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/export.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32 #include <linux/gfp.h>
33 #include <linux/socket.h>
34 #include <linux/compat.h>
38 * Attempt to steal a page from a pipe buffer. This should perhaps go into
39 * a vm helper function, it's already simplified quite a bit by the
40 * addition of remove_mapping(). If success is returned, the caller may
41 * attempt to reuse this page for another destination.
43 static int page_cache_pipe_buf_steal(struct pipe_inode_info
*pipe
,
44 struct pipe_buffer
*buf
)
46 struct page
*page
= buf
->page
;
47 struct address_space
*mapping
;
51 mapping
= page_mapping(page
);
53 WARN_ON(!PageUptodate(page
));
56 * At least for ext2 with nobh option, we need to wait on
57 * writeback completing on this page, since we'll remove it
58 * from the pagecache. Otherwise truncate wont wait on the
59 * page, allowing the disk blocks to be reused by someone else
60 * before we actually wrote our data to them. fs corruption
63 wait_on_page_writeback(page
);
65 if (page_has_private(page
) &&
66 !try_to_release_page(page
, GFP_KERNEL
))
70 * If we succeeded in removing the mapping, set LRU flag
73 if (remove_mapping(mapping
, page
)) {
74 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
80 * Raced with truncate or failed to remove page from current
81 * address space, unlock and return failure.
88 static void page_cache_pipe_buf_release(struct pipe_inode_info
*pipe
,
89 struct pipe_buffer
*buf
)
91 page_cache_release(buf
->page
);
92 buf
->flags
&= ~PIPE_BUF_FLAG_LRU
;
96 * Check whether the contents of buf is OK to access. Since the content
97 * is a page cache page, IO may be in flight.
99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info
*pipe
,
100 struct pipe_buffer
*buf
)
102 struct page
*page
= buf
->page
;
105 if (!PageUptodate(page
)) {
109 * Page got truncated/unhashed. This will cause a 0-byte
110 * splice, if this is the first page.
112 if (!page
->mapping
) {
118 * Uh oh, read-error from disk.
120 if (!PageUptodate(page
)) {
126 * Page is ok afterall, we are done.
137 const struct pipe_buf_operations page_cache_pipe_buf_ops
= {
139 .confirm
= page_cache_pipe_buf_confirm
,
140 .release
= page_cache_pipe_buf_release
,
141 .steal
= page_cache_pipe_buf_steal
,
142 .get
= generic_pipe_buf_get
,
145 static int user_page_pipe_buf_steal(struct pipe_inode_info
*pipe
,
146 struct pipe_buffer
*buf
)
148 if (!(buf
->flags
& PIPE_BUF_FLAG_GIFT
))
151 buf
->flags
|= PIPE_BUF_FLAG_LRU
;
152 return generic_pipe_buf_steal(pipe
, buf
);
155 static const struct pipe_buf_operations user_page_pipe_buf_ops
= {
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
,
163 static void wakeup_pipe_readers(struct pipe_inode_info
*pipe
)
166 if (waitqueue_active(&pipe
->wait
))
167 wake_up_interruptible(&pipe
->wait
);
168 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
172 * splice_to_pipe - fill passed data into a pipe
173 * @pipe: pipe to fill
177 * @spd contains a map of pages and len/offset tuples, along with
178 * the struct pipe_buf_operations associated with these pages. This
179 * function will link that data to the pipe.
182 ssize_t
splice_to_pipe(struct pipe_inode_info
*pipe
,
183 struct splice_pipe_desc
*spd
)
185 unsigned int spd_pages
= spd
->nr_pages
;
186 int ret
, do_wakeup
, page_nr
;
195 if (!pipe
->readers
) {
196 send_sig(SIGPIPE
, current
, 0);
202 if (pipe
->nrbufs
< pipe
->buffers
) {
203 int newbuf
= (pipe
->curbuf
+ pipe
->nrbufs
) & (pipe
->buffers
- 1);
204 struct pipe_buffer
*buf
= pipe
->bufs
+ newbuf
;
206 buf
->page
= spd
->pages
[page_nr
];
207 buf
->offset
= spd
->partial
[page_nr
].offset
;
208 buf
->len
= spd
->partial
[page_nr
].len
;
209 buf
->private = spd
->partial
[page_nr
].private;
211 if (spd
->flags
& SPLICE_F_GIFT
)
212 buf
->flags
|= PIPE_BUF_FLAG_GIFT
;
221 if (!--spd
->nr_pages
)
223 if (pipe
->nrbufs
< pipe
->buffers
)
229 if (spd
->flags
& SPLICE_F_NONBLOCK
) {
235 if (signal_pending(current
)) {
243 if (waitqueue_active(&pipe
->wait
))
244 wake_up_interruptible_sync(&pipe
->wait
);
245 kill_fasync(&pipe
->fasync_readers
, SIGIO
, POLL_IN
);
249 pipe
->waiting_writers
++;
251 pipe
->waiting_writers
--;
257 wakeup_pipe_readers(pipe
);
259 while (page_nr
< spd_pages
)
260 spd
->spd_release(spd
, page_nr
++);
264 EXPORT_SYMBOL_GPL(splice_to_pipe
);
266 void spd_release_page(struct splice_pipe_desc
*spd
, unsigned int i
)
268 page_cache_release(spd
->pages
[i
]);
272 * Check if we need to grow the arrays holding pages and partial page
275 int splice_grow_spd(const struct pipe_inode_info
*pipe
, struct splice_pipe_desc
*spd
)
277 unsigned int buffers
= ACCESS_ONCE(pipe
->buffers
);
279 spd
->nr_pages_max
= buffers
;
280 if (buffers
<= PIPE_DEF_BUFFERS
)
283 spd
->pages
= kmalloc(buffers
* sizeof(struct page
*), GFP_KERNEL
);
284 spd
->partial
= kmalloc(buffers
* sizeof(struct partial_page
), GFP_KERNEL
);
286 if (spd
->pages
&& spd
->partial
)
294 void splice_shrink_spd(struct splice_pipe_desc
*spd
)
296 if (spd
->nr_pages_max
<= PIPE_DEF_BUFFERS
)
304 __generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
305 struct pipe_inode_info
*pipe
, size_t len
,
308 struct address_space
*mapping
= in
->f_mapping
;
309 unsigned int loff
, nr_pages
, req_pages
;
310 struct page
*pages
[PIPE_DEF_BUFFERS
];
311 struct partial_page partial
[PIPE_DEF_BUFFERS
];
313 pgoff_t index
, end_index
;
316 struct splice_pipe_desc spd
= {
319 .nr_pages_max
= PIPE_DEF_BUFFERS
,
321 .ops
= &page_cache_pipe_buf_ops
,
322 .spd_release
= spd_release_page
,
325 if (splice_grow_spd(pipe
, &spd
))
328 index
= *ppos
>> PAGE_CACHE_SHIFT
;
329 loff
= *ppos
& ~PAGE_CACHE_MASK
;
330 req_pages
= (len
+ loff
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
331 nr_pages
= min(req_pages
, spd
.nr_pages_max
);
334 * Lookup the (hopefully) full range of pages we need.
336 spd
.nr_pages
= find_get_pages_contig(mapping
, index
, nr_pages
, spd
.pages
);
337 index
+= spd
.nr_pages
;
340 * If find_get_pages_contig() returned fewer pages than we needed,
341 * readahead/allocate the rest and fill in the holes.
343 if (spd
.nr_pages
< nr_pages
)
344 page_cache_sync_readahead(mapping
, &in
->f_ra
, in
,
345 index
, req_pages
- spd
.nr_pages
);
348 while (spd
.nr_pages
< nr_pages
) {
350 * Page could be there, find_get_pages_contig() breaks on
353 page
= find_get_page(mapping
, index
);
356 * page didn't exist, allocate one.
358 page
= page_cache_alloc_cold(mapping
);
362 error
= add_to_page_cache_lru(page
, mapping
, index
,
363 mapping_gfp_constraint(mapping
, GFP_KERNEL
));
364 if (unlikely(error
)) {
365 page_cache_release(page
);
366 if (error
== -EEXIST
)
371 * add_to_page_cache() locks the page, unlock it
372 * to avoid convoluting the logic below even more.
377 spd
.pages
[spd
.nr_pages
++] = page
;
382 * Now loop over the map and see if we need to start IO on any
383 * pages, fill in the partial map, etc.
385 index
= *ppos
>> PAGE_CACHE_SHIFT
;
386 nr_pages
= spd
.nr_pages
;
388 for (page_nr
= 0; page_nr
< nr_pages
; page_nr
++) {
389 unsigned int this_len
;
395 * this_len is the max we'll use from this page
397 this_len
= min_t(unsigned long, len
, PAGE_CACHE_SIZE
- loff
);
398 page
= spd
.pages
[page_nr
];
400 if (PageReadahead(page
))
401 page_cache_async_readahead(mapping
, &in
->f_ra
, in
,
402 page
, index
, req_pages
- page_nr
);
405 * If the page isn't uptodate, we may need to start io on it
407 if (!PageUptodate(page
)) {
411 * Page was truncated, or invalidated by the
412 * filesystem. Redo the find/create, but this time the
413 * page is kept locked, so there's no chance of another
414 * race with truncate/invalidate.
416 if (!page
->mapping
) {
418 page
= find_or_create_page(mapping
, index
,
419 mapping_gfp_mask(mapping
));
425 page_cache_release(spd
.pages
[page_nr
]);
426 spd
.pages
[page_nr
] = page
;
429 * page was already under io and is now done, great
431 if (PageUptodate(page
)) {
437 * need to read in the page
439 error
= mapping
->a_ops
->readpage(in
, page
);
440 if (unlikely(error
)) {
442 * We really should re-lookup the page here,
443 * but it complicates things a lot. Instead
444 * lets just do what we already stored, and
445 * we'll get it the next time we are called.
447 if (error
== AOP_TRUNCATED_PAGE
)
455 * i_size must be checked after PageUptodate.
457 isize
= i_size_read(mapping
->host
);
458 end_index
= (isize
- 1) >> PAGE_CACHE_SHIFT
;
459 if (unlikely(!isize
|| index
> end_index
))
463 * if this is the last page, see if we need to shrink
464 * the length and stop
466 if (end_index
== index
) {
470 * max good bytes in this page
472 plen
= ((isize
- 1) & ~PAGE_CACHE_MASK
) + 1;
477 * force quit after adding this page
479 this_len
= min(this_len
, plen
- loff
);
483 spd
.partial
[page_nr
].offset
= loff
;
484 spd
.partial
[page_nr
].len
= this_len
;
492 * Release any pages at the end, if we quit early. 'page_nr' is how far
493 * we got, 'nr_pages' is how many pages are in the map.
495 while (page_nr
< nr_pages
)
496 page_cache_release(spd
.pages
[page_nr
++]);
497 in
->f_ra
.prev_pos
= (loff_t
)index
<< PAGE_CACHE_SHIFT
;
500 error
= splice_to_pipe(pipe
, &spd
);
502 splice_shrink_spd(&spd
);
507 * generic_file_splice_read - splice data from file to a pipe
508 * @in: file to splice from
509 * @ppos: position in @in
510 * @pipe: pipe to splice to
511 * @len: number of bytes to splice
512 * @flags: splice modifier flags
515 * Will read pages from given file and fill them into a pipe. Can be
516 * used as long as the address_space operations for the source implements
520 ssize_t
generic_file_splice_read(struct file
*in
, loff_t
*ppos
,
521 struct pipe_inode_info
*pipe
, size_t len
,
527 if (IS_DAX(in
->f_mapping
->host
))
528 return default_file_splice_read(in
, ppos
, pipe
, len
, flags
);
530 isize
= i_size_read(in
->f_mapping
->host
);
531 if (unlikely(*ppos
>= isize
))
534 left
= isize
- *ppos
;
535 if (unlikely(left
< len
))
538 ret
= __generic_file_splice_read(in
, ppos
, pipe
, len
, flags
);
546 EXPORT_SYMBOL(generic_file_splice_read
);
548 static const struct pipe_buf_operations default_pipe_buf_ops
= {
550 .confirm
= generic_pipe_buf_confirm
,
551 .release
= generic_pipe_buf_release
,
552 .steal
= generic_pipe_buf_steal
,
553 .get
= generic_pipe_buf_get
,
556 static int generic_pipe_buf_nosteal(struct pipe_inode_info
*pipe
,
557 struct pipe_buffer
*buf
)
562 /* Pipe buffer operations for a socket and similar. */
563 const struct pipe_buf_operations nosteal_pipe_buf_ops
= {
565 .confirm
= generic_pipe_buf_confirm
,
566 .release
= generic_pipe_buf_release
,
567 .steal
= generic_pipe_buf_nosteal
,
568 .get
= generic_pipe_buf_get
,
570 EXPORT_SYMBOL(nosteal_pipe_buf_ops
);
572 static ssize_t
kernel_readv(struct file
*file
, const struct iovec
*vec
,
573 unsigned long vlen
, loff_t offset
)
581 /* The cast to a user pointer is valid due to the set_fs() */
582 res
= vfs_readv(file
, (const struct iovec __user
*)vec
, vlen
, &pos
);
588 ssize_t
kernel_write(struct file
*file
, const char *buf
, size_t count
,
596 /* The cast to a user pointer is valid due to the set_fs() */
597 res
= vfs_write(file
, (__force
const char __user
*)buf
, count
, &pos
);
602 EXPORT_SYMBOL(kernel_write
);
604 ssize_t
default_file_splice_read(struct file
*in
, loff_t
*ppos
,
605 struct pipe_inode_info
*pipe
, size_t len
,
608 unsigned int nr_pages
;
609 unsigned int nr_freed
;
611 struct page
*pages
[PIPE_DEF_BUFFERS
];
612 struct partial_page partial
[PIPE_DEF_BUFFERS
];
613 struct iovec
*vec
, __vec
[PIPE_DEF_BUFFERS
];
618 struct splice_pipe_desc spd
= {
621 .nr_pages_max
= PIPE_DEF_BUFFERS
,
623 .ops
= &default_pipe_buf_ops
,
624 .spd_release
= spd_release_page
,
627 if (splice_grow_spd(pipe
, &spd
))
632 if (spd
.nr_pages_max
> PIPE_DEF_BUFFERS
) {
633 vec
= kmalloc(spd
.nr_pages_max
* sizeof(struct iovec
), GFP_KERNEL
);
638 offset
= *ppos
& ~PAGE_CACHE_MASK
;
639 nr_pages
= (len
+ offset
+ PAGE_CACHE_SIZE
- 1) >> PAGE_CACHE_SHIFT
;
641 for (i
= 0; i
< nr_pages
&& i
< spd
.nr_pages_max
&& len
; i
++) {
644 page
= alloc_page(GFP_USER
);
649 this_len
= min_t(size_t, len
, PAGE_CACHE_SIZE
- offset
);
650 vec
[i
].iov_base
= (void __user
*) page_address(page
);
651 vec
[i
].iov_len
= this_len
;
658 res
= kernel_readv(in
, vec
, spd
.nr_pages
, *ppos
);
669 for (i
= 0; i
< spd
.nr_pages
; i
++) {
670 this_len
= min_t(size_t, vec
[i
].iov_len
, res
);
671 spd
.partial
[i
].offset
= 0;
672 spd
.partial
[i
].len
= this_len
;
674 __free_page(spd
.pages
[i
]);
680 spd
.nr_pages
-= nr_freed
;
682 res
= splice_to_pipe(pipe
, &spd
);
689 splice_shrink_spd(&spd
);
693 for (i
= 0; i
< spd
.nr_pages
; i
++)
694 __free_page(spd
.pages
[i
]);
699 EXPORT_SYMBOL(default_file_splice_read
);
702 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
703 * using sendpage(). Return the number of bytes sent.
705 static int pipe_to_sendpage(struct pipe_inode_info
*pipe
,
706 struct pipe_buffer
*buf
, struct splice_desc
*sd
)
708 struct file
*file
= sd
->u
.file
;
709 loff_t pos
= sd
->pos
;
712 if (!likely(file
->f_op
->sendpage
))
715 more
= (sd
->flags
& SPLICE_F_MORE
) ? MSG_MORE
: 0;
717 if (sd
->len
< sd
->total_len
&& pipe
->nrbufs
> 1)
718 more
|= MSG_SENDPAGE_NOTLAST
;
720 return file
->f_op
->sendpage(file
, buf
->page
, buf
->offset
,
721 sd
->len
, &pos
, more
);
724 static void wakeup_pipe_writers(struct pipe_inode_info
*pipe
)
727 if (waitqueue_active(&pipe
->wait
))
728 wake_up_interruptible(&pipe
->wait
);
729 kill_fasync(&pipe
->fasync_writers
, SIGIO
, POLL_OUT
);
733 * splice_from_pipe_feed - feed available data from a pipe to a file
734 * @pipe: pipe to splice from
735 * @sd: information to @actor
736 * @actor: handler that splices the data
739 * This function loops over the pipe and calls @actor to do the
740 * actual moving of a single struct pipe_buffer to the desired
741 * destination. It returns when there's no more buffers left in
742 * the pipe or if the requested number of bytes (@sd->total_len)
743 * have been copied. It returns a positive number (one) if the
744 * pipe needs to be filled with more data, zero if the required
745 * number of bytes have been copied and -errno on error.
747 * This, together with splice_from_pipe_{begin,end,next}, may be
748 * used to implement the functionality of __splice_from_pipe() when
749 * locking is required around copying the pipe buffers to the
752 static int splice_from_pipe_feed(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
757 while (pipe
->nrbufs
) {
758 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
759 const struct pipe_buf_operations
*ops
= buf
->ops
;
762 if (sd
->len
> sd
->total_len
)
763 sd
->len
= sd
->total_len
;
765 ret
= buf
->ops
->confirm(pipe
, buf
);
772 ret
= actor(pipe
, buf
, sd
);
779 sd
->num_spliced
+= ret
;
782 sd
->total_len
-= ret
;
786 ops
->release(pipe
, buf
);
787 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (pipe
->buffers
- 1);
790 sd
->need_wakeup
= true;
801 * splice_from_pipe_next - wait for some data to splice from
802 * @pipe: pipe to splice from
803 * @sd: information about the splice operation
806 * This function will wait for some data and return a positive
807 * value (one) if pipe buffers are available. It will return zero
808 * or -errno if no more data needs to be spliced.
810 static int splice_from_pipe_next(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
813 * Check for signal early to make process killable when there are
814 * always buffers available
816 if (signal_pending(current
))
819 while (!pipe
->nrbufs
) {
823 if (!pipe
->waiting_writers
&& sd
->num_spliced
)
826 if (sd
->flags
& SPLICE_F_NONBLOCK
)
829 if (signal_pending(current
))
832 if (sd
->need_wakeup
) {
833 wakeup_pipe_writers(pipe
);
834 sd
->need_wakeup
= false;
844 * splice_from_pipe_begin - start splicing from pipe
845 * @sd: information about the splice operation
848 * This function should be called before a loop containing
849 * splice_from_pipe_next() and splice_from_pipe_feed() to
850 * initialize the necessary fields of @sd.
852 static void splice_from_pipe_begin(struct splice_desc
*sd
)
855 sd
->need_wakeup
= false;
859 * splice_from_pipe_end - finish splicing from pipe
860 * @pipe: pipe to splice from
861 * @sd: information about the splice operation
864 * This function will wake up pipe writers if necessary. It should
865 * be called after a loop containing splice_from_pipe_next() and
866 * splice_from_pipe_feed().
868 static void splice_from_pipe_end(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
)
871 wakeup_pipe_writers(pipe
);
875 * __splice_from_pipe - splice data from a pipe to given actor
876 * @pipe: pipe to splice from
877 * @sd: information to @actor
878 * @actor: handler that splices the data
881 * This function does little more than loop over the pipe and call
882 * @actor to do the actual moving of a single struct pipe_buffer to
883 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
887 ssize_t
__splice_from_pipe(struct pipe_inode_info
*pipe
, struct splice_desc
*sd
,
892 splice_from_pipe_begin(sd
);
895 ret
= splice_from_pipe_next(pipe
, sd
);
897 ret
= splice_from_pipe_feed(pipe
, sd
, actor
);
899 splice_from_pipe_end(pipe
, sd
);
901 return sd
->num_spliced
? sd
->num_spliced
: ret
;
903 EXPORT_SYMBOL(__splice_from_pipe
);
906 * splice_from_pipe - splice data from a pipe to a file
907 * @pipe: pipe to splice from
908 * @out: file to splice to
909 * @ppos: position in @out
910 * @len: how many bytes to splice
911 * @flags: splice modifier flags
912 * @actor: handler that splices the data
915 * See __splice_from_pipe. This function locks the pipe inode,
916 * otherwise it's identical to __splice_from_pipe().
919 ssize_t
splice_from_pipe(struct pipe_inode_info
*pipe
, struct file
*out
,
920 loff_t
*ppos
, size_t len
, unsigned int flags
,
924 struct splice_desc sd
= {
932 ret
= __splice_from_pipe(pipe
, &sd
, actor
);
939 * iter_file_splice_write - splice data from a pipe to a file
941 * @out: file to write to
942 * @ppos: position in @out
943 * @len: number of bytes to splice
944 * @flags: splice modifier flags
947 * Will either move or copy pages (determined by @flags options) from
948 * the given pipe inode to the given file.
949 * This one is ->write_iter-based.
953 iter_file_splice_write(struct pipe_inode_info
*pipe
, struct file
*out
,
954 loff_t
*ppos
, size_t len
, unsigned int flags
)
956 struct splice_desc sd
= {
962 int nbufs
= pipe
->buffers
;
963 struct bio_vec
*array
= kcalloc(nbufs
, sizeof(struct bio_vec
),
967 if (unlikely(!array
))
972 splice_from_pipe_begin(&sd
);
973 while (sd
.total_len
) {
974 struct iov_iter from
;
978 ret
= splice_from_pipe_next(pipe
, &sd
);
982 if (unlikely(nbufs
< pipe
->buffers
)) {
984 nbufs
= pipe
->buffers
;
985 array
= kcalloc(nbufs
, sizeof(struct bio_vec
),
993 /* build the vector */
995 for (n
= 0, idx
= pipe
->curbuf
; left
&& n
< pipe
->nrbufs
; n
++, idx
++) {
996 struct pipe_buffer
*buf
= pipe
->bufs
+ idx
;
997 size_t this_len
= buf
->len
;
1002 if (idx
== pipe
->buffers
- 1)
1005 ret
= buf
->ops
->confirm(pipe
, buf
);
1006 if (unlikely(ret
)) {
1007 if (ret
== -ENODATA
)
1012 array
[n
].bv_page
= buf
->page
;
1013 array
[n
].bv_len
= this_len
;
1014 array
[n
].bv_offset
= buf
->offset
;
1018 iov_iter_bvec(&from
, ITER_BVEC
| WRITE
, array
, n
,
1019 sd
.total_len
- left
);
1020 ret
= vfs_iter_write(out
, &from
, &sd
.pos
);
1024 sd
.num_spliced
+= ret
;
1025 sd
.total_len
-= ret
;
1028 /* dismiss the fully eaten buffers, adjust the partial one */
1030 struct pipe_buffer
*buf
= pipe
->bufs
+ pipe
->curbuf
;
1031 if (ret
>= buf
->len
) {
1032 const struct pipe_buf_operations
*ops
= buf
->ops
;
1036 ops
->release(pipe
, buf
);
1037 pipe
->curbuf
= (pipe
->curbuf
+ 1) & (pipe
->buffers
- 1);
1040 sd
.need_wakeup
= true;
1050 splice_from_pipe_end(pipe
, &sd
);
1055 ret
= sd
.num_spliced
;
1060 EXPORT_SYMBOL(iter_file_splice_write
);
1062 static int write_pipe_buf(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1063 struct splice_desc
*sd
)
1067 loff_t tmp
= sd
->pos
;
1069 data
= kmap(buf
->page
);
1070 ret
= __kernel_write(sd
->u
.file
, data
+ buf
->offset
, sd
->len
, &tmp
);
1076 static ssize_t
default_file_splice_write(struct pipe_inode_info
*pipe
,
1077 struct file
*out
, loff_t
*ppos
,
1078 size_t len
, unsigned int flags
)
1082 ret
= splice_from_pipe(pipe
, out
, ppos
, len
, flags
, write_pipe_buf
);
1090 * generic_splice_sendpage - splice data from a pipe to a socket
1091 * @pipe: pipe to splice from
1092 * @out: socket to write to
1093 * @ppos: position in @out
1094 * @len: number of bytes to splice
1095 * @flags: splice modifier flags
1098 * Will send @len bytes from the pipe to a network socket. No data copying
1102 ssize_t
generic_splice_sendpage(struct pipe_inode_info
*pipe
, struct file
*out
,
1103 loff_t
*ppos
, size_t len
, unsigned int flags
)
1105 return splice_from_pipe(pipe
, out
, ppos
, len
, flags
, pipe_to_sendpage
);
1108 EXPORT_SYMBOL(generic_splice_sendpage
);
1111 * Attempt to initiate a splice from pipe to file.
1113 static long do_splice_from(struct pipe_inode_info
*pipe
, struct file
*out
,
1114 loff_t
*ppos
, size_t len
, unsigned int flags
)
1116 ssize_t (*splice_write
)(struct pipe_inode_info
*, struct file
*,
1117 loff_t
*, size_t, unsigned int);
1119 if (out
->f_op
->splice_write
)
1120 splice_write
= out
->f_op
->splice_write
;
1122 splice_write
= default_file_splice_write
;
1124 return splice_write(pipe
, out
, ppos
, len
, flags
);
1128 * Attempt to initiate a splice from a file to a pipe.
1130 static long do_splice_to(struct file
*in
, loff_t
*ppos
,
1131 struct pipe_inode_info
*pipe
, size_t len
,
1134 ssize_t (*splice_read
)(struct file
*, loff_t
*,
1135 struct pipe_inode_info
*, size_t, unsigned int);
1138 if (unlikely(!(in
->f_mode
& FMODE_READ
)))
1141 ret
= rw_verify_area(READ
, in
, ppos
, len
);
1142 if (unlikely(ret
< 0))
1145 if (in
->f_op
->splice_read
)
1146 splice_read
= in
->f_op
->splice_read
;
1148 splice_read
= default_file_splice_read
;
1150 return splice_read(in
, ppos
, pipe
, len
, flags
);
1154 * splice_direct_to_actor - splices data directly between two non-pipes
1155 * @in: file to splice from
1156 * @sd: actor information on where to splice to
1157 * @actor: handles the data splicing
1160 * This is a special case helper to splice directly between two
1161 * points, without requiring an explicit pipe. Internally an allocated
1162 * pipe is cached in the process, and reused during the lifetime of
1166 ssize_t
splice_direct_to_actor(struct file
*in
, struct splice_desc
*sd
,
1167 splice_direct_actor
*actor
)
1169 struct pipe_inode_info
*pipe
;
1176 * We require the input being a regular file, as we don't want to
1177 * randomly drop data for eg socket -> socket splicing. Use the
1178 * piped splicing for that!
1180 i_mode
= file_inode(in
)->i_mode
;
1181 if (unlikely(!S_ISREG(i_mode
) && !S_ISBLK(i_mode
)))
1185 * neither in nor out is a pipe, setup an internal pipe attached to
1186 * 'out' and transfer the wanted data from 'in' to 'out' through that
1188 pipe
= current
->splice_pipe
;
1189 if (unlikely(!pipe
)) {
1190 pipe
= alloc_pipe_info();
1195 * We don't have an immediate reader, but we'll read the stuff
1196 * out of the pipe right after the splice_to_pipe(). So set
1197 * PIPE_READERS appropriately.
1201 current
->splice_pipe
= pipe
;
1209 len
= sd
->total_len
;
1213 * Don't block on output, we have to drain the direct pipe.
1215 sd
->flags
&= ~SPLICE_F_NONBLOCK
;
1216 more
= sd
->flags
& SPLICE_F_MORE
;
1220 loff_t pos
= sd
->pos
, prev_pos
= pos
;
1222 ret
= do_splice_to(in
, &pos
, pipe
, len
, flags
);
1223 if (unlikely(ret
<= 0))
1227 sd
->total_len
= read_len
;
1230 * If more data is pending, set SPLICE_F_MORE
1231 * If this is the last data and SPLICE_F_MORE was not set
1232 * initially, clears it.
1235 sd
->flags
|= SPLICE_F_MORE
;
1237 sd
->flags
&= ~SPLICE_F_MORE
;
1239 * NOTE: nonblocking mode only applies to the input. We
1240 * must not do the output in nonblocking mode as then we
1241 * could get stuck data in the internal pipe:
1243 ret
= actor(pipe
, sd
);
1244 if (unlikely(ret
<= 0)) {
1253 if (ret
< read_len
) {
1254 sd
->pos
= prev_pos
+ ret
;
1260 pipe
->nrbufs
= pipe
->curbuf
= 0;
1266 * If we did an incomplete transfer we must release
1267 * the pipe buffers in question:
1269 for (i
= 0; i
< pipe
->buffers
; i
++) {
1270 struct pipe_buffer
*buf
= pipe
->bufs
+ i
;
1273 buf
->ops
->release(pipe
, buf
);
1283 EXPORT_SYMBOL(splice_direct_to_actor
);
1285 static int direct_splice_actor(struct pipe_inode_info
*pipe
,
1286 struct splice_desc
*sd
)
1288 struct file
*file
= sd
->u
.file
;
1290 return do_splice_from(pipe
, file
, sd
->opos
, sd
->total_len
,
1295 * do_splice_direct - splices data directly between two files
1296 * @in: file to splice from
1297 * @ppos: input file offset
1298 * @out: file to splice to
1299 * @opos: output file offset
1300 * @len: number of bytes to splice
1301 * @flags: splice modifier flags
1304 * For use by do_sendfile(). splice can easily emulate sendfile, but
1305 * doing it in the application would incur an extra system call
1306 * (splice in + splice out, as compared to just sendfile()). So this helper
1307 * can splice directly through a process-private pipe.
1310 long do_splice_direct(struct file
*in
, loff_t
*ppos
, struct file
*out
,
1311 loff_t
*opos
, size_t len
, unsigned int flags
)
1313 struct splice_desc sd
= {
1323 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
1326 if (unlikely(out
->f_flags
& O_APPEND
))
1329 ret
= rw_verify_area(WRITE
, out
, opos
, len
);
1330 if (unlikely(ret
< 0))
1333 ret
= splice_direct_to_actor(in
, &sd
, direct_splice_actor
);
1339 EXPORT_SYMBOL(do_splice_direct
);
1341 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1342 struct pipe_inode_info
*opipe
,
1343 size_t len
, unsigned int flags
);
1346 * Determine where to splice to/from.
1348 static long do_splice(struct file
*in
, loff_t __user
*off_in
,
1349 struct file
*out
, loff_t __user
*off_out
,
1350 size_t len
, unsigned int flags
)
1352 struct pipe_inode_info
*ipipe
;
1353 struct pipe_inode_info
*opipe
;
1357 ipipe
= get_pipe_info(in
);
1358 opipe
= get_pipe_info(out
);
1360 if (ipipe
&& opipe
) {
1361 if (off_in
|| off_out
)
1364 if (!(in
->f_mode
& FMODE_READ
))
1367 if (!(out
->f_mode
& FMODE_WRITE
))
1370 /* Splicing to self would be fun, but... */
1374 return splice_pipe_to_pipe(ipipe
, opipe
, len
, flags
);
1381 if (!(out
->f_mode
& FMODE_PWRITE
))
1383 if (copy_from_user(&offset
, off_out
, sizeof(loff_t
)))
1386 offset
= out
->f_pos
;
1389 if (unlikely(!(out
->f_mode
& FMODE_WRITE
)))
1392 if (unlikely(out
->f_flags
& O_APPEND
))
1395 ret
= rw_verify_area(WRITE
, out
, &offset
, len
);
1396 if (unlikely(ret
< 0))
1399 file_start_write(out
);
1400 ret
= do_splice_from(ipipe
, out
, &offset
, len
, flags
);
1401 file_end_write(out
);
1404 out
->f_pos
= offset
;
1405 else if (copy_to_user(off_out
, &offset
, sizeof(loff_t
)))
1415 if (!(in
->f_mode
& FMODE_PREAD
))
1417 if (copy_from_user(&offset
, off_in
, sizeof(loff_t
)))
1423 ret
= do_splice_to(in
, &offset
, opipe
, len
, flags
);
1427 else if (copy_to_user(off_in
, &offset
, sizeof(loff_t
)))
1437 * Map an iov into an array of pages and offset/length tupples. With the
1438 * partial_page structure, we can map several non-contiguous ranges into
1439 * our ones pages[] map instead of splitting that operation into pieces.
1440 * Could easily be exported as a generic helper for other users, in which
1441 * case one would probably want to add a 'max_nr_pages' parameter as well.
1443 static int get_iovec_page_array(const struct iovec __user
*iov
,
1444 unsigned int nr_vecs
, struct page
**pages
,
1445 struct partial_page
*partial
, bool aligned
,
1446 unsigned int pipe_buffers
)
1448 int buffers
= 0, error
= 0;
1451 unsigned long off
, npages
;
1458 if (copy_from_user(&entry
, iov
, sizeof(entry
)))
1461 base
= entry
.iov_base
;
1462 len
= entry
.iov_len
;
1465 * Sanity check this iovec. 0 read succeeds.
1471 if (!access_ok(VERIFY_READ
, base
, len
))
1475 * Get this base offset and number of pages, then map
1476 * in the user pages.
1478 off
= (unsigned long) base
& ~PAGE_MASK
;
1481 * If asked for alignment, the offset must be zero and the
1482 * length a multiple of the PAGE_SIZE.
1485 if (aligned
&& (off
|| len
& ~PAGE_MASK
))
1488 npages
= (off
+ len
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1489 if (npages
> pipe_buffers
- buffers
)
1490 npages
= pipe_buffers
- buffers
;
1492 error
= get_user_pages_fast((unsigned long)base
, npages
,
1493 0, &pages
[buffers
]);
1495 if (unlikely(error
<= 0))
1499 * Fill this contiguous range into the partial page map.
1501 for (i
= 0; i
< error
; i
++) {
1502 const int plen
= min_t(size_t, len
, PAGE_SIZE
- off
);
1504 partial
[buffers
].offset
= off
;
1505 partial
[buffers
].len
= plen
;
1513 * We didn't complete this iov, stop here since it probably
1514 * means we have to move some of this into a pipe to
1515 * be able to continue.
1521 * Don't continue if we mapped fewer pages than we asked for,
1522 * or if we mapped the max number of pages that we have
1525 if (error
< npages
|| buffers
== pipe_buffers
)
1538 static int pipe_to_user(struct pipe_inode_info
*pipe
, struct pipe_buffer
*buf
,
1539 struct splice_desc
*sd
)
1541 int n
= copy_page_to_iter(buf
->page
, buf
->offset
, sd
->len
, sd
->u
.data
);
1542 return n
== sd
->len
? n
: -EFAULT
;
1546 * For lack of a better implementation, implement vmsplice() to userspace
1547 * as a simple copy of the pipes pages to the user iov.
1549 static long vmsplice_to_user(struct file
*file
, const struct iovec __user
*uiov
,
1550 unsigned long nr_segs
, unsigned int flags
)
1552 struct pipe_inode_info
*pipe
;
1553 struct splice_desc sd
;
1555 struct iovec iovstack
[UIO_FASTIOV
];
1556 struct iovec
*iov
= iovstack
;
1557 struct iov_iter iter
;
1559 pipe
= get_pipe_info(file
);
1563 ret
= import_iovec(READ
, uiov
, nr_segs
,
1564 ARRAY_SIZE(iovstack
), &iov
, &iter
);
1568 sd
.total_len
= iov_iter_count(&iter
);
1576 ret
= __splice_from_pipe(pipe
, &sd
, pipe_to_user
);
1585 * vmsplice splices a user address range into a pipe. It can be thought of
1586 * as splice-from-memory, where the regular splice is splice-from-file (or
1587 * to file). In both cases the output is a pipe, naturally.
1589 static long vmsplice_to_pipe(struct file
*file
, const struct iovec __user
*iov
,
1590 unsigned long nr_segs
, unsigned int flags
)
1592 struct pipe_inode_info
*pipe
;
1593 struct page
*pages
[PIPE_DEF_BUFFERS
];
1594 struct partial_page partial
[PIPE_DEF_BUFFERS
];
1595 struct splice_pipe_desc spd
= {
1598 .nr_pages_max
= PIPE_DEF_BUFFERS
,
1600 .ops
= &user_page_pipe_buf_ops
,
1601 .spd_release
= spd_release_page
,
1605 pipe
= get_pipe_info(file
);
1609 if (splice_grow_spd(pipe
, &spd
))
1612 spd
.nr_pages
= get_iovec_page_array(iov
, nr_segs
, spd
.pages
,
1615 if (spd
.nr_pages
<= 0)
1618 ret
= splice_to_pipe(pipe
, &spd
);
1620 splice_shrink_spd(&spd
);
1625 * Note that vmsplice only really supports true splicing _from_ user memory
1626 * to a pipe, not the other way around. Splicing from user memory is a simple
1627 * operation that can be supported without any funky alignment restrictions
1628 * or nasty vm tricks. We simply map in the user memory and fill them into
1629 * a pipe. The reverse isn't quite as easy, though. There are two possible
1630 * solutions for that:
1632 * - memcpy() the data internally, at which point we might as well just
1633 * do a regular read() on the buffer anyway.
1634 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1635 * has restriction limitations on both ends of the pipe).
1637 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1640 SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct iovec __user
*, iov
,
1641 unsigned long, nr_segs
, unsigned int, flags
)
1646 if (unlikely(nr_segs
> UIO_MAXIOV
))
1648 else if (unlikely(!nr_segs
))
1654 if (f
.file
->f_mode
& FMODE_WRITE
)
1655 error
= vmsplice_to_pipe(f
.file
, iov
, nr_segs
, flags
);
1656 else if (f
.file
->f_mode
& FMODE_READ
)
1657 error
= vmsplice_to_user(f
.file
, iov
, nr_segs
, flags
);
1665 #ifdef CONFIG_COMPAT
1666 COMPAT_SYSCALL_DEFINE4(vmsplice
, int, fd
, const struct compat_iovec __user
*, iov32
,
1667 unsigned int, nr_segs
, unsigned int, flags
)
1670 struct iovec __user
*iov
;
1671 if (nr_segs
> UIO_MAXIOV
)
1673 iov
= compat_alloc_user_space(nr_segs
* sizeof(struct iovec
));
1674 for (i
= 0; i
< nr_segs
; i
++) {
1675 struct compat_iovec v
;
1676 if (get_user(v
.iov_base
, &iov32
[i
].iov_base
) ||
1677 get_user(v
.iov_len
, &iov32
[i
].iov_len
) ||
1678 put_user(compat_ptr(v
.iov_base
), &iov
[i
].iov_base
) ||
1679 put_user(v
.iov_len
, &iov
[i
].iov_len
))
1682 return sys_vmsplice(fd
, iov
, nr_segs
, flags
);
1686 SYSCALL_DEFINE6(splice
, int, fd_in
, loff_t __user
*, off_in
,
1687 int, fd_out
, loff_t __user
*, off_out
,
1688 size_t, len
, unsigned int, flags
)
1699 if (in
.file
->f_mode
& FMODE_READ
) {
1700 out
= fdget(fd_out
);
1702 if (out
.file
->f_mode
& FMODE_WRITE
)
1703 error
= do_splice(in
.file
, off_in
,
1715 * Make sure there's data to read. Wait for input if we can, otherwise
1716 * return an appropriate error.
1718 static int ipipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1723 * Check ->nrbufs without the inode lock first. This function
1724 * is speculative anyways, so missing one is ok.
1732 while (!pipe
->nrbufs
) {
1733 if (signal_pending(current
)) {
1739 if (!pipe
->waiting_writers
) {
1740 if (flags
& SPLICE_F_NONBLOCK
) {
1753 * Make sure there's writeable room. Wait for room if we can, otherwise
1754 * return an appropriate error.
1756 static int opipe_prep(struct pipe_inode_info
*pipe
, unsigned int flags
)
1761 * Check ->nrbufs without the inode lock first. This function
1762 * is speculative anyways, so missing one is ok.
1764 if (pipe
->nrbufs
< pipe
->buffers
)
1770 while (pipe
->nrbufs
>= pipe
->buffers
) {
1771 if (!pipe
->readers
) {
1772 send_sig(SIGPIPE
, current
, 0);
1776 if (flags
& SPLICE_F_NONBLOCK
) {
1780 if (signal_pending(current
)) {
1784 pipe
->waiting_writers
++;
1786 pipe
->waiting_writers
--;
1794 * Splice contents of ipipe to opipe.
1796 static int splice_pipe_to_pipe(struct pipe_inode_info
*ipipe
,
1797 struct pipe_inode_info
*opipe
,
1798 size_t len
, unsigned int flags
)
1800 struct pipe_buffer
*ibuf
, *obuf
;
1802 bool input_wakeup
= false;
1806 ret
= ipipe_prep(ipipe
, flags
);
1810 ret
= opipe_prep(opipe
, flags
);
1815 * Potential ABBA deadlock, work around it by ordering lock
1816 * grabbing by pipe info address. Otherwise two different processes
1817 * could deadlock (one doing tee from A -> B, the other from B -> A).
1819 pipe_double_lock(ipipe
, opipe
);
1822 if (!opipe
->readers
) {
1823 send_sig(SIGPIPE
, current
, 0);
1829 if (!ipipe
->nrbufs
&& !ipipe
->writers
)
1833 * Cannot make any progress, because either the input
1834 * pipe is empty or the output pipe is full.
1836 if (!ipipe
->nrbufs
|| opipe
->nrbufs
>= opipe
->buffers
) {
1837 /* Already processed some buffers, break */
1841 if (flags
& SPLICE_F_NONBLOCK
) {
1847 * We raced with another reader/writer and haven't
1848 * managed to process any buffers. A zero return
1849 * value means EOF, so retry instead.
1856 ibuf
= ipipe
->bufs
+ ipipe
->curbuf
;
1857 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (opipe
->buffers
- 1);
1858 obuf
= opipe
->bufs
+ nbuf
;
1860 if (len
>= ibuf
->len
) {
1862 * Simply move the whole buffer from ipipe to opipe
1867 ipipe
->curbuf
= (ipipe
->curbuf
+ 1) & (ipipe
->buffers
- 1);
1869 input_wakeup
= true;
1872 * Get a reference to this pipe buffer,
1873 * so we can copy the contents over.
1875 ibuf
->ops
->get(ipipe
, ibuf
);
1879 * Don't inherit the gift flag, we need to
1880 * prevent multiple steals of this page.
1882 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1886 ibuf
->offset
+= obuf
->len
;
1887 ibuf
->len
-= obuf
->len
;
1897 * If we put data in the output pipe, wakeup any potential readers.
1900 wakeup_pipe_readers(opipe
);
1903 wakeup_pipe_writers(ipipe
);
1909 * Link contents of ipipe to opipe.
1911 static int link_pipe(struct pipe_inode_info
*ipipe
,
1912 struct pipe_inode_info
*opipe
,
1913 size_t len
, unsigned int flags
)
1915 struct pipe_buffer
*ibuf
, *obuf
;
1916 int ret
= 0, i
= 0, nbuf
;
1919 * Potential ABBA deadlock, work around it by ordering lock
1920 * grabbing by pipe info address. Otherwise two different processes
1921 * could deadlock (one doing tee from A -> B, the other from B -> A).
1923 pipe_double_lock(ipipe
, opipe
);
1926 if (!opipe
->readers
) {
1927 send_sig(SIGPIPE
, current
, 0);
1934 * If we have iterated all input buffers or ran out of
1935 * output room, break.
1937 if (i
>= ipipe
->nrbufs
|| opipe
->nrbufs
>= opipe
->buffers
)
1940 ibuf
= ipipe
->bufs
+ ((ipipe
->curbuf
+ i
) & (ipipe
->buffers
-1));
1941 nbuf
= (opipe
->curbuf
+ opipe
->nrbufs
) & (opipe
->buffers
- 1);
1944 * Get a reference to this pipe buffer,
1945 * so we can copy the contents over.
1947 ibuf
->ops
->get(ipipe
, ibuf
);
1949 obuf
= opipe
->bufs
+ nbuf
;
1953 * Don't inherit the gift flag, we need to
1954 * prevent multiple steals of this page.
1956 obuf
->flags
&= ~PIPE_BUF_FLAG_GIFT
;
1958 if (obuf
->len
> len
)
1968 * return EAGAIN if we have the potential of some data in the
1969 * future, otherwise just return 0
1971 if (!ret
&& ipipe
->waiting_writers
&& (flags
& SPLICE_F_NONBLOCK
))
1978 * If we put data in the output pipe, wakeup any potential readers.
1981 wakeup_pipe_readers(opipe
);
1987 * This is a tee(1) implementation that works on pipes. It doesn't copy
1988 * any data, it simply references the 'in' pages on the 'out' pipe.
1989 * The 'flags' used are the SPLICE_F_* variants, currently the only
1990 * applicable one is SPLICE_F_NONBLOCK.
1992 static long do_tee(struct file
*in
, struct file
*out
, size_t len
,
1995 struct pipe_inode_info
*ipipe
= get_pipe_info(in
);
1996 struct pipe_inode_info
*opipe
= get_pipe_info(out
);
2000 * Duplicate the contents of ipipe to opipe without actually
2003 if (ipipe
&& opipe
&& ipipe
!= opipe
) {
2005 * Keep going, unless we encounter an error. The ipipe/opipe
2006 * ordering doesn't really matter.
2008 ret
= ipipe_prep(ipipe
, flags
);
2010 ret
= opipe_prep(opipe
, flags
);
2012 ret
= link_pipe(ipipe
, opipe
, len
, flags
);
2019 SYSCALL_DEFINE4(tee
, int, fdin
, int, fdout
, size_t, len
, unsigned int, flags
)
2030 if (in
.file
->f_mode
& FMODE_READ
) {
2031 struct fd out
= fdget(fdout
);
2033 if (out
.file
->f_mode
& FMODE_WRITE
)
2034 error
= do_tee(in
.file
, out
.file
,