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 inode *inode, struct page **pages,
140 int nr_pages, unsigned long offset,
141 unsigned long len, unsigned int flags)
143 struct pipe_inode_info *info;
144 int ret, do_wakeup, i;
150 mutex_lock(PIPE_MUTEX(*inode));
152 info = inode->i_pipe;
156 if (!PIPE_READERS(*inode)) {
157 send_sig(SIGPIPE, current, 0);
164 if (bufs < PIPE_BUFFERS) {
165 int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS - 1);
166 struct pipe_buffer *buf = info->bufs + newbuf;
167 struct page *page = pages[i++];
168 unsigned long this_len;
170 this_len = PAGE_CACHE_SIZE - offset;
175 buf->offset = offset;
177 buf->ops = &page_cache_pipe_buf_ops;
178 info->nrbufs = ++bufs;
188 if (bufs < PIPE_BUFFERS)
194 if (flags & SPLICE_F_NONBLOCK) {
200 if (signal_pending(current)) {
208 if (waitqueue_active(PIPE_WAIT(*inode)))
209 wake_up_interruptible_sync(PIPE_WAIT(*inode));
210 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO,
215 PIPE_WAITING_WRITERS(*inode)++;
217 PIPE_WAITING_WRITERS(*inode)--;
220 mutex_unlock(PIPE_MUTEX(*inode));
224 if (waitqueue_active(PIPE_WAIT(*inode)))
225 wake_up_interruptible(PIPE_WAIT(*inode));
226 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
230 page_cache_release(pages[i++]);
235 static int __generic_file_splice_read(struct file *in, struct inode *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
263 for (i = 0; i < nr_pages; i++, index++) {
265 * no page there, look one up / create it
267 page = find_or_create_page(mapping, index,
268 mapping_gfp_mask(mapping));
272 if (PageUptodate(page))
275 int error = mapping->a_ops->readpage(in, page);
277 if (unlikely(error)) {
278 page_cache_release(page);
286 return move_to_pipe(pipe, pages, i, offset, len, flags);
292 * generic_file_splice_read - splice data from file to a pipe
293 * @in: file to splice from
294 * @pipe: pipe to splice to
295 * @len: number of bytes to splice
296 * @flags: splice modifier flags
298 * Will read pages from given file and fill them into a pipe.
301 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
302 size_t len, unsigned int flags)
310 ret = __generic_file_splice_read(in, pipe, len, flags);
319 if (!(flags & SPLICE_F_NONBLOCK))
331 EXPORT_SYMBOL(generic_file_splice_read);
334 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
337 static int pipe_to_sendpage(struct pipe_inode_info *info,
338 struct pipe_buffer *buf, struct splice_desc *sd)
340 struct file *file = sd->file;
341 loff_t pos = sd->pos;
348 * sub-optimal, but we are limited by the pipe ->map. we don't
349 * need a kmap'ed buffer here, we just want to make sure we
350 * have the page pinned if the pipe page originates from the
353 ptr = buf->ops->map(file, info, buf);
357 offset = pos & ~PAGE_CACHE_MASK;
358 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
360 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
362 buf->ops->unmap(info, buf);
370 * This is a little more tricky than the file -> pipe splicing. There are
371 * basically three cases:
373 * - Destination page already exists in the address space and there
374 * are users of it. For that case we have no other option that
375 * copying the data. Tough luck.
376 * - Destination page already exists in the address space, but there
377 * are no users of it. Make sure it's uptodate, then drop it. Fall
378 * through to last case.
379 * - Destination page does not exist, we can add the pipe page to
380 * the page cache and avoid the copy.
382 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
383 * sd->flags), we attempt to migrate pages from the pipe to the output
384 * file address space page cache. This is possible if no one else has
385 * the pipe page referenced outside of the pipe and page cache. If
386 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
387 * a new page in the output file page cache and fill/dirty that.
389 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
390 struct splice_desc *sd)
392 struct file *file = sd->file;
393 struct address_space *mapping = file->f_mapping;
394 gfp_t gfp_mask = mapping_gfp_mask(mapping);
402 * make sure the data in this buffer is uptodate
404 src = buf->ops->map(file, info, buf);
408 index = sd->pos >> PAGE_CACHE_SHIFT;
409 offset = sd->pos & ~PAGE_CACHE_MASK;
412 * reuse buf page, if SPLICE_F_MOVE is set
414 if (sd->flags & SPLICE_F_MOVE) {
416 * If steal succeeds, buf->page is now pruned from the vm
417 * side (LRU and page cache) and we can reuse it.
419 if (buf->ops->steal(info, buf))
423 * this will also set the page locked
426 if (add_to_page_cache(page, mapping, index, gfp_mask))
429 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
434 page = find_or_create_page(mapping, index, gfp_mask);
439 * If the page is uptodate, it is also locked. If it isn't
440 * uptodate, we can mark it uptodate if we are filling the
441 * full page. Otherwise we need to read it in first...
443 if (!PageUptodate(page)) {
444 if (sd->len < PAGE_CACHE_SIZE) {
445 ret = mapping->a_ops->readpage(file, page);
451 if (!PageUptodate(page)) {
453 * page got invalidated, repeat
455 if (!page->mapping) {
457 page_cache_release(page);
464 WARN_ON(!PageLocked(page));
465 SetPageUptodate(page);
470 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
471 if (ret == AOP_TRUNCATED_PAGE) {
472 page_cache_release(page);
477 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
478 char *dst = kmap_atomic(page, KM_USER0);
480 memcpy(dst + offset, src + buf->offset, sd->len);
481 flush_dcache_page(page);
482 kunmap_atomic(dst, KM_USER0);
485 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
486 if (ret == AOP_TRUNCATED_PAGE) {
487 page_cache_release(page);
492 mark_page_accessed(page);
493 balance_dirty_pages_ratelimited(mapping);
495 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
496 page_cache_release(page);
500 buf->ops->unmap(info, buf);
504 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
505 struct splice_desc *);
508 * Pipe input worker. Most of this logic works like a regular pipe, the
509 * key here is the 'actor' worker passed in that actually moves the data
510 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
512 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
513 size_t len, unsigned int flags,
516 struct pipe_inode_info *info;
517 int ret, do_wakeup, err;
518 struct splice_desc sd;
528 mutex_lock(PIPE_MUTEX(*inode));
530 info = inode->i_pipe;
532 int bufs = info->nrbufs;
535 int curbuf = info->curbuf;
536 struct pipe_buffer *buf = info->bufs + curbuf;
537 struct pipe_buf_operations *ops = buf->ops;
540 if (sd.len > sd.total_len)
541 sd.len = sd.total_len;
543 err = actor(info, buf, &sd);
545 if (!ret && err != -ENODATA)
552 buf->offset += sd.len;
556 ops->release(info, buf);
557 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
558 info->curbuf = curbuf;
559 info->nrbufs = --bufs;
564 sd.total_len -= sd.len;
571 if (!PIPE_WRITERS(*inode))
573 if (!PIPE_WAITING_WRITERS(*inode)) {
578 if (flags & SPLICE_F_NONBLOCK) {
584 if (signal_pending(current)) {
592 if (waitqueue_active(PIPE_WAIT(*inode)))
593 wake_up_interruptible_sync(PIPE_WAIT(*inode));
594 kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
601 mutex_unlock(PIPE_MUTEX(*inode));
605 if (waitqueue_active(PIPE_WAIT(*inode)))
606 wake_up_interruptible(PIPE_WAIT(*inode));
607 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
610 mutex_lock(&out->f_mapping->host->i_mutex);
612 mutex_unlock(&out->f_mapping->host->i_mutex);
618 * generic_file_splice_write - splice data from a pipe to a file
620 * @out: file to write to
621 * @len: number of bytes to splice
622 * @flags: splice modifier flags
624 * Will either move or copy pages (determined by @flags options) from
625 * the given pipe inode to the given file.
628 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
629 size_t len, unsigned int flags)
631 struct address_space *mapping = out->f_mapping;
632 ssize_t ret = move_from_pipe(inode, out, len, flags, pipe_to_file);
635 * if file or inode is SYNC and we actually wrote some data, sync it
637 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
639 struct inode *inode = mapping->host;
642 mutex_lock(&inode->i_mutex);
643 err = generic_osync_inode(mapping->host, mapping,
644 OSYNC_METADATA|OSYNC_DATA);
645 mutex_unlock(&inode->i_mutex);
654 EXPORT_SYMBOL(generic_file_splice_write);
657 * generic_splice_sendpage - splice data from a pipe to a socket
659 * @out: socket to write to
660 * @len: number of bytes to splice
661 * @flags: splice modifier flags
663 * Will send @len bytes from the pipe to a network socket. No data copying
667 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
668 size_t len, unsigned int flags)
670 return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
673 EXPORT_SYMBOL(generic_splice_sendpage);
676 * Attempt to initiate a splice from pipe to file.
678 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
684 if (!out->f_op || !out->f_op->splice_write)
687 if (!(out->f_mode & FMODE_WRITE))
691 ret = rw_verify_area(WRITE, out, &pos, len);
692 if (unlikely(ret < 0))
695 return out->f_op->splice_write(pipe, out, len, flags);
699 * Attempt to initiate a splice from a file to a pipe.
701 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
704 loff_t pos, isize, left;
707 if (!in->f_op || !in->f_op->splice_read)
710 if (!(in->f_mode & FMODE_READ))
714 ret = rw_verify_area(READ, in, &pos, len);
715 if (unlikely(ret < 0))
718 isize = i_size_read(in->f_mapping->host);
719 if (unlikely(in->f_pos >= isize))
722 left = isize - in->f_pos;
726 return in->f_op->splice_read(in, pipe, len, flags);
730 * Determine where to splice to/from.
732 static long do_splice(struct file *in, struct file *out, size_t len,
737 pipe = in->f_dentry->d_inode;
739 return do_splice_from(pipe, out, len, flags);
741 pipe = out->f_dentry->d_inode;
743 return do_splice_to(in, pipe, len, flags);
748 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
751 struct file *in, *out;
752 int fput_in, fput_out;
758 in = fget_light(fdin, &fput_in);
760 if (in->f_mode & FMODE_READ) {
761 out = fget_light(fdout, &fput_out);
763 if (out->f_mode & FMODE_WRITE)
764 error = do_splice(in, out, len, flags);
765 fput_light(out, fput_out);
769 fput_light(in, fput_in);