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 (page_has_private(page) &&
63 !try_to_release_page(page, GFP_KERNEL))
67 * If we succeeded in removing the mapping, set LRU flag
70 if (remove_mapping(mapping, page)) {
71 buf->flags |= PIPE_BUF_FLAG_LRU;
77 * Raced with truncate or failed to remove page from current
78 * address space, unlock and return failure.
85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
86 struct pipe_buffer *buf)
88 page_cache_release(buf->page);
89 buf->flags &= ~PIPE_BUF_FLAG_LRU;
93 * Check whether the contents of buf is OK to access. Since the content
94 * is a page cache page, IO may be in flight.
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
97 struct pipe_buffer *buf)
99 struct page *page = buf->page;
102 if (!PageUptodate(page)) {
106 * Page got truncated/unhashed. This will cause a 0-byte
107 * splice, if this is the first page.
109 if (!page->mapping) {
115 * Uh oh, read-error from disk.
117 if (!PageUptodate(page)) {
123 * Page is ok afterall, we are done.
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
136 .map = generic_pipe_buf_map,
137 .unmap = generic_pipe_buf_unmap,
138 .confirm = page_cache_pipe_buf_confirm,
139 .release = page_cache_pipe_buf_release,
140 .steal = page_cache_pipe_buf_steal,
141 .get = generic_pipe_buf_get,
144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
145 struct pipe_buffer *buf)
147 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
150 buf->flags |= PIPE_BUF_FLAG_LRU;
151 return generic_pipe_buf_steal(pipe, buf);
154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
156 .map = generic_pipe_buf_map,
157 .unmap = generic_pipe_buf_unmap,
158 .confirm = generic_pipe_buf_confirm,
159 .release = page_cache_pipe_buf_release,
160 .steal = user_page_pipe_buf_steal,
161 .get = generic_pipe_buf_get,
165 * splice_to_pipe - fill passed data into a pipe
166 * @pipe: pipe to fill
170 * @spd contains a map of pages and len/offset tuples, along with
171 * the struct pipe_buf_operations associated with these pages. This
172 * function will link that data to the pipe.
175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
176 struct splice_pipe_desc *spd)
178 unsigned int spd_pages = spd->nr_pages;
179 int ret, do_wakeup, page_nr;
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--;
251 if (waitqueue_active(&pipe->wait))
252 wake_up_interruptible(&pipe->wait);
253 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
256 while (page_nr < spd_pages)
257 spd->spd_release(spd, page_nr++);
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
264 page_cache_release(spd->pages[i]);
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269 struct pipe_inode_info *pipe, size_t len,
272 struct address_space *mapping = in->f_mapping;
273 unsigned int loff, nr_pages, req_pages;
274 struct page *pages[PIPE_BUFFERS];
275 struct partial_page partial[PIPE_BUFFERS];
277 pgoff_t index, end_index;
280 struct splice_pipe_desc spd = {
284 .ops = &page_cache_pipe_buf_ops,
285 .spd_release = spd_release_page,
288 index = *ppos >> PAGE_CACHE_SHIFT;
289 loff = *ppos & ~PAGE_CACHE_MASK;
290 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
294 * Lookup the (hopefully) full range of pages we need.
296 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297 index += spd.nr_pages;
300 * If find_get_pages_contig() returned fewer pages than we needed,
301 * readahead/allocate the rest and fill in the holes.
303 if (spd.nr_pages < nr_pages)
304 page_cache_sync_readahead(mapping, &in->f_ra, in,
305 index, req_pages - spd.nr_pages);
308 while (spd.nr_pages < nr_pages) {
310 * Page could be there, find_get_pages_contig() breaks on
313 page = find_get_page(mapping, index);
316 * page didn't exist, allocate one.
318 page = page_cache_alloc_cold(mapping);
322 error = add_to_page_cache_lru(page, mapping, index,
323 mapping_gfp_mask(mapping));
324 if (unlikely(error)) {
325 page_cache_release(page);
326 if (error == -EEXIST)
331 * add_to_page_cache() locks the page, unlock it
332 * to avoid convoluting the logic below even more.
337 pages[spd.nr_pages++] = page;
342 * Now loop over the map and see if we need to start IO on any
343 * pages, fill in the partial map, etc.
345 index = *ppos >> PAGE_CACHE_SHIFT;
346 nr_pages = spd.nr_pages;
348 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349 unsigned int this_len;
355 * this_len is the max we'll use from this page
357 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358 page = pages[page_nr];
360 if (PageReadahead(page))
361 page_cache_async_readahead(mapping, &in->f_ra, in,
362 page, index, req_pages - page_nr);
365 * If the page isn't uptodate, we may need to start io on it
367 if (!PageUptodate(page)) {
371 * Page was truncated, or invalidated by the
372 * filesystem. Redo the find/create, but this time the
373 * page is kept locked, so there's no chance of another
374 * race with truncate/invalidate.
376 if (!page->mapping) {
378 page = find_or_create_page(mapping, index,
379 mapping_gfp_mask(mapping));
385 page_cache_release(pages[page_nr]);
386 pages[page_nr] = page;
389 * page was already under io and is now done, great
391 if (PageUptodate(page)) {
397 * need to read in the page
399 error = mapping->a_ops->readpage(in, page);
400 if (unlikely(error)) {
402 * We really should re-lookup the page here,
403 * but it complicates things a lot. Instead
404 * lets just do what we already stored, and
405 * we'll get it the next time we are called.
407 if (error == AOP_TRUNCATED_PAGE)
415 * i_size must be checked after PageUptodate.
417 isize = i_size_read(mapping->host);
418 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
419 if (unlikely(!isize || index > end_index))
423 * if this is the last page, see if we need to shrink
424 * the length and stop
426 if (end_index == index) {
430 * max good bytes in this page
432 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
437 * force quit after adding this page
439 this_len = min(this_len, plen - loff);
443 partial[page_nr].offset = loff;
444 partial[page_nr].len = this_len;
452 * Release any pages at the end, if we quit early. 'page_nr' is how far
453 * we got, 'nr_pages' is how many pages are in the map.
455 while (page_nr < nr_pages)
456 page_cache_release(pages[page_nr++]);
457 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
460 return splice_to_pipe(pipe, &spd);
466 * generic_file_splice_read - splice data from file to a pipe
467 * @in: file to splice from
468 * @ppos: position in @in
469 * @pipe: pipe to splice to
470 * @len: number of bytes to splice
471 * @flags: splice modifier flags
474 * Will read pages from given file and fill them into a pipe. Can be
475 * used as long as the address_space operations for the source implements
479 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
480 struct pipe_inode_info *pipe, size_t len,
486 isize = i_size_read(in->f_mapping->host);
487 if (unlikely(*ppos >= isize))
490 left = isize - *ppos;
491 if (unlikely(left < len))
494 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
502 EXPORT_SYMBOL(generic_file_splice_read);
504 static const struct pipe_buf_operations default_pipe_buf_ops = {
506 .map = generic_pipe_buf_map,
507 .unmap = generic_pipe_buf_unmap,
508 .confirm = generic_pipe_buf_confirm,
509 .release = generic_pipe_buf_release,
510 .steal = generic_pipe_buf_steal,
511 .get = generic_pipe_buf_get,
514 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
515 unsigned long vlen, loff_t offset)
523 /* The cast to a user pointer is valid due to the set_fs() */
524 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
530 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
538 /* The cast to a user pointer is valid due to the set_fs() */
539 res = vfs_write(file, (const char __user *)buf, count, &pos);
545 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
546 struct pipe_inode_info *pipe, size_t len,
549 unsigned int nr_pages;
550 unsigned int nr_freed;
552 struct page *pages[PIPE_BUFFERS];
553 struct partial_page partial[PIPE_BUFFERS];
554 struct iovec vec[PIPE_BUFFERS];
560 struct splice_pipe_desc spd = {
564 .ops = &default_pipe_buf_ops,
565 .spd_release = spd_release_page,
568 index = *ppos >> PAGE_CACHE_SHIFT;
569 offset = *ppos & ~PAGE_CACHE_MASK;
570 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
572 for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
575 page = alloc_page(GFP_USER);
580 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
581 vec[i].iov_base = (void __user *) page_address(page);
582 vec[i].iov_len = this_len;
589 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
600 for (i = 0; i < spd.nr_pages; i++) {
601 this_len = min_t(size_t, vec[i].iov_len, res);
602 partial[i].offset = 0;
603 partial[i].len = this_len;
605 __free_page(pages[i]);
611 spd.nr_pages -= nr_freed;
613 res = splice_to_pipe(pipe, &spd);
620 for (i = 0; i < spd.nr_pages; i++)
621 __free_page(pages[i]);
625 EXPORT_SYMBOL(default_file_splice_read);
628 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
629 * using sendpage(). Return the number of bytes sent.
631 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
632 struct pipe_buffer *buf, struct splice_desc *sd)
634 struct file *file = sd->u.file;
635 loff_t pos = sd->pos;
638 ret = buf->ops->confirm(pipe, buf);
640 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
641 if (file->f_op && file->f_op->sendpage)
642 ret = file->f_op->sendpage(file, buf->page, buf->offset,
643 sd->len, &pos, more);
652 * This is a little more tricky than the file -> pipe splicing. There are
653 * basically three cases:
655 * - Destination page already exists in the address space and there
656 * are users of it. For that case we have no other option that
657 * copying the data. Tough luck.
658 * - Destination page already exists in the address space, but there
659 * are no users of it. Make sure it's uptodate, then drop it. Fall
660 * through to last case.
661 * - Destination page does not exist, we can add the pipe page to
662 * the page cache and avoid the copy.
664 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
665 * sd->flags), we attempt to migrate pages from the pipe to the output
666 * file address space page cache. This is possible if no one else has
667 * the pipe page referenced outside of the pipe and page cache. If
668 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
669 * a new page in the output file page cache and fill/dirty that.
671 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
672 struct splice_desc *sd)
674 struct file *file = sd->u.file;
675 struct address_space *mapping = file->f_mapping;
676 unsigned int offset, this_len;
682 * make sure the data in this buffer is uptodate
684 ret = buf->ops->confirm(pipe, buf);
688 offset = sd->pos & ~PAGE_CACHE_MASK;
691 if (this_len + offset > PAGE_CACHE_SIZE)
692 this_len = PAGE_CACHE_SIZE - offset;
694 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
695 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
699 if (buf->page != page) {
701 * Careful, ->map() uses KM_USER0!
703 char *src = buf->ops->map(pipe, buf, 1);
704 char *dst = kmap_atomic(page, KM_USER1);
706 memcpy(dst + offset, src + buf->offset, this_len);
707 flush_dcache_page(page);
708 kunmap_atomic(dst, KM_USER1);
709 buf->ops->unmap(pipe, buf, src);
711 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
716 EXPORT_SYMBOL(pipe_to_file);
718 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
721 if (waitqueue_active(&pipe->wait))
722 wake_up_interruptible(&pipe->wait);
723 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
727 * splice_from_pipe_feed - feed available data from a pipe to a file
728 * @pipe: pipe to splice from
729 * @sd: information to @actor
730 * @actor: handler that splices the data
733 * This function loops over the pipe and calls @actor to do the
734 * actual moving of a single struct pipe_buffer to the desired
735 * destination. It returns when there's no more buffers left in
736 * the pipe or if the requested number of bytes (@sd->total_len)
737 * have been copied. It returns a positive number (one) if the
738 * pipe needs to be filled with more data, zero if the required
739 * number of bytes have been copied and -errno on error.
741 * This, together with splice_from_pipe_{begin,end,next}, may be
742 * used to implement the functionality of __splice_from_pipe() when
743 * locking is required around copying the pipe buffers to the
746 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
751 while (pipe->nrbufs) {
752 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
753 const struct pipe_buf_operations *ops = buf->ops;
756 if (sd->len > sd->total_len)
757 sd->len = sd->total_len;
759 ret = actor(pipe, buf, sd);
768 sd->num_spliced += ret;
771 sd->total_len -= ret;
775 ops->release(pipe, buf);
776 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
779 sd->need_wakeup = true;
788 EXPORT_SYMBOL(splice_from_pipe_feed);
791 * splice_from_pipe_next - wait for some data to splice from
792 * @pipe: pipe to splice from
793 * @sd: information about the splice operation
796 * This function will wait for some data and return a positive
797 * value (one) if pipe buffers are available. It will return zero
798 * or -errno if no more data needs to be spliced.
800 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
802 while (!pipe->nrbufs) {
806 if (!pipe->waiting_writers && sd->num_spliced)
809 if (sd->flags & SPLICE_F_NONBLOCK)
812 if (signal_pending(current))
815 if (sd->need_wakeup) {
816 wakeup_pipe_writers(pipe);
817 sd->need_wakeup = false;
825 EXPORT_SYMBOL(splice_from_pipe_next);
828 * splice_from_pipe_begin - start splicing from pipe
829 * @sd: information about the splice operation
832 * This function should be called before a loop containing
833 * splice_from_pipe_next() and splice_from_pipe_feed() to
834 * initialize the necessary fields of @sd.
836 void splice_from_pipe_begin(struct splice_desc *sd)
839 sd->need_wakeup = false;
841 EXPORT_SYMBOL(splice_from_pipe_begin);
844 * splice_from_pipe_end - finish splicing from pipe
845 * @pipe: pipe to splice from
846 * @sd: information about the splice operation
849 * This function will wake up pipe writers if necessary. It should
850 * be called after a loop containing splice_from_pipe_next() and
851 * splice_from_pipe_feed().
853 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
856 wakeup_pipe_writers(pipe);
858 EXPORT_SYMBOL(splice_from_pipe_end);
861 * __splice_from_pipe - splice data from a pipe to given actor
862 * @pipe: pipe to splice from
863 * @sd: information to @actor
864 * @actor: handler that splices the data
867 * This function does little more than loop over the pipe and call
868 * @actor to do the actual moving of a single struct pipe_buffer to
869 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
873 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
878 splice_from_pipe_begin(sd);
880 ret = splice_from_pipe_next(pipe, sd);
882 ret = splice_from_pipe_feed(pipe, sd, actor);
884 splice_from_pipe_end(pipe, sd);
886 return sd->num_spliced ? sd->num_spliced : ret;
888 EXPORT_SYMBOL(__splice_from_pipe);
891 * splice_from_pipe - splice data from a pipe to a file
892 * @pipe: pipe to splice from
893 * @out: file to splice to
894 * @ppos: position in @out
895 * @len: how many bytes to splice
896 * @flags: splice modifier flags
897 * @actor: handler that splices the data
900 * See __splice_from_pipe. This function locks the pipe inode,
901 * otherwise it's identical to __splice_from_pipe().
904 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
905 loff_t *ppos, size_t len, unsigned int flags,
909 struct splice_desc sd = {
917 ret = __splice_from_pipe(pipe, &sd, actor);
924 * generic_file_splice_write - splice data from a pipe to a file
926 * @out: file to write to
927 * @ppos: position in @out
928 * @len: number of bytes to splice
929 * @flags: splice modifier flags
932 * Will either move or copy pages (determined by @flags options) from
933 * the given pipe inode to the given file.
937 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
938 loff_t *ppos, size_t len, unsigned int flags)
940 struct address_space *mapping = out->f_mapping;
941 struct inode *inode = mapping->host;
942 struct splice_desc sd = {
952 splice_from_pipe_begin(&sd);
954 ret = splice_from_pipe_next(pipe, &sd);
958 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
959 ret = file_remove_suid(out);
961 file_update_time(out);
962 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
964 mutex_unlock(&inode->i_mutex);
966 splice_from_pipe_end(pipe, &sd);
971 ret = sd.num_spliced;
974 unsigned long nr_pages;
977 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
979 err = generic_write_sync(out, *ppos, ret);
984 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
990 EXPORT_SYMBOL(generic_file_splice_write);
992 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
993 struct splice_desc *sd)
998 ret = buf->ops->confirm(pipe, buf);
1002 data = buf->ops->map(pipe, buf, 0);
1003 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1004 buf->ops->unmap(pipe, buf, data);
1009 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1010 struct file *out, loff_t *ppos,
1011 size_t len, unsigned int flags)
1015 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1023 * generic_splice_sendpage - splice data from a pipe to a socket
1024 * @pipe: pipe to splice from
1025 * @out: socket to write to
1026 * @ppos: position in @out
1027 * @len: number of bytes to splice
1028 * @flags: splice modifier flags
1031 * Will send @len bytes from the pipe to a network socket. No data copying
1035 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1036 loff_t *ppos, size_t len, unsigned int flags)
1038 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1041 EXPORT_SYMBOL(generic_splice_sendpage);
1044 * Attempt to initiate a splice from pipe to file.
1046 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1047 loff_t *ppos, size_t len, unsigned int flags)
1049 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1050 loff_t *, size_t, unsigned int);
1053 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1056 if (unlikely(out->f_flags & O_APPEND))
1059 ret = rw_verify_area(WRITE, out, ppos, len);
1060 if (unlikely(ret < 0))
1063 if (out->f_op && out->f_op->splice_write)
1064 splice_write = out->f_op->splice_write;
1066 splice_write = default_file_splice_write;
1068 return splice_write(pipe, out, ppos, len, flags);
1072 * Attempt to initiate a splice from a file to a pipe.
1074 static long do_splice_to(struct file *in, loff_t *ppos,
1075 struct pipe_inode_info *pipe, size_t len,
1078 ssize_t (*splice_read)(struct file *, loff_t *,
1079 struct pipe_inode_info *, size_t, unsigned int);
1082 if (unlikely(!(in->f_mode & FMODE_READ)))
1085 ret = rw_verify_area(READ, in, ppos, len);
1086 if (unlikely(ret < 0))
1089 if (in->f_op && in->f_op->splice_read)
1090 splice_read = in->f_op->splice_read;
1092 splice_read = default_file_splice_read;
1094 return splice_read(in, ppos, pipe, len, flags);
1098 * splice_direct_to_actor - splices data directly between two non-pipes
1099 * @in: file to splice from
1100 * @sd: actor information on where to splice to
1101 * @actor: handles the data splicing
1104 * This is a special case helper to splice directly between two
1105 * points, without requiring an explicit pipe. Internally an allocated
1106 * pipe is cached in the process, and reused during the lifetime of
1110 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1111 splice_direct_actor *actor)
1113 struct pipe_inode_info *pipe;
1120 * We require the input being a regular file, as we don't want to
1121 * randomly drop data for eg socket -> socket splicing. Use the
1122 * piped splicing for that!
1124 i_mode = in->f_path.dentry->d_inode->i_mode;
1125 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1129 * neither in nor out is a pipe, setup an internal pipe attached to
1130 * 'out' and transfer the wanted data from 'in' to 'out' through that
1132 pipe = current->splice_pipe;
1133 if (unlikely(!pipe)) {
1134 pipe = alloc_pipe_info(NULL);
1139 * We don't have an immediate reader, but we'll read the stuff
1140 * out of the pipe right after the splice_to_pipe(). So set
1141 * PIPE_READERS appropriately.
1145 current->splice_pipe = pipe;
1153 len = sd->total_len;
1157 * Don't block on output, we have to drain the direct pipe.
1159 sd->flags &= ~SPLICE_F_NONBLOCK;
1163 loff_t pos = sd->pos, prev_pos = pos;
1165 ret = do_splice_to(in, &pos, pipe, len, flags);
1166 if (unlikely(ret <= 0))
1170 sd->total_len = read_len;
1173 * NOTE: nonblocking mode only applies to the input. We
1174 * must not do the output in nonblocking mode as then we
1175 * could get stuck data in the internal pipe:
1177 ret = actor(pipe, sd);
1178 if (unlikely(ret <= 0)) {
1187 if (ret < read_len) {
1188 sd->pos = prev_pos + ret;
1194 pipe->nrbufs = pipe->curbuf = 0;
1200 * If we did an incomplete transfer we must release
1201 * the pipe buffers in question:
1203 for (i = 0; i < PIPE_BUFFERS; i++) {
1204 struct pipe_buffer *buf = pipe->bufs + i;
1207 buf->ops->release(pipe, buf);
1217 EXPORT_SYMBOL(splice_direct_to_actor);
1219 static int direct_splice_actor(struct pipe_inode_info *pipe,
1220 struct splice_desc *sd)
1222 struct file *file = sd->u.file;
1224 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1229 * do_splice_direct - splices data directly between two files
1230 * @in: file to splice from
1231 * @ppos: input file offset
1232 * @out: file to splice to
1233 * @len: number of bytes to splice
1234 * @flags: splice modifier flags
1237 * For use by do_sendfile(). splice can easily emulate sendfile, but
1238 * doing it in the application would incur an extra system call
1239 * (splice in + splice out, as compared to just sendfile()). So this helper
1240 * can splice directly through a process-private pipe.
1243 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1244 size_t len, unsigned int flags)
1246 struct splice_desc sd = {
1255 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1262 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1263 struct pipe_inode_info *opipe,
1264 size_t len, unsigned int flags);
1266 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1267 * location, so checking ->i_pipe is not enough to verify that this is a
1270 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1272 if (S_ISFIFO(inode->i_mode))
1273 return inode->i_pipe;
1279 * Determine where to splice to/from.
1281 static long do_splice(struct file *in, loff_t __user *off_in,
1282 struct file *out, loff_t __user *off_out,
1283 size_t len, unsigned int flags)
1285 struct pipe_inode_info *ipipe;
1286 struct pipe_inode_info *opipe;
1287 loff_t offset, *off;
1290 ipipe = pipe_info(in->f_path.dentry->d_inode);
1291 opipe = pipe_info(out->f_path.dentry->d_inode);
1293 if (ipipe && opipe) {
1294 if (off_in || off_out)
1297 if (!(in->f_mode & FMODE_READ))
1300 if (!(out->f_mode & FMODE_WRITE))
1303 /* Splicing to self would be fun, but... */
1307 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1314 if (!out->f_op || !out->f_op->llseek ||
1315 out->f_op->llseek == no_llseek)
1317 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1323 ret = do_splice_from(ipipe, out, off, len, flags);
1325 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1335 if (!in->f_op || !in->f_op->llseek ||
1336 in->f_op->llseek == no_llseek)
1338 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1344 ret = do_splice_to(in, off, opipe, len, flags);
1346 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1356 * Map an iov into an array of pages and offset/length tupples. With the
1357 * partial_page structure, we can map several non-contiguous ranges into
1358 * our ones pages[] map instead of splitting that operation into pieces.
1359 * Could easily be exported as a generic helper for other users, in which
1360 * case one would probably want to add a 'max_nr_pages' parameter as well.
1362 static int get_iovec_page_array(const struct iovec __user *iov,
1363 unsigned int nr_vecs, struct page **pages,
1364 struct partial_page *partial, int aligned)
1366 int buffers = 0, error = 0;
1369 unsigned long off, npages;
1376 if (copy_from_user(&entry, iov, sizeof(entry)))
1379 base = entry.iov_base;
1380 len = entry.iov_len;
1383 * Sanity check this iovec. 0 read succeeds.
1389 if (!access_ok(VERIFY_READ, base, len))
1393 * Get this base offset and number of pages, then map
1394 * in the user pages.
1396 off = (unsigned long) base & ~PAGE_MASK;
1399 * If asked for alignment, the offset must be zero and the
1400 * length a multiple of the PAGE_SIZE.
1403 if (aligned && (off || len & ~PAGE_MASK))
1406 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1407 if (npages > PIPE_BUFFERS - buffers)
1408 npages = PIPE_BUFFERS - buffers;
1410 error = get_user_pages_fast((unsigned long)base, npages,
1411 0, &pages[buffers]);
1413 if (unlikely(error <= 0))
1417 * Fill this contiguous range into the partial page map.
1419 for (i = 0; i < error; i++) {
1420 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1422 partial[buffers].offset = off;
1423 partial[buffers].len = plen;
1431 * We didn't complete this iov, stop here since it probably
1432 * means we have to move some of this into a pipe to
1433 * be able to continue.
1439 * Don't continue if we mapped fewer pages than we asked for,
1440 * or if we mapped the max number of pages that we have
1443 if (error < npages || buffers == PIPE_BUFFERS)
1456 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1457 struct splice_desc *sd)
1462 ret = buf->ops->confirm(pipe, buf);
1467 * See if we can use the atomic maps, by prefaulting in the
1468 * pages and doing an atomic copy
1470 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1471 src = buf->ops->map(pipe, buf, 1);
1472 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1474 buf->ops->unmap(pipe, buf, src);
1482 * No dice, use slow non-atomic map and copy
1484 src = buf->ops->map(pipe, buf, 0);
1487 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1490 buf->ops->unmap(pipe, buf, src);
1493 sd->u.userptr += ret;
1498 * For lack of a better implementation, implement vmsplice() to userspace
1499 * as a simple copy of the pipes pages to the user iov.
1501 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1502 unsigned long nr_segs, unsigned int flags)
1504 struct pipe_inode_info *pipe;
1505 struct splice_desc sd;
1510 pipe = pipe_info(file->f_path.dentry->d_inode);
1522 * Get user address base and length for this iovec.
1524 error = get_user(base, &iov->iov_base);
1525 if (unlikely(error))
1527 error = get_user(len, &iov->iov_len);
1528 if (unlikely(error))
1532 * Sanity check this iovec. 0 read succeeds.
1536 if (unlikely(!base)) {
1541 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1549 sd.u.userptr = base;
1552 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1578 * vmsplice splices a user address range into a pipe. It can be thought of
1579 * as splice-from-memory, where the regular splice is splice-from-file (or
1580 * to file). In both cases the output is a pipe, naturally.
1582 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1583 unsigned long nr_segs, unsigned int flags)
1585 struct pipe_inode_info *pipe;
1586 struct page *pages[PIPE_BUFFERS];
1587 struct partial_page partial[PIPE_BUFFERS];
1588 struct splice_pipe_desc spd = {
1592 .ops = &user_page_pipe_buf_ops,
1593 .spd_release = spd_release_page,
1596 pipe = pipe_info(file->f_path.dentry->d_inode);
1600 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1601 flags & SPLICE_F_GIFT);
1602 if (spd.nr_pages <= 0)
1603 return spd.nr_pages;
1605 return splice_to_pipe(pipe, &spd);
1609 * Note that vmsplice only really supports true splicing _from_ user memory
1610 * to a pipe, not the other way around. Splicing from user memory is a simple
1611 * operation that can be supported without any funky alignment restrictions
1612 * or nasty vm tricks. We simply map in the user memory and fill them into
1613 * a pipe. The reverse isn't quite as easy, though. There are two possible
1614 * solutions for that:
1616 * - memcpy() the data internally, at which point we might as well just
1617 * do a regular read() on the buffer anyway.
1618 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1619 * has restriction limitations on both ends of the pipe).
1621 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1624 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1625 unsigned long, nr_segs, unsigned int, flags)
1631 if (unlikely(nr_segs > UIO_MAXIOV))
1633 else if (unlikely(!nr_segs))
1637 file = fget_light(fd, &fput);
1639 if (file->f_mode & FMODE_WRITE)
1640 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1641 else if (file->f_mode & FMODE_READ)
1642 error = vmsplice_to_user(file, iov, nr_segs, flags);
1644 fput_light(file, fput);
1650 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1651 int, fd_out, loff_t __user *, off_out,
1652 size_t, len, unsigned int, flags)
1655 struct file *in, *out;
1656 int fput_in, fput_out;
1662 in = fget_light(fd_in, &fput_in);
1664 if (in->f_mode & FMODE_READ) {
1665 out = fget_light(fd_out, &fput_out);
1667 if (out->f_mode & FMODE_WRITE)
1668 error = do_splice(in, off_in,
1671 fput_light(out, fput_out);
1675 fput_light(in, fput_in);
1682 * Make sure there's data to read. Wait for input if we can, otherwise
1683 * return an appropriate error.
1685 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1690 * Check ->nrbufs without the inode lock first. This function
1691 * is speculative anyways, so missing one is ok.
1699 while (!pipe->nrbufs) {
1700 if (signal_pending(current)) {
1706 if (!pipe->waiting_writers) {
1707 if (flags & SPLICE_F_NONBLOCK) {
1720 * Make sure there's writeable room. Wait for room if we can, otherwise
1721 * return an appropriate error.
1723 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1728 * Check ->nrbufs without the inode lock first. This function
1729 * is speculative anyways, so missing one is ok.
1731 if (pipe->nrbufs < PIPE_BUFFERS)
1737 while (pipe->nrbufs >= PIPE_BUFFERS) {
1738 if (!pipe->readers) {
1739 send_sig(SIGPIPE, current, 0);
1743 if (flags & SPLICE_F_NONBLOCK) {
1747 if (signal_pending(current)) {
1751 pipe->waiting_writers++;
1753 pipe->waiting_writers--;
1761 * Splice contents of ipipe to opipe.
1763 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1764 struct pipe_inode_info *opipe,
1765 size_t len, unsigned int flags)
1767 struct pipe_buffer *ibuf, *obuf;
1769 bool input_wakeup = false;
1773 ret = ipipe_prep(ipipe, flags);
1777 ret = opipe_prep(opipe, flags);
1782 * Potential ABBA deadlock, work around it by ordering lock
1783 * grabbing by pipe info address. Otherwise two different processes
1784 * could deadlock (one doing tee from A -> B, the other from B -> A).
1786 pipe_double_lock(ipipe, opipe);
1789 if (!opipe->readers) {
1790 send_sig(SIGPIPE, current, 0);
1796 if (!ipipe->nrbufs && !ipipe->writers)
1800 * Cannot make any progress, because either the input
1801 * pipe is empty or the output pipe is full.
1803 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1804 /* Already processed some buffers, break */
1808 if (flags & SPLICE_F_NONBLOCK) {
1814 * We raced with another reader/writer and haven't
1815 * managed to process any buffers. A zero return
1816 * value means EOF, so retry instead.
1823 ibuf = ipipe->bufs + ipipe->curbuf;
1824 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1825 obuf = opipe->bufs + nbuf;
1827 if (len >= ibuf->len) {
1829 * Simply move the whole buffer from ipipe to opipe
1834 ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1836 input_wakeup = true;
1839 * Get a reference to this pipe buffer,
1840 * so we can copy the contents over.
1842 ibuf->ops->get(ipipe, ibuf);
1846 * Don't inherit the gift flag, we need to
1847 * prevent multiple steals of this page.
1849 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1853 ibuf->offset += obuf->len;
1854 ibuf->len -= obuf->len;
1864 * If we put data in the output pipe, wakeup any potential readers.
1868 if (waitqueue_active(&opipe->wait))
1869 wake_up_interruptible(&opipe->wait);
1870 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1873 wakeup_pipe_writers(ipipe);
1879 * Link contents of ipipe to opipe.
1881 static int link_pipe(struct pipe_inode_info *ipipe,
1882 struct pipe_inode_info *opipe,
1883 size_t len, unsigned int flags)
1885 struct pipe_buffer *ibuf, *obuf;
1886 int ret = 0, i = 0, nbuf;
1889 * Potential ABBA deadlock, work around it by ordering lock
1890 * grabbing by pipe info address. Otherwise two different processes
1891 * could deadlock (one doing tee from A -> B, the other from B -> A).
1893 pipe_double_lock(ipipe, opipe);
1896 if (!opipe->readers) {
1897 send_sig(SIGPIPE, current, 0);
1904 * If we have iterated all input buffers or ran out of
1905 * output room, break.
1907 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1910 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1911 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1914 * Get a reference to this pipe buffer,
1915 * so we can copy the contents over.
1917 ibuf->ops->get(ipipe, ibuf);
1919 obuf = opipe->bufs + nbuf;
1923 * Don't inherit the gift flag, we need to
1924 * prevent multiple steals of this page.
1926 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1928 if (obuf->len > len)
1938 * return EAGAIN if we have the potential of some data in the
1939 * future, otherwise just return 0
1941 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1948 * If we put data in the output pipe, wakeup any potential readers.
1952 if (waitqueue_active(&opipe->wait))
1953 wake_up_interruptible(&opipe->wait);
1954 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1961 * This is a tee(1) implementation that works on pipes. It doesn't copy
1962 * any data, it simply references the 'in' pages on the 'out' pipe.
1963 * The 'flags' used are the SPLICE_F_* variants, currently the only
1964 * applicable one is SPLICE_F_NONBLOCK.
1966 static long do_tee(struct file *in, struct file *out, size_t len,
1969 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1970 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1974 * Duplicate the contents of ipipe to opipe without actually
1977 if (ipipe && opipe && ipipe != opipe) {
1979 * Keep going, unless we encounter an error. The ipipe/opipe
1980 * ordering doesn't really matter.
1982 ret = ipipe_prep(ipipe, flags);
1984 ret = opipe_prep(opipe, flags);
1986 ret = link_pipe(ipipe, opipe, len, flags);
1993 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2002 in = fget_light(fdin, &fput_in);
2004 if (in->f_mode & FMODE_READ) {
2006 struct file *out = fget_light(fdout, &fput_out);
2009 if (out->f_mode & FMODE_WRITE)
2010 error = do_tee(in, out, len, flags);
2011 fput_light(out, fput_out);
2014 fput_light(in, fput_in);