buf
)
96 struct page
*page
= buf
->page
;
99 if (!PageUptodate(page
)) {
103 * Page got truncated/unhashed. This will cause a 0-byte
104 * splice, if this is the first page.
106 if (!page
->mapping
) {
112 * Uh oh, read-error from disk.
114 if (!PageUptodate(page
)) {
120 * Page is ok afterall, we are done.
131 static const struct pipe_buf_operations page_cache_pipe_buf_ops
= {
133 .map
= generic_pipe_buf_map
,
134 .unmap
= generic_pipe_buf_unmap
,
135 .confirm
= page_cache_pipe_buf_confirm
,
136 .release
= page_cache_pipe_buf_release
,
137 .steal
= page_cache_pipe_buf_steal
,
138 .get
= generic_pipe_buf_get
,
141 static int user_page_pipe_buf_steal(struct pipe_inode_info
*pipe
,
142 struct pipe_buffer
*buf
)
144 if (!(buf
->flags
& PIPE_BUF_FLAG_GIFT
))
147 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
148 return generic_pipe_buf_steal(pipe
, buf
);
151 static const struct pipe_buf_operations user_page_pipe_buf_ops
= {
153 .map
= generic_pipe_buf_map
,
154 .unmap
= generic_pipe_buf_unmap
,
155 .confirm
= generic_pipe_buf_confirm
,
156 .release
= page_cache_pipe_buf_release
,
157 .steal
= user_page_pipe_buf_steal
,
158 .get
= generic_pipe_buf_get
,
162 * splice_to_pipe - fill passed data into a pipe
163 * @pipe: pipe to fill
167 * @spd contains a map of pages and len/offset tuples, along with
168 * the struct pipe_buf_operations associated with these pages. This
169 * function will link that data to the pipe.
172 ssize_t
splice_to_pipe(struct pipe_inode_info
*pipe
,
173 struct splice_pipe_desc
*spd
)
175 unsigned int spd_pages
= spd
->nr_pages
;
176 int ret
, do_wakeup
, page_nr
;
183 mutex_lock(&pipe
->inode
->i_mutex
);
186 if (!pipe
->readers
) {
187 send_sig(SIGPIPE
, current
, 0);
193 if (pipe
->nrbufs
< PIPE_BUFFERS
) {
194 int newbuf
= (pipe
->curbuf
+ pipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
195 struct pipe_buffer
*buf
= pipe
->bufs
+ newbuf
;
197 buf
->page
= spd
->pages
[page_nr
];
198 buf
->offset
= spd
->partial
[page_nr
].offset
;
199 buf
->len
= spd
->partial
[page_nr
].len
;
200 buf
->private = spd
->partial
[page_nr
].private;
202 if (spd
->flags
& SPLICE_F_GIFT
)
203 buf
->flags
|= PIPE_BUF_FLAG_GIFT
;
212 if (!--spd
->nr_pages
)
214 if (pipe
->nrbufs
< PIPE_BUFFERS
)
220 if (spd
->flags
& SPLICE_F_NONBLOCK
) {
226 if (signal_pending(current
)) {
234 if (waitqueue_active(&pipe
->wait
))
235 wake_up_interruptible_sync(&pipe
->wait
);
236 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
240 pipe
->waiting_writers
++;
242 pipe
->waiting_writers
--;
246 mutex_unlock(&pipe
->inode
->i_mutex
);
250 if (waitqueue_active(&pipe
->wait
))
251 wake_up_interruptible(&pipe
->wait
);
252 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
256 while (page_nr
< spd_pages
)
257 page_cache_release(spd
->pages
[page_nr
++]);
263 __generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
264 struct pipe_inode_info
*pipe
, size_t len
,
267 struct address_space
*mapping
= in
->f_mapping
;
268 unsigned int loff
, nr_pages
, req_pages
;
269 struct page
*pages
[PIPE_BUFFERS
];
270 struct partial_page partial
[PIPE_BUFFERS
];
272 pgoff_t index
, end_index
;
275 struct splice_pipe_desc spd
= {
279 .ops
= &page_cache_pipe_buf_ops
,
282 index
= *ppos
>> PAGE_CACHE_SHIFT
;
283 loff
= *ppos
& ~PAGE_CACHE_MASK
;
284 req_pages
= (len
+ loff
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
285 nr_pages
= min(req_pages
, (unsigned)PIPE_BUFFERS
);
288 * Lookup the (hopefully) full range of pages we need.
290 spd
.nr_pages
= find_get_pages_contig(mapping
, index
, nr_pages
, pages
);
291 index
+= spd
.nr_pages
;
294 * If find_get_pages_contig() returned fewer pages than we needed,
295 * readahead/allocate the rest and fill in the holes.
297 if (spd
.nr_pages
< nr_pages
)
298 page_cache_sync_readahead(mapping
, &in
->f_ra
, in
,
299 index
, req_pages
- spd
.nr_pages
);
302 while (spd
.nr_pages
< nr_pages
) {
304 * Page could be there, find_get_pages_contig() breaks on
307 page
= find_get_page(mapping
, index
);
310 * page didn't exist, allocate one.
312 page
= page_cache_alloc_cold(mapping
);
316 error
= add_to_page_cache_lru(page
, mapping
, index
,
318 if (unlikely(error
)) {
319 page_cache_release(page
);
320 if (error
== -EEXIST
)
325 * add_to_page_cache() locks the page, unlock it
326 * to avoid convoluting the logic below even more.
331 pages
[spd
.nr_pages
++] = page
;
336 * Now loop over the map and see if we need to start IO on any
337 * pages, fill in the partial map, etc.
339 index
= *ppos
>> PAGE_CACHE_SHIFT
;
340 nr_pages
= spd
.nr_pages
;
342 for (page_nr
= 0; page_nr
< nr_pages
; page_nr
++) {
343 unsigned int this_len
;
349 * this_len is the max we'll use from this page
351 this_len
= min_t(unsigned long, len
, PAGE_CACHE_SIZE
- loff
);
352 page
= pages
[page_nr
];
354 if (PageReadahead(page
))
355 page_cache_async_readahead(mapping
, &in
->f_ra
, in
,
356 page
, index
, req_pages
- page_nr
);
359 * If the page isn't uptodate, we may need to start io on it
361 if (!PageUptodate(page
)) {
363 * If in nonblock mode then dont block on waiting
364 * for an in-flight io page
366 if (flags
& SPLICE_F_NONBLOCK
) {
367 if (TestSetPageLocked(page
))
373 * page was truncated, stop here. if this isn't the
374 * first page, we'll just complete what we already
377 if (!page
->mapping
) {
382 * page was already under io and is now done, great
384 if (PageUptodate(page
)) {
390 * need to read in the page
392 error
= mapping
->a_ops
->readpage(in
, page
);
393 if (unlikely(error
)) {
395 * We really should re-lookup the page here,
396 * but it complicates things a lot. Instead
397 * lets just do what we already stored, and
398 * we'll get it the next time we are called.
400 if (error
== AOP_TRUNCATED_PAGE
)
408 * i_size must be checked after PageUptodate.
410 isize
= i_size_read(mapping
->host
);
411 end_index
= (isize
- 1) >> PAGE_CACHE_SHIFT
;
412 if (unlikely(!isize
|| index
> end_index
))
416 * if this is the last page, see if we need to shrink
417 * the length and stop
419 if (end_index
== index
) {
423 * max good bytes in this page
425 plen
= ((isize
- 1) & ~PAGE_CACHE_MASK
) + 1;
430 * force quit after adding this page
432 this_len
= min(this_len
, plen
- loff
);
436 partial
[page_nr
].offset
= loff
;
437 partial
[page_nr
].len
= this_len
;
445 * Release any pages at the end, if we quit early. 'page_nr' is how far
446 * we got, 'nr_pages' is how many pages are in the map.
448 while (page_nr
< nr_pages
)
449 page_cache_release(pages
[page_nr
++]);
450 in
->f_ra
.prev_index
= index
;
453 return splice_to_pipe(pipe
, &spd
);
459 * generic_file_splice_read - splice data from file to a pipe
460 * @in: file to splice from
461 * @ppos: position in @in
462 * @pipe: pipe to splice to
463 * @len: number of bytes to splice
464 * @flags: splice modifier flags
467 * Will read pages from given file and fill them into a pipe. Can be
468 * used as long as the address_space operations for the source implements
472 ssize_t
generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
473 struct pipe_inode_info
*pipe
, size_t len
,
480 isize
= i_size_read(in
->f_mapping
->host
);
481 if (unlikely(*ppos
>= isize
))
484 left
= isize
- *ppos
;
485 if (unlikely(left
< len
))
490 while (len
&& !spliced
) {
491 ret
= __generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
498 if (flags
& SPLICE_F_NONBLOCK
) {
515 EXPORT_SYMBOL(generic_file_splice_read
);
518 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
519 * using sendpage(). Return the number of bytes sent.
521 static int pipe_to_sendpage(struct pipe_inode_info
*pipe
,
522 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
524 struct file
*file
= sd
->u
.file
;
525 loff_t pos
= sd
->pos
;
528 ret
= buf
->ops
->confirm(pipe
, buf
);
530 more
= (sd
->flags
& SPLICE_F_MORE
) || sd
->len
< sd
->total_len
;
532 ret
= file
->f_op
->sendpage(file
, buf
->page
, buf
->offset
,
533 sd
->len
, &pos
, more
);
540 * This is a little more tricky than the file -> pipe splicing. There are
541 * basically three cases:
543 * - Destination page already exists in the address space and there
544 * are users of it. For that case we have no other option that
545 * copying the data. Tough luck.
546 * - Destination page already exists in the address space, but there
547 * are no users of it. Make sure it's uptodate, then drop it. Fall
548 * through to last case.
549 * - Destination page does not exist, we can add the pipe page to
550 * the page cache and avoid the copy.
552 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
553 * sd->flags), we attempt to migrate pages from the pipe to the output
554 * file address space page cache. This is possible if no one else has
555 * the pipe page referenced outside of the pipe and page cache. If
556 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
557 * a new page in the output file page cache and fill/dirty that.
559 static int pipe_to_file(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
560 struct splice_desc
*sd
)
562 struct file
*file
= sd
->u
.file
;
563 struct address_space
*mapping
= file
->f_mapping
;
564 unsigned int offset
, this_len
;
570 * make sure the data in this buffer is uptodate
572 ret
= buf
->ops
->confirm(pipe
, buf
);
576 index
= sd
->pos
>> PAGE_CACHE_SHIFT
;
577 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
580 if (this_len
+ offset
> PAGE_CACHE_SIZE
)
581 this_len
= PAGE_CACHE_SIZE
- offset
;
584 page
= find_lock_page(mapping
, index
);
587 page
= page_cache_alloc_cold(mapping
);
592 * This will also lock the page
594 ret
= add_to_page_cache_lru(page
, mapping
, index
,
600 ret
= mapping
->a_ops
->prepare_write(file
, page
, offset
, offset
+this_len
);
602 loff_t isize
= i_size_read(mapping
->host
);
604 if (ret
!= AOP_TRUNCATED_PAGE
)
606 page_cache_release(page
);
607 if (ret
== AOP_TRUNCATED_PAGE
)
611 * prepare_write() may have instantiated a few blocks
612 * outside i_size. Trim these off again.
614 if (sd
->pos
+ this_len
> isize
)
615 vmtruncate(mapping
->host
, isize
);
620 if (buf
->page
!= page
) {
622 * Careful, ->map() uses KM_USER0!
624 char *src
= buf
->ops
->map(pipe
, buf
, 1);
625 char *dst
= kmap_atomic(page
, KM_USER1
);
627 memcpy(dst
+ offset
, src
+ buf
->offset
, this_len
);
628 flush_dcache_page(page
);
629 kunmap_atomic(dst
, KM_USER1
);
630 buf
->ops
->unmap(pipe
, buf
, src
);
633 ret
= mapping
->a_ops
->commit_write(file
, page
, offset
, offset
+this_len
);
635 if (ret
== AOP_TRUNCATED_PAGE
) {
636 page_cache_release(page
);
642 * Partial write has happened, so 'ret' already initialized by
643 * number of bytes written, Where is nothing we have to do here.
648 * Return the number of bytes written and mark page as
649 * accessed, we are now done!
651 mark_page_accessed(page
);
655 page_cache_release(page
);
661 * __splice_from_pipe - splice data from a pipe to given actor
662 * @pipe: pipe to splice from
663 * @sd: information to @actor
664 * @actor: handler that splices the data
667 * This function does little more than loop over the pipe and call
668 * @actor to do the actual moving of a single struct pipe_buffer to
669 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
673 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
676 int ret
, do_wakeup
, err
;
683 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
684 const struct pipe_buf_operations
*ops
= buf
->ops
;
687 if (sd
->len
> sd
->total_len
)
688 sd
->len
= sd
->total_len
;
690 err
= actor(pipe
, buf
, sd
);
692 if (!ret
&& err
!= -ENODATA
)
704 sd
->total_len
-= err
;
710 ops
->release(pipe
, buf
);
711 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (PIPE_BUFFERS
- 1);
725 if (!pipe
->waiting_writers
) {
730 if (sd
->flags
& SPLICE_F_NONBLOCK
) {
736 if (signal_pending(current
)) {
744 if (waitqueue_active(&pipe
->wait
))
745 wake_up_interruptible_sync(&pipe
->wait
);
746 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
755 if (waitqueue_active(&pipe
->wait
))
756 wake_up_interruptible(&pipe
->wait
);
757 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
762 EXPORT_SYMBOL(__splice_from_pipe
);
765 * splice_from_pipe - splice data from a pipe to a file
766 * @pipe: pipe to splice from
767 * @out: file to splice to
768 * @ppos: position in @out
769 * @len: how many bytes to splice
770 * @flags: splice modifier flags
771 * @actor: handler that splices the data
774 * See __splice_from_pipe. This function locks the input and output inodes,
775 * otherwise it's identical to __splice_from_pipe().
778 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
779 loff_t
*ppos
, size_t len
, unsigned int flags
,
783 struct inode
*inode
= out
->f_mapping
->host
;
784 struct splice_desc sd
= {
792 * The actor worker might be calling ->prepare_write and
793 * ->commit_write. Most of the time, these expect i_mutex to
794 * be held. Since this may result in an ABBA deadlock with
795 * pipe->inode, we have to order lock acquiry here.
797 inode_double_lock(inode
, pipe
->inode
);
798 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
799 inode_double_unlock(inode
, pipe
->inode
);
805 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
807 * @out: file to write to
808 * @ppos: position in @out
809 * @len: number of bytes to splice
810 * @flags: splice modifier flags
813 * Will either move or copy pages (determined by @flags options) from
814 * the given pipe inode to the given file. The caller is responsible
815 * for acquiring i_mutex on both inodes.
819 generic_file_splice_write_nolock(struct pipe_inode_info
*pipe
, struct file
*out
,
820 loff_t
*ppos
, size_t len
, unsigned int flags
)
822 struct address_space
*mapping
= out
->f_mapping
;
823 struct inode
*inode
= mapping
->host
;
824 struct splice_desc sd
= {
833 err
= remove_suid(out
->f_path
.dentry
);
837 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_file
);
839 unsigned long nr_pages
;
842 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
845 * If file or inode is SYNC and we actually wrote some data,
848 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
849 err
= generic_osync_inode(inode
, mapping
,
850 OSYNC_METADATA
|OSYNC_DATA
);
855 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
861 EXPORT_SYMBOL(generic_file_splice_write_nolock
);
864 * generic_file_splice_write - splice data from a pipe to a file
866 * @out: file to write to
867 * @ppos: position in @out
868 * @len: number of bytes to splice
869 * @flags: splice modifier flags
872 * Will either move or copy pages (determined by @flags options) from
873 * the given pipe inode to the given file.
877 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
878 loff_t
*ppos
, size_t len
, unsigned int flags
)
880 struct address_space
*mapping
= out
->f_mapping
;
881 struct inode
*inode
= mapping
->host
;
885 err
= should_remove_suid(out
->f_path
.dentry
);
887 mutex_lock(&inode
->i_mutex
);
888 err
= __remove_suid(out
->f_path
.dentry
, err
);
889 mutex_unlock(&inode
->i_mutex
);
894 ret
= splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_file
);
896 unsigned long nr_pages
;
899 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
902 * If file or inode is SYNC and we actually wrote some data,
905 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
906 mutex_lock(&inode
->i_mutex
);
907 err
= generic_osync_inode(inode
, mapping
,
908 OSYNC_METADATA
|OSYNC_DATA
);
909 mutex_unlock(&inode
->i_mutex
);
914 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
920 EXPORT_SYMBOL(generic_file_splice_write
);
923 * generic_splice_sendpage - splice data from a pipe to a socket
924 * @pipe: pipe to splice from
925 * @out: socket to write to
926 * @ppos: position in @out
927 * @len: number of bytes to splice
928 * @flags: splice modifier flags
931 * Will send @len bytes from the pipe to a network socket. No data copying
935 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
936 loff_t
*ppos
, size_t len
, unsigned int flags
)
938 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
941 EXPORT_SYMBOL(generic_splice_sendpage
);
944 * Attempt to initiate a splice from pipe to file.
946 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
947 loff_t
*ppos
, size_t len
, unsigned int flags
)
951 if (unlikely(!out
->f_op
|| !out
->f_op
->splice_write
))
954 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
957 ret
= rw_verify_area(WRITE
, out
, ppos
, len
);
958 if (unlikely(ret
< 0))
961 ret
= security_file_permission(out
, MAY_WRITE
);
962 if (unlikely(ret
< 0))
965 return out
->f_op
->splice_write(pipe
, out
, ppos
, len
, flags
);
969 * Attempt to initiate a splice from a file to a pipe.
971 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
972 struct pipe_inode_info
*pipe
, size_t len
,
977 if (unlikely(!in
->f_op
|| !in
->f_op
->splice_read
))
980 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
983 ret
= rw_verify_area(READ
, in
, ppos
, len
);
984 if (unlikely(ret
< 0))
987 ret
= security_file_permission(in
, MAY_READ
);
988 if (unlikely(ret
< 0))
991 return in
->f_op
->splice_read(in
, ppos
, pipe
, len
, flags
);
995 * splice_direct_to_actor - splices data directly between two non-pipes
996 * @in: file to splice from
997 * @sd: actor information on where to splice to
998 * @actor: handles the data splicing
1001 * This is a special case helper to splice directly between two
1002 * points, without requiring an explicit pipe. Internally an allocated
1003 * pipe is cached in the process, and reused during the lifetime of
1007 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
1008 splice_direct_actor
*actor
)
1010 struct pipe_inode_info
*pipe
;
1017 * We require the input being a regular file, as we don't want to
1018 * randomly drop data for eg socket -> socket splicing. Use the
1019 * piped splicing for that!
1021 i_mode
= in
->f_path
.dentry
->d_inode
->i_mode
;
1022 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
1026 * neither in nor out is a pipe, setup an internal pipe attached to
1027 * 'out' and transfer the wanted data from 'in' to 'out' through that
1029 pipe
= current
->splice_pipe
;
1030 if (unlikely(!pipe
)) {
1031 pipe
= alloc_pipe_info(NULL
);
1036 * We don't have an immediate reader, but we'll read the stuff
1037 * out of the pipe right after the splice_to_pipe(). So set
1038 * PIPE_READERS appropriately.
1042 current
->splice_pipe
= pipe
;
1050 len
= sd
->total_len
;
1054 * Don't block on output, we have to drain the direct pipe.
1056 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1060 loff_t pos
= sd
->pos
;
1062 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1063 if (unlikely(ret
<= 0))
1067 sd
->total_len
= read_len
;
1070 * NOTE: nonblocking mode only applies to the input. We
1071 * must not do the output in nonblocking mode as then we
1072 * could get stuck data in the internal pipe:
1074 ret
= actor(pipe
, sd
);
1075 if (unlikely(ret
<= 0))
1086 pipe
->nrbufs
= pipe
->curbuf
= 0;
1091 * If we did an incomplete transfer we must release
1092 * the pipe buffers in question:
1094 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
1095 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1098 buf
->ops
->release(pipe
, buf
);
1102 pipe
->nrbufs
= pipe
->curbuf
= 0;
1105 * If we transferred some data, return the number of bytes:
1113 EXPORT_SYMBOL(splice_direct_to_actor
);
1115 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1116 struct splice_desc
*sd
)
1118 struct file
*file
= sd
->u
.file
;
1120 return do_splice_from(pipe
, file
, &sd
->pos
, sd
->total_len
, sd
->flags
);
1124 * do_splice_direct - splices data directly between two files
1125 * @in: file to splice from
1126 * @ppos: input file offset
1127 * @out: file to splice to
1128 * @len: number of bytes to splice
1129 * @flags: splice modifier flags
1132 * For use by do_sendfile(). splice can easily emulate sendfile, but
1133 * doing it in the application would incur an extra system call
1134 * (splice in + splice out, as compared to just sendfile()). So this helper
1135 * can splice directly through a process-private pipe.
1138 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1139 size_t len
, unsigned int flags
)
1141 struct splice_desc sd
= {
1150 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1158 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1159 * location, so checking ->i_pipe is not enough to verify that this is a
1162 static inline struct pipe_inode_info
*pipe_info(struct inode
*inode
)
1164 if (S_ISFIFO(inode
->i_mode
))
1165 return inode
->i_pipe
;
1171 * Determine where to splice to/from.
1173 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1174 struct file
*out
, loff_t __user
*off_out
,
1175 size_t len
, unsigned int flags
)
1177 struct pipe_inode_info
*pipe
;
1178 loff_t offset
, *off
;
1181 pipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1186 if (out
->f_op
->llseek
== no_llseek
)
1188 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1194 ret
= do_splice_from(pipe
, out
, off
, len
, flags
);
1196 if (off_out
&& copy_to_user(off_out
, off
, sizeof(loff_t
)))
1202 pipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1207 if (in
->f_op
->llseek
== no_llseek
)
1209 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1215 ret
= do_splice_to(in
, off
, pipe
, len
, flags
);
1217 if (off_in
&& copy_to_user(off_in
, off
, sizeof(loff_t
)))
1227 * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1228 * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1229 * for writing) and page faulting on the user memory pointed to by src.
1230 * This assumes that we will very rarely hit the partial != 0 path, or this
1231 * will not be a win.
1233 static int copy_from_user_mmap_sem(void *dst
, const void __user
*src
, size_t n
)
1237 if (!access_ok(VERIFY_READ
, src
, n
))
1240 pagefault_disable();
1241 partial
= __copy_from_user_inatomic(dst
, src
, n
);
1245 * Didn't copy everything, drop the mmap_sem and do a faulting copy
1247 if (unlikely(partial
)) {
1248 up_read(¤t
->mm
->mmap_sem
);
1249 partial
= copy_from_user(dst
, src
, n
);
1250 down_read(¤t
->mm
->mmap_sem
);
1257 * Map an iov into an array of pages and offset/length tupples. With the
1258 * partial_page structure, we can map several non-contiguous ranges into
1259 * our ones pages[] map instead of splitting that operation into pieces.
1260 * Could easily be exported as a generic helper for other users, in which
1261 * case one would probably want to add a 'max_nr_pages' parameter as well.
1263 static int get_iovec_page_array(const struct iovec __user
*iov
,
1264 unsigned int nr_vecs
, struct page
**pages
,
1265 struct partial_page
*partial
, int aligned
)
1267 int buffers
= 0, error
= 0;
1269 down_read(¤t
->mm
->mmap_sem
);
1272 unsigned long off
, npages
;
1279 if (copy_from_user_mmap_sem(&entry
, iov
, sizeof(entry
)))
1282 base
= entry
.iov_base
;
1283 len
= entry
.iov_len
;
1286 * Sanity check this iovec. 0 read succeeds.
1292 if (!access_ok(VERIFY_READ
, base
, len
))
1296 * Get this base offset and number of pages, then map
1297 * in the user pages.
1299 off
= (unsigned long) base
& ~PAGE_MASK
;
1302 * If asked for alignment, the offset must be zero and the
1303 * length a multiple of the PAGE_SIZE.
1306 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1309 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1310 if (npages
> PIPE_BUFFERS
- buffers
)
1311 npages
= PIPE_BUFFERS
- buffers
;
1313 error
= get_user_pages(current
, current
->mm
,
1314 (unsigned long) base
, npages
, 0, 0,
1315 &pages
[buffers
], NULL
);
1317 if (unlikely(error
<= 0))
1321 * Fill this contiguous range into the partial page map.
1323 for (i
= 0; i
< error
; i
++) {
1324 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1326 partial
[buffers
].offset
= off
;
1327 partial
[buffers
].len
= plen
;
1335 * We didn't complete this iov, stop here since it probably
1336 * means we have to move some of this into a pipe to
1337 * be able to continue.
1343 * Don't continue if we mapped fewer pages than we asked for,
1344 * or if we mapped the max number of pages that we have
1347 if (error
< npages
|| buffers
== PIPE_BUFFERS
)
1354 up_read(¤t
->mm
->mmap_sem
);
1362 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1363 struct splice_desc
*sd
)
1368 ret
= buf
->ops
->confirm(pipe
, buf
);
1373 * See if we can use the atomic maps, by prefaulting in the
1374 * pages and doing an atomic copy
1376 if (!fault_in_pages_writeable(sd
->u
.userptr
, sd
->len
)) {
1377 src
= buf
->ops
->map(pipe
, buf
, 1);
1378 ret
= __copy_to_user_inatomic(sd
->u
.userptr
, src
+ buf
->offset
,
1380 buf
->ops
->unmap(pipe
, buf
, src
);
1388 * No dice, use slow non-atomic map and copy
1390 src
= buf
->ops
->map(pipe
, buf
, 0);
1393 if (copy_to_user(sd
->u
.userptr
, src
+ buf
->offset
, sd
->len
))
1396 buf
->ops
->unmap(pipe
, buf
, src
);
1399 sd
->u
.userptr
+= ret
;
1404 * For lack of a better implementation, implement vmsplice() to userspace
1405 * as a simple copy of the pipes pages to the user iov.
1407 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*iov
,
1408 unsigned long nr_segs
, unsigned int flags
)
1410 struct pipe_inode_info
*pipe
;
1411 struct splice_desc sd
;
1416 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1421 mutex_lock(&pipe
->inode
->i_mutex
);
1429 * Get user address base and length for this iovec.
1431 error
= get_user(base
, &iov
->iov_base
);
1432 if (unlikely(error
))
1434 error
= get_user(len
, &iov
->iov_len
);
1435 if (unlikely(error
))
1439 * Sanity check this iovec. 0 read succeeds.
1443 if (unlikely(!base
)) {
1448 if (unlikely(!access_ok(VERIFY_WRITE
, base
, len
))) {
1456 sd
.u
.userptr
= base
;
1459 size
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1477 mutex_unlock(&pipe
->inode
->i_mutex
);
1486 * vmsplice splices a user address range into a pipe. It can be thought of
1487 * as splice-from-memory, where the regular splice is splice-from-file (or
1488 * to file). In both cases the output is a pipe, naturally.
1490 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1491 unsigned long nr_segs
, unsigned int flags
)
1493 struct pipe_inode_info
*pipe
;
1494 struct page
*pages
[PIPE_BUFFERS
];
1495 struct partial_page partial
[PIPE_BUFFERS
];
1496 struct splice_pipe_desc spd
= {
1500 .ops
= &user_page_pipe_buf_ops
,
1503 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1507 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, pages
, partial
,
1508 flags
& SPLICE_F_GIFT
);
1509 if (spd
.nr_pages
<= 0)
1510 return spd
.nr_pages
;
1512 return splice_to_pipe(pipe
, &spd
);
1516 * Note that vmsplice only really supports true splicing _from_ user memory
1517 * to a pipe, not the other way around. Splicing from user memory is a simple
1518 * operation that can be supported without any funky alignment restrictions
1519 * or nasty vm tricks. We simply map in the user memory and fill them into
1520 * a pipe. The reverse isn't quite as easy, though. There are two possible
1521 * solutions for that:
1523 * - memcpy() the data internally, at which point we might as well just
1524 * do a regular read() on the buffer anyway.
1525 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1526 * has restriction limitations on both ends of the pipe).
1528 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1531 asmlinkage
long sys_vmsplice(int fd
, const struct iovec __user
*iov
,
1532 unsigned long nr_segs
, unsigned int flags
)
1538 if (unlikely(nr_segs
> UIO_MAXIOV
))
1540 else if (unlikely(!nr_segs
))
1544 file
= fget_light(fd
, &fput
);
1546 if (file
->f_mode
& FMODE_WRITE
)
1547 error
= vmsplice_to_pipe(file
, iov
, nr_segs
, flags
);
1548 else if (file
->f_mode
& FMODE_READ
)
1549 error
= vmsplice_to_user(file
, iov
, nr_segs
, flags
);
1551 fput_light(file
, fput
);
1557 asmlinkage
long sys_splice(int fd_in
, loff_t __user
*off_in
,
1558 int fd_out
, loff_t __user
*off_out
,
1559 size_t len
, unsigned int flags
)
1562 struct file
*in
, *out
;
1563 int fput_in
, fput_out
;
1569 in
= fget_light(fd_in
, &fput_in
);
1571 if (in
->f_mode
& FMODE_READ
) {
1572 out
= fget_light(fd_out
, &fput_out
);
1574 if (out
->f_mode
& FMODE_WRITE
)
1575 error
= do_splice(in
, off_in
,
1578 fput_light(out
, fput_out
);
1582 fput_light(in
, fput_in
);
1589 * Make sure there's data to read. Wait for input if we can, otherwise
1590 * return an appropriate error.
1592 static int link_ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1597 * Check ->nrbufs without the inode lock first. This function
1598 * is speculative anyways, so missing one is ok.
1604 mutex_lock(&pipe
->inode
->i_mutex
);
1606 while (!pipe
->nrbufs
) {
1607 if (signal_pending(current
)) {
1613 if (!pipe
->waiting_writers
) {
1614 if (flags
& SPLICE_F_NONBLOCK
) {
1622 mutex_unlock(&pipe
->inode
->i_mutex
);
1627 * Make sure there's writeable room. Wait for room if we can, otherwise
1628 * return an appropriate error.
1630 static int link_opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1635 * Check ->nrbufs without the inode lock first. This function
1636 * is speculative anyways, so missing one is ok.
1638 if (pipe
->nrbufs
< PIPE_BUFFERS
)
1642 mutex_lock(&pipe
->inode
->i_mutex
);
1644 while (pipe
->nrbufs
>= PIPE_BUFFERS
) {
1645 if (!pipe
->readers
) {
1646 send_sig(SIGPIPE
, current
, 0);
1650 if (flags
& SPLICE_F_NONBLOCK
) {
1654 if (signal_pending(current
)) {
1658 pipe
->waiting_writers
++;
1660 pipe
->waiting_writers
--;
1663 mutex_unlock(&pipe
->inode
->i_mutex
);
1668 * Link contents of ipipe to opipe.
1670 static int link_pipe(struct pipe_inode_info
*ipipe
,
1671 struct pipe_inode_info
*opipe
,
1672 size_t len
, unsigned int flags
)
1674 struct pipe_buffer
*ibuf
, *obuf
;
1675 int ret
= 0, i
= 0, nbuf
;
1678 * Potential ABBA deadlock, work around it by ordering lock
1679 * grabbing by inode address. Otherwise two different processes
1680 * could deadlock (one doing tee from A -> B, the other from B -> A).
1682 inode_double_lock(ipipe
->inode
, opipe
->inode
);
1685 if (!opipe
->readers
) {
1686 send_sig(SIGPIPE
, current
, 0);
1693 * If we have iterated all input buffers or ran out of
1694 * output room, break.
1696 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= PIPE_BUFFERS
)
1699 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (PIPE_BUFFERS
- 1));
1700 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
1703 * Get a reference to this pipe buffer,
1704 * so we can copy the contents over.
1706 ibuf
->ops
->get(ipipe
, ibuf
);
1708 obuf
= opipe
->bufs
+ nbuf
;
1712 * Don't inherit the gift flag, we need to
1713 * prevent multiple steals of this page.
1715 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1717 if (obuf
->len
> len
)
1726 inode_double_unlock(ipipe
->inode
, opipe
->inode
);
1729 * If we put data in the output pipe, wakeup any potential readers.
1733 if (waitqueue_active(&opipe
->wait
))
1734 wake_up_interruptible(&opipe
->wait
);
1735 kill_fasync(&opipe
->fasync_readers
, SIGIO
, POLL_IN
);
1742 * This is a tee(1) implementation that works on pipes. It doesn't copy
1743 * any data, it simply references the 'in' pages on the 'out' pipe.
1744 * The 'flags' used are the SPLICE_F_* variants, currently the only
1745 * applicable one is SPLICE_F_NONBLOCK.
1747 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1750 struct pipe_inode_info
*ipipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1751 struct pipe_inode_info
*opipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1755 * Duplicate the contents of ipipe to opipe without actually
1758 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
1760 * Keep going, unless we encounter an error. The ipipe/opipe
1761 * ordering doesn't really matter.
1763 ret
= link_ipipe_prep(ipipe
, flags
);
1765 ret
= link_opipe_prep(opipe
, flags
);
1767 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
1768 if (!ret
&& (flags
& SPLICE_F_NONBLOCK
))
1777 asmlinkage
long sys_tee(int fdin
, int fdout
, size_t len
, unsigned int flags
)
1786 in
= fget_light(fdin
, &fput_in
);
1788 if (in
->f_mode
& FMODE_READ
) {
1790 struct file
*out
= fget_light(fdout
, &fput_out
);
1793 if (out
->f_mode
& FMODE_WRITE
)
1794 error
= do_tee(in
, out
, len
, flags
);
1795 fput_light(out
, fput_out
);
1798 fput_light(in
, fput_in
);