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 and fixing the initial implementation
15 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
16 * Copyright (C) 2005 Linus Torvalds <torvalds@osdl.org>
20 #include <linux/file.h>
21 #include <linux/pagemap.h>
22 #include <linux/pipe_fs_i.h>
23 #include <linux/mm_inline.h>
24 #include <linux/swap.h>
25 #include <linux/writeback.h>
26 #include <linux/buffer_head.h>
27 #include <linux/module.h>
28 #include <linux/syscalls.h>
31 * Passed to the actors
34 unsigned int len, total_len; /* current and remaining length */
35 unsigned int flags; /* splice flags */
36 struct file *file; /* file to read/write */
37 loff_t pos; /* file position */
41 * Attempt to steal a page from a pipe buffer. This should perhaps go into
42 * a vm helper function, it's already simplified quite a bit by the
43 * addition of remove_mapping(). If success is returned, the caller may
44 * attempt to reuse this page for another destination.
46 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
47 struct pipe_buffer *buf)
49 struct page *page = buf->page;
50 struct address_space *mapping = page_mapping(page);
52 WARN_ON(!PageLocked(page));
53 WARN_ON(!PageUptodate(page));
56 * At least for ext2 with nobh option, we need to wait on writeback
57 * completing on this page, since we'll remove it from the pagecache.
58 * Otherwise truncate wont wait on the page, allowing the disk
59 * blocks to be reused by someone else before we actually wrote our
60 * data to them. fs corruption ensues.
62 wait_on_page_writeback(page);
64 if (PagePrivate(page))
65 try_to_release_page(page, mapping_gfp_mask(mapping));
67 if (!remove_mapping(mapping, page))
70 buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
74 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
75 struct pipe_buffer *buf)
77 page_cache_release(buf->page);
79 buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
82 static void *page_cache_pipe_buf_map(struct file *file,
83 struct pipe_inode_info *info,
84 struct pipe_buffer *buf)
86 struct page *page = buf->page;
89 if (!PageUptodate(page)) {
93 * Page got truncated/unhashed. This will cause a 0-byte
94 * splice, if this is the first page
102 * uh oh, read-error from disk
104 if (!PageUptodate(page)) {
110 * page is ok afterall, fall through to mapping
121 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
122 struct pipe_buffer *buf)
127 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
129 .map = page_cache_pipe_buf_map,
130 .unmap = page_cache_pipe_buf_unmap,
131 .release = page_cache_pipe_buf_release,
132 .steal = page_cache_pipe_buf_steal,
136 * Pipe output worker. This sets up our pipe format with the page cache
137 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
139 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
140 int nr_pages, unsigned long offset,
141 unsigned long len, unsigned int flags)
143 int ret, do_wakeup, i;
150 mutex_lock(&pipe->inode->i_mutex);
153 if (!pipe->readers) {
154 send_sig(SIGPIPE, current, 0);
160 if (pipe->nrbufs < PIPE_BUFFERS) {
161 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
162 struct pipe_buffer *buf = pipe->bufs + newbuf;
163 struct page *page = pages[i++];
164 unsigned long this_len;
166 this_len = PAGE_CACHE_SIZE - offset;
171 buf->offset = offset;
173 buf->ops = &page_cache_pipe_buf_ops;
185 if (pipe->nrbufs < PIPE_BUFFERS)
191 if (flags & SPLICE_F_NONBLOCK) {
197 if (signal_pending(current)) {
205 if (waitqueue_active(&pipe->wait))
206 wake_up_interruptible_sync(&pipe->wait);
207 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
211 pipe->waiting_writers++;
213 pipe->waiting_writers--;
217 mutex_unlock(&pipe->inode->i_mutex);
221 if (waitqueue_active(&pipe->wait))
222 wake_up_interruptible(&pipe->wait);
223 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
227 page_cache_release(pages[i++]);
233 __generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
234 size_t len, unsigned int flags)
236 struct address_space *mapping = in->f_mapping;
237 unsigned int offset, nr_pages;
238 struct page *pages[PIPE_BUFFERS];
243 index = in->f_pos >> PAGE_CACHE_SHIFT;
244 offset = in->f_pos & ~PAGE_CACHE_MASK;
245 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
247 if (nr_pages > PIPE_BUFFERS)
248 nr_pages = PIPE_BUFFERS;
251 * initiate read-ahead on this page range. however, don't call into
252 * read-ahead if this is a non-zero offset (we are likely doing small
253 * chunk splice and the page is already there) for a single page.
255 if (!offset || nr_pages > 1)
256 do_page_cache_readahead(mapping, in, index, nr_pages);
259 * now fill in the holes
262 for (i = 0; i < nr_pages; i++, index++) {
265 * lookup the page for this index
267 page = find_get_page(mapping, index);
270 * If in nonblock mode then dont block on
271 * readpage (we've kicked readahead so there
272 * will be asynchronous progress):
274 if (flags & SPLICE_F_NONBLOCK)
278 * page didn't exist, allocate one
280 page = page_cache_alloc_cold(mapping);
284 error = add_to_page_cache_lru(page, mapping, index,
285 mapping_gfp_mask(mapping));
286 if (unlikely(error)) {
287 page_cache_release(page);
295 * If the page isn't uptodate, we may need to start io on it
297 if (!PageUptodate(page)) {
301 * page was truncated, stop here. if this isn't the
302 * first page, we'll just complete what we already
305 if (!page->mapping) {
307 page_cache_release(page);
311 * page was already under io and is now done, great
313 if (PageUptodate(page)) {
320 * need to read in the page
322 error = mapping->a_ops->readpage(in, page);
324 if (unlikely(error)) {
325 page_cache_release(page);
326 if (error == AOP_TRUNCATED_PAGE)
336 return move_to_pipe(pipe, pages, i, offset, len, flags);
342 * generic_file_splice_read - splice data from file to a pipe
343 * @in: file to splice from
344 * @pipe: pipe to splice to
345 * @len: number of bytes to splice
346 * @flags: splice modifier flags
348 * Will read pages from given file and fill them into a pipe.
350 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
351 size_t len, unsigned int flags)
360 ret = __generic_file_splice_read(in, pipe, len, flags);
369 if (!(flags & SPLICE_F_NONBLOCK))
381 EXPORT_SYMBOL(generic_file_splice_read);
384 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
387 static int pipe_to_sendpage(struct pipe_inode_info *info,
388 struct pipe_buffer *buf, struct splice_desc *sd)
390 struct file *file = sd->file;
391 loff_t pos = sd->pos;
398 * sub-optimal, but we are limited by the pipe ->map. we don't
399 * need a kmap'ed buffer here, we just want to make sure we
400 * have the page pinned if the pipe page originates from the
403 ptr = buf->ops->map(file, info, buf);
407 offset = pos & ~PAGE_CACHE_MASK;
408 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
410 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
412 buf->ops->unmap(info, buf);
420 * This is a little more tricky than the file -> pipe splicing. There are
421 * basically three cases:
423 * - Destination page already exists in the address space and there
424 * are users of it. For that case we have no other option that
425 * copying the data. Tough luck.
426 * - Destination page already exists in the address space, but there
427 * are no users of it. Make sure it's uptodate, then drop it. Fall
428 * through to last case.
429 * - Destination page does not exist, we can add the pipe page to
430 * the page cache and avoid the copy.
432 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
433 * sd->flags), we attempt to migrate pages from the pipe to the output
434 * file address space page cache. This is possible if no one else has
435 * the pipe page referenced outside of the pipe and page cache. If
436 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
437 * a new page in the output file page cache and fill/dirty that.
439 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
440 struct splice_desc *sd)
442 struct file *file = sd->file;
443 struct address_space *mapping = file->f_mapping;
444 gfp_t gfp_mask = mapping_gfp_mask(mapping);
452 * make sure the data in this buffer is uptodate
454 src = buf->ops->map(file, info, buf);
458 index = sd->pos >> PAGE_CACHE_SHIFT;
459 offset = sd->pos & ~PAGE_CACHE_MASK;
462 * reuse buf page, if SPLICE_F_MOVE is set
464 if (sd->flags & SPLICE_F_MOVE) {
466 * If steal succeeds, buf->page is now pruned from the vm
467 * side (LRU and page cache) and we can reuse it.
469 if (buf->ops->steal(info, buf))
473 * this will also set the page locked
476 if (add_to_page_cache(page, mapping, index, gfp_mask))
479 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
484 page = find_or_create_page(mapping, index, gfp_mask);
489 * If the page is uptodate, it is also locked. If it isn't
490 * uptodate, we can mark it uptodate if we are filling the
491 * full page. Otherwise we need to read it in first...
493 if (!PageUptodate(page)) {
494 if (sd->len < PAGE_CACHE_SIZE) {
495 ret = mapping->a_ops->readpage(file, page);
501 if (!PageUptodate(page)) {
503 * page got invalidated, repeat
505 if (!page->mapping) {
507 page_cache_release(page);
514 WARN_ON(!PageLocked(page));
515 SetPageUptodate(page);
520 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
521 if (ret == AOP_TRUNCATED_PAGE) {
522 page_cache_release(page);
527 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
528 char *dst = kmap_atomic(page, KM_USER0);
530 memcpy(dst + offset, src + buf->offset, sd->len);
531 flush_dcache_page(page);
532 kunmap_atomic(dst, KM_USER0);
535 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
536 if (ret == AOP_TRUNCATED_PAGE) {
537 page_cache_release(page);
542 mark_page_accessed(page);
543 balance_dirty_pages_ratelimited(mapping);
545 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
546 page_cache_release(page);
550 buf->ops->unmap(info, buf);
554 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
555 struct splice_desc *);
558 * Pipe input worker. Most of this logic works like a regular pipe, the
559 * key here is the 'actor' worker passed in that actually moves the data
560 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
562 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
563 size_t len, unsigned int flags,
566 int ret, do_wakeup, err;
567 struct splice_desc sd;
578 mutex_lock(&pipe->inode->i_mutex);
582 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
583 struct pipe_buf_operations *ops = buf->ops;
586 if (sd.len > sd.total_len)
587 sd.len = sd.total_len;
589 err = actor(pipe, buf, &sd);
591 if (!ret && err != -ENODATA)
598 buf->offset += sd.len;
602 ops->release(pipe, buf);
603 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
610 sd.total_len -= sd.len;
619 if (!pipe->waiting_writers) {
624 if (flags & SPLICE_F_NONBLOCK) {
630 if (signal_pending(current)) {
638 if (waitqueue_active(&pipe->wait))
639 wake_up_interruptible_sync(&pipe->wait);
640 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
648 mutex_unlock(&pipe->inode->i_mutex);
652 if (waitqueue_active(&pipe->wait))
653 wake_up_interruptible(&pipe->wait);
654 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
663 * generic_file_splice_write - splice data from a pipe to a file
665 * @out: file to write to
666 * @len: number of bytes to splice
667 * @flags: splice modifier flags
669 * Will either move or copy pages (determined by @flags options) from
670 * the given pipe inode to the given file.
674 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
675 size_t len, unsigned int flags)
677 struct address_space *mapping = out->f_mapping;
680 ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
683 * if file or inode is SYNC and we actually wrote some data, sync it
685 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
687 struct inode *inode = mapping->host;
690 mutex_lock(&inode->i_mutex);
691 err = generic_osync_inode(mapping->host, mapping,
692 OSYNC_METADATA|OSYNC_DATA);
693 mutex_unlock(&inode->i_mutex);
702 EXPORT_SYMBOL(generic_file_splice_write);
705 * generic_splice_sendpage - splice data from a pipe to a socket
707 * @out: socket to write to
708 * @len: number of bytes to splice
709 * @flags: splice modifier flags
711 * Will send @len bytes from the pipe to a network socket. No data copying
715 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
716 size_t len, unsigned int flags)
718 return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
721 EXPORT_SYMBOL(generic_splice_sendpage);
724 * Attempt to initiate a splice from pipe to file.
726 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
727 size_t len, unsigned int flags)
732 if (unlikely(!out->f_op || !out->f_op->splice_write))
735 if (unlikely(!(out->f_mode & FMODE_WRITE)))
740 ret = rw_verify_area(WRITE, out, &pos, len);
741 if (unlikely(ret < 0))
744 return out->f_op->splice_write(pipe, out, len, flags);
748 * Attempt to initiate a splice from a file to a pipe.
750 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
751 size_t len, unsigned int flags)
753 loff_t pos, isize, left;
756 if (unlikely(!in->f_op || !in->f_op->splice_read))
759 if (unlikely(!(in->f_mode & FMODE_READ)))
764 ret = rw_verify_area(READ, in, &pos, len);
765 if (unlikely(ret < 0))
768 isize = i_size_read(in->f_mapping->host);
769 if (unlikely(in->f_pos >= isize))
772 left = isize - in->f_pos;
773 if (unlikely(left < len))
776 return in->f_op->splice_read(in, pipe, len, flags);
779 long do_splice_direct(struct file *in, struct file *out, size_t len,
782 struct pipe_inode_info *pipe;
788 * We require the input being a regular file, as we don't want to
789 * randomly drop data for eg socket -> socket splicing. Use the
790 * piped splicing for that!
792 i_mode = in->f_dentry->d_inode->i_mode;
793 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
797 * neither in nor out is a pipe, setup an internal pipe attached to
798 * 'out' and transfer the wanted data from 'in' to 'out' through that
800 pipe = current->splice_pipe;
801 if (unlikely(!pipe)) {
802 pipe = alloc_pipe_info(NULL);
807 * We don't have an immediate reader, but we'll read the stuff
808 * out of the pipe right after the move_to_pipe(). So set
809 * PIPE_READERS appropriately.
813 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, 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, 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;
900 pipe = in->f_dentry->d_inode->i_pipe;
905 if (out->f_op->llseek == no_llseek)
907 if (copy_from_user(&out->f_pos, off_out,
912 return do_splice_from(pipe, out, len, flags);
915 pipe = out->f_dentry->d_inode->i_pipe;
920 if (in->f_op->llseek == no_llseek)
922 if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
926 return do_splice_to(in, pipe, len, flags);
932 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
933 int fd_out, loff_t __user *off_out,
934 size_t len, unsigned int flags)
937 struct file *in, *out;
938 int fput_in, fput_out;
944 in = fget_light(fd_in, &fput_in);
946 if (in->f_mode & FMODE_READ) {
947 out = fget_light(fd_out, &fput_out);
949 if (out->f_mode & FMODE_WRITE)
950 error = do_splice(in, off_in,
953 fput_light(out, fput_out);
957 fput_light(in, fput_in);