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, loff_t *ppos,
235 struct pipe_inode_info *pipe, size_t len,
238 struct address_space *mapping = in->f_mapping;
239 unsigned int offset, nr_pages;
240 struct page *pages[PIPE_BUFFERS];
245 index = *ppos >> PAGE_CACHE_SHIFT;
246 offset = *ppos & ~PAGE_CACHE_MASK;
247 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
249 if (nr_pages > PIPE_BUFFERS)
250 nr_pages = PIPE_BUFFERS;
253 * Initiate read-ahead on this page range. however, don't call into
254 * read-ahead if this is a non-zero offset (we are likely doing small
255 * chunk splice and the page is already there) for a single page.
257 if (!offset || nr_pages > 1)
258 do_page_cache_readahead(mapping, in, index, nr_pages);
261 * Now fill in the holes:
264 for (i = 0; i < nr_pages; i++, index++) {
267 * lookup the page for this index
269 page = find_get_page(mapping, index);
272 * If in nonblock mode then dont block on
273 * readpage (we've kicked readahead so there
274 * will be asynchronous progress):
276 if (flags & SPLICE_F_NONBLOCK)
280 * page didn't exist, allocate one
282 page = page_cache_alloc_cold(mapping);
286 error = add_to_page_cache_lru(page, mapping, index,
287 mapping_gfp_mask(mapping));
288 if (unlikely(error)) {
289 page_cache_release(page);
297 * If the page isn't uptodate, we may need to start io on it
299 if (!PageUptodate(page)) {
303 * page was truncated, stop here. if this isn't the
304 * first page, we'll just complete what we already
307 if (!page->mapping) {
309 page_cache_release(page);
313 * page was already under io and is now done, great
315 if (PageUptodate(page)) {
322 * need to read in the page
324 error = mapping->a_ops->readpage(in, page);
326 if (unlikely(error)) {
327 page_cache_release(page);
328 if (error == AOP_TRUNCATED_PAGE)
338 return move_to_pipe(pipe, pages, i, offset, len, flags);
344 * generic_file_splice_read - splice data from file to a pipe
345 * @in: file to splice from
346 * @pipe: pipe to splice to
347 * @len: number of bytes to splice
348 * @flags: splice modifier flags
350 * Will read pages from given file and fill them into a pipe.
352 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
353 struct pipe_inode_info *pipe, size_t len,
363 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
372 if (!(flags & SPLICE_F_NONBLOCK))
384 EXPORT_SYMBOL(generic_file_splice_read);
387 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
390 static int pipe_to_sendpage(struct pipe_inode_info *info,
391 struct pipe_buffer *buf, struct splice_desc *sd)
393 struct file *file = sd->file;
394 loff_t pos = sd->pos;
401 * Sub-optimal, but we are limited by the pipe ->map. We don't
402 * need a kmap'ed buffer here, we just want to make sure we
403 * have the page pinned if the pipe page originates from the
406 ptr = buf->ops->map(file, info, buf);
410 offset = pos & ~PAGE_CACHE_MASK;
411 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
413 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
415 buf->ops->unmap(info, buf);
423 * This is a little more tricky than the file -> pipe splicing. There are
424 * basically three cases:
426 * - Destination page already exists in the address space and there
427 * are users of it. For that case we have no other option that
428 * copying the data. Tough luck.
429 * - Destination page already exists in the address space, but there
430 * are no users of it. Make sure it's uptodate, then drop it. Fall
431 * through to last case.
432 * - Destination page does not exist, we can add the pipe page to
433 * the page cache and avoid the copy.
435 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
436 * sd->flags), we attempt to migrate pages from the pipe to the output
437 * file address space page cache. This is possible if no one else has
438 * the pipe page referenced outside of the pipe and page cache. If
439 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
440 * a new page in the output file page cache and fill/dirty that.
442 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
443 struct splice_desc *sd)
445 struct file *file = sd->file;
446 struct address_space *mapping = file->f_mapping;
447 gfp_t gfp_mask = mapping_gfp_mask(mapping);
455 * make sure the data in this buffer is uptodate
457 src = buf->ops->map(file, info, buf);
461 index = sd->pos >> PAGE_CACHE_SHIFT;
462 offset = sd->pos & ~PAGE_CACHE_MASK;
465 * Reuse buf page, if SPLICE_F_MOVE is set.
467 if (sd->flags & SPLICE_F_MOVE) {
469 * If steal succeeds, buf->page is now pruned from the vm
470 * side (LRU and page cache) and we can reuse it.
472 if (buf->ops->steal(info, buf))
476 * this will also set the page locked
479 if (add_to_page_cache(page, mapping, index, gfp_mask))
482 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
487 page = find_or_create_page(mapping, index, gfp_mask);
492 * If the page is uptodate, it is also locked. If it isn't
493 * uptodate, we can mark it uptodate if we are filling the
494 * full page. Otherwise we need to read it in first...
496 if (!PageUptodate(page)) {
497 if (sd->len < PAGE_CACHE_SIZE) {
498 ret = mapping->a_ops->readpage(file, page);
504 if (!PageUptodate(page)) {
506 * Page got invalidated, repeat.
508 if (!page->mapping) {
510 page_cache_release(page);
517 WARN_ON(!PageLocked(page));
518 SetPageUptodate(page);
523 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
524 if (ret == AOP_TRUNCATED_PAGE) {
525 page_cache_release(page);
530 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
531 char *dst = kmap_atomic(page, KM_USER0);
533 memcpy(dst + offset, src + buf->offset, sd->len);
534 flush_dcache_page(page);
535 kunmap_atomic(dst, KM_USER0);
538 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
539 if (ret == AOP_TRUNCATED_PAGE) {
540 page_cache_release(page);
545 mark_page_accessed(page);
546 balance_dirty_pages_ratelimited(mapping);
548 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
549 page_cache_release(page);
553 buf->ops->unmap(info, buf);
557 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
558 struct splice_desc *);
561 * Pipe input worker. Most of this logic works like a regular pipe, the
562 * key here is the 'actor' worker passed in that actually moves the data
563 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
565 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
566 loff_t *ppos, size_t len, unsigned int flags,
569 int ret, do_wakeup, err;
570 struct splice_desc sd;
581 mutex_lock(&pipe->inode->i_mutex);
585 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
586 struct pipe_buf_operations *ops = buf->ops;
589 if (sd.len > sd.total_len)
590 sd.len = sd.total_len;
592 err = actor(pipe, buf, &sd);
594 if (!ret && err != -ENODATA)
601 buf->offset += sd.len;
606 ops->release(pipe, buf);
607 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
614 sd.total_len -= sd.len;
623 if (!pipe->waiting_writers) {
628 if (flags & SPLICE_F_NONBLOCK) {
634 if (signal_pending(current)) {
642 if (waitqueue_active(&pipe->wait))
643 wake_up_interruptible_sync(&pipe->wait);
644 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
652 mutex_unlock(&pipe->inode->i_mutex);
656 if (waitqueue_active(&pipe->wait))
657 wake_up_interruptible(&pipe->wait);
658 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 loff_t *ppos, size_t len, unsigned int flags)
679 struct address_space *mapping = out->f_mapping;
682 ret = move_from_pipe(pipe, out, ppos, 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 loff_t *ppos, size_t len, unsigned int flags)
720 return move_from_pipe(pipe, out, ppos, 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 loff_t *ppos, size_t len, unsigned int flags)
733 if (unlikely(!out->f_op || !out->f_op->splice_write))
736 if (unlikely(!(out->f_mode & FMODE_WRITE)))
739 ret = rw_verify_area(WRITE, out, ppos, len);
740 if (unlikely(ret < 0))
743 return out->f_op->splice_write(pipe, out, ppos, len, flags);
747 * Attempt to initiate a splice from a file to a pipe.
749 static long do_splice_to(struct file *in, loff_t *ppos,
750 struct pipe_inode_info *pipe, size_t len,
756 if (unlikely(!in->f_op || !in->f_op->splice_read))
759 if (unlikely(!(in->f_mode & FMODE_READ)))
762 ret = rw_verify_area(READ, in, ppos, len);
763 if (unlikely(ret < 0))
766 isize = i_size_read(in->f_mapping->host);
767 if (unlikely(*ppos >= isize))
770 left = isize - *ppos;
771 if (unlikely(left < len))
774 return in->f_op->splice_read(in, ppos, pipe, len, flags);
777 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
778 size_t len, unsigned int flags)
780 struct pipe_inode_info *pipe;
787 * We require the input being a regular file, as we don't want to
788 * randomly drop data for eg socket -> socket splicing. Use the
789 * piped splicing for that!
791 i_mode = in->f_dentry->d_inode->i_mode;
792 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
796 * neither in nor out is a pipe, setup an internal pipe attached to
797 * 'out' and transfer the wanted data from 'in' to 'out' through that
799 pipe = current->splice_pipe;
800 if (unlikely(!pipe)) {
801 pipe = alloc_pipe_info(NULL);
806 * We don't have an immediate reader, but we'll read the stuff
807 * out of the pipe right after the move_to_pipe(). So set
808 * PIPE_READERS appropriately.
812 current->splice_pipe = pipe;
823 size_t read_len, max_read_len;
826 * Do at most PIPE_BUFFERS pages worth of transfer:
828 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
830 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
831 if (unlikely(ret < 0))
837 * NOTE: nonblocking mode only applies to the input. We
838 * must not do the output in nonblocking mode as then we
839 * could get stuck data in the internal pipe:
841 ret = do_splice_from(pipe, out, &out_off, read_len,
842 flags & ~SPLICE_F_NONBLOCK);
843 if (unlikely(ret < 0))
850 * In nonblocking mode, if we got back a short read then
851 * that was due to either an IO error or due to the
852 * pagecache entry not being there. In the IO error case
853 * the _next_ splice attempt will produce a clean IO error
854 * return value (not a short read), so in both cases it's
855 * correct to break out of the loop here:
857 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
861 pipe->nrbufs = pipe->curbuf = 0;
867 * If we did an incomplete transfer we must release
868 * the pipe buffers in question:
870 for (i = 0; i < PIPE_BUFFERS; i++) {
871 struct pipe_buffer *buf = pipe->bufs + i;
874 buf->ops->release(pipe, buf);
878 pipe->nrbufs = pipe->curbuf = 0;
881 * If we transferred some data, return the number of bytes:
889 EXPORT_SYMBOL(do_splice_direct);
892 * Determine where to splice to/from.
894 static long do_splice(struct file *in, loff_t __user *off_in,
895 struct file *out, loff_t __user *off_out,
896 size_t len, unsigned int flags)
898 struct pipe_inode_info *pipe;
901 pipe = in->f_dentry->d_inode->i_pipe;
906 if (out->f_op->llseek == no_llseek)
908 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
914 return do_splice_from(pipe, out, off, len, flags);
917 pipe = out->f_dentry->d_inode->i_pipe;
922 if (in->f_op->llseek == no_llseek)
924 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
930 return do_splice_to(in, off, pipe, len, flags);
936 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
937 int fd_out, loff_t __user *off_out,
938 size_t len, unsigned int flags)
941 struct file *in, *out;
942 int fput_in, fput_out;
948 in = fget_light(fd_in, &fput_in);
950 if (in->f_mode & FMODE_READ) {
951 out = fget_light(fd_out, &fput_out);
953 if (out->f_mode & FMODE_WRITE)
954 error = do_splice(in, off_in,
957 fput_light(out, fput_out);
961 fput_light(in, fput_in);