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@suse.de>
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/pipe_fs_i.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
32 * Passed to the actors
35 unsigned int len, total_len; /* current and remaining length */
36 unsigned int flags; /* splice flags */
37 struct file *file; /* file to read/write */
38 loff_t pos; /* file position */
42 * Attempt to steal a page from a pipe buffer. This should perhaps go into
43 * a vm helper function, it's already simplified quite a bit by the
44 * addition of remove_mapping(). If success is returned, the caller may
45 * attempt to reuse this page for another destination.
47 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
48 struct pipe_buffer *buf)
50 struct page *page = buf->page;
51 struct address_space *mapping = page_mapping(page);
53 WARN_ON(!PageLocked(page));
54 WARN_ON(!PageUptodate(page));
57 * At least for ext2 with nobh option, we need to wait on writeback
58 * completing on this page, since we'll remove it from the pagecache.
59 * Otherwise truncate wont wait on the page, allowing the disk
60 * blocks to be reused by someone else before we actually wrote our
61 * data to them. fs corruption ensues.
63 wait_on_page_writeback(page);
65 if (PagePrivate(page))
66 try_to_release_page(page, mapping_gfp_mask(mapping));
68 if (!remove_mapping(mapping, page))
71 buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
75 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
76 struct pipe_buffer *buf)
78 page_cache_release(buf->page);
80 buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
83 static void *page_cache_pipe_buf_map(struct file *file,
84 struct pipe_inode_info *info,
85 struct pipe_buffer *buf)
87 struct page *page = buf->page;
90 if (!PageUptodate(page)) {
94 * Page got truncated/unhashed. This will cause a 0-byte
95 * splice, if this is the first page.
103 * Uh oh, read-error from disk.
105 if (!PageUptodate(page)) {
111 * Page is ok afterall, fall through to mapping.
122 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
123 struct pipe_buffer *buf)
128 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
130 .map = page_cache_pipe_buf_map,
131 .unmap = page_cache_pipe_buf_unmap,
132 .release = page_cache_pipe_buf_release,
133 .steal = page_cache_pipe_buf_steal,
137 * Pipe output worker. This sets up our pipe format with the page cache
138 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
140 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
141 int nr_pages, unsigned long offset,
142 unsigned long len, unsigned int flags)
144 int ret, do_wakeup, i;
151 mutex_lock(&pipe->inode->i_mutex);
154 if (!pipe->readers) {
155 send_sig(SIGPIPE, current, 0);
161 if (pipe->nrbufs < PIPE_BUFFERS) {
162 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
163 struct pipe_buffer *buf = pipe->bufs + newbuf;
164 struct page *page = pages[i++];
165 unsigned long this_len;
167 this_len = PAGE_CACHE_SIZE - offset;
172 buf->offset = offset;
174 buf->ops = &page_cache_pipe_buf_ops;
186 if (pipe->nrbufs < PIPE_BUFFERS)
192 if (flags & SPLICE_F_NONBLOCK) {
198 if (signal_pending(current)) {
206 if (waitqueue_active(&pipe->wait))
207 wake_up_interruptible_sync(&pipe->wait);
208 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
212 pipe->waiting_writers++;
214 pipe->waiting_writers--;
218 mutex_unlock(&pipe->inode->i_mutex);
222 if (waitqueue_active(&pipe->wait))
223 wake_up_interruptible(&pipe->wait);
224 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
228 page_cache_release(pages[i++]);
234 __generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
235 size_t len, unsigned int flags)
237 struct address_space *mapping = in->f_mapping;
238 unsigned int offset, nr_pages;
239 struct page *pages[PIPE_BUFFERS];
244 index = in->f_pos >> PAGE_CACHE_SHIFT;
245 offset = in->f_pos & ~PAGE_CACHE_MASK;
246 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
248 if (nr_pages > PIPE_BUFFERS)
249 nr_pages = PIPE_BUFFERS;
252 * Initiate read-ahead on this page range. however, don't call into
253 * read-ahead if this is a non-zero offset (we are likely doing small
254 * chunk splice and the page is already there) for a single page.
256 if (!offset || nr_pages > 1)
257 do_page_cache_readahead(mapping, in, index, nr_pages);
260 * Now fill in the holes:
263 for (i = 0; i < nr_pages; i++, index++) {
266 * lookup the page for this index
268 page = find_get_page(mapping, index);
271 * If in nonblock mode then dont block on
272 * readpage (we've kicked readahead so there
273 * will be asynchronous progress):
275 if (flags & SPLICE_F_NONBLOCK)
279 * page didn't exist, allocate one
281 page = page_cache_alloc_cold(mapping);
285 error = add_to_page_cache_lru(page, mapping, index,
286 mapping_gfp_mask(mapping));
287 if (unlikely(error)) {
288 page_cache_release(page);
296 * If the page isn't uptodate, we may need to start io on it
298 if (!PageUptodate(page)) {
302 * page was truncated, stop here. if this isn't the
303 * first page, we'll just complete what we already
306 if (!page->mapping) {
308 page_cache_release(page);
312 * page was already under io and is now done, great
314 if (PageUptodate(page)) {
321 * need to read in the page
323 error = mapping->a_ops->readpage(in, page);
325 if (unlikely(error)) {
326 page_cache_release(page);
327 if (error == AOP_TRUNCATED_PAGE)
337 return move_to_pipe(pipe, pages, i, offset, len, flags);
343 * generic_file_splice_read - splice data from file to a pipe
344 * @in: file to splice from
345 * @pipe: pipe to splice to
346 * @len: number of bytes to splice
347 * @flags: splice modifier flags
349 * Will read pages from given file and fill them into a pipe.
351 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
352 size_t len, unsigned int flags)
361 ret = __generic_file_splice_read(in, pipe, len, flags);
370 if (!(flags & SPLICE_F_NONBLOCK))
382 EXPORT_SYMBOL(generic_file_splice_read);
385 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
388 static int pipe_to_sendpage(struct pipe_inode_info *info,
389 struct pipe_buffer *buf, struct splice_desc *sd)
391 struct file *file = sd->file;
392 loff_t pos = sd->pos;
399 * Sub-optimal, but we are limited by the pipe ->map. We don't
400 * need a kmap'ed buffer here, we just want to make sure we
401 * have the page pinned if the pipe page originates from the
404 ptr = buf->ops->map(file, info, buf);
408 offset = pos & ~PAGE_CACHE_MASK;
409 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
411 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
413 buf->ops->unmap(info, buf);
421 * This is a little more tricky than the file -> pipe splicing. There are
422 * basically three cases:
424 * - Destination page already exists in the address space and there
425 * are users of it. For that case we have no other option that
426 * copying the data. Tough luck.
427 * - Destination page already exists in the address space, but there
428 * are no users of it. Make sure it's uptodate, then drop it. Fall
429 * through to last case.
430 * - Destination page does not exist, we can add the pipe page to
431 * the page cache and avoid the copy.
433 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
434 * sd->flags), we attempt to migrate pages from the pipe to the output
435 * file address space page cache. This is possible if no one else has
436 * the pipe page referenced outside of the pipe and page cache. If
437 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
438 * a new page in the output file page cache and fill/dirty that.
440 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
441 struct splice_desc *sd)
443 struct file *file = sd->file;
444 struct address_space *mapping = file->f_mapping;
445 gfp_t gfp_mask = mapping_gfp_mask(mapping);
453 * make sure the data in this buffer is uptodate
455 src = buf->ops->map(file, info, buf);
459 index = sd->pos >> PAGE_CACHE_SHIFT;
460 offset = sd->pos & ~PAGE_CACHE_MASK;
463 * Reuse buf page, if SPLICE_F_MOVE is set.
465 if (sd->flags & SPLICE_F_MOVE) {
467 * If steal succeeds, buf->page is now pruned from the vm
468 * side (LRU and page cache) and we can reuse it.
470 if (buf->ops->steal(info, buf))
474 * this will also set the page locked
477 if (add_to_page_cache(page, mapping, index, gfp_mask))
480 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
485 page = find_or_create_page(mapping, index, gfp_mask);
490 * If the page is uptodate, it is also locked. If it isn't
491 * uptodate, we can mark it uptodate if we are filling the
492 * full page. Otherwise we need to read it in first...
494 if (!PageUptodate(page)) {
495 if (sd->len < PAGE_CACHE_SIZE) {
496 ret = mapping->a_ops->readpage(file, page);
502 if (!PageUptodate(page)) {
504 * Page got invalidated, repeat.
506 if (!page->mapping) {
508 page_cache_release(page);
515 WARN_ON(!PageLocked(page));
516 SetPageUptodate(page);
521 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
522 if (ret == AOP_TRUNCATED_PAGE) {
523 page_cache_release(page);
528 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
529 char *dst = kmap_atomic(page, KM_USER0);
531 memcpy(dst + offset, src + buf->offset, sd->len);
532 flush_dcache_page(page);
533 kunmap_atomic(dst, KM_USER0);
536 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
537 if (ret == AOP_TRUNCATED_PAGE) {
538 page_cache_release(page);
543 mark_page_accessed(page);
544 balance_dirty_pages_ratelimited(mapping);
546 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
547 page_cache_release(page);
551 buf->ops->unmap(info, buf);
555 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
556 struct splice_desc *);
559 * Pipe input worker. Most of this logic works like a regular pipe, the
560 * key here is the 'actor' worker passed in that actually moves the data
561 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
563 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
564 size_t len, unsigned int flags,
567 int ret, do_wakeup, err;
568 struct splice_desc sd;
579 mutex_lock(&pipe->inode->i_mutex);
583 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
584 struct pipe_buf_operations *ops = buf->ops;
587 if (sd.len > sd.total_len)
588 sd.len = sd.total_len;
590 err = actor(pipe, buf, &sd);
592 if (!ret && err != -ENODATA)
599 buf->offset += sd.len;
604 ops->release(pipe, buf);
605 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
612 sd.total_len -= sd.len;
621 if (!pipe->waiting_writers) {
626 if (flags & SPLICE_F_NONBLOCK) {
632 if (signal_pending(current)) {
640 if (waitqueue_active(&pipe->wait))
641 wake_up_interruptible_sync(&pipe->wait);
642 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
650 mutex_unlock(&pipe->inode->i_mutex);
654 if (waitqueue_active(&pipe->wait))
655 wake_up_interruptible(&pipe->wait);
656 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
665 * generic_file_splice_write - splice data from a pipe to a file
667 * @out: file to write to
668 * @len: number of bytes to splice
669 * @flags: splice modifier flags
671 * Will either move or copy pages (determined by @flags options) from
672 * the given pipe inode to the given file.
676 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
677 size_t len, unsigned int flags)
679 struct address_space *mapping = out->f_mapping;
682 ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
685 * If file or inode is SYNC and we actually wrote some data, sync it.
687 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
689 struct inode *inode = mapping->host;
692 mutex_lock(&inode->i_mutex);
693 err = generic_osync_inode(mapping->host, mapping,
694 OSYNC_METADATA|OSYNC_DATA);
695 mutex_unlock(&inode->i_mutex);
704 EXPORT_SYMBOL(generic_file_splice_write);
707 * generic_splice_sendpage - splice data from a pipe to a socket
709 * @out: socket to write to
710 * @len: number of bytes to splice
711 * @flags: splice modifier flags
713 * Will send @len bytes from the pipe to a network socket. No data copying
717 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
718 size_t len, unsigned int flags)
720 return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
723 EXPORT_SYMBOL(generic_splice_sendpage);
726 * Attempt to initiate a splice from pipe to file.
728 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
729 size_t len, unsigned int flags)
734 if (unlikely(!out->f_op || !out->f_op->splice_write))
737 if (unlikely(!(out->f_mode & FMODE_WRITE)))
742 ret = rw_verify_area(WRITE, out, &pos, len);
743 if (unlikely(ret < 0))
746 return out->f_op->splice_write(pipe, out, len, flags);
750 * Attempt to initiate a splice from a file to a pipe.
752 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
753 size_t len, unsigned int flags)
755 loff_t pos, isize, left;
758 if (unlikely(!in->f_op || !in->f_op->splice_read))
761 if (unlikely(!(in->f_mode & FMODE_READ)))
766 ret = rw_verify_area(READ, in, &pos, len);
767 if (unlikely(ret < 0))
770 isize = i_size_read(in->f_mapping->host);
771 if (unlikely(in->f_pos >= isize))
774 left = isize - in->f_pos;
775 if (unlikely(left < len))
778 return in->f_op->splice_read(in, pipe, len, flags);
781 long do_splice_direct(struct file *in, struct file *out, size_t len,
784 struct pipe_inode_info *pipe;
790 * We require the input being a regular file, as we don't want to
791 * randomly drop data for eg socket -> socket splicing. Use the
792 * piped splicing for that!
794 i_mode = in->f_dentry->d_inode->i_mode;
795 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
799 * neither in nor out is a pipe, setup an internal pipe attached to
800 * 'out' and transfer the wanted data from 'in' to 'out' through that
802 pipe = current->splice_pipe;
803 if (unlikely(!pipe)) {
804 pipe = alloc_pipe_info(NULL);
809 * We don't have an immediate reader, but we'll read the stuff
810 * out of the pipe right after the move_to_pipe(). So set
811 * PIPE_READERS appropriately.
815 current->splice_pipe = pipe;
825 size_t read_len, max_read_len;
828 * Do at most PIPE_BUFFERS pages worth of transfer:
830 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
832 ret = do_splice_to(in, pipe, max_read_len, flags);
833 if (unlikely(ret < 0))
839 * NOTE: nonblocking mode only applies to the input. We
840 * must not do the output in nonblocking mode as then we
841 * could get stuck data in the internal pipe:
843 ret = do_splice_from(pipe, out, read_len,
844 flags & ~SPLICE_F_NONBLOCK);
845 if (unlikely(ret < 0))
852 * In nonblocking mode, if we got back a short read then
853 * that was due to either an IO error or due to the
854 * pagecache entry not being there. In the IO error case
855 * the _next_ splice attempt will produce a clean IO error
856 * return value (not a short read), so in both cases it's
857 * correct to break out of the loop here:
859 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
863 pipe->nrbufs = pipe->curbuf = 0;
869 * If we did an incomplete transfer we must release
870 * the pipe buffers in question:
872 for (i = 0; i < PIPE_BUFFERS; i++) {
873 struct pipe_buffer *buf = pipe->bufs + i;
876 buf->ops->release(pipe, buf);
880 pipe->nrbufs = pipe->curbuf = 0;
883 * If we transferred some data, return the number of bytes:
891 EXPORT_SYMBOL(do_splice_direct);
894 * Determine where to splice to/from.
896 static long do_splice(struct file *in, loff_t __user *off_in,
897 struct file *out, loff_t __user *off_out,
898 size_t len, unsigned int flags)
900 struct pipe_inode_info *pipe;
902 pipe = in->f_dentry->d_inode->i_pipe;
907 if (out->f_op->llseek == no_llseek)
909 if (copy_from_user(&out->f_pos, off_out,
914 return do_splice_from(pipe, out, len, flags);
917 pipe = out->f_dentry->d_inode->i_pipe;
922 if (in->f_op->llseek == no_llseek)
924 if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
928 return do_splice_to(in, pipe, len, flags);
934 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
935 int fd_out, loff_t __user *off_out,
936 size_t len, unsigned int flags)
939 struct file *in, *out;
940 int fput_in, fput_out;
946 in = fget_light(fd_in, &fput_in);
948 if (in->f_mode & FMODE_READ) {
949 out = fget_light(fd_out, &fput_out);
951 if (out->f_mode & FMODE_WRITE)
952 error = do_splice(in, off_in,
955 fput_light(out, fput_out);
959 fput_light(in, fput_in);