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
255 do_page_cache_readahead(mapping, in, index, nr_pages);
258 * now fill in the holes
260 for (i = 0; i < nr_pages; i++, index++) {
262 * no page there, look one up / create it
264 page = find_or_create_page(mapping, index,
265 mapping_gfp_mask(mapping));
269 if (PageUptodate(page))
272 int error = mapping->a_ops->readpage(in, page);
274 if (unlikely(error)) {
275 page_cache_release(page);
283 return move_to_pipe(pipe, pages, i, offset, len, flags);
289 * generic_file_splice_read - splice data from file to a pipe
290 * @in: file to splice from
291 * @pipe: pipe to splice to
292 * @len: number of bytes to splice
293 * @flags: splice modifier flags
295 * Will read pages from given file and fill them into a pipe.
298 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
299 size_t len, unsigned int flags)
307 ret = __generic_file_splice_read(in, pipe, len, flags);
316 if (!(flags & SPLICE_F_NONBLOCK))
328 EXPORT_SYMBOL(generic_file_splice_read);
331 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
334 static int pipe_to_sendpage(struct pipe_inode_info *info,
335 struct pipe_buffer *buf, struct splice_desc *sd)
337 struct file *file = sd->file;
338 loff_t pos = sd->pos;
345 * sub-optimal, but we are limited by the pipe ->map. we don't
346 * need a kmap'ed buffer here, we just want to make sure we
347 * have the page pinned if the pipe page originates from the
350 ptr = buf->ops->map(file, info, buf);
354 offset = pos & ~PAGE_CACHE_MASK;
355 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
357 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
359 buf->ops->unmap(info, buf);
367 * This is a little more tricky than the file -> pipe splicing. There are
368 * basically three cases:
370 * - Destination page already exists in the address space and there
371 * are users of it. For that case we have no other option that
372 * copying the data. Tough luck.
373 * - Destination page already exists in the address space, but there
374 * are no users of it. Make sure it's uptodate, then drop it. Fall
375 * through to last case.
376 * - Destination page does not exist, we can add the pipe page to
377 * the page cache and avoid the copy.
379 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
380 * sd->flags), we attempt to migrate pages from the pipe to the output
381 * file address space page cache. This is possible if no one else has
382 * the pipe page referenced outside of the pipe and page cache. If
383 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
384 * a new page in the output file page cache and fill/dirty that.
386 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
387 struct splice_desc *sd)
389 struct file *file = sd->file;
390 struct address_space *mapping = file->f_mapping;
391 gfp_t gfp_mask = mapping_gfp_mask(mapping);
399 * make sure the data in this buffer is uptodate
401 src = buf->ops->map(file, info, buf);
405 index = sd->pos >> PAGE_CACHE_SHIFT;
406 offset = sd->pos & ~PAGE_CACHE_MASK;
409 * reuse buf page, if SPLICE_F_MOVE is set
411 if (sd->flags & SPLICE_F_MOVE) {
413 * If steal succeeds, buf->page is now pruned from the vm
414 * side (LRU and page cache) and we can reuse it.
416 if (buf->ops->steal(info, buf))
420 * this will also set the page locked
423 if (add_to_page_cache(page, mapping, index, gfp_mask))
426 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
431 page = find_or_create_page(mapping, index, gfp_mask);
436 * If the page is uptodate, it is also locked. If it isn't
437 * uptodate, we can mark it uptodate if we are filling the
438 * full page. Otherwise we need to read it in first...
440 if (!PageUptodate(page)) {
441 if (sd->len < PAGE_CACHE_SIZE) {
442 ret = mapping->a_ops->readpage(file, page);
448 if (!PageUptodate(page)) {
450 * page got invalidated, repeat
452 if (!page->mapping) {
454 page_cache_release(page);
461 WARN_ON(!PageLocked(page));
462 SetPageUptodate(page);
467 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
468 if (ret == AOP_TRUNCATED_PAGE) {
469 page_cache_release(page);
474 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
475 char *dst = kmap_atomic(page, KM_USER0);
477 memcpy(dst + offset, src + buf->offset, sd->len);
478 flush_dcache_page(page);
479 kunmap_atomic(dst, KM_USER0);
482 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
483 if (ret == AOP_TRUNCATED_PAGE) {
484 page_cache_release(page);
489 mark_page_accessed(page);
490 balance_dirty_pages_ratelimited(mapping);
492 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
493 page_cache_release(page);
497 buf->ops->unmap(info, buf);
501 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
502 struct splice_desc *);
505 * Pipe input worker. Most of this logic works like a regular pipe, the
506 * key here is the 'actor' worker passed in that actually moves the data
507 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
509 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
510 size_t len, unsigned int flags,
513 struct pipe_inode_info *info;
514 int ret, do_wakeup, err;
515 struct splice_desc sd;
525 mutex_lock(PIPE_MUTEX(*inode));
527 info = inode->i_pipe;
529 int bufs = info->nrbufs;
532 int curbuf = info->curbuf;
533 struct pipe_buffer *buf = info->bufs + curbuf;
534 struct pipe_buf_operations *ops = buf->ops;
537 if (sd.len > sd.total_len)
538 sd.len = sd.total_len;
540 err = actor(info, buf, &sd);
542 if (!ret && err != -ENODATA)
549 buf->offset += sd.len;
553 ops->release(info, buf);
554 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
555 info->curbuf = curbuf;
556 info->nrbufs = --bufs;
561 sd.total_len -= sd.len;
568 if (!PIPE_WRITERS(*inode))
570 if (!PIPE_WAITING_WRITERS(*inode)) {
575 if (flags & SPLICE_F_NONBLOCK) {
581 if (signal_pending(current)) {
589 if (waitqueue_active(PIPE_WAIT(*inode)))
590 wake_up_interruptible_sync(PIPE_WAIT(*inode));
591 kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
598 mutex_unlock(PIPE_MUTEX(*inode));
602 if (waitqueue_active(PIPE_WAIT(*inode)))
603 wake_up_interruptible(PIPE_WAIT(*inode));
604 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
607 mutex_lock(&out->f_mapping->host->i_mutex);
609 mutex_unlock(&out->f_mapping->host->i_mutex);
615 * generic_file_splice_write - splice data from a pipe to a file
617 * @out: file to write to
618 * @len: number of bytes to splice
619 * @flags: splice modifier flags
621 * Will either move or copy pages (determined by @flags options) from
622 * the given pipe inode to the given file.
625 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
626 size_t len, unsigned int flags)
628 struct address_space *mapping = out->f_mapping;
629 ssize_t ret = move_from_pipe(inode, out, len, flags, pipe_to_file);
632 * if file or inode is SYNC and we actually wrote some data, sync it
634 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
636 struct inode *inode = mapping->host;
639 mutex_lock(&inode->i_mutex);
640 err = generic_osync_inode(mapping->host, mapping,
641 OSYNC_METADATA|OSYNC_DATA);
642 mutex_unlock(&inode->i_mutex);
651 EXPORT_SYMBOL(generic_file_splice_write);
654 * generic_splice_sendpage - splice data from a pipe to a socket
656 * @out: socket to write to
657 * @len: number of bytes to splice
658 * @flags: splice modifier flags
660 * Will send @len bytes from the pipe to a network socket. No data copying
664 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
665 size_t len, unsigned int flags)
667 return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
670 EXPORT_SYMBOL(generic_splice_sendpage);
673 * Attempt to initiate a splice from pipe to file.
675 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
681 if (!out->f_op || !out->f_op->splice_write)
684 if (!(out->f_mode & FMODE_WRITE))
688 ret = rw_verify_area(WRITE, out, &pos, len);
689 if (unlikely(ret < 0))
692 return out->f_op->splice_write(pipe, out, len, flags);
696 * Attempt to initiate a splice from a file to a pipe.
698 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
701 loff_t pos, isize, left;
704 if (!in->f_op || !in->f_op->splice_read)
707 if (!(in->f_mode & FMODE_READ))
711 ret = rw_verify_area(READ, in, &pos, len);
712 if (unlikely(ret < 0))
715 isize = i_size_read(in->f_mapping->host);
716 if (unlikely(in->f_pos >= isize))
719 left = isize - in->f_pos;
723 return in->f_op->splice_read(in, pipe, len, flags);
727 * Determine where to splice to/from.
729 static long do_splice(struct file *in, struct file *out, size_t len,
734 pipe = in->f_dentry->d_inode;
736 return do_splice_from(pipe, out, len, flags);
738 pipe = out->f_dentry->d_inode;
740 return do_splice_to(in, pipe, len, flags);
745 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
748 struct file *in, *out;
749 int fput_in, fput_out;
755 in = fget_light(fdin, &fput_in);
757 if (in->f_mode & FMODE_READ) {
758 out = fget_light(fdout, &fput_out);
760 if (out->f_mode & FMODE_WRITE)
761 error = do_splice(in, out, len, flags);
762 fput_light(out, fput_out);
766 fput_light(in, fput_in);