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@kernel.dk>
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/splice.h>
24 #include <linux/memcontrol.h>
25 #include <linux/mm_inline.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/buffer_head.h>
29 #include <linux/module.h>
30 #include <linux/syscalls.h>
31 #include <linux/uio.h>
32 #include <linux/security.h>
35 * Attempt to steal a page from a pipe buffer. This should perhaps go into
36 * a vm helper function, it's already simplified quite a bit by the
37 * addition of remove_mapping(). If success is returned, the caller may
38 * attempt to reuse this page for another destination.
40 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
41 struct pipe_buffer *buf)
43 struct page *page = buf->page;
44 struct address_space *mapping;
48 mapping = page_mapping(page);
50 WARN_ON(!PageUptodate(page));
53 * At least for ext2 with nobh option, we need to wait on
54 * writeback completing on this page, since we'll remove it
55 * from the pagecache. Otherwise truncate wont wait on the
56 * page, allowing the disk blocks to be reused by someone else
57 * before we actually wrote our data to them. fs corruption
60 wait_on_page_writeback(page);
62 if (page_has_private(page) &&
63 !try_to_release_page(page, GFP_KERNEL))
67 * If we succeeded in removing the mapping, set LRU flag
70 if (remove_mapping(mapping, page)) {
71 buf->flags |= PIPE_BUF_FLAG_LRU;
77 * Raced with truncate or failed to remove page from current
78 * address space, unlock and return failure.
85 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
86 struct pipe_buffer *buf)
88 page_cache_release(buf->page);
89 buf->flags &= ~PIPE_BUF_FLAG_LRU;
93 * Check whether the contents of buf is OK to access. Since the content
94 * is a page cache page, IO may be in flight.
96 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
97 struct pipe_buffer *buf)
99 struct page *page = buf->page;
102 if (!PageUptodate(page)) {
106 * Page got truncated/unhashed. This will cause a 0-byte
107 * splice, if this is the first page.
109 if (!page->mapping) {
115 * Uh oh, read-error from disk.
117 if (!PageUptodate(page)) {
123 * Page is ok afterall, we are done.
134 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
136 .map = generic_pipe_buf_map,
137 .unmap = generic_pipe_buf_unmap,
138 .confirm = page_cache_pipe_buf_confirm,
139 .release = page_cache_pipe_buf_release,
140 .steal = page_cache_pipe_buf_steal,
141 .get = generic_pipe_buf_get,
144 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
145 struct pipe_buffer *buf)
147 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
150 buf->flags |= PIPE_BUF_FLAG_LRU;
151 return generic_pipe_buf_steal(pipe, buf);
154 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
156 .map = generic_pipe_buf_map,
157 .unmap = generic_pipe_buf_unmap,
158 .confirm = generic_pipe_buf_confirm,
159 .release = page_cache_pipe_buf_release,
160 .steal = user_page_pipe_buf_steal,
161 .get = generic_pipe_buf_get,
165 * splice_to_pipe - fill passed data into a pipe
166 * @pipe: pipe to fill
170 * @spd contains a map of pages and len/offset tuples, along with
171 * the struct pipe_buf_operations associated with these pages. This
172 * function will link that data to the pipe.
175 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
176 struct splice_pipe_desc *spd)
178 unsigned int spd_pages = spd->nr_pages;
179 int ret, do_wakeup, page_nr;
188 if (!pipe->readers) {
189 send_sig(SIGPIPE, current, 0);
195 if (pipe->nrbufs < PIPE_BUFFERS) {
196 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
197 struct pipe_buffer *buf = pipe->bufs + newbuf;
199 buf->page = spd->pages[page_nr];
200 buf->offset = spd->partial[page_nr].offset;
201 buf->len = spd->partial[page_nr].len;
202 buf->private = spd->partial[page_nr].private;
204 if (spd->flags & SPLICE_F_GIFT)
205 buf->flags |= PIPE_BUF_FLAG_GIFT;
214 if (!--spd->nr_pages)
216 if (pipe->nrbufs < PIPE_BUFFERS)
222 if (spd->flags & SPLICE_F_NONBLOCK) {
228 if (signal_pending(current)) {
236 if (waitqueue_active(&pipe->wait))
237 wake_up_interruptible_sync(&pipe->wait);
238 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
242 pipe->waiting_writers++;
244 pipe->waiting_writers--;
251 if (waitqueue_active(&pipe->wait))
252 wake_up_interruptible(&pipe->wait);
253 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
256 while (page_nr < spd_pages)
257 spd->spd_release(spd, page_nr++);
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
264 page_cache_release(spd->pages[i]);
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269 struct pipe_inode_info *pipe, size_t len,
272 struct address_space *mapping = in->f_mapping;
273 unsigned int loff, nr_pages, req_pages;
274 struct page *pages[PIPE_BUFFERS];
275 struct partial_page partial[PIPE_BUFFERS];
277 pgoff_t index, end_index;
280 struct splice_pipe_desc spd = {
284 .ops = &page_cache_pipe_buf_ops,
285 .spd_release = spd_release_page,
288 index = *ppos >> PAGE_CACHE_SHIFT;
289 loff = *ppos & ~PAGE_CACHE_MASK;
290 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
294 * Lookup the (hopefully) full range of pages we need.
296 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297 index += spd.nr_pages;
300 * If find_get_pages_contig() returned fewer pages than we needed,
301 * readahead/allocate the rest and fill in the holes.
303 if (spd.nr_pages < nr_pages)
304 page_cache_sync_readahead(mapping, &in->f_ra, in,
305 index, req_pages - spd.nr_pages);
308 while (spd.nr_pages < nr_pages) {
310 * Page could be there, find_get_pages_contig() breaks on
313 page = find_get_page(mapping, index);
316 * page didn't exist, allocate one.
318 page = page_cache_alloc_cold(mapping);
322 error = add_to_page_cache_lru(page, mapping, index,
323 mapping_gfp_mask(mapping));
324 if (unlikely(error)) {
325 page_cache_release(page);
326 if (error == -EEXIST)
331 * add_to_page_cache() locks the page, unlock it
332 * to avoid convoluting the logic below even more.
337 pages[spd.nr_pages++] = page;
342 * Now loop over the map and see if we need to start IO on any
343 * pages, fill in the partial map, etc.
345 index = *ppos >> PAGE_CACHE_SHIFT;
346 nr_pages = spd.nr_pages;
348 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349 unsigned int this_len;
355 * this_len is the max we'll use from this page
357 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358 page = pages[page_nr];
360 if (PageReadahead(page))
361 page_cache_async_readahead(mapping, &in->f_ra, in,
362 page, index, req_pages - page_nr);
365 * If the page isn't uptodate, we may need to start io on it
367 if (!PageUptodate(page)) {
369 * If in nonblock mode then dont block on waiting
370 * for an in-flight io page
372 if (flags & SPLICE_F_NONBLOCK) {
373 if (!trylock_page(page)) {
381 * Page was truncated, or invalidated by the
382 * filesystem. Redo the find/create, but this time the
383 * page is kept locked, so there's no chance of another
384 * race with truncate/invalidate.
386 if (!page->mapping) {
388 page = find_or_create_page(mapping, index,
389 mapping_gfp_mask(mapping));
395 page_cache_release(pages[page_nr]);
396 pages[page_nr] = page;
399 * page was already under io and is now done, great
401 if (PageUptodate(page)) {
407 * need to read in the page
409 error = mapping->a_ops->readpage(in, page);
410 if (unlikely(error)) {
412 * We really should re-lookup the page here,
413 * but it complicates things a lot. Instead
414 * lets just do what we already stored, and
415 * we'll get it the next time we are called.
417 if (error == AOP_TRUNCATED_PAGE)
425 * i_size must be checked after PageUptodate.
427 isize = i_size_read(mapping->host);
428 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
429 if (unlikely(!isize || index > end_index))
433 * if this is the last page, see if we need to shrink
434 * the length and stop
436 if (end_index == index) {
440 * max good bytes in this page
442 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
447 * force quit after adding this page
449 this_len = min(this_len, plen - loff);
453 partial[page_nr].offset = loff;
454 partial[page_nr].len = this_len;
462 * Release any pages at the end, if we quit early. 'page_nr' is how far
463 * we got, 'nr_pages' is how many pages are in the map.
465 while (page_nr < nr_pages)
466 page_cache_release(pages[page_nr++]);
467 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
470 return splice_to_pipe(pipe, &spd);
476 * generic_file_splice_read - splice data from file to a pipe
477 * @in: file to splice from
478 * @ppos: position in @in
479 * @pipe: pipe to splice to
480 * @len: number of bytes to splice
481 * @flags: splice modifier flags
484 * Will read pages from given file and fill them into a pipe. Can be
485 * used as long as the address_space operations for the source implements
489 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
490 struct pipe_inode_info *pipe, size_t len,
496 isize = i_size_read(in->f_mapping->host);
497 if (unlikely(*ppos >= isize))
500 left = isize - *ppos;
501 if (unlikely(left < len))
504 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
510 EXPORT_SYMBOL(generic_file_splice_read);
512 static const struct pipe_buf_operations default_pipe_buf_ops = {
514 .map = generic_pipe_buf_map,
515 .unmap = generic_pipe_buf_unmap,
516 .confirm = generic_pipe_buf_confirm,
517 .release = generic_pipe_buf_release,
518 .steal = generic_pipe_buf_steal,
519 .get = generic_pipe_buf_get,
522 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
523 unsigned long vlen, loff_t offset)
531 /* The cast to a user pointer is valid due to the set_fs() */
532 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
538 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
546 /* The cast to a user pointer is valid due to the set_fs() */
547 res = vfs_write(file, (const char __user *)buf, count, &pos);
553 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
554 struct pipe_inode_info *pipe, size_t len,
557 unsigned int nr_pages;
558 unsigned int nr_freed;
560 struct page *pages[PIPE_BUFFERS];
561 struct partial_page partial[PIPE_BUFFERS];
562 struct iovec vec[PIPE_BUFFERS];
568 struct splice_pipe_desc spd = {
572 .ops = &default_pipe_buf_ops,
573 .spd_release = spd_release_page,
576 index = *ppos >> PAGE_CACHE_SHIFT;
577 offset = *ppos & ~PAGE_CACHE_MASK;
578 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
580 for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
583 page = alloc_page(GFP_USER);
588 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
589 vec[i].iov_base = (void __user *) page_address(page);
590 vec[i].iov_len = this_len;
597 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
608 for (i = 0; i < spd.nr_pages; i++) {
609 this_len = min_t(size_t, vec[i].iov_len, res);
610 partial[i].offset = 0;
611 partial[i].len = this_len;
613 __free_page(pages[i]);
619 spd.nr_pages -= nr_freed;
621 res = splice_to_pipe(pipe, &spd);
628 for (i = 0; i < spd.nr_pages; i++)
629 __free_page(pages[i]);
633 EXPORT_SYMBOL(default_file_splice_read);
636 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
637 * using sendpage(). Return the number of bytes sent.
639 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
640 struct pipe_buffer *buf, struct splice_desc *sd)
642 struct file *file = sd->u.file;
643 loff_t pos = sd->pos;
646 ret = buf->ops->confirm(pipe, buf);
648 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
650 ret = file->f_op->sendpage(file, buf->page, buf->offset,
651 sd->len, &pos, more);
658 * This is a little more tricky than the file -> pipe splicing. There are
659 * basically three cases:
661 * - Destination page already exists in the address space and there
662 * are users of it. For that case we have no other option that
663 * copying the data. Tough luck.
664 * - Destination page already exists in the address space, but there
665 * are no users of it. Make sure it's uptodate, then drop it. Fall
666 * through to last case.
667 * - Destination page does not exist, we can add the pipe page to
668 * the page cache and avoid the copy.
670 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
671 * sd->flags), we attempt to migrate pages from the pipe to the output
672 * file address space page cache. This is possible if no one else has
673 * the pipe page referenced outside of the pipe and page cache. If
674 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
675 * a new page in the output file page cache and fill/dirty that.
677 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
678 struct splice_desc *sd)
680 struct file *file = sd->u.file;
681 struct address_space *mapping = file->f_mapping;
682 unsigned int offset, this_len;
688 * make sure the data in this buffer is uptodate
690 ret = buf->ops->confirm(pipe, buf);
694 offset = sd->pos & ~PAGE_CACHE_MASK;
697 if (this_len + offset > PAGE_CACHE_SIZE)
698 this_len = PAGE_CACHE_SIZE - offset;
700 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
701 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
705 if (buf->page != page) {
707 * Careful, ->map() uses KM_USER0!
709 char *src = buf->ops->map(pipe, buf, 1);
710 char *dst = kmap_atomic(page, KM_USER1);
712 memcpy(dst + offset, src + buf->offset, this_len);
713 flush_dcache_page(page);
714 kunmap_atomic(dst, KM_USER1);
715 buf->ops->unmap(pipe, buf, src);
717 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
722 EXPORT_SYMBOL(pipe_to_file);
724 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
727 if (waitqueue_active(&pipe->wait))
728 wake_up_interruptible(&pipe->wait);
729 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
733 * splice_from_pipe_feed - feed available data from a pipe to a file
734 * @pipe: pipe to splice from
735 * @sd: information to @actor
736 * @actor: handler that splices the data
739 * This function loops over the pipe and calls @actor to do the
740 * actual moving of a single struct pipe_buffer to the desired
741 * destination. It returns when there's no more buffers left in
742 * the pipe or if the requested number of bytes (@sd->total_len)
743 * have been copied. It returns a positive number (one) if the
744 * pipe needs to be filled with more data, zero if the required
745 * number of bytes have been copied and -errno on error.
747 * This, together with splice_from_pipe_{begin,end,next}, may be
748 * used to implement the functionality of __splice_from_pipe() when
749 * locking is required around copying the pipe buffers to the
752 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
757 while (pipe->nrbufs) {
758 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
759 const struct pipe_buf_operations *ops = buf->ops;
762 if (sd->len > sd->total_len)
763 sd->len = sd->total_len;
765 ret = actor(pipe, buf, sd);
774 sd->num_spliced += ret;
777 sd->total_len -= ret;
781 ops->release(pipe, buf);
782 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
785 sd->need_wakeup = true;
794 EXPORT_SYMBOL(splice_from_pipe_feed);
797 * splice_from_pipe_next - wait for some data to splice from
798 * @pipe: pipe to splice from
799 * @sd: information about the splice operation
802 * This function will wait for some data and return a positive
803 * value (one) if pipe buffers are available. It will return zero
804 * or -errno if no more data needs to be spliced.
806 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
808 while (!pipe->nrbufs) {
812 if (!pipe->waiting_writers && sd->num_spliced)
815 if (sd->flags & SPLICE_F_NONBLOCK)
818 if (signal_pending(current))
821 if (sd->need_wakeup) {
822 wakeup_pipe_writers(pipe);
823 sd->need_wakeup = false;
831 EXPORT_SYMBOL(splice_from_pipe_next);
834 * splice_from_pipe_begin - start splicing from pipe
835 * @sd: information about the splice operation
838 * This function should be called before a loop containing
839 * splice_from_pipe_next() and splice_from_pipe_feed() to
840 * initialize the necessary fields of @sd.
842 void splice_from_pipe_begin(struct splice_desc *sd)
845 sd->need_wakeup = false;
847 EXPORT_SYMBOL(splice_from_pipe_begin);
850 * splice_from_pipe_end - finish splicing from pipe
851 * @pipe: pipe to splice from
852 * @sd: information about the splice operation
855 * This function will wake up pipe writers if necessary. It should
856 * be called after a loop containing splice_from_pipe_next() and
857 * splice_from_pipe_feed().
859 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
862 wakeup_pipe_writers(pipe);
864 EXPORT_SYMBOL(splice_from_pipe_end);
867 * __splice_from_pipe - splice data from a pipe to given actor
868 * @pipe: pipe to splice from
869 * @sd: information to @actor
870 * @actor: handler that splices the data
873 * This function does little more than loop over the pipe and call
874 * @actor to do the actual moving of a single struct pipe_buffer to
875 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
879 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
884 splice_from_pipe_begin(sd);
886 ret = splice_from_pipe_next(pipe, sd);
888 ret = splice_from_pipe_feed(pipe, sd, actor);
890 splice_from_pipe_end(pipe, sd);
892 return sd->num_spliced ? sd->num_spliced : ret;
894 EXPORT_SYMBOL(__splice_from_pipe);
897 * splice_from_pipe - splice data from a pipe to a file
898 * @pipe: pipe to splice from
899 * @out: file to splice to
900 * @ppos: position in @out
901 * @len: how many bytes to splice
902 * @flags: splice modifier flags
903 * @actor: handler that splices the data
906 * See __splice_from_pipe. This function locks the pipe inode,
907 * otherwise it's identical to __splice_from_pipe().
910 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
911 loff_t *ppos, size_t len, unsigned int flags,
915 struct splice_desc sd = {
923 ret = __splice_from_pipe(pipe, &sd, actor);
930 * generic_file_splice_write - splice data from a pipe to a file
932 * @out: file to write to
933 * @ppos: position in @out
934 * @len: number of bytes to splice
935 * @flags: splice modifier flags
938 * Will either move or copy pages (determined by @flags options) from
939 * the given pipe inode to the given file.
943 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
944 loff_t *ppos, size_t len, unsigned int flags)
946 struct address_space *mapping = out->f_mapping;
947 struct inode *inode = mapping->host;
948 struct splice_desc sd = {
958 splice_from_pipe_begin(&sd);
960 ret = splice_from_pipe_next(pipe, &sd);
964 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
965 ret = file_remove_suid(out);
967 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
968 mutex_unlock(&inode->i_mutex);
970 splice_from_pipe_end(pipe, &sd);
975 ret = sd.num_spliced;
978 unsigned long nr_pages;
981 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
984 * If file or inode is SYNC and we actually wrote some data,
987 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
990 mutex_lock(&inode->i_mutex);
991 err = generic_osync_inode(inode, mapping,
992 OSYNC_METADATA|OSYNC_DATA);
993 mutex_unlock(&inode->i_mutex);
998 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1004 EXPORT_SYMBOL(generic_file_splice_write);
1006 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1007 struct splice_desc *sd)
1012 ret = buf->ops->confirm(pipe, buf);
1016 data = buf->ops->map(pipe, buf, 0);
1017 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1018 buf->ops->unmap(pipe, buf, data);
1023 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1024 struct file *out, loff_t *ppos,
1025 size_t len, unsigned int flags)
1029 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1037 * generic_splice_sendpage - splice data from a pipe to a socket
1038 * @pipe: pipe to splice from
1039 * @out: socket to write to
1040 * @ppos: position in @out
1041 * @len: number of bytes to splice
1042 * @flags: splice modifier flags
1045 * Will send @len bytes from the pipe to a network socket. No data copying
1049 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1050 loff_t *ppos, size_t len, unsigned int flags)
1052 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1055 EXPORT_SYMBOL(generic_splice_sendpage);
1058 * Attempt to initiate a splice from pipe to file.
1060 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1061 loff_t *ppos, size_t len, unsigned int flags)
1063 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1064 loff_t *, size_t, unsigned int);
1067 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1070 if (unlikely(out->f_flags & O_APPEND))
1073 ret = rw_verify_area(WRITE, out, ppos, len);
1074 if (unlikely(ret < 0))
1077 splice_write = out->f_op->splice_write;
1079 splice_write = default_file_splice_write;
1081 return splice_write(pipe, out, ppos, len, flags);
1085 * Attempt to initiate a splice from a file to a pipe.
1087 static long do_splice_to(struct file *in, loff_t *ppos,
1088 struct pipe_inode_info *pipe, size_t len,
1091 ssize_t (*splice_read)(struct file *, loff_t *,
1092 struct pipe_inode_info *, size_t, unsigned int);
1095 if (unlikely(!(in->f_mode & FMODE_READ)))
1098 ret = rw_verify_area(READ, in, ppos, len);
1099 if (unlikely(ret < 0))
1102 splice_read = in->f_op->splice_read;
1104 splice_read = default_file_splice_read;
1106 return splice_read(in, ppos, pipe, len, flags);
1110 * splice_direct_to_actor - splices data directly between two non-pipes
1111 * @in: file to splice from
1112 * @sd: actor information on where to splice to
1113 * @actor: handles the data splicing
1116 * This is a special case helper to splice directly between two
1117 * points, without requiring an explicit pipe. Internally an allocated
1118 * pipe is cached in the process, and reused during the lifetime of
1122 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1123 splice_direct_actor *actor)
1125 struct pipe_inode_info *pipe;
1132 * We require the input being a regular file, as we don't want to
1133 * randomly drop data for eg socket -> socket splicing. Use the
1134 * piped splicing for that!
1136 i_mode = in->f_path.dentry->d_inode->i_mode;
1137 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1141 * neither in nor out is a pipe, setup an internal pipe attached to
1142 * 'out' and transfer the wanted data from 'in' to 'out' through that
1144 pipe = current->splice_pipe;
1145 if (unlikely(!pipe)) {
1146 pipe = alloc_pipe_info(NULL);
1151 * We don't have an immediate reader, but we'll read the stuff
1152 * out of the pipe right after the splice_to_pipe(). So set
1153 * PIPE_READERS appropriately.
1157 current->splice_pipe = pipe;
1165 len = sd->total_len;
1169 * Don't block on output, we have to drain the direct pipe.
1171 sd->flags &= ~SPLICE_F_NONBLOCK;
1175 loff_t pos = sd->pos, prev_pos = pos;
1177 ret = do_splice_to(in, &pos, pipe, len, flags);
1178 if (unlikely(ret <= 0))
1182 sd->total_len = read_len;
1185 * NOTE: nonblocking mode only applies to the input. We
1186 * must not do the output in nonblocking mode as then we
1187 * could get stuck data in the internal pipe:
1189 ret = actor(pipe, sd);
1190 if (unlikely(ret <= 0)) {
1199 if (ret < read_len) {
1200 sd->pos = prev_pos + ret;
1206 pipe->nrbufs = pipe->curbuf = 0;
1212 * If we did an incomplete transfer we must release
1213 * the pipe buffers in question:
1215 for (i = 0; i < PIPE_BUFFERS; i++) {
1216 struct pipe_buffer *buf = pipe->bufs + i;
1219 buf->ops->release(pipe, buf);
1229 EXPORT_SYMBOL(splice_direct_to_actor);
1231 static int direct_splice_actor(struct pipe_inode_info *pipe,
1232 struct splice_desc *sd)
1234 struct file *file = sd->u.file;
1236 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1240 * do_splice_direct - splices data directly between two files
1241 * @in: file to splice from
1242 * @ppos: input file offset
1243 * @out: file to splice to
1244 * @len: number of bytes to splice
1245 * @flags: splice modifier flags
1248 * For use by do_sendfile(). splice can easily emulate sendfile, but
1249 * doing it in the application would incur an extra system call
1250 * (splice in + splice out, as compared to just sendfile()). So this helper
1251 * can splice directly through a process-private pipe.
1254 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1255 size_t len, unsigned int flags)
1257 struct splice_desc sd = {
1266 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1273 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1274 struct pipe_inode_info *opipe,
1275 size_t len, unsigned int flags);
1277 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1278 * location, so checking ->i_pipe is not enough to verify that this is a
1281 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1283 if (S_ISFIFO(inode->i_mode))
1284 return inode->i_pipe;
1290 * Determine where to splice to/from.
1292 static long do_splice(struct file *in, loff_t __user *off_in,
1293 struct file *out, loff_t __user *off_out,
1294 size_t len, unsigned int flags)
1296 struct pipe_inode_info *ipipe;
1297 struct pipe_inode_info *opipe;
1298 loff_t offset, *off;
1301 ipipe = pipe_info(in->f_path.dentry->d_inode);
1302 opipe = pipe_info(out->f_path.dentry->d_inode);
1304 if (ipipe && opipe) {
1305 if (off_in || off_out)
1308 if (!(in->f_mode & FMODE_READ))
1311 if (!(out->f_mode & FMODE_WRITE))
1314 /* Splicing to self would be fun, but... */
1318 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1325 if (out->f_op->llseek == no_llseek)
1327 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1333 ret = do_splice_from(ipipe, out, off, len, flags);
1335 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1345 if (in->f_op->llseek == no_llseek)
1347 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1353 ret = do_splice_to(in, off, opipe, len, flags);
1355 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1365 * Map an iov into an array of pages and offset/length tupples. With the
1366 * partial_page structure, we can map several non-contiguous ranges into
1367 * our ones pages[] map instead of splitting that operation into pieces.
1368 * Could easily be exported as a generic helper for other users, in which
1369 * case one would probably want to add a 'max_nr_pages' parameter as well.
1371 static int get_iovec_page_array(const struct iovec __user *iov,
1372 unsigned int nr_vecs, struct page **pages,
1373 struct partial_page *partial, int aligned)
1375 int buffers = 0, error = 0;
1378 unsigned long off, npages;
1385 if (copy_from_user(&entry, iov, sizeof(entry)))
1388 base = entry.iov_base;
1389 len = entry.iov_len;
1392 * Sanity check this iovec. 0 read succeeds.
1398 if (!access_ok(VERIFY_READ, base, len))
1402 * Get this base offset and number of pages, then map
1403 * in the user pages.
1405 off = (unsigned long) base & ~PAGE_MASK;
1408 * If asked for alignment, the offset must be zero and the
1409 * length a multiple of the PAGE_SIZE.
1412 if (aligned && (off || len & ~PAGE_MASK))
1415 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1416 if (npages > PIPE_BUFFERS - buffers)
1417 npages = PIPE_BUFFERS - buffers;
1419 error = get_user_pages_fast((unsigned long)base, npages,
1420 0, &pages[buffers]);
1422 if (unlikely(error <= 0))
1426 * Fill this contiguous range into the partial page map.
1428 for (i = 0; i < error; i++) {
1429 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1431 partial[buffers].offset = off;
1432 partial[buffers].len = plen;
1440 * We didn't complete this iov, stop here since it probably
1441 * means we have to move some of this into a pipe to
1442 * be able to continue.
1448 * Don't continue if we mapped fewer pages than we asked for,
1449 * or if we mapped the max number of pages that we have
1452 if (error < npages || buffers == PIPE_BUFFERS)
1465 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1466 struct splice_desc *sd)
1471 ret = buf->ops->confirm(pipe, buf);
1476 * See if we can use the atomic maps, by prefaulting in the
1477 * pages and doing an atomic copy
1479 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1480 src = buf->ops->map(pipe, buf, 1);
1481 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1483 buf->ops->unmap(pipe, buf, src);
1491 * No dice, use slow non-atomic map and copy
1493 src = buf->ops->map(pipe, buf, 0);
1496 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1499 buf->ops->unmap(pipe, buf, src);
1502 sd->u.userptr += ret;
1507 * For lack of a better implementation, implement vmsplice() to userspace
1508 * as a simple copy of the pipes pages to the user iov.
1510 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1511 unsigned long nr_segs, unsigned int flags)
1513 struct pipe_inode_info *pipe;
1514 struct splice_desc sd;
1519 pipe = pipe_info(file->f_path.dentry->d_inode);
1531 * Get user address base and length for this iovec.
1533 error = get_user(base, &iov->iov_base);
1534 if (unlikely(error))
1536 error = get_user(len, &iov->iov_len);
1537 if (unlikely(error))
1541 * Sanity check this iovec. 0 read succeeds.
1545 if (unlikely(!base)) {
1550 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1558 sd.u.userptr = base;
1561 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1587 * vmsplice splices a user address range into a pipe. It can be thought of
1588 * as splice-from-memory, where the regular splice is splice-from-file (or
1589 * to file). In both cases the output is a pipe, naturally.
1591 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1592 unsigned long nr_segs, unsigned int flags)
1594 struct pipe_inode_info *pipe;
1595 struct page *pages[PIPE_BUFFERS];
1596 struct partial_page partial[PIPE_BUFFERS];
1597 struct splice_pipe_desc spd = {
1601 .ops = &user_page_pipe_buf_ops,
1602 .spd_release = spd_release_page,
1605 pipe = pipe_info(file->f_path.dentry->d_inode);
1609 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1610 flags & SPLICE_F_GIFT);
1611 if (spd.nr_pages <= 0)
1612 return spd.nr_pages;
1614 return splice_to_pipe(pipe, &spd);
1618 * Note that vmsplice only really supports true splicing _from_ user memory
1619 * to a pipe, not the other way around. Splicing from user memory is a simple
1620 * operation that can be supported without any funky alignment restrictions
1621 * or nasty vm tricks. We simply map in the user memory and fill them into
1622 * a pipe. The reverse isn't quite as easy, though. There are two possible
1623 * solutions for that:
1625 * - memcpy() the data internally, at which point we might as well just
1626 * do a regular read() on the buffer anyway.
1627 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1628 * has restriction limitations on both ends of the pipe).
1630 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1633 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1634 unsigned long, nr_segs, unsigned int, flags)
1640 if (unlikely(nr_segs > UIO_MAXIOV))
1642 else if (unlikely(!nr_segs))
1646 file = fget_light(fd, &fput);
1648 if (file->f_mode & FMODE_WRITE)
1649 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1650 else if (file->f_mode & FMODE_READ)
1651 error = vmsplice_to_user(file, iov, nr_segs, flags);
1653 fput_light(file, fput);
1659 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1660 int, fd_out, loff_t __user *, off_out,
1661 size_t, len, unsigned int, flags)
1664 struct file *in, *out;
1665 int fput_in, fput_out;
1671 in = fget_light(fd_in, &fput_in);
1673 if (in->f_mode & FMODE_READ) {
1674 out = fget_light(fd_out, &fput_out);
1676 if (out->f_mode & FMODE_WRITE)
1677 error = do_splice(in, off_in,
1680 fput_light(out, fput_out);
1684 fput_light(in, fput_in);
1691 * Make sure there's data to read. Wait for input if we can, otherwise
1692 * return an appropriate error.
1694 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1699 * Check ->nrbufs without the inode lock first. This function
1700 * is speculative anyways, so missing one is ok.
1708 while (!pipe->nrbufs) {
1709 if (signal_pending(current)) {
1715 if (!pipe->waiting_writers) {
1716 if (flags & SPLICE_F_NONBLOCK) {
1729 * Make sure there's writeable room. Wait for room if we can, otherwise
1730 * return an appropriate error.
1732 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1737 * Check ->nrbufs without the inode lock first. This function
1738 * is speculative anyways, so missing one is ok.
1740 if (pipe->nrbufs < PIPE_BUFFERS)
1746 while (pipe->nrbufs >= PIPE_BUFFERS) {
1747 if (!pipe->readers) {
1748 send_sig(SIGPIPE, current, 0);
1752 if (flags & SPLICE_F_NONBLOCK) {
1756 if (signal_pending(current)) {
1760 pipe->waiting_writers++;
1762 pipe->waiting_writers--;
1770 * Splice contents of ipipe to opipe.
1772 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1773 struct pipe_inode_info *opipe,
1774 size_t len, unsigned int flags)
1776 struct pipe_buffer *ibuf, *obuf;
1778 bool input_wakeup = false;
1782 ret = ipipe_prep(ipipe, flags);
1786 ret = opipe_prep(opipe, flags);
1791 * Potential ABBA deadlock, work around it by ordering lock
1792 * grabbing by pipe info address. Otherwise two different processes
1793 * could deadlock (one doing tee from A -> B, the other from B -> A).
1795 pipe_double_lock(ipipe, opipe);
1798 if (!opipe->readers) {
1799 send_sig(SIGPIPE, current, 0);
1805 if (!ipipe->nrbufs && !ipipe->writers)
1809 * Cannot make any progress, because either the input
1810 * pipe is empty or the output pipe is full.
1812 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1813 /* Already processed some buffers, break */
1817 if (flags & SPLICE_F_NONBLOCK) {
1823 * We raced with another reader/writer and haven't
1824 * managed to process any buffers. A zero return
1825 * value means EOF, so retry instead.
1832 ibuf = ipipe->bufs + ipipe->curbuf;
1833 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1834 obuf = opipe->bufs + nbuf;
1836 if (len >= ibuf->len) {
1838 * Simply move the whole buffer from ipipe to opipe
1843 ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1845 input_wakeup = true;
1848 * Get a reference to this pipe buffer,
1849 * so we can copy the contents over.
1851 ibuf->ops->get(ipipe, ibuf);
1855 * Don't inherit the gift flag, we need to
1856 * prevent multiple steals of this page.
1858 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1862 ibuf->offset += obuf->len;
1863 ibuf->len -= obuf->len;
1873 * If we put data in the output pipe, wakeup any potential readers.
1877 if (waitqueue_active(&opipe->wait))
1878 wake_up_interruptible(&opipe->wait);
1879 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1882 wakeup_pipe_writers(ipipe);
1888 * Link contents of ipipe to opipe.
1890 static int link_pipe(struct pipe_inode_info *ipipe,
1891 struct pipe_inode_info *opipe,
1892 size_t len, unsigned int flags)
1894 struct pipe_buffer *ibuf, *obuf;
1895 int ret = 0, i = 0, nbuf;
1898 * Potential ABBA deadlock, work around it by ordering lock
1899 * grabbing by pipe info address. Otherwise two different processes
1900 * could deadlock (one doing tee from A -> B, the other from B -> A).
1902 pipe_double_lock(ipipe, opipe);
1905 if (!opipe->readers) {
1906 send_sig(SIGPIPE, current, 0);
1913 * If we have iterated all input buffers or ran out of
1914 * output room, break.
1916 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1919 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1920 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1923 * Get a reference to this pipe buffer,
1924 * so we can copy the contents over.
1926 ibuf->ops->get(ipipe, ibuf);
1928 obuf = opipe->bufs + nbuf;
1932 * Don't inherit the gift flag, we need to
1933 * prevent multiple steals of this page.
1935 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1937 if (obuf->len > len)
1947 * return EAGAIN if we have the potential of some data in the
1948 * future, otherwise just return 0
1950 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1957 * If we put data in the output pipe, wakeup any potential readers.
1961 if (waitqueue_active(&opipe->wait))
1962 wake_up_interruptible(&opipe->wait);
1963 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1970 * This is a tee(1) implementation that works on pipes. It doesn't copy
1971 * any data, it simply references the 'in' pages on the 'out' pipe.
1972 * The 'flags' used are the SPLICE_F_* variants, currently the only
1973 * applicable one is SPLICE_F_NONBLOCK.
1975 static long do_tee(struct file *in, struct file *out, size_t len,
1978 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1979 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1983 * Duplicate the contents of ipipe to opipe without actually
1986 if (ipipe && opipe && ipipe != opipe) {
1988 * Keep going, unless we encounter an error. The ipipe/opipe
1989 * ordering doesn't really matter.
1991 ret = ipipe_prep(ipipe, flags);
1993 ret = opipe_prep(opipe, flags);
1995 ret = link_pipe(ipipe, opipe, len, flags);
2002 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2011 in = fget_light(fdin, &fput_in);
2013 if (in->f_mode & FMODE_READ) {
2015 struct file *out = fget_light(fdout, &fput_out);
2018 if (out->f_mode & FMODE_WRITE)
2019 error = do_tee(in, out, len, flags);
2020 fput_light(out, fput_out);
2023 fput_light(in, fput_in);