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;
90 if (!PageUptodate(page)) {
97 return ERR_PTR(-ENODATA);
100 return kmap(buf->page);
103 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
104 struct pipe_buffer *buf)
106 unlock_page(buf->page);
110 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
112 .map = page_cache_pipe_buf_map,
113 .unmap = page_cache_pipe_buf_unmap,
114 .release = page_cache_pipe_buf_release,
115 .steal = page_cache_pipe_buf_steal,
119 * Pipe output worker. This sets up our pipe format with the page cache
120 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
122 static ssize_t move_to_pipe(struct inode *inode, struct page **pages,
123 int nr_pages, unsigned long offset,
124 unsigned long len, unsigned int flags)
126 struct pipe_inode_info *info;
127 int ret, do_wakeup, i;
133 mutex_lock(PIPE_MUTEX(*inode));
135 info = inode->i_pipe;
139 if (!PIPE_READERS(*inode)) {
140 send_sig(SIGPIPE, current, 0);
147 if (bufs < PIPE_BUFFERS) {
148 int newbuf = (info->curbuf + bufs) & (PIPE_BUFFERS - 1);
149 struct pipe_buffer *buf = info->bufs + newbuf;
150 struct page *page = pages[i++];
151 unsigned long this_len;
153 this_len = PAGE_CACHE_SIZE - offset;
158 buf->offset = offset;
160 buf->ops = &page_cache_pipe_buf_ops;
161 info->nrbufs = ++bufs;
171 if (bufs < PIPE_BUFFERS)
177 if (flags & SPLICE_F_NONBLOCK) {
183 if (signal_pending(current)) {
190 wake_up_interruptible_sync(PIPE_WAIT(*inode));
191 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO,
196 PIPE_WAITING_WRITERS(*inode)++;
198 PIPE_WAITING_WRITERS(*inode)--;
201 mutex_unlock(PIPE_MUTEX(*inode));
204 wake_up_interruptible(PIPE_WAIT(*inode));
205 kill_fasync(PIPE_FASYNC_READERS(*inode), SIGIO, POLL_IN);
209 page_cache_release(pages[i++]);
214 static int __generic_file_splice_read(struct file *in, struct inode *pipe,
215 size_t len, unsigned int flags)
217 struct address_space *mapping = in->f_mapping;
218 unsigned int offset, nr_pages;
219 struct page *pages[PIPE_BUFFERS], *shadow[PIPE_BUFFERS];
224 index = in->f_pos >> PAGE_CACHE_SHIFT;
225 offset = in->f_pos & ~PAGE_CACHE_MASK;
226 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
228 if (nr_pages > PIPE_BUFFERS)
229 nr_pages = PIPE_BUFFERS;
232 * initiate read-ahead on this page range
234 do_page_cache_readahead(mapping, in, index, nr_pages);
237 * Get as many pages from the page cache as possible..
238 * Start IO on the page cache entries we create (we
239 * can assume that any pre-existing ones we find have
240 * already had IO started on them).
242 i = find_get_pages(mapping, index, nr_pages, pages);
245 * common case - we found all pages and they are contiguous,
248 if (i && (pages[i - 1]->index == index + i - 1))
252 * fill shadow[] with pages at the right locations, so we only
255 memset(shadow, 0, nr_pages * sizeof(struct page *));
256 for (j = 0; j < i; j++)
257 shadow[pages[j]->index - index] = pages[j];
260 * now fill in the holes
262 for (i = 0, pidx = index; i < nr_pages; pidx++, i++) {
269 * no page there, look one up / create it
271 page = find_or_create_page(mapping, pidx,
272 mapping_gfp_mask(mapping));
276 if (PageUptodate(page))
279 error = mapping->a_ops->readpage(in, page);
281 if (unlikely(error)) {
282 page_cache_release(page);
290 for (i = 0; i < nr_pages; i++) {
292 page_cache_release(shadow[i]);
297 memcpy(pages, shadow, i * sizeof(struct page *));
300 * Now we splice them into the pipe..
303 return move_to_pipe(pipe, pages, i, offset, len, flags);
307 * generic_file_splice_read - splice data from file to a pipe
308 * @in: file to splice from
309 * @pipe: pipe to splice to
310 * @len: number of bytes to splice
311 * @flags: splice modifier flags
313 * Will read pages from given file and fill them into a pipe.
316 ssize_t generic_file_splice_read(struct file *in, struct inode *pipe,
317 size_t len, unsigned int flags)
325 ret = __generic_file_splice_read(in, pipe, len, flags);
334 if (!(flags & SPLICE_F_NONBLOCK))
346 EXPORT_SYMBOL(generic_file_splice_read);
349 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
352 static int pipe_to_sendpage(struct pipe_inode_info *info,
353 struct pipe_buffer *buf, struct splice_desc *sd)
355 struct file *file = sd->file;
356 loff_t pos = sd->pos;
363 * sub-optimal, but we are limited by the pipe ->map. we don't
364 * need a kmap'ed buffer here, we just want to make sure we
365 * have the page pinned if the pipe page originates from the
368 ptr = buf->ops->map(file, info, buf);
372 offset = pos & ~PAGE_CACHE_MASK;
373 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
375 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
377 buf->ops->unmap(info, buf);
385 * This is a little more tricky than the file -> pipe splicing. There are
386 * basically three cases:
388 * - Destination page already exists in the address space and there
389 * are users of it. For that case we have no other option that
390 * copying the data. Tough luck.
391 * - Destination page already exists in the address space, but there
392 * are no users of it. Make sure it's uptodate, then drop it. Fall
393 * through to last case.
394 * - Destination page does not exist, we can add the pipe page to
395 * the page cache and avoid the copy.
397 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
398 * sd->flags), we attempt to migrate pages from the pipe to the output
399 * file address space page cache. This is possible if no one else has
400 * the pipe page referenced outside of the pipe and page cache. If
401 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
402 * a new page in the output file page cache and fill/dirty that.
404 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
405 struct splice_desc *sd)
407 struct file *file = sd->file;
408 struct address_space *mapping = file->f_mapping;
409 gfp_t gfp_mask = mapping_gfp_mask(mapping);
417 * after this, page will be locked and unmapped
419 src = buf->ops->map(file, info, buf);
423 index = sd->pos >> PAGE_CACHE_SHIFT;
424 offset = sd->pos & ~PAGE_CACHE_MASK;
427 * reuse buf page, if SPLICE_F_MOVE is set
429 if (sd->flags & SPLICE_F_MOVE) {
431 * If steal succeeds, buf->page is now pruned from the vm
432 * side (LRU and page cache) and we can reuse it.
434 if (buf->ops->steal(info, buf))
438 if (add_to_page_cache(page, mapping, index, gfp_mask))
441 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
446 page = find_or_create_page(mapping, index, gfp_mask);
451 * If the page is uptodate, it is also locked. If it isn't
452 * uptodate, we can mark it uptodate if we are filling the
453 * full page. Otherwise we need to read it in first...
455 if (!PageUptodate(page)) {
456 if (sd->len < PAGE_CACHE_SIZE) {
457 ret = mapping->a_ops->readpage(file, page);
463 if (!PageUptodate(page)) {
465 * page got invalidated, repeat
467 if (!page->mapping) {
469 page_cache_release(page);
476 WARN_ON(!PageLocked(page));
477 SetPageUptodate(page);
482 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
483 if (ret == AOP_TRUNCATED_PAGE) {
484 page_cache_release(page);
489 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
490 char *dst = kmap_atomic(page, KM_USER0);
492 memcpy(dst + offset, src + buf->offset, sd->len);
493 flush_dcache_page(page);
494 kunmap_atomic(dst, KM_USER0);
497 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
498 if (ret == AOP_TRUNCATED_PAGE) {
499 page_cache_release(page);
504 balance_dirty_pages_ratelimited(mapping);
506 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
507 page_cache_release(page);
510 buf->ops->unmap(info, buf);
514 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
515 struct splice_desc *);
518 * Pipe input worker. Most of this logic works like a regular pipe, the
519 * key here is the 'actor' worker passed in that actually moves the data
520 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
522 static ssize_t move_from_pipe(struct inode *inode, struct file *out,
523 size_t len, unsigned int flags,
526 struct pipe_inode_info *info;
527 int ret, do_wakeup, err;
528 struct splice_desc sd;
538 mutex_lock(PIPE_MUTEX(*inode));
540 info = inode->i_pipe;
542 int bufs = info->nrbufs;
545 int curbuf = info->curbuf;
546 struct pipe_buffer *buf = info->bufs + curbuf;
547 struct pipe_buf_operations *ops = buf->ops;
550 if (sd.len > sd.total_len)
551 sd.len = sd.total_len;
553 err = actor(info, buf, &sd);
555 if (!ret && err != -ENODATA)
562 buf->offset += sd.len;
566 ops->release(info, buf);
567 curbuf = (curbuf + 1) & (PIPE_BUFFERS - 1);
568 info->curbuf = curbuf;
569 info->nrbufs = --bufs;
574 sd.total_len -= sd.len;
581 if (!PIPE_WRITERS(*inode))
583 if (!PIPE_WAITING_WRITERS(*inode)) {
588 if (flags & SPLICE_F_NONBLOCK) {
594 if (signal_pending(current)) {
601 wake_up_interruptible_sync(PIPE_WAIT(*inode));
602 kill_fasync(PIPE_FASYNC_WRITERS(*inode),SIGIO,POLL_OUT);
609 mutex_unlock(PIPE_MUTEX(*inode));
612 wake_up_interruptible(PIPE_WAIT(*inode));
613 kill_fasync(PIPE_FASYNC_WRITERS(*inode), SIGIO, POLL_OUT);
616 mutex_lock(&out->f_mapping->host->i_mutex);
618 mutex_unlock(&out->f_mapping->host->i_mutex);
624 * generic_file_splice_write - splice data from a pipe to a file
626 * @out: file to write to
627 * @len: number of bytes to splice
628 * @flags: splice modifier flags
630 * Will either move or copy pages (determined by @flags options) from
631 * the given pipe inode to the given file.
634 ssize_t generic_file_splice_write(struct inode *inode, struct file *out,
635 size_t len, unsigned int flags)
637 struct address_space *mapping = out->f_mapping;
638 ssize_t ret = move_from_pipe(inode, out, len, flags, pipe_to_file);
641 * if file or inode is SYNC and we actually wrote some data, sync it
643 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
645 struct inode *inode = mapping->host;
648 mutex_lock(&inode->i_mutex);
649 err = generic_osync_inode(mapping->host, mapping,
650 OSYNC_METADATA|OSYNC_DATA);
651 mutex_unlock(&inode->i_mutex);
660 EXPORT_SYMBOL(generic_file_splice_write);
663 * generic_splice_sendpage - splice data from a pipe to a socket
665 * @out: socket to write to
666 * @len: number of bytes to splice
667 * @flags: splice modifier flags
669 * Will send @len bytes from the pipe to a network socket. No data copying
673 ssize_t generic_splice_sendpage(struct inode *inode, struct file *out,
674 size_t len, unsigned int flags)
676 return move_from_pipe(inode, out, len, flags, pipe_to_sendpage);
679 EXPORT_SYMBOL(generic_splice_sendpage);
682 * Attempt to initiate a splice from pipe to file.
684 static long do_splice_from(struct inode *pipe, struct file *out, size_t len,
690 if (!out->f_op || !out->f_op->splice_write)
693 if (!(out->f_mode & FMODE_WRITE))
697 ret = rw_verify_area(WRITE, out, &pos, len);
698 if (unlikely(ret < 0))
701 return out->f_op->splice_write(pipe, out, len, flags);
705 * Attempt to initiate a splice from a file to a pipe.
707 static long do_splice_to(struct file *in, struct inode *pipe, size_t len,
710 loff_t pos, isize, left;
713 if (!in->f_op || !in->f_op->splice_read)
716 if (!(in->f_mode & FMODE_READ))
720 ret = rw_verify_area(READ, in, &pos, len);
721 if (unlikely(ret < 0))
724 isize = i_size_read(in->f_mapping->host);
725 if (unlikely(in->f_pos >= isize))
728 left = isize - in->f_pos;
732 return in->f_op->splice_read(in, pipe, len, flags);
736 * Determine where to splice to/from.
738 static long do_splice(struct file *in, struct file *out, size_t len,
743 pipe = in->f_dentry->d_inode;
745 return do_splice_from(pipe, out, len, flags);
747 pipe = out->f_dentry->d_inode;
749 return do_splice_to(in, pipe, len, flags);
754 asmlinkage long sys_splice(int fdin, int fdout, size_t len, unsigned int flags)
757 struct file *in, *out;
758 int fput_in, fput_out;
764 in = fget_light(fdin, &fput_in);
766 if (in->f_mode & FMODE_READ) {
767 out = fget_light(fdout, &fput_out);
769 if (out->f_mode & FMODE_WRITE)
770 error = do_splice(in, out, len, flags);
771 fput_light(out, fput_out);
775 fput_light(in, fput_in);