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
,
603 static void wakeup_pipe_writers(struct pipe_inode_info
*pipe
)
606 if (waitqueue_active(&pipe
->wait
))
607 wake_up_interruptible(&pipe
->wait
);
608 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
612 * splice_from_pipe_feed - feed available data from a pipe to a file
613 * @pipe: pipe to splice from
614 * @sd: information to @actor
615 * @actor: handler that splices the data
619 * This function loops over the pipe and calls @actor to do the
620 * actual moving of a single struct pipe_buffer to the desired
621 * destination. It returns when there's no more buffers left in
622 * the pipe or if the requested number of bytes (@sd->total_len)
623 * have been copied. It returns a positive number (one) if the
624 * pipe needs to be filled with more data, zero if the required
625 * number of bytes have been copied and -errno on error.
627 * This, together with splice_from_pipe_{begin,end,next}, may be
628 * used to implement the functionality of __splice_from_pipe() when
629 * locking is required around copying the pipe buffers to the
632 int splice_from_pipe_feed(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
637 while (pipe
->nrbufs
) {
638 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
639 const struct pipe_buf_operations
*ops
= buf
->ops
;
642 if (sd
->len
> sd
->total_len
)
643 sd
->len
= sd
->total_len
;
645 ret
= actor(pipe
, buf
, sd
);
654 sd
->num_spliced
+= ret
;
657 sd
->total_len
-= ret
;
661 ops
->release(pipe
, buf
);
662 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (PIPE_BUFFERS
- 1);
665 sd
->need_wakeup
= true;
674 EXPORT_SYMBOL(splice_from_pipe_feed
);
677 * splice_from_pipe_next - wait for some data to splice from
678 * @pipe: pipe to splice from
679 * @sd: information about the splice operation
682 * This function will wait for some data and return a positive
683 * value (one) if pipe buffers are available. It will return zero
684 * or -errno if no more data needs to be spliced.
686 int splice_from_pipe_next(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
688 while (!pipe
->nrbufs
) {
692 if (!pipe
->waiting_writers
&& sd
->num_spliced
)
695 if (sd
->flags
& SPLICE_F_NONBLOCK
)
698 if (signal_pending(current
))
701 if (sd
->need_wakeup
) {
702 wakeup_pipe_writers(pipe
);
703 sd
->need_wakeup
= false;
711 EXPORT_SYMBOL(splice_from_pipe_next
);
714 * splice_from_pipe_begin - start splicing from pipe
715 * @pipe: pipe to splice from
718 * This function should be called before a loop containing
719 * splice_from_pipe_next() and splice_from_pipe_feed() to
720 * initialize the necessary fields of @sd.
722 void splice_from_pipe_begin(struct splice_desc
*sd
)
725 sd
->need_wakeup
= false;
727 EXPORT_SYMBOL(splice_from_pipe_begin
);
730 * splice_from_pipe_end - finish splicing from pipe
731 * @pipe: pipe to splice from
732 * @sd: information about the splice operation
735 * This function will wake up pipe writers if necessary. It should
736 * be called after a loop containing splice_from_pipe_next() and
737 * splice_from_pipe_feed().
739 void splice_from_pipe_end(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
742 wakeup_pipe_writers(pipe
);
744 EXPORT_SYMBOL(splice_from_pipe_end
);
747 * __splice_from_pipe - splice data from a pipe to given actor
748 * @pipe: pipe to splice from
749 * @sd: information to @actor
750 * @actor: handler that splices the data
753 * This function does little more than loop over the pipe and call
754 * @actor to do the actual moving of a single struct pipe_buffer to
755 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
759 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
764 splice_from_pipe_begin(sd
);
766 ret
= splice_from_pipe_next(pipe
, sd
);
768 ret
= splice_from_pipe_feed(pipe
, sd
, actor
);
770 splice_from_pipe_end(pipe
, sd
);
772 return sd
->num_spliced
? sd
->num_spliced
: ret
;
774 EXPORT_SYMBOL(__splice_from_pipe
);
777 * splice_from_pipe - splice data from a pipe to a file
778 * @pipe: pipe to splice from
779 * @out: file to splice to
780 * @ppos: position in @out
781 * @len: how many bytes to splice
782 * @flags: splice modifier flags
783 * @actor: handler that splices the data
786 * See __splice_from_pipe. This function locks the pipe inode,
787 * otherwise it's identical to __splice_from_pipe().
790 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
791 loff_t
*ppos
, size_t len
, unsigned int flags
,
795 struct splice_desc sd
= {
803 mutex_lock(&pipe
->inode
->i_mutex
);
804 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
806 mutex_unlock(&pipe
->inode
->i_mutex
);
812 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
814 * @out: file to write to
815 * @ppos: position in @out
816 * @len: number of bytes to splice
817 * @flags: splice modifier flags
820 * Will either move or copy pages (determined by @flags options) from
821 * the given pipe inode to the given file. The caller is responsible
822 * for acquiring i_mutex on both inodes.
826 generic_file_splice_write_nolock(struct pipe_inode_info
*pipe
, struct file
*out
,
827 loff_t
*ppos
, size_t len
, unsigned int flags
)
829 struct address_space
*mapping
= out
->f_mapping
;
830 struct inode
*inode
= mapping
->host
;
831 struct splice_desc sd
= {
840 err
= file_remove_suid(out
);
844 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_file
);
846 unsigned long nr_pages
;
849 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
852 * If file or inode is SYNC and we actually wrote some data,
855 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
856 err
= generic_osync_inode(inode
, mapping
,
857 OSYNC_METADATA
|OSYNC_DATA
);
862 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
868 EXPORT_SYMBOL(generic_file_splice_write_nolock
);
871 * generic_file_splice_write - splice data from a pipe to a file
873 * @out: file to write to
874 * @ppos: position in @out
875 * @len: number of bytes to splice
876 * @flags: splice modifier flags
879 * Will either move or copy pages (determined by @flags options) from
880 * the given pipe inode to the given file.
884 generic_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
885 loff_t
*ppos
, size_t len
, unsigned int flags
)
887 struct address_space
*mapping
= out
->f_mapping
;
888 struct inode
*inode
= mapping
->host
;
889 struct splice_desc sd
= {
897 WARN_ON(S_ISFIFO(inode
->i_mode
));
898 mutex_lock_nested(&inode
->i_mutex
, I_MUTEX_PARENT
);
899 ret
= file_remove_suid(out
);
902 mutex_lock_nested(&pipe
->inode
->i_mutex
, I_MUTEX_CHILD
);
903 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_file
);
905 mutex_unlock(&pipe
->inode
->i_mutex
);
907 mutex_unlock(&inode
->i_mutex
);
909 unsigned long nr_pages
;
912 nr_pages
= (ret
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
915 * If file or inode is SYNC and we actually wrote some data,
918 if (unlikely((out
->f_flags
& O_SYNC
) || IS_SYNC(inode
))) {
921 mutex_lock(&inode
->i_mutex
);
922 err
= generic_osync_inode(inode
, mapping
,
923 OSYNC_METADATA
|OSYNC_DATA
);
924 mutex_unlock(&inode
->i_mutex
);
929 balance_dirty_pages_ratelimited_nr(mapping
, nr_pages
);
935 EXPORT_SYMBOL(generic_file_splice_write
);
938 * generic_splice_sendpage - splice data from a pipe to a socket
939 * @pipe: pipe to splice from
940 * @out: socket to write to
941 * @ppos: position in @out
942 * @len: number of bytes to splice
943 * @flags: splice modifier flags
946 * Will send @len bytes from the pipe to a network socket. No data copying
950 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
951 loff_t
*ppos
, size_t len
, unsigned int flags
)
953 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
956 EXPORT_SYMBOL(generic_splice_sendpage
);
959 * Attempt to initiate a splice from pipe to file.
961 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
962 loff_t
*ppos
, size_t len
, unsigned int flags
)
966 if (unlikely(!out
->f_op
|| !out
->f_op
->splice_write
))
969 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
972 if (unlikely(out
->f_flags
& O_APPEND
))
975 ret
= rw_verify_area(WRITE
, out
, ppos
, len
);
976 if (unlikely(ret
< 0))
979 return out
->f_op
->splice_write(pipe
, out
, ppos
, len
, flags
);
983 * Attempt to initiate a splice from a file to a pipe.
985 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
986 struct pipe_inode_info
*pipe
, size_t len
,
991 if (unlikely(!in
->f_op
|| !in
->f_op
->splice_read
))
994 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
997 ret
= rw_verify_area(READ
, in
, ppos
, len
);
998 if (unlikely(ret
< 0))
1001 return in
->f_op
->splice_read(in
, ppos
, pipe
, len
, flags
);
1005 * splice_direct_to_actor - splices data directly between two non-pipes
1006 * @in: file to splice from
1007 * @sd: actor information on where to splice to
1008 * @actor: handles the data splicing
1011 * This is a special case helper to splice directly between two
1012 * points, without requiring an explicit pipe. Internally an allocated
1013 * pipe is cached in the process, and reused during the lifetime of
1017 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
1018 splice_direct_actor
*actor
)
1020 struct pipe_inode_info
*pipe
;
1027 * We require the input being a regular file, as we don't want to
1028 * randomly drop data for eg socket -> socket splicing. Use the
1029 * piped splicing for that!
1031 i_mode
= in
->f_path
.dentry
->d_inode
->i_mode
;
1032 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
1036 * neither in nor out is a pipe, setup an internal pipe attached to
1037 * 'out' and transfer the wanted data from 'in' to 'out' through that
1039 pipe
= current
->splice_pipe
;
1040 if (unlikely(!pipe
)) {
1041 pipe
= alloc_pipe_info(NULL
);
1046 * We don't have an immediate reader, but we'll read the stuff
1047 * out of the pipe right after the splice_to_pipe(). So set
1048 * PIPE_READERS appropriately.
1052 current
->splice_pipe
= pipe
;
1060 len
= sd
->total_len
;
1064 * Don't block on output, we have to drain the direct pipe.
1066 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1070 loff_t pos
= sd
->pos
, prev_pos
= pos
;
1072 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1073 if (unlikely(ret
<= 0))
1077 sd
->total_len
= read_len
;
1080 * NOTE: nonblocking mode only applies to the input. We
1081 * must not do the output in nonblocking mode as then we
1082 * could get stuck data in the internal pipe:
1084 ret
= actor(pipe
, sd
);
1085 if (unlikely(ret
<= 0)) {
1094 if (ret
< read_len
) {
1095 sd
->pos
= prev_pos
+ ret
;
1101 pipe
->nrbufs
= pipe
->curbuf
= 0;
1107 * If we did an incomplete transfer we must release
1108 * the pipe buffers in question:
1110 for (i
= 0; i
< PIPE_BUFFERS
; i
++) {
1111 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1114 buf
->ops
->release(pipe
, buf
);
1124 EXPORT_SYMBOL(splice_direct_to_actor
);
1126 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1127 struct splice_desc
*sd
)
1129 struct file
*file
= sd
->u
.file
;
1131 return do_splice_from(pipe
, file
, &sd
->pos
, sd
->total_len
, sd
->flags
);
1135 * do_splice_direct - splices data directly between two files
1136 * @in: file to splice from
1137 * @ppos: input file offset
1138 * @out: file to splice to
1139 * @len: number of bytes to splice
1140 * @flags: splice modifier flags
1143 * For use by do_sendfile(). splice can easily emulate sendfile, but
1144 * doing it in the application would incur an extra system call
1145 * (splice in + splice out, as compared to just sendfile()). So this helper
1146 * can splice directly through a process-private pipe.
1149 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1150 size_t len
, unsigned int flags
)
1152 struct splice_desc sd
= {
1161 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1169 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1170 * location, so checking ->i_pipe is not enough to verify that this is a
1173 static inline struct pipe_inode_info
*pipe_info(struct inode
*inode
)
1175 if (S_ISFIFO(inode
->i_mode
))
1176 return inode
->i_pipe
;
1182 * Determine where to splice to/from.
1184 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1185 struct file
*out
, loff_t __user
*off_out
,
1186 size_t len
, unsigned int flags
)
1188 struct pipe_inode_info
*pipe
;
1189 loff_t offset
, *off
;
1192 pipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1197 if (out
->f_op
->llseek
== no_llseek
)
1199 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1205 ret
= do_splice_from(pipe
, out
, off
, len
, flags
);
1207 if (off_out
&& copy_to_user(off_out
, off
, sizeof(loff_t
)))
1213 pipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1218 if (in
->f_op
->llseek
== no_llseek
)
1220 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1226 ret
= do_splice_to(in
, off
, pipe
, len
, flags
);
1228 if (off_in
&& copy_to_user(off_in
, off
, sizeof(loff_t
)))
1238 * Map an iov into an array of pages and offset/length tupples. With the
1239 * partial_page structure, we can map several non-contiguous ranges into
1240 * our ones pages[] map instead of splitting that operation into pieces.
1241 * Could easily be exported as a generic helper for other users, in which
1242 * case one would probably want to add a 'max_nr_pages' parameter as well.
1244 static int get_iovec_page_array(const struct iovec __user
*iov
,
1245 unsigned int nr_vecs
, struct page
**pages
,
1246 struct partial_page
*partial
, int aligned
)
1248 int buffers
= 0, error
= 0;
1251 unsigned long off
, npages
;
1258 if (copy_from_user(&entry
, iov
, sizeof(entry
)))
1261 base
= entry
.iov_base
;
1262 len
= entry
.iov_len
;
1265 * Sanity check this iovec. 0 read succeeds.
1271 if (!access_ok(VERIFY_READ
, base
, len
))
1275 * Get this base offset and number of pages, then map
1276 * in the user pages.
1278 off
= (unsigned long) base
& ~PAGE_MASK
;
1281 * If asked for alignment, the offset must be zero and the
1282 * length a multiple of the PAGE_SIZE.
1285 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1288 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1289 if (npages
> PIPE_BUFFERS
- buffers
)
1290 npages
= PIPE_BUFFERS
- buffers
;
1292 error
= get_user_pages_fast((unsigned long)base
, npages
,
1293 0, &pages
[buffers
]);
1295 if (unlikely(error
<= 0))
1299 * Fill this contiguous range into the partial page map.
1301 for (i
= 0; i
< error
; i
++) {
1302 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1304 partial
[buffers
].offset
= off
;
1305 partial
[buffers
].len
= plen
;
1313 * We didn't complete this iov, stop here since it probably
1314 * means we have to move some of this into a pipe to
1315 * be able to continue.
1321 * Don't continue if we mapped fewer pages than we asked for,
1322 * or if we mapped the max number of pages that we have
1325 if (error
< npages
|| buffers
== PIPE_BUFFERS
)
1338 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1339 struct splice_desc
*sd
)
1344 ret
= buf
->ops
->confirm(pipe
, buf
);
1349 * See if we can use the atomic maps, by prefaulting in the
1350 * pages and doing an atomic copy
1352 if (!fault_in_pages_writeable(sd
->u
.userptr
, sd
->len
)) {
1353 src
= buf
->ops
->map(pipe
, buf
, 1);
1354 ret
= __copy_to_user_inatomic(sd
->u
.userptr
, src
+ buf
->offset
,
1356 buf
->ops
->unmap(pipe
, buf
, src
);
1364 * No dice, use slow non-atomic map and copy
1366 src
= buf
->ops
->map(pipe
, buf
, 0);
1369 if (copy_to_user(sd
->u
.userptr
, src
+ buf
->offset
, sd
->len
))
1372 buf
->ops
->unmap(pipe
, buf
, src
);
1375 sd
->u
.userptr
+= ret
;
1380 * For lack of a better implementation, implement vmsplice() to userspace
1381 * as a simple copy of the pipes pages to the user iov.
1383 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*iov
,
1384 unsigned long nr_segs
, unsigned int flags
)
1386 struct pipe_inode_info
*pipe
;
1387 struct splice_desc sd
;
1392 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1397 mutex_lock(&pipe
->inode
->i_mutex
);
1405 * Get user address base and length for this iovec.
1407 error
= get_user(base
, &iov
->iov_base
);
1408 if (unlikely(error
))
1410 error
= get_user(len
, &iov
->iov_len
);
1411 if (unlikely(error
))
1415 * Sanity check this iovec. 0 read succeeds.
1419 if (unlikely(!base
)) {
1424 if (unlikely(!access_ok(VERIFY_WRITE
, base
, len
))) {
1432 sd
.u
.userptr
= base
;
1435 size
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1453 mutex_unlock(&pipe
->inode
->i_mutex
);
1462 * vmsplice splices a user address range into a pipe. It can be thought of
1463 * as splice-from-memory, where the regular splice is splice-from-file (or
1464 * to file). In both cases the output is a pipe, naturally.
1466 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1467 unsigned long nr_segs
, unsigned int flags
)
1469 struct pipe_inode_info
*pipe
;
1470 struct page
*pages
[PIPE_BUFFERS
];
1471 struct partial_page partial
[PIPE_BUFFERS
];
1472 struct splice_pipe_desc spd
= {
1476 .ops
= &user_page_pipe_buf_ops
,
1477 .spd_release
= spd_release_page
,
1480 pipe
= pipe_info(file
->f_path
.dentry
->d_inode
);
1484 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, pages
, partial
,
1485 flags
& SPLICE_F_GIFT
);
1486 if (spd
.nr_pages
<= 0)
1487 return spd
.nr_pages
;
1489 return splice_to_pipe(pipe
, &spd
);
1493 * Note that vmsplice only really supports true splicing _from_ user memory
1494 * to a pipe, not the other way around. Splicing from user memory is a simple
1495 * operation that can be supported without any funky alignment restrictions
1496 * or nasty vm tricks. We simply map in the user memory and fill them into
1497 * a pipe. The reverse isn't quite as easy, though. There are two possible
1498 * solutions for that:
1500 * - memcpy() the data internally, at which point we might as well just
1501 * do a regular read() on the buffer anyway.
1502 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1503 * has restriction limitations on both ends of the pipe).
1505 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1508 SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct iovec __user
*, iov
,
1509 unsigned long, nr_segs
, unsigned int, flags
)
1515 if (unlikely(nr_segs
> UIO_MAXIOV
))
1517 else if (unlikely(!nr_segs
))
1521 file
= fget_light(fd
, &fput
);
1523 if (file
->f_mode
& FMODE_WRITE
)
1524 error
= vmsplice_to_pipe(file
, iov
, nr_segs
, flags
);
1525 else if (file
->f_mode
& FMODE_READ
)
1526 error
= vmsplice_to_user(file
, iov
, nr_segs
, flags
);
1528 fput_light(file
, fput
);
1534 SYSCALL_DEFINE6(splice
, int, fd_in
, loff_t __user
*, off_in
,
1535 int, fd_out
, loff_t __user
*, off_out
,
1536 size_t, len
, unsigned int, flags
)
1539 struct file
*in
, *out
;
1540 int fput_in
, fput_out
;
1546 in
= fget_light(fd_in
, &fput_in
);
1548 if (in
->f_mode
& FMODE_READ
) {
1549 out
= fget_light(fd_out
, &fput_out
);
1551 if (out
->f_mode
& FMODE_WRITE
)
1552 error
= do_splice(in
, off_in
,
1555 fput_light(out
, fput_out
);
1559 fput_light(in
, fput_in
);
1566 * Make sure there's data to read. Wait for input if we can, otherwise
1567 * return an appropriate error.
1569 static int link_ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1574 * Check ->nrbufs without the inode lock first. This function
1575 * is speculative anyways, so missing one is ok.
1581 mutex_lock(&pipe
->inode
->i_mutex
);
1583 while (!pipe
->nrbufs
) {
1584 if (signal_pending(current
)) {
1590 if (!pipe
->waiting_writers
) {
1591 if (flags
& SPLICE_F_NONBLOCK
) {
1599 mutex_unlock(&pipe
->inode
->i_mutex
);
1604 * Make sure there's writeable room. Wait for room if we can, otherwise
1605 * return an appropriate error.
1607 static int link_opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1612 * Check ->nrbufs without the inode lock first. This function
1613 * is speculative anyways, so missing one is ok.
1615 if (pipe
->nrbufs
< PIPE_BUFFERS
)
1619 mutex_lock(&pipe
->inode
->i_mutex
);
1621 while (pipe
->nrbufs
>= PIPE_BUFFERS
) {
1622 if (!pipe
->readers
) {
1623 send_sig(SIGPIPE
, current
, 0);
1627 if (flags
& SPLICE_F_NONBLOCK
) {
1631 if (signal_pending(current
)) {
1635 pipe
->waiting_writers
++;
1637 pipe
->waiting_writers
--;
1640 mutex_unlock(&pipe
->inode
->i_mutex
);
1645 * Link contents of ipipe to opipe.
1647 static int link_pipe(struct pipe_inode_info
*ipipe
,
1648 struct pipe_inode_info
*opipe
,
1649 size_t len
, unsigned int flags
)
1651 struct pipe_buffer
*ibuf
, *obuf
;
1652 int ret
= 0, i
= 0, nbuf
;
1655 * Potential ABBA deadlock, work around it by ordering lock
1656 * grabbing by inode address. Otherwise two different processes
1657 * could deadlock (one doing tee from A -> B, the other from B -> A).
1659 inode_double_lock(ipipe
->inode
, opipe
->inode
);
1662 if (!opipe
->readers
) {
1663 send_sig(SIGPIPE
, current
, 0);
1670 * If we have iterated all input buffers or ran out of
1671 * output room, break.
1673 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= PIPE_BUFFERS
)
1676 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (PIPE_BUFFERS
- 1));
1677 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (PIPE_BUFFERS
- 1);
1680 * Get a reference to this pipe buffer,
1681 * so we can copy the contents over.
1683 ibuf
->ops
->get(ipipe
, ibuf
);
1685 obuf
= opipe
->bufs
+ nbuf
;
1689 * Don't inherit the gift flag, we need to
1690 * prevent multiple steals of this page.
1692 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1694 if (obuf
->len
> len
)
1704 * return EAGAIN if we have the potential of some data in the
1705 * future, otherwise just return 0
1707 if (!ret
&& ipipe
->waiting_writers
&& (flags
& SPLICE_F_NONBLOCK
))
1710 inode_double_unlock(ipipe
->inode
, opipe
->inode
);
1713 * If we put data in the output pipe, wakeup any potential readers.
1717 if (waitqueue_active(&opipe
->wait
))
1718 wake_up_interruptible(&opipe
->wait
);
1719 kill_fasync(&opipe
->fasync_readers
, SIGIO
, POLL_IN
);
1726 * This is a tee(1) implementation that works on pipes. It doesn't copy
1727 * any data, it simply references the 'in' pages on the 'out' pipe.
1728 * The 'flags' used are the SPLICE_F_* variants, currently the only
1729 * applicable one is SPLICE_F_NONBLOCK.
1731 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1734 struct pipe_inode_info
*ipipe
= pipe_info(in
->f_path
.dentry
->d_inode
);
1735 struct pipe_inode_info
*opipe
= pipe_info(out
->f_path
.dentry
->d_inode
);
1739 * Duplicate the contents of ipipe to opipe without actually
1742 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
1744 * Keep going, unless we encounter an error. The ipipe/opipe
1745 * ordering doesn't really matter.
1747 ret
= link_ipipe_prep(ipipe
, flags
);
1749 ret
= link_opipe_prep(opipe
, flags
);
1751 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
1758 SYSCALL_DEFINE4(tee
, int, fdin
, int, fdout
, size_t, len
, unsigned int, flags
)
1767 in
= fget_light(fdin
, &fput_in
);
1769 if (in
->f_mode
& FMODE_READ
) {
1771 struct file
*out
= fget_light(fdout
, &fput_out
);
1774 if (out
->f_mode
& FMODE_WRITE
)
1775 error
= do_tee(in
, out
, len
, flags
);
1776 fput_light(out
, fput_out
);
1779 fput_light(in
, fput_in
);