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 (PagePrivate(page
) && !try_to_release_page(page
, GFP_KERNEL
))
66 * If we succeeded in removing the mapping, set LRU flag
69 if (remove_mapping(mapping
, page
)) {
70 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
76 * Raced with truncate or failed to remove page from current
77 * address space, unlock and return failure.
84 static void page_cache_pipe_buf_release(struct pipe_inode_info
*pipe
,
85 struct pipe_buffer
*buf
)
87 page_cache_release(buf
->page
);
88 buf
->flags
&= ~PIPE_BUF_FLAG_LRU
;
92 * Check whether the contents of buf is OK to access. Since the content
93 * is a page cache page, IO may be in flight.
95 static int page_cache_pipe_buf_confirm(struct pipe_inode_info
*pipe
,
96 struct pipe_buffer
*buf
)
98 struct page
*page
= buf
->page
;
101 if (!PageUptodate(page
)) {
105 * Page got truncated/unhashed. This will cause a 0-byte
106 * splice, if this is the first page.
108 if (!page
->mapping
) {
114 * Uh oh, read-error from disk.
116 if (!PageUptodate(page
)) {
122 * Page is ok afterall, we are done.
133 static const struct pipe_buf_operations page_cache_pipe_buf_ops
= {
135 .map
= generic_pipe_buf_map
,
136 .unmap
= generic_pipe_buf_unmap
,
137 .confirm
= page_cache_pipe_buf_confirm
,
138 .release
= page_cache_pipe_buf_release
,
139 .steal
= page_cache_pipe_buf_steal
,
140 .get
= generic_pipe_buf_get
,
143 static int user_page_pipe_buf_steal(struct pipe_inode_info
*pipe
,
144 struct pipe_buffer
*buf
)
146 if (!(buf
->flags
& PIPE_BUF_FLAG_GIFT
))
149 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
150 return generic_pipe_buf_steal(pipe
, buf
);
153 static const struct pipe_buf_operations user_page_pipe_buf_ops
= {
155 .map
= generic_pipe_buf_map
,
156 .unmap
= generic_pipe_buf_unmap
,
157 .confirm
= generic_pipe_buf_confirm
,
158 .release
= page_cache_pipe_buf_release
,
159 .steal
= user_page_pipe_buf_steal
,
160 .get
= generic_pipe_buf_get
,
164 * splice_to_pipe - fill passed data into a pipe
165 * @pipe: pipe to fill
169 * @spd contains a map of pages and len/offset tuples, along with
170 * the struct pipe_buf_operations associated with these pages. This
171 * function will link that data to the pipe.
174 ssize_t
splice_to_pipe(struct pipe_inode_info
*pipe
,
175 struct splice_pipe_desc
*spd
)
177 unsigned int spd_pages
= spd
->nr_pages
;
178 int ret
, do_wakeup
, page_nr
;
185 mutex_lock(&pipe
->inode
->i_mutex
);
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
--;
248 mutex_unlock(&pipe
->inode
->i_mutex
);
252 if (waitqueue_active(&pipe
->wait
))
253 wake_up_interruptible(&pipe
->wait
);
254 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
258 while (page_nr
< spd_pages
)
259 spd
->spd_release(spd
, page_nr
++);
264 static void spd_release_page(struct splice_pipe_desc
*spd
, unsigned int i
)
266 page_cache_release(spd
->pages
[i
]);
270 __generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
271 struct pipe_inode_info
*pipe
, size_t len
,
274 struct address_space
*mapping
= in
->f_mapping
;
275 unsigned int loff
, nr_pages
, req_pages
;
276 struct page
*pages
[PIPE_BUFFERS
];
277 struct partial_page partial
[PIPE_BUFFERS
];
279 pgoff_t index
, end_index
;
282 struct splice_pipe_desc spd
= {
286 .ops
= &page_cache_pipe_buf_ops
,
287 .spd_release
= spd_release_page
,
290 index
= *ppos
>> PAGE_CACHE_SHIFT
;
291 loff
= *ppos
& ~PAGE_CACHE_MASK
;
292 req_pages
= (len
+ loff
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
293 nr_pages
= min(req_pages
, (unsigned)PIPE_BUFFERS
);
296 * Lookup the (hopefully) full range of pages we need.
298 spd
.nr_pages
= find_get_pages_contig(mapping
, index
, nr_pages
, pages
);
299 index
+= spd
.nr_pages
;
302 * If find_get_pages_contig() returned fewer pages than we needed,
303 * readahead/allocate the rest and fill in the holes.
305 if (spd
.nr_pages
< nr_pages
)
306 page_cache_sync_readahead(mapping
, &in
->f_ra
, in
,
307 index
, req_pages
- spd
.nr_pages
);
310 while (spd
.nr_pages
< nr_pages
) {
312 * Page could be there, find_get_pages_contig() breaks on
315 page
= find_get_page(mapping
, index
);
318 * page didn't exist, allocate one.
320 page
= page_cache_alloc_cold(mapping
);
324 error
= add_to_page_cache_lru(page
, mapping
, index
,
325 mapping_gfp_mask(mapping
));
326 if (unlikely(error
)) {
327 page_cache_release(page
);
328 if (error
== -EEXIST
)
333 * add_to_page_cache() locks the page, unlock it
334 * to avoid convoluting the logic below even more.
339 pages
[spd
.nr_pages
++] = page
;
344 * Now loop over the map and see if we need to start IO on any
345 * pages, fill in the partial map, etc.
347 index
= *ppos
>> PAGE_CACHE_SHIFT
;
348 nr_pages
= spd
.nr_pages
;
350 for (page_nr
= 0; page_nr
< nr_pages
; page_nr
++) {
351 unsigned int this_len
;
357 * this_len is the max we'll use from this page
359 this_len
= min_t(unsigned long, len
, PAGE_CACHE_SIZE
- loff
);
360 page
= pages
[page_nr
];
362 if (PageReadahead(page
))
363 page_cache_async_readahead(mapping
, &in
->f_ra
, in
,
364 page
, index
, req_pages
- page_nr
);
367 * If the page isn't uptodate, we may need to start io on it
369 if (!PageUptodate(page
)) {
371 * If in nonblock mode then dont block on waiting
372 * for an in-flight io page
374 if (flags
& SPLICE_F_NONBLOCK
) {
375 if (!trylock_page(page
)) {
383 * Page was truncated, or invalidated by the
384 * filesystem. Redo the find/create, but this time the
385 * page is kept locked, so there's no chance of another
386 * race with truncate/invalidate.
388 if (!page
->mapping
) {
390 page
= find_or_create_page(mapping
, index
,
391 mapping_gfp_mask(mapping
));
397 page_cache_release(pages
[page_nr
]);
398 pages
[page_nr
] = page
;
401 * page was already under io and is now done, great
403 if (PageUptodate(page
)) {
409 * need to read in the page
411 error
= mapping
->a_ops
->readpage(in
, page
);
412 if (unlikely(error
)) {
414 * We really should re-lookup the page here,
415 * but it complicates things a lot. Instead
416 * lets just do what we already stored, and
417 * we'll get it the next time we are called.
419 if (error
== AOP_TRUNCATED_PAGE
)
427 * i_size must be checked after PageUptodate.
429 isize
= i_size_read(mapping
->host
);
430 end_index
= (isize
- 1) >> PAGE_CACHE_SHIFT
;
431 if (unlikely(!isize
|| index
> end_index
))
435 * if this is the last page, see if we need to shrink
436 * the length and stop
438 if (end_index
== index
) {
442 * max good bytes in this page
444 plen
= ((isize
- 1) & ~PAGE_CACHE_MASK
) + 1;
449 * force quit after adding this page
451 this_len
= min(this_len
, plen
- loff
);
455 partial
[page_nr
].offset
= loff
;
456 partial
[page_nr
].len
= this_len
;
464 * Release any pages at the end, if we quit early. 'page_nr' is how far
465 * we got, 'nr_pages' is how many pages are in the map.
467 while (page_nr
< nr_pages
)
468 page_cache_release(pages
[page_nr
++]);
469 in
->f_ra
.prev_pos
= (loff_t
)index
<< PAGE_CACHE_SHIFT
;
472 return splice_to_pipe(pipe
, &spd
);
478 * generic_file_splice_read - splice data from file to a pipe
479 * @in: file to splice from
480 * @ppos: position in @in
481 * @pipe: pipe to splice to
482 * @len: number of bytes to splice
483 * @flags: splice modifier flags
486 * Will read pages from given file and fill them into a pipe. Can be
487 * used as long as the address_space operations for the source implements
491 ssize_t
generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
492 struct pipe_inode_info
*pipe
, size_t len
,
498 isize
= i_size_read(in
->f_mapping
->host
);
499 if (unlikely(*ppos
>= isize
))
502 left
= isize
- *ppos
;
503 if (unlikely(left
< len
))
506 ret
= __generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
513 EXPORT_SYMBOL(generic_file_splice_read
);
516 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
517 * using sendpage(). Return the number of bytes sent.
519 static int pipe_to_sendpage(struct pipe_inode_info
*pipe
,
520 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
522 struct file
*file
= sd
->u
.file
;
523 loff_t pos
= sd
->pos
;
526 ret
= buf
->ops
->confirm(pipe
, buf
);
528 more
= (sd
->flags
& SPLICE_F_MORE
) || sd
->len
< sd
->total_len
;
530 ret
= file
->f_op
->sendpage(file
, buf
->page
, buf
->offset
,
531 sd
->len
, &pos
, more
);
538 * This is a little more tricky than the file -> pipe splicing. There are
539 * basically three cases:
541 * - Destination page already exists in the address space and there
542 * are users of it. For that case we have no other option that
543 * copying the data. Tough luck.
544 * - Destination page already exists in the address space, but there
545 * are no users of it. Make sure it's uptodate, then drop it. Fall
546 * through to last case.
547 * - Destination page does not exist, we can add the pipe page to
548 * the page cache and avoid the copy.
550 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
551 * sd->flags), we attempt to migrate pages from the pipe to the output
552 * file address space page cache. This is possible if no one else has
553 * the pipe page referenced outside of the pipe and page cache. If
554 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
555 * a new page in the output file page cache and fill/dirty that.
557 static int pipe_to_file(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
558 struct splice_desc
*sd
)
560 struct file
*file
= sd
->u
.file
;
561 struct address_space
*mapping
= file
->f_mapping
;
562 unsigned int offset
, this_len
;
568 * make sure the data in this buffer is uptodate
570 ret
= buf
->ops
->confirm(pipe
, buf
);
574 offset
= sd
->pos
& ~PAGE_CACHE_MASK
;
577 if (this_len
+ offset
> PAGE_CACHE_SIZE
)
578 this_len
= PAGE_CACHE_SIZE
- offset
;
580 ret
= pagecache_write_begin(file
, mapping
, sd
->pos
, this_len
,
581 AOP_FLAG_UNINTERRUPTIBLE
, &page
, &fsdata
);
585 if (buf
->page
!= page
) {
587 * Careful, ->map() uses KM_USER0!
589 char *src
= buf
->ops
->map(pipe
, buf
, 1);
590 char *dst
= kmap_atomic(page
, KM_USER1
);
592 memcpy(dst
+ offset
, src
+ buf
->offset
, this_len
);
593 flush_dcache_page(page
);
594 kunmap_atomic(dst
, KM_USER1
);
595 buf
->ops
->unmap(pipe
, buf
, src
);
597 ret
= pagecache_write_end(file
, mapping
, sd
->pos
, this_len
, this_len
,
604 * __splice_from_pipe - splice data from a pipe to given actor
605 * @pipe: pipe to splice from
606 * @sd: information to @actor
607 * @actor: handler that splices the data
610 * This function does little more than loop over the pipe and call
611 * @actor to do the actual moving of a single struct pipe_buffer to
612 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
616 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
619 int ret
, do_wakeup
, err
;
626 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
627 const struct pipe_buf_operations
*ops
= buf
->ops
;
630 if (sd
->len
> sd
->total_len
)
631 sd
->len
= sd
->total_len
;
633 err
= actor(pipe
, buf
, sd
);
635 if (!ret
&& err
!= -ENODATA
)
647 sd
->total_len
-= err
;
653 ops
->release(pipe
, buf
);
654 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (PIPE_BUFFERS
- 1);
668 if (!pipe
->waiting_writers
) {
673 if (sd
->flags
& SPLICE_F_NONBLOCK
) {
679 if (signal_pending(current
)) {
687 if (waitqueue_active(&pipe
->wait
))
688 wake_up_interruptible_sync(&pipe
->wait
);
689 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
698 if (waitqueue_active(&pipe
->wait
))
699 wake_up_interruptible(&pipe
->wait
);
700 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
705 EXPORT_SYMBOL(__splice_from_pipe
);
708 * splice_from_pipe - splice data from a pipe to a file
709 * @pipe: pipe to splice from
710 * @out: file to splice to
711 * @ppos: position in @out
712 * @len: how many bytes to splice
713 * @flags: splice modifier flags
714 * @actor: handler that splices the data
717 * See __splice_from_pipe. This function locks the input and output inodes,
718 * otherwise it's identical to __splice_from_pipe().
721 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
722 loff_t
*ppos
, size_t len
, unsigned int flags
,
726 struct inode
*inode
= out
->f_mapping
->host
;
727 struct splice_desc sd
= {
735 * The actor worker might be calling ->write_begin and
736 * ->write_end. Most of the time, these expect i_mutex to
737 * be held. Since this may result in an ABBA deadlock with
738 * pipe->inode, we have to order lock acquiry here.
740 * Outer lock must be inode->i_mutex, as pipe_wait() will
741 * release and reacquire pipe->inode->i_mutex, AND inode must
744 WARN_ON(S_ISFIFO(inode
->i_mode
));
745 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_PARENT
);
747 mutex_lock_nested(&pipe
->inode
->i_mutex
, I_MUTEX_CHILD
);
748 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
750 mutex_unlock(&pipe
->inode
->i_mutex
);
751 mutex_unlock(&inode
->i_mutex
);
757 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
759 * @out: file to write to
760 * @ppos: position in @out
761 * @len: number of bytes to splice
762 * @flags: splice modifier flags
765 * Will either move or copy pages (determined by @flags options) from
766 * the given pipe inode to the given file. The caller is responsible
767 * for acquiring i_mutex on both inodes.
771 generic_file_splice_write_nolock(struct pipe_inode_info
*pipe
, struct file
*out
,
772 loff_t
*ppos
, size_t len
, unsigned int flags
)
774 struct address_space
*mapping
= out
->f_mapping
;
775 struct inode
*inode
= mapping
->host
;
776 struct splice_desc sd
= {
785 err
= file_remove_suid(out
);
789 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_file
);
791 unsigned long nr_pages
;
794 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
797 * If file or inode is SYNC and we actually wrote some data,
800 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
801 err
= generic_osync_inode(inode
, mapping
,
802 OSYNC_METADATA
|OSYNC_DATA
);
807 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
813 EXPORT_SYMBOL(generic_file_splice_write_nolock
);
816 * generic_file_splice_write - splice data from a pipe to a file
818 * @out: file to write to
819 * @ppos: position in @out
820 * @len: number of bytes to splice
821 * @flags: splice modifier flags
824 * Will either move or copy pages (determined by @flags options) from
825 * the given pipe inode to the given file.
829 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
830 loff_t
*ppos
, size_t len
, unsigned int flags
)
832 struct address_space
*mapping
= out
->f_mapping
;
833 struct inode
*inode
= mapping
->host
;
834 struct splice_desc sd
= {
842 WARN_ON(S_ISFIFO(inode
->i_mode
));
843 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_PARENT
);
844 ret
= file_remove_suid(out
);
847 mutex_lock_nested(&pipe
->inode
->i_mutex
, I_MUTEX_CHILD
);
848 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_file
);
850 mutex_unlock(&pipe
->inode
->i_mutex
);
852 mutex_unlock(&inode
->i_mutex
);
854 unsigned long nr_pages
;
857 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
860 * If file or inode is SYNC and we actually wrote some data,
863 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
866 mutex_lock(&inode
->i_mutex
);
867 err
= generic_osync_inode(inode
, mapping
,
868 OSYNC_METADATA
|OSYNC_DATA
);
869 mutex_unlock(&inode
->i_mutex
);
874 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
880 EXPORT_SYMBOL(generic_file_splice_write
);
883 * generic_splice_sendpage - splice data from a pipe to a socket
884 * @pipe: pipe to splice from
885 * @out: socket to write to
886 * @ppos: position in @out
887 * @len: number of bytes to splice
888 * @flags: splice modifier flags
891 * Will send @len bytes from the pipe to a network socket. No data copying
895 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
896 loff_t
*ppos
, size_t len
, unsigned int flags
)
898 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
901 EXPORT_SYMBOL(generic_splice_sendpage
);
904 * Attempt to initiate a splice from pipe to file.
906 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
907 loff_t
*ppos
, size_t len
, unsigned int flags
)
911 if (unlikely(!out
->f_op
|| !out
->f_op
->splice_write
))
914 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
917 if (unlikely(out
->f_flags
& O_APPEND
))
920 ret
= rw_verify_area(WRITE
, out
, ppos
, len
);
921 if (unlikely(ret
< 0))
924 return out
->f_op
->splice_write(pipe
, out
, ppos
, len
, flags
);
928 * Attempt to initiate a splice from a file to a pipe.
930 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
931 struct pipe_inode_info
*pipe
, size_t len
,
936 if (unlikely(!in
->f_op
|| !in
->f_op
->splice_read
))
939 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
942 ret
= rw_verify_area(READ
, in
, ppos
, len
);
943 if (unlikely(ret
< 0))
946 return in
->f_op
->splice_read(in
, ppos
, pipe
, len
, flags
);
950 * splice_direct_to_actor - splices data directly between two non-pipes
951 * @in: file to splice from
952 * @sd: actor information on where to splice to
953 * @actor: handles the data splicing
956 * This is a special case helper to splice directly between two
957 * points, without requiring an explicit pipe. Internally an allocated
958 * pipe is cached in the process, and reused during the lifetime of
962 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
963 splice_direct_actor
*actor
)
965 struct pipe_inode_info
*pipe
;
972 * We require the input being a regular file, as we don't want to
973 * randomly drop data for eg socket -> socket splicing. Use the
974 * piped splicing for that!
976 i_mode
= in
->f_path
.dentry
->d_inode
->i_mode
;
977 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
981 * neither in nor out is a pipe, setup an internal pipe attached to
982 * 'out' and transfer the wanted data from 'in' to 'out' through that
984 pipe
= current
->splice_pipe
;
985 if (unlikely(!pipe
)) {
986 pipe
= alloc_pipe_info(NULL
);
991 * We don't have an immediate reader, but we'll read the stuff
992 * out of the pipe right after the splice_to_pipe(). So set
993 * PIPE_READERS appropriately.
997 current
->splice_pipe
= pipe
;
1005 len
= sd
->total_len
;
1009 * Don't block on output, we have to drain the direct pipe.
1011 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1015 loff_t pos
= sd
->pos
, prev_pos
= pos
;
1017 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1018 if (unlikely(ret
<= 0))
1022 sd
->total_len
= read_len
;
1025 * NOTE: nonblocking mode only applies to the input. We
1026 * must not do the output in nonblocking mode as then we
1027 * could get stuck data in the internal pipe:
1029 ret
= actor(pipe
, sd
);
1030 if (unlikely(ret
<= 0)) {
1039 if (ret
< read_len
) {
1040 sd
->pos
= prev_pos
+ ret
;
1046 pipe
->nrbufs
= pipe
->curbuf
= 0;
1052 * If we did an incomplete transfer we must release
1053 * the pipe buffers in question:
1055 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
1056 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1059 buf
->ops
->release(pipe
, buf
);
1069 EXPORT_SYMBOL(splice_direct_to_actor
);
1071 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1072 struct splice_desc
*sd
)
1074 struct file
*file
= sd
->u
.file
;
1076 return do_splice_from(pipe
, file
, &sd
->pos
, sd
->total_len
, sd
->flags
);
1080 * do_splice_direct - splices data directly between two files
1081 * @in: file to splice from
1082 * @ppos: input file offset
1083 * @out: file to splice to
1084 * @len: number of bytes to splice
1085 * @flags: splice modifier flags
1088 * For use by do_sendfile(). splice can easily emulate sendfile, but
1089 * doing it in the application would incur an extra system call
1090 * (splice in + splice out, as compared to just sendfile()). So this helper
1091 * can splice directly through a process-private pipe.
1094 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1095 size_t len
, unsigned int flags
)
1097 struct splice_desc sd
= {
1106 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1114 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1115 * location, so checking ->i_pipe is not enough to verify that this is a
1118 static inline struct pipe_inode_info
*pipe_info(struct inode
*inode
)
1120 if (S_ISFIFO(inode
->i_mode
))
1121 return inode
->i_pipe
;
1127 * Determine where to splice to/from.
1129 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1130 struct file
*out
, loff_t __user
*off_out
,
1131 size_t len
, unsigned int flags
)
1133 struct pipe_inode_info
*pipe
;
1134 loff_t offset
, *off
;
1137 pipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1142 if (out
->f_op
->llseek
== no_llseek
)
1144 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1150 ret
= do_splice_from(pipe
, out
, off
, len
, flags
);
1152 if (off_out
&& copy_to_user(off_out
, off
, sizeof(loff_t
)))
1158 pipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1163 if (in
->f_op
->llseek
== no_llseek
)
1165 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1171 ret
= do_splice_to(in
, off
, pipe
, len
, flags
);
1173 if (off_in
&& copy_to_user(off_in
, off
, sizeof(loff_t
)))
1183 * Map an iov into an array of pages and offset/length tupples. With the
1184 * partial_page structure, we can map several non-contiguous ranges into
1185 * our ones pages[] map instead of splitting that operation into pieces.
1186 * Could easily be exported as a generic helper for other users, in which
1187 * case one would probably want to add a 'max_nr_pages' parameter as well.
1189 static int get_iovec_page_array(const struct iovec __user
*iov
,
1190 unsigned int nr_vecs
, struct page
**pages
,
1191 struct partial_page
*partial
, int aligned
)
1193 int buffers
= 0, error
= 0;
1196 unsigned long off
, npages
;
1203 if (copy_from_user(&entry
, iov
, sizeof(entry
)))
1206 base
= entry
.iov_base
;
1207 len
= entry
.iov_len
;
1210 * Sanity check this iovec. 0 read succeeds.
1216 if (!access_ok(VERIFY_READ
, base
, len
))
1220 * Get this base offset and number of pages, then map
1221 * in the user pages.
1223 off
= (unsigned long) base
& ~PAGE_MASK
;
1226 * If asked for alignment, the offset must be zero and the
1227 * length a multiple of the PAGE_SIZE.
1230 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1233 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1234 if (npages
> PIPE_BUFFERS
- buffers
)
1235 npages
= PIPE_BUFFERS
- buffers
;
1237 error
= get_user_pages_fast((unsigned long)base
, npages
,
1238 0, &pages
[buffers
]);
1240 if (unlikely(error
<= 0))
1244 * Fill this contiguous range into the partial page map.
1246 for (i
= 0; i
< error
; i
++) {
1247 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1249 partial
[buffers
].offset
= off
;
1250 partial
[buffers
].len
= plen
;
1258 * We didn't complete this iov, stop here since it probably
1259 * means we have to move some of this into a pipe to
1260 * be able to continue.
1266 * Don't continue if we mapped fewer pages than we asked for,
1267 * or if we mapped the max number of pages that we have
1270 if (error
< npages
|| buffers
== PIPE_BUFFERS
)
1283 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1284 struct splice_desc
*sd
)
1289 ret
= buf
->ops
->confirm(pipe
, buf
);
1294 * See if we can use the atomic maps, by prefaulting in the
1295 * pages and doing an atomic copy
1297 if (!fault_in_pages_writeable(sd
->u
.userptr
, sd
->len
)) {
1298 src
= buf
->ops
->map(pipe
, buf
, 1);
1299 ret
= __copy_to_user_inatomic(sd
->u
.userptr
, src
+ buf
->offset
,
1301 buf
->ops
->unmap(pipe
, buf
, src
);
1309 * No dice, use slow non-atomic map and copy
1311 src
= buf
->ops
->map(pipe
, buf
, 0);
1314 if (copy_to_user(sd
->u
.userptr
, src
+ buf
->offset
, sd
->len
))
1317 buf
->ops
->unmap(pipe
, buf
, src
);
1320 sd
->u
.userptr
+= ret
;
1325 * For lack of a better implementation, implement vmsplice() to userspace
1326 * as a simple copy of the pipes pages to the user iov.
1328 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*iov
,
1329 unsigned long nr_segs
, unsigned int flags
)
1331 struct pipe_inode_info
*pipe
;
1332 struct splice_desc sd
;
1337 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1342 mutex_lock(&pipe
->inode
->i_mutex
);
1350 * Get user address base and length for this iovec.
1352 error
= get_user(base
, &iov
->iov_base
);
1353 if (unlikely(error
))
1355 error
= get_user(len
, &iov
->iov_len
);
1356 if (unlikely(error
))
1360 * Sanity check this iovec. 0 read succeeds.
1364 if (unlikely(!base
)) {
1369 if (unlikely(!access_ok(VERIFY_WRITE
, base
, len
))) {
1377 sd
.u
.userptr
= base
;
1380 size
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1398 mutex_unlock(&pipe
->inode
->i_mutex
);
1407 * vmsplice splices a user address range into a pipe. It can be thought of
1408 * as splice-from-memory, where the regular splice is splice-from-file (or
1409 * to file). In both cases the output is a pipe, naturally.
1411 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1412 unsigned long nr_segs
, unsigned int flags
)
1414 struct pipe_inode_info
*pipe
;
1415 struct page
*pages
[PIPE_BUFFERS
];
1416 struct partial_page partial
[PIPE_BUFFERS
];
1417 struct splice_pipe_desc spd
= {
1421 .ops
= &user_page_pipe_buf_ops
,
1422 .spd_release
= spd_release_page
,
1425 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1429 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, pages
, partial
,
1430 flags
& SPLICE_F_GIFT
);
1431 if (spd
.nr_pages
<= 0)
1432 return spd
.nr_pages
;
1434 return splice_to_pipe(pipe
, &spd
);
1438 * Note that vmsplice only really supports true splicing _from_ user memory
1439 * to a pipe, not the other way around. Splicing from user memory is a simple
1440 * operation that can be supported without any funky alignment restrictions
1441 * or nasty vm tricks. We simply map in the user memory and fill them into
1442 * a pipe. The reverse isn't quite as easy, though. There are two possible
1443 * solutions for that:
1445 * - memcpy() the data internally, at which point we might as well just
1446 * do a regular read() on the buffer anyway.
1447 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1448 * has restriction limitations on both ends of the pipe).
1450 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1453 SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct iovec __user
*, iov
,
1454 unsigned long, nr_segs
, unsigned int, flags
)
1460 if (unlikely(nr_segs
> UIO_MAXIOV
))
1462 else if (unlikely(!nr_segs
))
1466 file
= fget_light(fd
, &fput
);
1468 if (file
->f_mode
& FMODE_WRITE
)
1469 error
= vmsplice_to_pipe(file
, iov
, nr_segs
, flags
);
1470 else if (file
->f_mode
& FMODE_READ
)
1471 error
= vmsplice_to_user(file
, iov
, nr_segs
, flags
);
1473 fput_light(file
, fput
);
1479 SYSCALL_DEFINE6(splice
, int, fd_in
, loff_t __user
*, off_in
,
1480 int, fd_out
, loff_t __user
*, off_out
,
1481 size_t, len
, unsigned int, flags
)
1484 struct file
*in
, *out
;
1485 int fput_in
, fput_out
;
1491 in
= fget_light(fd_in
, &fput_in
);
1493 if (in
->f_mode
& FMODE_READ
) {
1494 out
= fget_light(fd_out
, &fput_out
);
1496 if (out
->f_mode
& FMODE_WRITE
)
1497 error
= do_splice(in
, off_in
,
1500 fput_light(out
, fput_out
);
1504 fput_light(in
, fput_in
);
1511 * Make sure there's data to read. Wait for input if we can, otherwise
1512 * return an appropriate error.
1514 static int link_ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1519 * Check ->nrbufs without the inode lock first. This function
1520 * is speculative anyways, so missing one is ok.
1526 mutex_lock(&pipe
->inode
->i_mutex
);
1528 while (!pipe
->nrbufs
) {
1529 if (signal_pending(current
)) {
1535 if (!pipe
->waiting_writers
) {
1536 if (flags
& SPLICE_F_NONBLOCK
) {
1544 mutex_unlock(&pipe
->inode
->i_mutex
);
1549 * Make sure there's writeable room. Wait for room if we can, otherwise
1550 * return an appropriate error.
1552 static int link_opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1557 * Check ->nrbufs without the inode lock first. This function
1558 * is speculative anyways, so missing one is ok.
1560 if (pipe
->nrbufs
< PIPE_BUFFERS
)
1564 mutex_lock(&pipe
->inode
->i_mutex
);
1566 while (pipe
->nrbufs
>= PIPE_BUFFERS
) {
1567 if (!pipe
->readers
) {
1568 send_sig(SIGPIPE
, current
, 0);
1572 if (flags
& SPLICE_F_NONBLOCK
) {
1576 if (signal_pending(current
)) {
1580 pipe
->waiting_writers
++;
1582 pipe
->waiting_writers
--;
1585 mutex_unlock(&pipe
->inode
->i_mutex
);
1590 * Link contents of ipipe to opipe.
1592 static int link_pipe(struct pipe_inode_info
*ipipe
,
1593 struct pipe_inode_info
*opipe
,
1594 size_t len
, unsigned int flags
)
1596 struct pipe_buffer
*ibuf
, *obuf
;
1597 int ret
= 0, i
= 0, nbuf
;
1600 * Potential ABBA deadlock, work around it by ordering lock
1601 * grabbing by inode address. Otherwise two different processes
1602 * could deadlock (one doing tee from A -> B, the other from B -> A).
1604 inode_double_lock(ipipe
->inode
, opipe
->inode
);
1607 if (!opipe
->readers
) {
1608 send_sig(SIGPIPE
, current
, 0);
1615 * If we have iterated all input buffers or ran out of
1616 * output room, break.
1618 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= PIPE_BUFFERS
)
1621 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (PIPE_BUFFERS
- 1));
1622 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
1625 * Get a reference to this pipe buffer,
1626 * so we can copy the contents over.
1628 ibuf
->ops
->get(ipipe
, ibuf
);
1630 obuf
= opipe
->bufs
+ nbuf
;
1634 * Don't inherit the gift flag, we need to
1635 * prevent multiple steals of this page.
1637 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1639 if (obuf
->len
> len
)
1649 * return EAGAIN if we have the potential of some data in the
1650 * future, otherwise just return 0
1652 if (!ret
&& ipipe
->waiting_writers
&& (flags
& SPLICE_F_NONBLOCK
))
1655 inode_double_unlock(ipipe
->inode
, opipe
->inode
);
1658 * If we put data in the output pipe, wakeup any potential readers.
1662 if (waitqueue_active(&opipe
->wait
))
1663 wake_up_interruptible(&opipe
->wait
);
1664 kill_fasync(&opipe
->fasync_readers
, SIGIO
, POLL_IN
);
1671 * This is a tee(1) implementation that works on pipes. It doesn't copy
1672 * any data, it simply references the 'in' pages on the 'out' pipe.
1673 * The 'flags' used are the SPLICE_F_* variants, currently the only
1674 * applicable one is SPLICE_F_NONBLOCK.
1676 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1679 struct pipe_inode_info
*ipipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1680 struct pipe_inode_info
*opipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1684 * Duplicate the contents of ipipe to opipe without actually
1687 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
1689 * Keep going, unless we encounter an error. The ipipe/opipe
1690 * ordering doesn't really matter.
1692 ret
= link_ipipe_prep(ipipe
, flags
);
1694 ret
= link_opipe_prep(opipe
, flags
);
1696 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
1703 SYSCALL_DEFINE4(tee
, int, fdin
, int, fdout
, size_t, len
, unsigned int, flags
)
1712 in
= fget_light(fdin
, &fput_in
);
1714 if (in
->f_mode
& FMODE_READ
) {
1716 struct file
*out
= fget_light(fdout
, &fput_out
);
1719 if (out
->f_mode
& FMODE_WRITE
)
1720 error
= do_tee(in
, out
, len
, flags
);
1721 fput_light(out
, fput_out
);
1724 fput_light(in
, fput_in
);