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
));
55 if (PagePrivate(page
))
56 try_to_release_page(page
, mapping_gfp_mask(mapping
));
58 if (!remove_mapping(mapping
, page
))
62 struct zone
*zone
= page_zone(page
);
64 spin_lock_irq(&zone
->lru_lock
);
65 BUG_ON(!PageLRU(page
));
67 del_page_from_lru(zone
, page
);
68 spin_unlock_irq(&zone
->lru_lock
);
74 static void page_cache_pipe_buf_release(struct pipe_inode_info
*info
,
75 struct pipe_buffer
*buf
)
77 page_cache_release(buf
->page
);
81 static void *page_cache_pipe_buf_map(struct file
*file
,
82 struct pipe_inode_info
*info
,
83 struct pipe_buffer
*buf
)
85 struct page
*page
= buf
->page
;
89 if (!PageUptodate(page
)) {
96 return ERR_PTR(-ENODATA
);
99 return kmap(buf
->page
);
102 static void page_cache_pipe_buf_unmap(struct pipe_inode_info
*info
,
103 struct pipe_buffer
*buf
)
105 unlock_page(buf
->page
);
109 static struct pipe_buf_operations page_cache_pipe_buf_ops
= {
111 .map
= page_cache_pipe_buf_map
,
112 .unmap
= page_cache_pipe_buf_unmap
,
113 .release
= page_cache_pipe_buf_release
,
114 .steal
= page_cache_pipe_buf_steal
,
118 * Pipe output worker. This sets up our pipe format with the page cache
119 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
121 static ssize_t
move_to_pipe(struct inode
*inode
, struct page
**pages
,
122 int nr_pages
, unsigned long offset
,
123 unsigned long len
, unsigned int flags
)
125 struct pipe_inode_info
*info
;
126 int ret
, do_wakeup
, i
;
132 mutex_lock(PIPE_MUTEX(*inode
));
134 info
= inode
->i_pipe
;
138 if (!PIPE_READERS(*inode
)) {
139 send_sig(SIGPIPE
, current
, 0);
146 if (bufs
< PIPE_BUFFERS
) {
147 int newbuf
= (info
->curbuf
+ bufs
) & (PIPE_BUFFERS
- 1);
148 struct pipe_buffer
*buf
= info
->bufs
+ newbuf
;
149 struct page
*page
= pages
[i
++];
150 unsigned long this_len
;
152 this_len
= PAGE_CACHE_SIZE
- offset
;
157 buf
->offset
= offset
;
159 buf
->ops
= &page_cache_pipe_buf_ops
;
160 info
->nrbufs
= ++bufs
;
170 if (bufs
< PIPE_BUFFERS
)
176 if (flags
& SPLICE_F_NONBLOCK
) {
182 if (signal_pending(current
)) {
189 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
190 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
,
195 PIPE_WAITING_WRITERS(*inode
)++;
197 PIPE_WAITING_WRITERS(*inode
)--;
200 mutex_unlock(PIPE_MUTEX(*inode
));
203 wake_up_interruptible(PIPE_WAIT(*inode
));
204 kill_fasync(PIPE_FASYNC_READERS(*inode
), SIGIO
, POLL_IN
);
208 page_cache_release(pages
[i
++]);
213 static int __generic_file_splice_read(struct file
*in
, struct inode
*pipe
,
214 size_t len
, unsigned int flags
)
216 struct address_space
*mapping
= in
->f_mapping
;
217 unsigned int offset
, nr_pages
;
218 struct page
*pages
[PIPE_BUFFERS
], *shadow
[PIPE_BUFFERS
];
223 index
= in
->f_pos
>> PAGE_CACHE_SHIFT
;
224 offset
= in
->f_pos
& ~PAGE_CACHE_MASK
;
225 nr_pages
= (len
+ offset
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
227 if (nr_pages
> PIPE_BUFFERS
)
228 nr_pages
= PIPE_BUFFERS
;
231 * initiate read-ahead on this page range
233 do_page_cache_readahead(mapping
, in
, index
, nr_pages
);
236 * Get as many pages from the page cache as possible..
237 * Start IO on the page cache entries we create (we
238 * can assume that any pre-existing ones we find have
239 * already had IO started on them).
241 i
= find_get_pages(mapping
, index
, nr_pages
, pages
);
244 * common case - we found all pages and they are contiguous,
247 if (i
&& (pages
[i
- 1]->index
== index
+ i
- 1))
251 * fill shadow[] with pages at the right locations, so we only
254 memset(shadow
, 0, nr_pages
* sizeof(struct page
*));
255 for (j
= 0; j
< i
; j
++)
256 shadow
[pages
[j
]->index
- index
] = pages
[j
];
259 * now fill in the holes
261 for (i
= 0, pidx
= index
; i
< nr_pages
; pidx
++, i
++) {
268 * no page there, look one up / create it
270 page
= find_or_create_page(mapping
, pidx
,
271 mapping_gfp_mask(mapping
));
275 if (PageUptodate(page
))
278 error
= mapping
->a_ops
->readpage(in
, page
);
280 if (unlikely(error
)) {
281 page_cache_release(page
);
289 for (i
= 0; i
< nr_pages
; i
++) {
291 page_cache_release(shadow
[i
]);
296 memcpy(pages
, shadow
, i
* sizeof(struct page
*));
299 * Now we splice them into the pipe..
302 return move_to_pipe(pipe
, pages
, i
, offset
, len
, flags
);
306 * generic_file_splice_read - splice data from file to a pipe
307 * @in: file to splice from
308 * @pipe: pipe to splice to
309 * @len: number of bytes to splice
310 * @flags: splice modifier flags
312 * Will read pages from given file and fill them into a pipe.
315 ssize_t
generic_file_splice_read(struct file
*in
, struct inode
*pipe
,
316 size_t len
, unsigned int flags
)
324 ret
= __generic_file_splice_read(in
, pipe
, len
, flags
);
333 if (!(flags
& SPLICE_F_NONBLOCK
))
346 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
349 static int pipe_to_sendpage(struct pipe_inode_info
*info
,
350 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
352 struct file
*file
= sd
->file
;
353 loff_t pos
= sd
->pos
;
359 * sub-optimal, but we are limited by the pipe ->map. we don't
360 * need a kmap'ed buffer here, we just want to make sure we
361 * have the page pinned if the pipe page originates from the
364 ptr
= buf
->ops
->map(file
, info
, buf
);
368 offset
= pos
& ~PAGE_CACHE_MASK
;
370 ret
= file
->f_op
->sendpage(file
, buf
->page
, offset
, sd
->len
, &pos
,
371 sd
->len
< sd
->total_len
);
373 buf
->ops
->unmap(info
, buf
);
381 * This is a little more tricky than the file -> pipe splicing. There are
382 * basically three cases:
384 * - Destination page already exists in the address space and there
385 * are users of it. For that case we have no other option that
386 * copying the data. Tough luck.
387 * - Destination page already exists in the address space, but there
388 * are no users of it. Make sure it's uptodate, then drop it. Fall
389 * through to last case.
390 * - Destination page does not exist, we can add the pipe page to
391 * the page cache and avoid the copy.
393 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
394 * sd->flags), we attempt to migrate pages from the pipe to the output
395 * file address space page cache. This is possible if no one else has
396 * the pipe page referenced outside of the pipe and page cache. If
397 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
398 * a new page in the output file page cache and fill/dirty that.
400 static int pipe_to_file(struct pipe_inode_info
*info
, struct pipe_buffer
*buf
,
401 struct splice_desc
*sd
)
403 struct file
*file
= sd
->file
;
404 struct address_space
*mapping
= file
->f_mapping
;
412 * after this, page will be locked and unmapped
414 src
= buf
->ops
->map(file
, info
, buf
);
418 index
= sd
->pos
>> PAGE_CACHE_SHIFT
;
419 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
423 * reuse buf page, if SPLICE_F_MOVE is set
425 if (sd
->flags
& SPLICE_F_MOVE
) {
427 * If steal succeeds, buf->page is now pruned from the vm
428 * side (LRU and page cache) and we can reuse it.
430 if (buf
->ops
->steal(info
, buf
))
435 if (add_to_page_cache_lru(page
, mapping
, index
,
436 mapping_gfp_mask(mapping
)))
441 page
= find_or_create_page(mapping
, index
,
442 mapping_gfp_mask(mapping
));
447 * If the page is uptodate, it is also locked. If it isn't
448 * uptodate, we can mark it uptodate if we are filling the
449 * full page. Otherwise we need to read it in first...
451 if (!PageUptodate(page
)) {
452 if (sd
->len
< PAGE_CACHE_SIZE
) {
453 ret
= mapping
->a_ops
->readpage(file
, page
);
459 if (!PageUptodate(page
)) {
461 * page got invalidated, repeat
463 if (!page
->mapping
) {
465 page_cache_release(page
);
472 WARN_ON(!PageLocked(page
));
473 SetPageUptodate(page
);
478 ret
= mapping
->a_ops
->prepare_write(file
, page
, 0, sd
->len
);
479 if (ret
== AOP_TRUNCATED_PAGE
) {
480 page_cache_release(page
);
486 char *dst
= kmap_atomic(page
, KM_USER0
);
488 memcpy(dst
+ offset
, src
+ buf
->offset
, sd
->len
);
489 flush_dcache_page(page
);
490 kunmap_atomic(dst
, KM_USER0
);
493 ret
= mapping
->a_ops
->commit_write(file
, page
, 0, sd
->len
);
494 if (ret
== AOP_TRUNCATED_PAGE
) {
495 page_cache_release(page
);
500 balance_dirty_pages_ratelimited(mapping
);
503 page_cache_release(page
);
506 buf
->ops
->unmap(info
, buf
);
510 typedef int (splice_actor
)(struct pipe_inode_info
*, struct pipe_buffer
*,
511 struct splice_desc
*);
514 * Pipe input worker. Most of this logic works like a regular pipe, the
515 * key here is the 'actor' worker passed in that actually moves the data
516 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
518 static ssize_t
move_from_pipe(struct inode
*inode
, struct file
*out
,
519 size_t len
, unsigned int flags
,
522 struct pipe_inode_info
*info
;
523 int ret
, do_wakeup
, err
;
524 struct splice_desc sd
;
534 mutex_lock(PIPE_MUTEX(*inode
));
536 info
= inode
->i_pipe
;
538 int bufs
= info
->nrbufs
;
541 int curbuf
= info
->curbuf
;
542 struct pipe_buffer
*buf
= info
->bufs
+ curbuf
;
543 struct pipe_buf_operations
*ops
= buf
->ops
;
546 if (sd
.len
> sd
.total_len
)
547 sd
.len
= sd
.total_len
;
549 err
= actor(info
, buf
, &sd
);
551 if (!ret
&& err
!= -ENODATA
)
558 buf
->offset
+= sd
.len
;
562 ops
->release(info
, buf
);
563 curbuf
= (curbuf
+ 1) & (PIPE_BUFFERS
- 1);
564 info
->curbuf
= curbuf
;
565 info
->nrbufs
= --bufs
;
570 sd
.total_len
-= sd
.len
;
577 if (!PIPE_WRITERS(*inode
))
579 if (!PIPE_WAITING_WRITERS(*inode
)) {
584 if (flags
& SPLICE_F_NONBLOCK
) {
590 if (signal_pending(current
)) {
597 wake_up_interruptible_sync(PIPE_WAIT(*inode
));
598 kill_fasync(PIPE_FASYNC_WRITERS(*inode
),SIGIO
,POLL_OUT
);
605 mutex_unlock(PIPE_MUTEX(*inode
));
608 wake_up_interruptible(PIPE_WAIT(*inode
));
609 kill_fasync(PIPE_FASYNC_WRITERS(*inode
), SIGIO
, POLL_OUT
);
612 mutex_lock(&out
->f_mapping
->host
->i_mutex
);
614 mutex_unlock(&out
->f_mapping
->host
->i_mutex
);
620 * generic_file_splice_write - splice data from a pipe to a file
622 * @out: file to write to
623 * @len: number of bytes to splice
624 * @flags: splice modifier flags
626 * Will either move or copy pages (determined by @flags options) from
627 * the given pipe inode to the given file.
630 ssize_t
generic_file_splice_write(struct inode
*inode
, struct file
*out
,
631 size_t len
, unsigned int flags
)
633 struct address_space
*mapping
= out
->f_mapping
;
634 ssize_t ret
= move_from_pipe(inode
, out
, len
, flags
, pipe_to_file
);
637 * if file or inode is SYNC and we actually wrote some data, sync it
639 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(mapping
->host
))
641 struct inode
*inode
= mapping
->host
;
644 mutex_lock(&inode
->i_mutex
);
645 err
= generic_osync_inode(mapping
->host
, mapping
,
646 OSYNC_METADATA
|OSYNC_DATA
);
647 mutex_unlock(&inode
->i_mutex
);
657 * generic_splice_sendpage - splice data from a pipe to a socket
659 * @out: socket to write to
660 * @len: number of bytes to splice
661 * @flags: splice modifier flags
663 * Will send @len bytes from the pipe to a network socket. No data copying
667 ssize_t
generic_splice_sendpage(struct inode
*inode
, struct file
*out
,
668 size_t len
, unsigned int flags
)
670 return move_from_pipe(inode
, out
, len
, flags
, pipe_to_sendpage
);
673 EXPORT_SYMBOL(generic_file_splice_write
);
674 EXPORT_SYMBOL(generic_file_splice_read
);
677 * Attempt to initiate a splice from pipe to file.
679 static long do_splice_from(struct inode
*pipe
, struct file
*out
, size_t len
,
685 if (!out
->f_op
|| !out
->f_op
->splice_write
)
688 if (!(out
->f_mode
& FMODE_WRITE
))
692 ret
= rw_verify_area(WRITE
, out
, &pos
, len
);
693 if (unlikely(ret
< 0))
696 return out
->f_op
->splice_write(pipe
, out
, len
, flags
);
700 * Attempt to initiate a splice from a file to a pipe.
702 static long do_splice_to(struct file
*in
, struct inode
*pipe
, size_t len
,
705 loff_t pos
, isize
, left
;
708 if (!in
->f_op
|| !in
->f_op
->splice_read
)
711 if (!(in
->f_mode
& FMODE_READ
))
715 ret
= rw_verify_area(READ
, in
, &pos
, len
);
716 if (unlikely(ret
< 0))
719 isize
= i_size_read(in
->f_mapping
->host
);
720 if (unlikely(in
->f_pos
>= isize
))
723 left
= isize
- in
->f_pos
;
727 return in
->f_op
->splice_read(in
, pipe
, len
, flags
);
731 * Determine where to splice to/from.
733 static long do_splice(struct file
*in
, struct file
*out
, size_t len
,
738 pipe
= in
->f_dentry
->d_inode
;
740 return do_splice_from(pipe
, out
, len
, flags
);
742 pipe
= out
->f_dentry
->d_inode
;
744 return do_splice_to(in
, pipe
, len
, flags
);
749 asmlinkage
long sys_splice(int fdin
, int fdout
, size_t len
, unsigned int flags
)
752 struct file
*in
, *out
;
753 int fput_in
, fput_out
;
759 in
= fget_light(fdin
, &fput_in
);
761 if (in
->f_mode
& FMODE_READ
) {
762 out
= fget_light(fdout
, &fput_out
);
764 if (out
->f_mode
& FMODE_WRITE
)
765 error
= do_splice(in
, out
, len
, flags
);
766 fput_light(out
, fput_out
);
770 fput_light(in
, fput_in
);