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, network, direct splicing, etc and
13 * fixing lots of bugs.
15 * Copyright (C) 2005-2006 Jens Axboe <axboe@suse.de>
16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/pipe_fs_i.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
32 * Passed to the actors
35 unsigned int len, total_len; /* current and remaining length */
36 unsigned int flags; /* splice flags */
37 struct file *file; /* file to read/write */
38 loff_t pos; /* file position */
42 * Attempt to steal a page from a pipe buffer. This should perhaps go into
43 * a vm helper function, it's already simplified quite a bit by the
44 * addition of remove_mapping(). If success is returned, the caller may
45 * attempt to reuse this page for another destination.
47 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
48 struct pipe_buffer *buf)
50 struct page *page = buf->page;
51 struct address_space *mapping = page_mapping(page);
53 WARN_ON(!PageLocked(page));
54 WARN_ON(!PageUptodate(page));
57 * At least for ext2 with nobh option, we need to wait on writeback
58 * completing on this page, since we'll remove it from the pagecache.
59 * Otherwise truncate wont wait on the page, allowing the disk
60 * blocks to be reused by someone else before we actually wrote our
61 * data to them. fs corruption ensues.
63 wait_on_page_writeback(page);
65 if (PagePrivate(page))
66 try_to_release_page(page, mapping_gfp_mask(mapping));
68 if (!remove_mapping(mapping, page))
71 buf->flags |= PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU;
75 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
76 struct pipe_buffer *buf)
78 page_cache_release(buf->page);
80 buf->flags &= ~(PIPE_BUF_FLAG_STOLEN | PIPE_BUF_FLAG_LRU);
83 static void *page_cache_pipe_buf_map(struct file *file,
84 struct pipe_inode_info *info,
85 struct pipe_buffer *buf)
87 struct page *page = buf->page;
90 if (!PageUptodate(page)) {
94 * Page got truncated/unhashed. This will cause a 0-byte
95 * splice, if this is the first page.
103 * Uh oh, read-error from disk.
105 if (!PageUptodate(page)) {
111 * Page is ok afterall, fall through to mapping.
122 static void page_cache_pipe_buf_unmap(struct pipe_inode_info *info,
123 struct pipe_buffer *buf)
128 static void page_cache_pipe_buf_get(struct pipe_inode_info *info,
129 struct pipe_buffer *buf)
131 page_cache_get(buf->page);
134 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
136 .map = page_cache_pipe_buf_map,
137 .unmap = page_cache_pipe_buf_unmap,
138 .release = page_cache_pipe_buf_release,
139 .steal = page_cache_pipe_buf_steal,
140 .get = page_cache_pipe_buf_get,
144 * Pipe output worker. This sets up our pipe format with the page cache
145 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
147 static ssize_t move_to_pipe(struct pipe_inode_info *pipe, struct page **pages,
148 int nr_pages, unsigned long offset,
149 unsigned long len, unsigned int flags)
151 int ret, do_wakeup, i;
158 mutex_lock(&pipe->inode->i_mutex);
161 if (!pipe->readers) {
162 send_sig(SIGPIPE, current, 0);
168 if (pipe->nrbufs < PIPE_BUFFERS) {
169 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
170 struct pipe_buffer *buf = pipe->bufs + newbuf;
171 struct page *page = pages[i++];
172 unsigned long this_len;
174 this_len = PAGE_CACHE_SIZE - offset;
179 buf->offset = offset;
181 buf->ops = &page_cache_pipe_buf_ops;
193 if (pipe->nrbufs < PIPE_BUFFERS)
199 if (flags & SPLICE_F_NONBLOCK) {
205 if (signal_pending(current)) {
213 if (waitqueue_active(&pipe->wait))
214 wake_up_interruptible_sync(&pipe->wait);
215 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
219 pipe->waiting_writers++;
221 pipe->waiting_writers--;
225 mutex_unlock(&pipe->inode->i_mutex);
229 if (waitqueue_active(&pipe->wait))
230 wake_up_interruptible(&pipe->wait);
231 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
235 page_cache_release(pages[i++]);
241 __generic_file_splice_read(struct file *in, loff_t *ppos,
242 struct pipe_inode_info *pipe, size_t len,
245 struct address_space *mapping = in->f_mapping;
246 unsigned int offset, nr_pages;
247 struct page *pages[PIPE_BUFFERS];
252 index = *ppos >> PAGE_CACHE_SHIFT;
253 offset = *ppos & ~PAGE_CACHE_MASK;
254 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
256 if (nr_pages > PIPE_BUFFERS)
257 nr_pages = PIPE_BUFFERS;
260 * Initiate read-ahead on this page range. however, don't call into
261 * read-ahead if this is a non-zero offset (we are likely doing small
262 * chunk splice and the page is already there) for a single page.
264 if (!offset || nr_pages > 1)
265 do_page_cache_readahead(mapping, in, index, nr_pages);
268 * Now fill in the holes:
271 for (i = 0; i < nr_pages; i++, index++) {
274 * lookup the page for this index
276 page = find_get_page(mapping, index);
279 * If in nonblock mode then dont block on
280 * readpage (we've kicked readahead so there
281 * will be asynchronous progress):
283 if (flags & SPLICE_F_NONBLOCK)
287 * page didn't exist, allocate one
289 page = page_cache_alloc_cold(mapping);
293 error = add_to_page_cache_lru(page, mapping, index,
294 mapping_gfp_mask(mapping));
295 if (unlikely(error)) {
296 page_cache_release(page);
304 * If the page isn't uptodate, we may need to start io on it
306 if (!PageUptodate(page)) {
310 * page was truncated, stop here. if this isn't the
311 * first page, we'll just complete what we already
314 if (!page->mapping) {
316 page_cache_release(page);
320 * page was already under io and is now done, great
322 if (PageUptodate(page)) {
329 * need to read in the page
331 error = mapping->a_ops->readpage(in, page);
333 if (unlikely(error)) {
334 page_cache_release(page);
335 if (error == AOP_TRUNCATED_PAGE)
345 return move_to_pipe(pipe, pages, i, offset, len, flags);
351 * generic_file_splice_read - splice data from file to a pipe
352 * @in: file to splice from
353 * @pipe: pipe to splice to
354 * @len: number of bytes to splice
355 * @flags: splice modifier flags
357 * Will read pages from given file and fill them into a pipe.
359 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
360 struct pipe_inode_info *pipe, size_t len,
370 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
379 if (!(flags & SPLICE_F_NONBLOCK))
391 EXPORT_SYMBOL(generic_file_splice_read);
394 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
397 static int pipe_to_sendpage(struct pipe_inode_info *info,
398 struct pipe_buffer *buf, struct splice_desc *sd)
400 struct file *file = sd->file;
401 loff_t pos = sd->pos;
408 * Sub-optimal, but we are limited by the pipe ->map. We don't
409 * need a kmap'ed buffer here, we just want to make sure we
410 * have the page pinned if the pipe page originates from the
413 ptr = buf->ops->map(file, info, buf);
417 offset = pos & ~PAGE_CACHE_MASK;
418 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
420 ret = file->f_op->sendpage(file, buf->page, offset, sd->len, &pos,more);
422 buf->ops->unmap(info, buf);
430 * This is a little more tricky than the file -> pipe splicing. There are
431 * basically three cases:
433 * - Destination page already exists in the address space and there
434 * are users of it. For that case we have no other option that
435 * copying the data. Tough luck.
436 * - Destination page already exists in the address space, but there
437 * are no users of it. Make sure it's uptodate, then drop it. Fall
438 * through to last case.
439 * - Destination page does not exist, we can add the pipe page to
440 * the page cache and avoid the copy.
442 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
443 * sd->flags), we attempt to migrate pages from the pipe to the output
444 * file address space page cache. This is possible if no one else has
445 * the pipe page referenced outside of the pipe and page cache. If
446 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
447 * a new page in the output file page cache and fill/dirty that.
449 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
450 struct splice_desc *sd)
452 struct file *file = sd->file;
453 struct address_space *mapping = file->f_mapping;
454 gfp_t gfp_mask = mapping_gfp_mask(mapping);
462 * make sure the data in this buffer is uptodate
464 src = buf->ops->map(file, info, buf);
468 index = sd->pos >> PAGE_CACHE_SHIFT;
469 offset = sd->pos & ~PAGE_CACHE_MASK;
472 * Reuse buf page, if SPLICE_F_MOVE is set.
474 if (sd->flags & SPLICE_F_MOVE) {
476 * If steal succeeds, buf->page is now pruned from the vm
477 * side (LRU and page cache) and we can reuse it.
479 if (buf->ops->steal(info, buf))
483 * this will also set the page locked
486 if (add_to_page_cache(page, mapping, index, gfp_mask))
489 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
494 page = find_or_create_page(mapping, index, gfp_mask);
499 * If the page is uptodate, it is also locked. If it isn't
500 * uptodate, we can mark it uptodate if we are filling the
501 * full page. Otherwise we need to read it in first...
503 if (!PageUptodate(page)) {
504 if (sd->len < PAGE_CACHE_SIZE) {
505 ret = mapping->a_ops->readpage(file, page);
511 if (!PageUptodate(page)) {
513 * Page got invalidated, repeat.
515 if (!page->mapping) {
517 page_cache_release(page);
524 WARN_ON(!PageLocked(page));
525 SetPageUptodate(page);
530 ret = mapping->a_ops->prepare_write(file, page, 0, sd->len);
531 if (ret == AOP_TRUNCATED_PAGE) {
532 page_cache_release(page);
537 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
538 char *dst = kmap_atomic(page, KM_USER0);
540 memcpy(dst + offset, src + buf->offset, sd->len);
541 flush_dcache_page(page);
542 kunmap_atomic(dst, KM_USER0);
545 ret = mapping->a_ops->commit_write(file, page, 0, sd->len);
546 if (ret == AOP_TRUNCATED_PAGE) {
547 page_cache_release(page);
552 mark_page_accessed(page);
553 balance_dirty_pages_ratelimited(mapping);
555 if (!(buf->flags & PIPE_BUF_FLAG_STOLEN)) {
556 page_cache_release(page);
560 buf->ops->unmap(info, buf);
564 typedef int (splice_actor)(struct pipe_inode_info *, struct pipe_buffer *,
565 struct splice_desc *);
568 * Pipe input worker. Most of this logic works like a regular pipe, the
569 * key here is the 'actor' worker passed in that actually moves the data
570 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
572 static ssize_t move_from_pipe(struct pipe_inode_info *pipe, struct file *out,
573 loff_t *ppos, size_t len, unsigned int flags,
576 int ret, do_wakeup, err;
577 struct splice_desc sd;
588 mutex_lock(&pipe->inode->i_mutex);
592 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
593 struct pipe_buf_operations *ops = buf->ops;
596 if (sd.len > sd.total_len)
597 sd.len = sd.total_len;
599 err = actor(pipe, buf, &sd);
601 if (!ret && err != -ENODATA)
608 buf->offset += sd.len;
613 ops->release(pipe, buf);
614 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
621 sd.total_len -= sd.len;
630 if (!pipe->waiting_writers) {
635 if (flags & SPLICE_F_NONBLOCK) {
641 if (signal_pending(current)) {
649 if (waitqueue_active(&pipe->wait))
650 wake_up_interruptible_sync(&pipe->wait);
651 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
659 mutex_unlock(&pipe->inode->i_mutex);
663 if (waitqueue_active(&pipe->wait))
664 wake_up_interruptible(&pipe->wait);
665 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
672 * generic_file_splice_write - splice data from a pipe to a file
674 * @out: file to write to
675 * @len: number of bytes to splice
676 * @flags: splice modifier flags
678 * Will either move or copy pages (determined by @flags options) from
679 * the given pipe inode to the given file.
683 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
684 loff_t *ppos, size_t len, unsigned int flags)
686 struct address_space *mapping = out->f_mapping;
689 ret = move_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
692 * If file or inode is SYNC and we actually wrote some data, sync it.
694 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(mapping->host))
696 struct inode *inode = mapping->host;
699 mutex_lock(&inode->i_mutex);
700 err = generic_osync_inode(mapping->host, mapping,
701 OSYNC_METADATA|OSYNC_DATA);
702 mutex_unlock(&inode->i_mutex);
711 EXPORT_SYMBOL(generic_file_splice_write);
714 * generic_splice_sendpage - splice data from a pipe to a socket
716 * @out: socket to write to
717 * @len: number of bytes to splice
718 * @flags: splice modifier flags
720 * Will send @len bytes from the pipe to a network socket. No data copying
724 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
725 loff_t *ppos, size_t len, unsigned int flags)
727 return move_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
730 EXPORT_SYMBOL(generic_splice_sendpage);
733 * Attempt to initiate a splice from pipe to file.
735 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
736 loff_t *ppos, size_t len, unsigned int flags)
740 if (unlikely(!out->f_op || !out->f_op->splice_write))
743 if (unlikely(!(out->f_mode & FMODE_WRITE)))
746 ret = rw_verify_area(WRITE, out, ppos, len);
747 if (unlikely(ret < 0))
750 return out->f_op->splice_write(pipe, out, ppos, len, flags);
754 * Attempt to initiate a splice from a file to a pipe.
756 static long do_splice_to(struct file *in, loff_t *ppos,
757 struct pipe_inode_info *pipe, size_t len,
763 if (unlikely(!in->f_op || !in->f_op->splice_read))
766 if (unlikely(!(in->f_mode & FMODE_READ)))
769 ret = rw_verify_area(READ, in, ppos, len);
770 if (unlikely(ret < 0))
773 isize = i_size_read(in->f_mapping->host);
774 if (unlikely(*ppos >= isize))
777 left = isize - *ppos;
778 if (unlikely(left < len))
781 return in->f_op->splice_read(in, ppos, pipe, len, flags);
784 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
785 size_t len, unsigned int flags)
787 struct pipe_inode_info *pipe;
794 * We require the input being a regular file, as we don't want to
795 * randomly drop data for eg socket -> socket splicing. Use the
796 * piped splicing for that!
798 i_mode = in->f_dentry->d_inode->i_mode;
799 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
803 * neither in nor out is a pipe, setup an internal pipe attached to
804 * 'out' and transfer the wanted data from 'in' to 'out' through that
806 pipe = current->splice_pipe;
807 if (unlikely(!pipe)) {
808 pipe = alloc_pipe_info(NULL);
813 * We don't have an immediate reader, but we'll read the stuff
814 * out of the pipe right after the move_to_pipe(). So set
815 * PIPE_READERS appropriately.
819 current->splice_pipe = pipe;
830 size_t read_len, max_read_len;
833 * Do at most PIPE_BUFFERS pages worth of transfer:
835 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
837 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
838 if (unlikely(ret < 0))
844 * NOTE: nonblocking mode only applies to the input. We
845 * must not do the output in nonblocking mode as then we
846 * could get stuck data in the internal pipe:
848 ret = do_splice_from(pipe, out, &out_off, read_len,
849 flags & ~SPLICE_F_NONBLOCK);
850 if (unlikely(ret < 0))
857 * In nonblocking mode, if we got back a short read then
858 * that was due to either an IO error or due to the
859 * pagecache entry not being there. In the IO error case
860 * the _next_ splice attempt will produce a clean IO error
861 * return value (not a short read), so in both cases it's
862 * correct to break out of the loop here:
864 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
868 pipe->nrbufs = pipe->curbuf = 0;
874 * If we did an incomplete transfer we must release
875 * the pipe buffers in question:
877 for (i = 0; i < PIPE_BUFFERS; i++) {
878 struct pipe_buffer *buf = pipe->bufs + i;
881 buf->ops->release(pipe, buf);
885 pipe->nrbufs = pipe->curbuf = 0;
888 * If we transferred some data, return the number of bytes:
896 EXPORT_SYMBOL(do_splice_direct);
899 * Determine where to splice to/from.
901 static long do_splice(struct file *in, loff_t __user *off_in,
902 struct file *out, loff_t __user *off_out,
903 size_t len, unsigned int flags)
905 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(&offset, off_out, sizeof(loff_t)))
921 return do_splice_from(pipe, out, off, len, flags);
924 pipe = out->f_dentry->d_inode->i_pipe;
929 if (in->f_op->llseek == no_llseek)
931 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
937 return do_splice_to(in, off, pipe, len, flags);
943 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
944 int fd_out, loff_t __user *off_out,
945 size_t len, unsigned int flags)
948 struct file *in, *out;
949 int fput_in, fput_out;
955 in = fget_light(fd_in, &fput_in);
957 if (in->f_mode & FMODE_READ) {
958 out = fget_light(fd_out, &fput_out);
960 if (out->f_mode & FMODE_WRITE)
961 error = do_splice(in, off_in,
964 fput_light(out, fput_out);
968 fput_light(in, fput_in);
975 * Link contents of ipipe to opipe.
977 static int link_pipe(struct pipe_inode_info *ipipe,
978 struct pipe_inode_info *opipe,
979 size_t len, unsigned int flags)
981 struct pipe_buffer *ibuf, *obuf;
982 int ret = 0, do_wakeup = 0, i;
985 * Potential ABBA deadlock, work around it by ordering lock
986 * grabbing by inode address. Otherwise two different processes
987 * could deadlock (one doing tee from A -> B, the other from B -> A).
989 if (ipipe->inode < opipe->inode) {
990 mutex_lock(&ipipe->inode->i_mutex);
991 mutex_lock(&opipe->inode->i_mutex);
993 mutex_lock(&opipe->inode->i_mutex);
994 mutex_lock(&ipipe->inode->i_mutex);
998 if (!opipe->readers) {
999 send_sig(SIGPIPE, current, 0);
1004 if (ipipe->nrbufs - i) {
1005 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1008 * If we have room, fill this buffer
1010 if (opipe->nrbufs < PIPE_BUFFERS) {
1011 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1014 * Get a reference to this pipe buffer,
1015 * so we can copy the contents over.
1017 ibuf->ops->get(ipipe, ibuf);
1019 obuf = opipe->bufs + nbuf;
1022 if (obuf->len > len)
1032 if (opipe->nrbufs < PIPE_BUFFERS)
1037 * We have input available, but no output room.
1038 * If we already copied data, return that.
1040 if (flags & SPLICE_F_NONBLOCK) {
1045 if (signal_pending(current)) {
1052 if (waitqueue_active(&opipe->wait))
1053 wake_up_interruptible(&opipe->wait);
1054 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1058 opipe->waiting_writers++;
1060 opipe->waiting_writers--;
1065 * No input buffers, do the usual checks for available
1066 * writers and blocking and wait if necessary
1068 if (!ipipe->writers)
1070 if (!ipipe->waiting_writers) {
1074 if (flags & SPLICE_F_NONBLOCK) {
1079 if (signal_pending(current)) {
1085 if (waitqueue_active(&ipipe->wait))
1086 wake_up_interruptible_sync(&ipipe->wait);
1087 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1092 mutex_unlock(&ipipe->inode->i_mutex);
1093 mutex_unlock(&opipe->inode->i_mutex);
1097 if (waitqueue_active(&opipe->wait))
1098 wake_up_interruptible(&opipe->wait);
1099 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1106 * This is a tee(1) implementation that works on pipes. It doesn't copy
1107 * any data, it simply references the 'in' pages on the 'out' pipe.
1108 * The 'flags' used are the SPLICE_F_* variants, currently the only
1109 * applicable one is SPLICE_F_NONBLOCK.
1111 static long do_tee(struct file *in, struct file *out, size_t len,
1114 struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1115 struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1118 * Link ipipe to the two output pipes, consuming as we go along.
1121 return link_pipe(ipipe, opipe, len, flags);
1126 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1135 in = fget_light(fdin, &fput_in);
1137 if (in->f_mode & FMODE_READ) {
1139 struct file *out = fget_light(fdout, &fput_out);
1142 if (out->f_mode & FMODE_WRITE)
1143 error = do_tee(in, out, len, flags);
1144 fput_light(out, fput_out);
1147 fput_light(in, fput_in);