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);
155 if (!pipe->readers) {
156 send_sig(SIGPIPE, current, 0);
163 if (bufs < PIPE_BUFFERS) {
164 int newbuf = (pipe->curbuf + bufs) & (PIPE_BUFFERS - 1);
165 struct pipe_buffer *buf = pipe->bufs + newbuf;
166 struct page *page = pages[i++];
167 unsigned long this_len;
169 this_len = PAGE_CACHE_SIZE - offset;
174 buf->offset = offset;
176 buf->ops = &page_cache_pipe_buf_ops;
177 pipe->nrbufs = ++bufs;
187 if (bufs < PIPE_BUFFERS)
193 if (flags & SPLICE_F_NONBLOCK) {
199 if (signal_pending(current)) {
207 if (waitqueue_active(&pipe->wait))
208 wake_up_interruptible_sync(&pipe->wait);
209 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
213 pipe->waiting_writers++;
215 pipe->waiting_writers--;
219 mutex_unlock(&pipe->inode->i_mutex);
223 if (waitqueue_active(&pipe->wait))
224 wake_up_interruptible(&pipe->wait);
225 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
229 page_cache_release(pages[i++]);
235 __generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
236 size_t len, unsigned int flags)
238 struct address_space *mapping = in->f_mapping;
239 unsigned int offset, nr_pages;
240 struct page *pages[PIPE_BUFFERS];
245 index = in->f_pos >> PAGE_CACHE_SHIFT;
246 offset = in->f_pos & ~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.
353 ssize_t generic_file_splice_read(struct file *in, struct pipe_inode_info *pipe,
354 size_t len, unsigned int flags)
363 ret = __generic_file_splice_read(in, 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 size_t len, unsigned int flags,
569 int ret, do_wakeup, err;
570 struct splice_desc sd;
581 mutex_lock(&pipe->inode->i_mutex);
584 int bufs = pipe->nrbufs;
587 int curbuf = pipe->curbuf;
588 struct pipe_buffer *buf = pipe->bufs + curbuf;
589 struct pipe_buf_operations *ops = buf->ops;
592 if (sd.len > sd.total_len)
593 sd.len = sd.total_len;
595 err = actor(pipe, buf, &sd);
597 if (!ret && err != -ENODATA)
604 buf->offset += sd.len;
608 ops->release(pipe, buf);
609 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
610 pipe->curbuf = curbuf;
611 pipe->nrbufs = --bufs;
616 sd.total_len -= sd.len;
625 if (!pipe->waiting_writers) {
630 if (flags & SPLICE_F_NONBLOCK) {
636 if (signal_pending(current)) {
644 if (waitqueue_active(&pipe->wait))
645 wake_up_interruptible_sync(&pipe->wait);
646 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
654 mutex_unlock(&pipe->inode->i_mutex);
658 if (waitqueue_active(&pipe->wait))
659 wake_up_interruptible(&pipe->wait);
660 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
663 mutex_lock(&out->f_mapping->host->i_mutex);
665 mutex_unlock(&out->f_mapping->host->i_mutex);
671 * generic_file_splice_write - splice data from a pipe to a file
673 * @out: file to write to
674 * @len: number of bytes to splice
675 * @flags: splice modifier flags
677 * Will either move or copy pages (determined by @flags options) from
678 * the given pipe inode to the given file.
682 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
683 size_t len, unsigned int flags)
685 struct address_space *mapping = out->f_mapping;
688 ret = move_from_pipe(pipe, out, len, flags, pipe_to_file);
691 * if file or inode is SYNC and we actually wrote some data, sync it
693 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
695 struct inode *inode = mapping->host;
698 mutex_lock(&inode->i_mutex);
699 err = generic_osync_inode(mapping->host, mapping,
700 OSYNC_METADATA|OSYNC_DATA);
701 mutex_unlock(&inode->i_mutex);
710 EXPORT_SYMBOL(generic_file_splice_write);
713 * generic_splice_sendpage - splice data from a pipe to a socket
715 * @out: socket to write to
716 * @len: number of bytes to splice
717 * @flags: splice modifier flags
719 * Will send @len bytes from the pipe to a network socket. No data copying
723 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
724 size_t len, unsigned int flags)
726 return move_from_pipe(pipe, out, len, flags, pipe_to_sendpage);
729 EXPORT_SYMBOL(generic_splice_sendpage);
732 * Attempt to initiate a splice from pipe to file.
734 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
735 size_t len, unsigned int flags)
740 if (!out->f_op || !out->f_op->splice_write)
743 if (!(out->f_mode & FMODE_WRITE))
748 ret = rw_verify_area(WRITE, out, &pos, len);
749 if (unlikely(ret < 0))
752 return out->f_op->splice_write(pipe, out, len, flags);
756 * Attempt to initiate a splice from a file to a pipe.
758 static long do_splice_to(struct file *in, struct pipe_inode_info *pipe,
759 size_t len, unsigned int flags)
761 loff_t pos, isize, left;
764 if (!in->f_op || !in->f_op->splice_read)
767 if (!(in->f_mode & FMODE_READ))
772 ret = rw_verify_area(READ, in, &pos, len);
773 if (unlikely(ret < 0))
776 isize = i_size_read(in->f_mapping->host);
777 if (unlikely(in->f_pos >= isize))
780 left = isize - in->f_pos;
784 return in->f_op->splice_read(in, pipe, len, flags);
787 long do_splice_direct(struct file *in, struct file *out, size_t len,
790 struct pipe_inode_info *pipe;
796 * We require the input being a regular file, as we don't want to
797 * randomly drop data for eg socket -> socket splicing. Use the
798 * piped splicing for that!
800 i_mode = in->f_dentry->d_inode->i_mode;
801 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
805 * neither in nor out is a pipe, setup an internal pipe attached to
806 * 'out' and transfer the wanted data from 'in' to 'out' through that
808 pipe = current->splice_pipe;
810 pipe = alloc_pipe_info(NULL);
815 * We don't have an immediate reader, but we'll read the stuff
816 * out of the pipe right after the move_to_pipe(). So set
817 * PIPE_READERS appropriately.
821 current->splice_pipe = pipe;
831 size_t read_len, max_read_len;
834 * Do at most PIPE_BUFFERS pages worth of transfer:
836 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
838 ret = do_splice_to(in, pipe, max_read_len, flags);
839 if (unlikely(ret < 0))
845 * NOTE: nonblocking mode only applies to the input. We
846 * must not do the output in nonblocking mode as then we
847 * could get stuck data in the internal pipe:
849 ret = do_splice_from(pipe, out, read_len,
850 flags & ~SPLICE_F_NONBLOCK);
851 if (unlikely(ret < 0))
858 * In nonblocking mode, if we got back a short read then
859 * that was due to either an IO error or due to the
860 * pagecache entry not being there. In the IO error case
861 * the _next_ splice attempt will produce a clean IO error
862 * return value (not a short read), so in both cases it's
863 * correct to break out of the loop here:
865 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
869 pipe->nrbufs = pipe->curbuf = 0;
875 * If we did an incomplete transfer we must release
876 * the pipe buffers in question:
878 for (i = 0; i < PIPE_BUFFERS; i++) {
879 struct pipe_buffer *buf = pipe->bufs + i;
882 buf->ops->release(pipe, buf);
886 pipe->nrbufs = pipe->curbuf = 0;
889 * If we transferred some data, return the number of bytes:
897 EXPORT_SYMBOL(do_splice_direct);
900 * Determine where to splice to/from.
902 static long do_splice(struct file *in, loff_t __user *off_in,
903 struct file *out, loff_t __user *off_out,
904 size_t len, unsigned int flags)
906 struct pipe_inode_info *pipe;
908 pipe = in->f_dentry->d_inode->i_pipe;
913 if (out->f_op->llseek == no_llseek)
915 if (copy_from_user(&out->f_pos, off_out,
920 return do_splice_from(pipe, out, len, flags);
923 pipe = out->f_dentry->d_inode->i_pipe;
928 if (in->f_op->llseek == no_llseek)
930 if (copy_from_user(&in->f_pos, off_in, sizeof(loff_t)))
934 return do_splice_to(in, pipe, len, flags);
940 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
941 int fd_out, loff_t __user *off_out,
942 size_t len, unsigned int flags)
945 struct file *in, *out;
946 int fput_in, fput_out;
952 in = fget_light(fd_in, &fput_in);
954 if (in->f_mode & FMODE_READ) {
955 out = fget_light(fd_out, &fput_out);
957 if (out->f_mode & FMODE_WRITE)
958 error = do_splice(in, off_in,
961 fput_light(out, fput_out);
965 fput_light(in, fput_in);