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 and fixing the initial implementation
15 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
16 * Copyright (C) 2005 Linus Torvalds <torvalds@osdl.org>
20 #include <linux/file.h>
21 #include <linux/pagemap.h>
22 #include <linux/pipe_fs_i.h>
23 #include <linux/mm_inline.h>
24 #include <linux/swap.h>
25 #include <linux/writeback.h>
26 #include <linux/buffer_head.h>
27 #include <linux/module.h>
28 #include <linux/syscalls.h>
31 * Passed to the actors
34 unsigned int len
, total_len
; /* current and remaining length */
35 unsigned int flags
; /* splice flags */
36 struct file
*file
; /* file to read/write */
37 loff_t pos
; /* file position */
41 * Attempt to steal a page from a pipe buffer. This should perhaps go into
42 * a vm helper function, it's already simplified quite a bit by the
43 * addition of remove_mapping(). If success is returned, the caller may
44 * attempt to reuse this page for another destination.
46 static int page_cache_pipe_buf_steal(struct pipe_inode_info
*info
,
47 struct pipe_buffer
*buf
)
49 struct page
*page
= buf
->page
;
50 struct address_space
*mapping
= page_mapping(page
);
52 WARN_ON(!PageLocked(page
));
53 WARN_ON(!PageUptodate(page
));
56 * At least for ext2 with nobh option, we need to wait on writeback
57 * completing on this page, since we'll remove it from the pagecache.
58 * Otherwise truncate wont wait on the page, allowing the disk
59 * blocks to be reused by someone else before we actually wrote our
60 * data to them. fs corruption ensues.
62 wait_on_page_writeback(page
);
64 if (PagePrivate(page
))
65 try_to_release_page(page
, mapping_gfp_mask(mapping
));
67 if (!remove_mapping(mapping
, page
))
70 buf
->flags
|= PIPE_BUF_FLAG_STOLEN
| PIPE_BUF_FLAG_LRU
;
74 static void page_cache_pipe_buf_release(struct pipe_inode_info
*info
,
75 struct pipe_buffer
*buf
)
77 page_cache_release(buf
->page
);
79 buf
->flags
&= ~(PIPE_BUF_FLAG_STOLEN
| PIPE_BUF_FLAG_LRU
);
82 static void *page_cache_pipe_buf_map(struct file
*file
,
83 struct pipe_inode_info
*info
,
84 struct pipe_buffer
*buf
)
86 struct page
*page
= buf
->page
;
89 if (!PageUptodate(page
)) {
93 * Page got truncated/unhashed. This will cause a 0-byte
94 * splice, if this is the first page
102 * uh oh, read-error from disk
104 if (!PageUptodate(page
)) {
110 * page is ok afterall, fall through to mapping
121 static void page_cache_pipe_buf_unmap(struct pipe_inode_info
*info
,
122 struct pipe_buffer
*buf
)
127 static struct pipe_buf_operations page_cache_pipe_buf_ops
= {
129 .map
= page_cache_pipe_buf_map
,
130 .unmap
= page_cache_pipe_buf_unmap
,
131 .release
= page_cache_pipe_buf_release
,
132 .steal
= page_cache_pipe_buf_steal
,
136 * Pipe output worker. This sets up our pipe format with the page cache
137 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
139 static ssize_t
move_to_pipe(struct pipe_inode_info
*pipe
, struct page
**pages
,
140 int nr_pages
, unsigned long offset
,
141 unsigned long len
, unsigned int flags
)
143 int ret
, do_wakeup
, i
;
150 mutex_lock(&pipe
->inode
->i_mutex
);
155 if (!pipe
->readers
) {
156 send_sig(SIGPIPE
, current
, 0);
163 if (bufs
< PIPE_BUFFERS
) {
164 int newbuf
= (pipe
->curbuf
+ bufs
) & (PIPE_BUFFERS
- 1);
165 struct pipe_buffer
*buf
= pipe
->bufs
+ newbuf
;
166 struct page
*page
= pages
[i
++];
167 unsigned long this_len
;
169 this_len
= PAGE_CACHE_SIZE
- offset
;
174 buf
->offset
= offset
;
176 buf
->ops
= &page_cache_pipe_buf_ops
;
177 pipe
->nrbufs
= ++bufs
;
187 if (bufs
< PIPE_BUFFERS
)
193 if (flags
& SPLICE_F_NONBLOCK
) {
199 if (signal_pending(current
)) {
207 if (waitqueue_active(&pipe
->wait
))
208 wake_up_interruptible_sync(&pipe
->wait
);
209 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
213 pipe
->waiting_writers
++;
215 pipe
->waiting_writers
--;
219 mutex_unlock(&pipe
->inode
->i_mutex
);
223 if (waitqueue_active(&pipe
->wait
))
224 wake_up_interruptible(&pipe
->wait
);
225 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
229 page_cache_release(pages
[i
++]);
235 __generic_file_splice_read(struct file
*in
, struct pipe_inode_info
*pipe
,
236 size_t len
, unsigned int flags
)
238 struct address_space
*mapping
= in
->f_mapping
;
239 unsigned int offset
, nr_pages
;
240 struct page
*pages
[PIPE_BUFFERS
];
245 index
= in
->f_pos
>> PAGE_CACHE_SHIFT
;
246 offset
= in
->f_pos
& ~PAGE_CACHE_MASK
;
247 nr_pages
= (len
+ offset
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
249 if (nr_pages
> PIPE_BUFFERS
)
250 nr_pages
= PIPE_BUFFERS
;
253 * initiate read-ahead on this page range. however, don't call into
254 * read-ahead if this is a non-zero offset (we are likely doing small
255 * chunk splice and the page is already there) for a single page.
257 if (!offset
|| nr_pages
> 1)
258 do_page_cache_readahead(mapping
, in
, index
, nr_pages
);
261 * now fill in the holes
263 for (i
= 0; i
< nr_pages
; i
++, index
++) {
265 * no page there, look one up / create it
267 page
= find_or_create_page(mapping
, index
,
268 mapping_gfp_mask(mapping
));
272 if (PageUptodate(page
))
275 int error
= mapping
->a_ops
->readpage(in
, page
);
277 if (unlikely(error
)) {
278 page_cache_release(page
);
286 return move_to_pipe(pipe
, pages
, i
, offset
, len
, flags
);
292 * generic_file_splice_read - splice data from file to a pipe
293 * @in: file to splice from
294 * @pipe: pipe to splice to
295 * @len: number of bytes to splice
296 * @flags: splice modifier flags
298 * Will read pages from given file and fill them into a pipe.
301 ssize_t
generic_file_splice_read(struct file
*in
, struct pipe_inode_info
*pipe
,
302 size_t len
, unsigned int flags
)
311 ret
= __generic_file_splice_read(in
, pipe
, len
, flags
);
320 if (!(flags
& SPLICE_F_NONBLOCK
))
332 EXPORT_SYMBOL(generic_file_splice_read
);
335 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
338 static int pipe_to_sendpage(struct pipe_inode_info
*info
,
339 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
341 struct file
*file
= sd
->file
;
342 loff_t pos
= sd
->pos
;
349 * sub-optimal, but we are limited by the pipe ->map. we don't
350 * need a kmap'ed buffer here, we just want to make sure we
351 * have the page pinned if the pipe page originates from the
354 ptr
= buf
->ops
->map(file
, info
, buf
);
358 offset
= pos
& ~PAGE_CACHE_MASK
;
359 more
= (sd
->flags
& SPLICE_F_MORE
) || sd
->len
< sd
->total_len
;
361 ret
= file
->f_op
->sendpage(file
, buf
->page
, offset
, sd
->len
, &pos
,more
);
363 buf
->ops
->unmap(info
, buf
);
371 * This is a little more tricky than the file -> pipe splicing. There are
372 * basically three cases:
374 * - Destination page already exists in the address space and there
375 * are users of it. For that case we have no other option that
376 * copying the data. Tough luck.
377 * - Destination page already exists in the address space, but there
378 * are no users of it. Make sure it's uptodate, then drop it. Fall
379 * through to last case.
380 * - Destination page does not exist, we can add the pipe page to
381 * the page cache and avoid the copy.
383 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
384 * sd->flags), we attempt to migrate pages from the pipe to the output
385 * file address space page cache. This is possible if no one else has
386 * the pipe page referenced outside of the pipe and page cache. If
387 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
388 * a new page in the output file page cache and fill/dirty that.
390 static int pipe_to_file(struct pipe_inode_info
*info
, struct pipe_buffer
*buf
,
391 struct splice_desc
*sd
)
393 struct file
*file
= sd
->file
;
394 struct address_space
*mapping
= file
->f_mapping
;
395 gfp_t gfp_mask
= mapping_gfp_mask(mapping
);
403 * make sure the data in this buffer is uptodate
405 src
= buf
->ops
->map(file
, info
, buf
);
409 index
= sd
->pos
>> PAGE_CACHE_SHIFT
;
410 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
413 * reuse buf page, if SPLICE_F_MOVE is set
415 if (sd
->flags
& SPLICE_F_MOVE
) {
417 * If steal succeeds, buf->page is now pruned from the vm
418 * side (LRU and page cache) and we can reuse it.
420 if (buf
->ops
->steal(info
, buf
))
424 * this will also set the page locked
427 if (add_to_page_cache(page
, mapping
, index
, gfp_mask
))
430 if (!(buf
->flags
& PIPE_BUF_FLAG_LRU
))
435 page
= find_or_create_page(mapping
, index
, gfp_mask
);
440 * If the page is uptodate, it is also locked. If it isn't
441 * uptodate, we can mark it uptodate if we are filling the
442 * full page. Otherwise we need to read it in first...
444 if (!PageUptodate(page
)) {
445 if (sd
->len
< PAGE_CACHE_SIZE
) {
446 ret
= mapping
->a_ops
->readpage(file
, page
);
452 if (!PageUptodate(page
)) {
454 * page got invalidated, repeat
456 if (!page
->mapping
) {
458 page_cache_release(page
);
465 WARN_ON(!PageLocked(page
));
466 SetPageUptodate(page
);
471 ret
= mapping
->a_ops
->prepare_write(file
, page
, 0, sd
->len
);
472 if (ret
== AOP_TRUNCATED_PAGE
) {
473 page_cache_release(page
);
478 if (!(buf
->flags
& PIPE_BUF_FLAG_STOLEN
)) {
479 char *dst
= kmap_atomic(page
, KM_USER0
);
481 memcpy(dst
+ offset
, src
+ buf
->offset
, sd
->len
);
482 flush_dcache_page(page
);
483 kunmap_atomic(dst
, KM_USER0
);
486 ret
= mapping
->a_ops
->commit_write(file
, page
, 0, sd
->len
);
487 if (ret
== AOP_TRUNCATED_PAGE
) {
488 page_cache_release(page
);
493 mark_page_accessed(page
);
494 balance_dirty_pages_ratelimited(mapping
);
496 if (!(buf
->flags
& PIPE_BUF_FLAG_STOLEN
)) {
497 page_cache_release(page
);
501 buf
->ops
->unmap(info
, buf
);
505 typedef int (splice_actor
)(struct pipe_inode_info
*, struct pipe_buffer
*,
506 struct splice_desc
*);
509 * Pipe input worker. Most of this logic works like a regular pipe, the
510 * key here is the 'actor' worker passed in that actually moves the data
511 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
513 static ssize_t
move_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
514 size_t len
, unsigned int flags
,
517 int ret
, do_wakeup
, err
;
518 struct splice_desc sd
;
529 mutex_lock(&pipe
->inode
->i_mutex
);
532 int bufs
= pipe
->nrbufs
;
535 int curbuf
= pipe
->curbuf
;
536 struct pipe_buffer
*buf
= pipe
->bufs
+ curbuf
;
537 struct pipe_buf_operations
*ops
= buf
->ops
;
540 if (sd
.len
> sd
.total_len
)
541 sd
.len
= sd
.total_len
;
543 err
= actor(pipe
, buf
, &sd
);
545 if (!ret
&& err
!= -ENODATA
)
552 buf
->offset
+= sd
.len
;
556 ops
->release(pipe
, buf
);
557 curbuf
= (curbuf
+ 1) & (PIPE_BUFFERS
- 1);
558 pipe
->curbuf
= curbuf
;
559 pipe
->nrbufs
= --bufs
;
564 sd
.total_len
-= sd
.len
;
573 if (!pipe
->waiting_writers
) {
578 if (flags
& SPLICE_F_NONBLOCK
) {
584 if (signal_pending(current
)) {
592 if (waitqueue_active(&pipe
->wait
))
593 wake_up_interruptible_sync(&pipe
->wait
);
594 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
602 mutex_unlock(&pipe
->inode
->i_mutex
);
606 if (waitqueue_active(&pipe
->wait
))
607 wake_up_interruptible(&pipe
->wait
);
608 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
611 mutex_lock(&out
->f_mapping
->host
->i_mutex
);
613 mutex_unlock(&out
->f_mapping
->host
->i_mutex
);
619 * generic_file_splice_write - splice data from a pipe to a file
621 * @out: file to write to
622 * @len: number of bytes to splice
623 * @flags: splice modifier flags
625 * Will either move or copy pages (determined by @flags options) from
626 * the given pipe inode to the given file.
630 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
631 size_t len
, unsigned int flags
)
633 struct address_space
*mapping
= out
->f_mapping
;
636 ret
= move_from_pipe(pipe
, out
, len
, flags
, pipe_to_file
);
639 * if file or inode is SYNC and we actually wrote some data, sync it
641 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(mapping
->host
))
643 struct inode
*inode
= mapping
->host
;
646 mutex_lock(&inode
->i_mutex
);
647 err
= generic_osync_inode(mapping
->host
, mapping
,
648 OSYNC_METADATA
|OSYNC_DATA
);
649 mutex_unlock(&inode
->i_mutex
);
658 EXPORT_SYMBOL(generic_file_splice_write
);
661 * generic_splice_sendpage - splice data from a pipe to a socket
663 * @out: socket to write to
664 * @len: number of bytes to splice
665 * @flags: splice modifier flags
667 * Will send @len bytes from the pipe to a network socket. No data copying
671 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
672 size_t len
, unsigned int flags
)
674 return move_from_pipe(pipe
, out
, len
, flags
, pipe_to_sendpage
);
677 EXPORT_SYMBOL(generic_splice_sendpage
);
680 * Attempt to initiate a splice from pipe to file.
682 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
683 size_t len
, unsigned int flags
)
688 if (!out
->f_op
|| !out
->f_op
->splice_write
)
691 if (!(out
->f_mode
& FMODE_WRITE
))
696 ret
= rw_verify_area(WRITE
, out
, &pos
, len
);
697 if (unlikely(ret
< 0))
700 return out
->f_op
->splice_write(pipe
, out
, len
, flags
);
704 * Attempt to initiate a splice from a file to a pipe.
706 static long do_splice_to(struct file
*in
, struct pipe_inode_info
*pipe
,
707 size_t len
, unsigned int flags
)
709 loff_t pos
, isize
, left
;
712 if (!in
->f_op
|| !in
->f_op
->splice_read
)
715 if (!(in
->f_mode
& FMODE_READ
))
720 ret
= rw_verify_area(READ
, in
, &pos
, len
);
721 if (unlikely(ret
< 0))
724 isize
= i_size_read(in
->f_mapping
->host
);
725 if (unlikely(in
->f_pos
>= isize
))
728 left
= isize
- in
->f_pos
;
732 return in
->f_op
->splice_read(in
, pipe
, len
, flags
);
735 long do_splice_direct(struct file
*in
, struct file
*out
, size_t len
,
738 struct pipe_inode_info
*pipe
;
744 * We require the input being a regular file, as we don't want to
745 * randomly drop data for eg socket -> socket splicing. Use the
746 * piped splicing for that!
748 i_mode
= in
->f_dentry
->d_inode
->i_mode
;
749 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
753 * neither in nor out is a pipe, setup an internal pipe attached to
754 * 'out' and transfer the wanted data from 'in' to 'out' through that
756 pipe
= current
->splice_pipe
;
758 pipe
= alloc_pipe_info(NULL
);
763 * We don't have an immediate reader, but we'll read the stuff
764 * out of the pipe right after the move_to_pipe(). So set
765 * PIPE_READERS appropriately.
769 current
->splice_pipe
= pipe
;
779 size_t read_len
, max_read_len
;
782 * Do at most PIPE_BUFFERS pages worth of transfer:
784 max_read_len
= min(len
, (size_t)(PIPE_BUFFERS
*PAGE_SIZE
));
786 ret
= do_splice_to(in
, pipe
, max_read_len
, flags
);
787 if (unlikely(ret
< 0))
793 * NOTE: nonblocking mode only applies to the input. We
794 * must not do the output in nonblocking mode as then we
795 * could get stuck data in the internal pipe:
797 ret
= do_splice_from(pipe
, out
, read_len
,
798 flags
& ~SPLICE_F_NONBLOCK
);
799 if (unlikely(ret
< 0))
806 * In nonblocking mode, if we got back a short read then
807 * that was due to either an IO error or due to the
808 * pagecache entry not being there. In the IO error case
809 * the _next_ splice attempt will produce a clean IO error
810 * return value (not a short read), so in both cases it's
811 * correct to break out of the loop here:
813 if ((flags
& SPLICE_F_NONBLOCK
) && (read_len
< max_read_len
))
817 pipe
->nrbufs
= pipe
->curbuf
= 0;
823 * If we did an incomplete transfer we must release
824 * the pipe buffers in question:
826 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
827 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
830 buf
->ops
->release(pipe
, buf
);
834 pipe
->nrbufs
= pipe
->curbuf
= 0;
837 * If we transferred some data, return the number of bytes:
845 EXPORT_SYMBOL(do_splice_direct
);
848 * Determine where to splice to/from.
850 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
851 struct file
*out
, loff_t __user
*off_out
,
852 size_t len
, unsigned int flags
)
854 struct pipe_inode_info
*pipe
;
856 pipe
= in
->f_dentry
->d_inode
->i_pipe
;
861 if (out
->f_op
->llseek
== no_llseek
)
863 if (copy_from_user(&out
->f_pos
, off_out
,
868 return do_splice_from(pipe
, out
, len
, flags
);
871 pipe
= out
->f_dentry
->d_inode
->i_pipe
;
876 if (in
->f_op
->llseek
== no_llseek
)
878 if (copy_from_user(&in
->f_pos
, off_in
, sizeof(loff_t
)))
882 return do_splice_to(in
, pipe
, len
, flags
);
888 asmlinkage
long sys_splice(int fd_in
, loff_t __user
*off_in
,
889 int fd_out
, loff_t __user
*off_out
,
890 size_t len
, unsigned int flags
)
893 struct file
*in
, *out
;
894 int fput_in
, fput_out
;
900 in
= fget_light(fd_in
, &fput_in
);
902 if (in
->f_mode
& FMODE_READ
) {
903 out
= fget_light(fd_out
, &fput_out
);
905 if (out
->f_mode
& FMODE_WRITE
)
906 error
= do_splice(in
, off_in
,
909 fput_light(out
, fput_out
);
913 fput_light(in
, fput_in
);