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_writev(struct file *file, const struct iovec *vec,
539 unsigned long vlen, loff_t *ppos)
546 /* The cast to a user pointer is valid due to the set_fs() */
547 res = vfs_writev(file, (const struct iovec __user *)vec, vlen, ppos);
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);
606 for (i = 0; i < spd.nr_pages; i++) {
607 this_len = min_t(size_t, vec[i].iov_len, res);
608 partial[i].offset = 0;
609 partial[i].len = this_len;
611 __free_page(pages[i]);
617 spd.nr_pages -= nr_freed;
619 res = splice_to_pipe(pipe, &spd);
626 for (i = 0; i < spd.nr_pages; i++)
627 __free_page(pages[i]);
631 EXPORT_SYMBOL(default_file_splice_read);
634 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
635 * using sendpage(). Return the number of bytes sent.
637 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
638 struct pipe_buffer *buf, struct splice_desc *sd)
640 struct file *file = sd->u.file;
641 loff_t pos = sd->pos;
644 ret = buf->ops->confirm(pipe, buf);
646 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
648 ret = file->f_op->sendpage(file, buf->page, buf->offset,
649 sd->len, &pos, more);
656 * This is a little more tricky than the file -> pipe splicing. There are
657 * basically three cases:
659 * - Destination page already exists in the address space and there
660 * are users of it. For that case we have no other option that
661 * copying the data. Tough luck.
662 * - Destination page already exists in the address space, but there
663 * are no users of it. Make sure it's uptodate, then drop it. Fall
664 * through to last case.
665 * - Destination page does not exist, we can add the pipe page to
666 * the page cache and avoid the copy.
668 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
669 * sd->flags), we attempt to migrate pages from the pipe to the output
670 * file address space page cache. This is possible if no one else has
671 * the pipe page referenced outside of the pipe and page cache. If
672 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
673 * a new page in the output file page cache and fill/dirty that.
675 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
676 struct splice_desc *sd)
678 struct file *file = sd->u.file;
679 struct address_space *mapping = file->f_mapping;
680 unsigned int offset, this_len;
686 * make sure the data in this buffer is uptodate
688 ret = buf->ops->confirm(pipe, buf);
692 offset = sd->pos & ~PAGE_CACHE_MASK;
695 if (this_len + offset > PAGE_CACHE_SIZE)
696 this_len = PAGE_CACHE_SIZE - offset;
698 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
699 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
703 if (buf->page != page) {
705 * Careful, ->map() uses KM_USER0!
707 char *src = buf->ops->map(pipe, buf, 1);
708 char *dst = kmap_atomic(page, KM_USER1);
710 memcpy(dst + offset, src + buf->offset, this_len);
711 flush_dcache_page(page);
712 kunmap_atomic(dst, KM_USER1);
713 buf->ops->unmap(pipe, buf, src);
715 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
720 EXPORT_SYMBOL(pipe_to_file);
722 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
725 if (waitqueue_active(&pipe->wait))
726 wake_up_interruptible(&pipe->wait);
727 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
731 * splice_from_pipe_feed - feed available data from a pipe to a file
732 * @pipe: pipe to splice from
733 * @sd: information to @actor
734 * @actor: handler that splices the data
737 * This function loops over the pipe and calls @actor to do the
738 * actual moving of a single struct pipe_buffer to the desired
739 * destination. It returns when there's no more buffers left in
740 * the pipe or if the requested number of bytes (@sd->total_len)
741 * have been copied. It returns a positive number (one) if the
742 * pipe needs to be filled with more data, zero if the required
743 * number of bytes have been copied and -errno on error.
745 * This, together with splice_from_pipe_{begin,end,next}, may be
746 * used to implement the functionality of __splice_from_pipe() when
747 * locking is required around copying the pipe buffers to the
750 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
755 while (pipe->nrbufs) {
756 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
757 const struct pipe_buf_operations *ops = buf->ops;
760 if (sd->len > sd->total_len)
761 sd->len = sd->total_len;
763 ret = actor(pipe, buf, sd);
772 sd->num_spliced += ret;
775 sd->total_len -= ret;
779 ops->release(pipe, buf);
780 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
783 sd->need_wakeup = true;
792 EXPORT_SYMBOL(splice_from_pipe_feed);
795 * splice_from_pipe_next - wait for some data to splice from
796 * @pipe: pipe to splice from
797 * @sd: information about the splice operation
800 * This function will wait for some data and return a positive
801 * value (one) if pipe buffers are available. It will return zero
802 * or -errno if no more data needs to be spliced.
804 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
806 while (!pipe->nrbufs) {
810 if (!pipe->waiting_writers && sd->num_spliced)
813 if (sd->flags & SPLICE_F_NONBLOCK)
816 if (signal_pending(current))
819 if (sd->need_wakeup) {
820 wakeup_pipe_writers(pipe);
821 sd->need_wakeup = false;
829 EXPORT_SYMBOL(splice_from_pipe_next);
832 * splice_from_pipe_begin - start splicing from pipe
833 * @sd: information about the splice operation
836 * This function should be called before a loop containing
837 * splice_from_pipe_next() and splice_from_pipe_feed() to
838 * initialize the necessary fields of @sd.
840 void splice_from_pipe_begin(struct splice_desc *sd)
843 sd->need_wakeup = false;
845 EXPORT_SYMBOL(splice_from_pipe_begin);
848 * splice_from_pipe_end - finish splicing from pipe
849 * @pipe: pipe to splice from
850 * @sd: information about the splice operation
853 * This function will wake up pipe writers if necessary. It should
854 * be called after a loop containing splice_from_pipe_next() and
855 * splice_from_pipe_feed().
857 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
860 wakeup_pipe_writers(pipe);
862 EXPORT_SYMBOL(splice_from_pipe_end);
865 * __splice_from_pipe - splice data from a pipe to given actor
866 * @pipe: pipe to splice from
867 * @sd: information to @actor
868 * @actor: handler that splices the data
871 * This function does little more than loop over the pipe and call
872 * @actor to do the actual moving of a single struct pipe_buffer to
873 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
877 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
882 splice_from_pipe_begin(sd);
884 ret = splice_from_pipe_next(pipe, sd);
886 ret = splice_from_pipe_feed(pipe, sd, actor);
888 splice_from_pipe_end(pipe, sd);
890 return sd->num_spliced ? sd->num_spliced : ret;
892 EXPORT_SYMBOL(__splice_from_pipe);
895 * splice_from_pipe - splice data from a pipe to a file
896 * @pipe: pipe to splice from
897 * @out: file to splice to
898 * @ppos: position in @out
899 * @len: how many bytes to splice
900 * @flags: splice modifier flags
901 * @actor: handler that splices the data
904 * See __splice_from_pipe. This function locks the pipe inode,
905 * otherwise it's identical to __splice_from_pipe().
908 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
909 loff_t *ppos, size_t len, unsigned int flags,
913 struct splice_desc sd = {
921 ret = __splice_from_pipe(pipe, &sd, actor);
928 * generic_file_splice_write - splice data from a pipe to a file
930 * @out: file to write to
931 * @ppos: position in @out
932 * @len: number of bytes to splice
933 * @flags: splice modifier flags
936 * Will either move or copy pages (determined by @flags options) from
937 * the given pipe inode to the given file.
941 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
942 loff_t *ppos, size_t len, unsigned int flags)
944 struct address_space *mapping = out->f_mapping;
945 struct inode *inode = mapping->host;
946 struct splice_desc sd = {
956 splice_from_pipe_begin(&sd);
958 ret = splice_from_pipe_next(pipe, &sd);
962 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
963 ret = file_remove_suid(out);
965 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
966 mutex_unlock(&inode->i_mutex);
968 splice_from_pipe_end(pipe, &sd);
973 ret = sd.num_spliced;
976 unsigned long nr_pages;
979 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
982 * If file or inode is SYNC and we actually wrote some data,
985 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
988 mutex_lock(&inode->i_mutex);
989 err = generic_osync_inode(inode, mapping,
990 OSYNC_METADATA|OSYNC_DATA);
991 mutex_unlock(&inode->i_mutex);
996 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1002 EXPORT_SYMBOL(generic_file_splice_write);
1004 static struct pipe_buffer *nth_pipe_buf(struct pipe_inode_info *pipe, int n)
1006 return &pipe->bufs[(pipe->curbuf + n) % PIPE_BUFFERS];
1009 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1010 struct file *out, loff_t *ppos,
1011 size_t len, unsigned int flags)
1014 ssize_t total_len = 0;
1019 struct pipe_buffer *buf;
1020 void *data[PIPE_BUFFERS];
1021 struct iovec vec[PIPE_BUFFERS];
1022 unsigned int nr_pages = 0;
1023 unsigned int write_len = 0;
1024 unsigned int now_len = len;
1025 unsigned int this_len;
1028 BUG_ON(pipe->nrbufs > PIPE_BUFFERS);
1029 for (i = 0; i < pipe->nrbufs && now_len; i++) {
1030 buf = nth_pipe_buf(pipe, i);
1032 ret = buf->ops->confirm(pipe, buf);
1036 data[i] = buf->ops->map(pipe, buf, 0);
1037 this_len = min(buf->len, now_len);
1038 vec[i].iov_base = (void __user *) data[i] + buf->offset;
1039 vec[i].iov_len = this_len;
1040 now_len -= this_len;
1041 write_len += this_len;
1046 ret = kernel_writev(out, vec, nr_pages, ppos);
1055 for (i = 0; i < nr_pages; i++) {
1056 buf = nth_pipe_buf(pipe, i);
1057 buf->ops->unmap(pipe, buf, data[i]);
1060 this_len = min_t(unsigned, vec[i].iov_len, ret);
1061 buf->offset += this_len;
1062 buf->len -= this_len;
1070 while (pipe->nrbufs) {
1071 const struct pipe_buf_operations *ops;
1073 buf = nth_pipe_buf(pipe, 0);
1079 ops->release(pipe, buf);
1080 pipe->curbuf = (pipe->curbuf + 1) % PIPE_BUFFERS;
1090 if (!pipe->waiting_writers) {
1095 if (flags & SPLICE_F_NONBLOCK) {
1100 if (signal_pending(current)) {
1106 wakeup_pipe_writers(pipe);
1115 wakeup_pipe_writers(pipe);
1117 return total_len ? total_len : ret;
1121 * generic_splice_sendpage - splice data from a pipe to a socket
1122 * @pipe: pipe to splice from
1123 * @out: socket to write to
1124 * @ppos: position in @out
1125 * @len: number of bytes to splice
1126 * @flags: splice modifier flags
1129 * Will send @len bytes from the pipe to a network socket. No data copying
1133 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1134 loff_t *ppos, size_t len, unsigned int flags)
1136 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1139 EXPORT_SYMBOL(generic_splice_sendpage);
1142 * Attempt to initiate a splice from pipe to file.
1144 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1145 loff_t *ppos, size_t len, unsigned int flags)
1147 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1148 loff_t *, size_t, unsigned int);
1151 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1154 if (unlikely(out->f_flags & O_APPEND))
1157 ret = rw_verify_area(WRITE, out, ppos, len);
1158 if (unlikely(ret < 0))
1161 splice_write = out->f_op->splice_write;
1163 splice_write = default_file_splice_write;
1165 return splice_write(pipe, out, ppos, len, flags);
1169 * Attempt to initiate a splice from a file to a pipe.
1171 static long do_splice_to(struct file *in, loff_t *ppos,
1172 struct pipe_inode_info *pipe, size_t len,
1175 ssize_t (*splice_read)(struct file *, loff_t *,
1176 struct pipe_inode_info *, size_t, unsigned int);
1179 if (unlikely(!(in->f_mode & FMODE_READ)))
1182 ret = rw_verify_area(READ, in, ppos, len);
1183 if (unlikely(ret < 0))
1186 splice_read = in->f_op->splice_read;
1188 splice_read = default_file_splice_read;
1190 return splice_read(in, ppos, pipe, len, flags);
1194 * splice_direct_to_actor - splices data directly between two non-pipes
1195 * @in: file to splice from
1196 * @sd: actor information on where to splice to
1197 * @actor: handles the data splicing
1200 * This is a special case helper to splice directly between two
1201 * points, without requiring an explicit pipe. Internally an allocated
1202 * pipe is cached in the process, and reused during the lifetime of
1206 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1207 splice_direct_actor *actor)
1209 struct pipe_inode_info *pipe;
1216 * We require the input being a regular file, as we don't want to
1217 * randomly drop data for eg socket -> socket splicing. Use the
1218 * piped splicing for that!
1220 i_mode = in->f_path.dentry->d_inode->i_mode;
1221 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1225 * neither in nor out is a pipe, setup an internal pipe attached to
1226 * 'out' and transfer the wanted data from 'in' to 'out' through that
1228 pipe = current->splice_pipe;
1229 if (unlikely(!pipe)) {
1230 pipe = alloc_pipe_info(NULL);
1235 * We don't have an immediate reader, but we'll read the stuff
1236 * out of the pipe right after the splice_to_pipe(). So set
1237 * PIPE_READERS appropriately.
1241 current->splice_pipe = pipe;
1249 len = sd->total_len;
1253 * Don't block on output, we have to drain the direct pipe.
1255 sd->flags &= ~SPLICE_F_NONBLOCK;
1259 loff_t pos = sd->pos, prev_pos = pos;
1261 ret = do_splice_to(in, &pos, pipe, len, flags);
1262 if (unlikely(ret <= 0))
1266 sd->total_len = read_len;
1269 * NOTE: nonblocking mode only applies to the input. We
1270 * must not do the output in nonblocking mode as then we
1271 * could get stuck data in the internal pipe:
1273 ret = actor(pipe, sd);
1274 if (unlikely(ret <= 0)) {
1283 if (ret < read_len) {
1284 sd->pos = prev_pos + ret;
1290 pipe->nrbufs = pipe->curbuf = 0;
1296 * If we did an incomplete transfer we must release
1297 * the pipe buffers in question:
1299 for (i = 0; i < PIPE_BUFFERS; i++) {
1300 struct pipe_buffer *buf = pipe->bufs + i;
1303 buf->ops->release(pipe, buf);
1313 EXPORT_SYMBOL(splice_direct_to_actor);
1315 static int direct_splice_actor(struct pipe_inode_info *pipe,
1316 struct splice_desc *sd)
1318 struct file *file = sd->u.file;
1320 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1324 * do_splice_direct - splices data directly between two files
1325 * @in: file to splice from
1326 * @ppos: input file offset
1327 * @out: file to splice to
1328 * @len: number of bytes to splice
1329 * @flags: splice modifier flags
1332 * For use by do_sendfile(). splice can easily emulate sendfile, but
1333 * doing it in the application would incur an extra system call
1334 * (splice in + splice out, as compared to just sendfile()). So this helper
1335 * can splice directly through a process-private pipe.
1338 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1339 size_t len, unsigned int flags)
1341 struct splice_desc sd = {
1350 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1357 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1358 struct pipe_inode_info *opipe,
1359 size_t len, unsigned int flags);
1361 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1362 * location, so checking ->i_pipe is not enough to verify that this is a
1365 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1367 if (S_ISFIFO(inode->i_mode))
1368 return inode->i_pipe;
1374 * Determine where to splice to/from.
1376 static long do_splice(struct file *in, loff_t __user *off_in,
1377 struct file *out, loff_t __user *off_out,
1378 size_t len, unsigned int flags)
1380 struct pipe_inode_info *ipipe;
1381 struct pipe_inode_info *opipe;
1382 loff_t offset, *off;
1385 ipipe = pipe_info(in->f_path.dentry->d_inode);
1386 opipe = pipe_info(out->f_path.dentry->d_inode);
1388 if (ipipe && opipe) {
1389 if (off_in || off_out)
1392 if (!(in->f_mode & FMODE_READ))
1395 if (!(out->f_mode & FMODE_WRITE))
1398 /* Splicing to self would be fun, but... */
1402 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1409 if (out->f_op->llseek == no_llseek)
1411 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1417 ret = do_splice_from(ipipe, out, off, len, flags);
1419 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1429 if (in->f_op->llseek == no_llseek)
1431 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1437 ret = do_splice_to(in, off, opipe, len, flags);
1439 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1449 * Map an iov into an array of pages and offset/length tupples. With the
1450 * partial_page structure, we can map several non-contiguous ranges into
1451 * our ones pages[] map instead of splitting that operation into pieces.
1452 * Could easily be exported as a generic helper for other users, in which
1453 * case one would probably want to add a 'max_nr_pages' parameter as well.
1455 static int get_iovec_page_array(const struct iovec __user *iov,
1456 unsigned int nr_vecs, struct page **pages,
1457 struct partial_page *partial, int aligned)
1459 int buffers = 0, error = 0;
1462 unsigned long off, npages;
1469 if (copy_from_user(&entry, iov, sizeof(entry)))
1472 base = entry.iov_base;
1473 len = entry.iov_len;
1476 * Sanity check this iovec. 0 read succeeds.
1482 if (!access_ok(VERIFY_READ, base, len))
1486 * Get this base offset and number of pages, then map
1487 * in the user pages.
1489 off = (unsigned long) base & ~PAGE_MASK;
1492 * If asked for alignment, the offset must be zero and the
1493 * length a multiple of the PAGE_SIZE.
1496 if (aligned && (off || len & ~PAGE_MASK))
1499 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1500 if (npages > PIPE_BUFFERS - buffers)
1501 npages = PIPE_BUFFERS - buffers;
1503 error = get_user_pages_fast((unsigned long)base, npages,
1504 0, &pages[buffers]);
1506 if (unlikely(error <= 0))
1510 * Fill this contiguous range into the partial page map.
1512 for (i = 0; i < error; i++) {
1513 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1515 partial[buffers].offset = off;
1516 partial[buffers].len = plen;
1524 * We didn't complete this iov, stop here since it probably
1525 * means we have to move some of this into a pipe to
1526 * be able to continue.
1532 * Don't continue if we mapped fewer pages than we asked for,
1533 * or if we mapped the max number of pages that we have
1536 if (error < npages || buffers == PIPE_BUFFERS)
1549 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1550 struct splice_desc *sd)
1555 ret = buf->ops->confirm(pipe, buf);
1560 * See if we can use the atomic maps, by prefaulting in the
1561 * pages and doing an atomic copy
1563 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1564 src = buf->ops->map(pipe, buf, 1);
1565 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1567 buf->ops->unmap(pipe, buf, src);
1575 * No dice, use slow non-atomic map and copy
1577 src = buf->ops->map(pipe, buf, 0);
1580 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1583 buf->ops->unmap(pipe, buf, src);
1586 sd->u.userptr += ret;
1591 * For lack of a better implementation, implement vmsplice() to userspace
1592 * as a simple copy of the pipes pages to the user iov.
1594 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1595 unsigned long nr_segs, unsigned int flags)
1597 struct pipe_inode_info *pipe;
1598 struct splice_desc sd;
1603 pipe = pipe_info(file->f_path.dentry->d_inode);
1615 * Get user address base and length for this iovec.
1617 error = get_user(base, &iov->iov_base);
1618 if (unlikely(error))
1620 error = get_user(len, &iov->iov_len);
1621 if (unlikely(error))
1625 * Sanity check this iovec. 0 read succeeds.
1629 if (unlikely(!base)) {
1634 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1642 sd.u.userptr = base;
1645 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1671 * vmsplice splices a user address range into a pipe. It can be thought of
1672 * as splice-from-memory, where the regular splice is splice-from-file (or
1673 * to file). In both cases the output is a pipe, naturally.
1675 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1676 unsigned long nr_segs, unsigned int flags)
1678 struct pipe_inode_info *pipe;
1679 struct page *pages[PIPE_BUFFERS];
1680 struct partial_page partial[PIPE_BUFFERS];
1681 struct splice_pipe_desc spd = {
1685 .ops = &user_page_pipe_buf_ops,
1686 .spd_release = spd_release_page,
1689 pipe = pipe_info(file->f_path.dentry->d_inode);
1693 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1694 flags & SPLICE_F_GIFT);
1695 if (spd.nr_pages <= 0)
1696 return spd.nr_pages;
1698 return splice_to_pipe(pipe, &spd);
1702 * Note that vmsplice only really supports true splicing _from_ user memory
1703 * to a pipe, not the other way around. Splicing from user memory is a simple
1704 * operation that can be supported without any funky alignment restrictions
1705 * or nasty vm tricks. We simply map in the user memory and fill them into
1706 * a pipe. The reverse isn't quite as easy, though. There are two possible
1707 * solutions for that:
1709 * - memcpy() the data internally, at which point we might as well just
1710 * do a regular read() on the buffer anyway.
1711 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1712 * has restriction limitations on both ends of the pipe).
1714 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1717 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1718 unsigned long, nr_segs, unsigned int, flags)
1724 if (unlikely(nr_segs > UIO_MAXIOV))
1726 else if (unlikely(!nr_segs))
1730 file = fget_light(fd, &fput);
1732 if (file->f_mode & FMODE_WRITE)
1733 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1734 else if (file->f_mode & FMODE_READ)
1735 error = vmsplice_to_user(file, iov, nr_segs, flags);
1737 fput_light(file, fput);
1743 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1744 int, fd_out, loff_t __user *, off_out,
1745 size_t, len, unsigned int, flags)
1748 struct file *in, *out;
1749 int fput_in, fput_out;
1755 in = fget_light(fd_in, &fput_in);
1757 if (in->f_mode & FMODE_READ) {
1758 out = fget_light(fd_out, &fput_out);
1760 if (out->f_mode & FMODE_WRITE)
1761 error = do_splice(in, off_in,
1764 fput_light(out, fput_out);
1768 fput_light(in, fput_in);
1775 * Make sure there's data to read. Wait for input if we can, otherwise
1776 * return an appropriate error.
1778 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1783 * Check ->nrbufs without the inode lock first. This function
1784 * is speculative anyways, so missing one is ok.
1792 while (!pipe->nrbufs) {
1793 if (signal_pending(current)) {
1799 if (!pipe->waiting_writers) {
1800 if (flags & SPLICE_F_NONBLOCK) {
1813 * Make sure there's writeable room. Wait for room if we can, otherwise
1814 * return an appropriate error.
1816 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1821 * Check ->nrbufs without the inode lock first. This function
1822 * is speculative anyways, so missing one is ok.
1824 if (pipe->nrbufs < PIPE_BUFFERS)
1830 while (pipe->nrbufs >= PIPE_BUFFERS) {
1831 if (!pipe->readers) {
1832 send_sig(SIGPIPE, current, 0);
1836 if (flags & SPLICE_F_NONBLOCK) {
1840 if (signal_pending(current)) {
1844 pipe->waiting_writers++;
1846 pipe->waiting_writers--;
1854 * Splice contents of ipipe to opipe.
1856 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1857 struct pipe_inode_info *opipe,
1858 size_t len, unsigned int flags)
1860 struct pipe_buffer *ibuf, *obuf;
1862 bool input_wakeup = false;
1866 ret = ipipe_prep(ipipe, flags);
1870 ret = opipe_prep(opipe, flags);
1875 * Potential ABBA deadlock, work around it by ordering lock
1876 * grabbing by pipe info address. Otherwise two different processes
1877 * could deadlock (one doing tee from A -> B, the other from B -> A).
1879 pipe_double_lock(ipipe, opipe);
1882 if (!opipe->readers) {
1883 send_sig(SIGPIPE, current, 0);
1889 if (!ipipe->nrbufs && !ipipe->writers)
1893 * Cannot make any progress, because either the input
1894 * pipe is empty or the output pipe is full.
1896 if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
1897 /* Already processed some buffers, break */
1901 if (flags & SPLICE_F_NONBLOCK) {
1907 * We raced with another reader/writer and haven't
1908 * managed to process any buffers. A zero return
1909 * value means EOF, so retry instead.
1916 ibuf = ipipe->bufs + ipipe->curbuf;
1917 nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
1918 obuf = opipe->bufs + nbuf;
1920 if (len >= ibuf->len) {
1922 * Simply move the whole buffer from ipipe to opipe
1927 ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
1929 input_wakeup = true;
1932 * Get a reference to this pipe buffer,
1933 * so we can copy the contents over.
1935 ibuf->ops->get(ipipe, ibuf);
1939 * Don't inherit the gift flag, we need to
1940 * prevent multiple steals of this page.
1942 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1946 ibuf->offset += obuf->len;
1947 ibuf->len -= obuf->len;
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);
1966 wakeup_pipe_writers(ipipe);
1972 * Link contents of ipipe to opipe.
1974 static int link_pipe(struct pipe_inode_info *ipipe,
1975 struct pipe_inode_info *opipe,
1976 size_t len, unsigned int flags)
1978 struct pipe_buffer *ibuf, *obuf;
1979 int ret = 0, i = 0, nbuf;
1982 * Potential ABBA deadlock, work around it by ordering lock
1983 * grabbing by pipe info address. Otherwise two different processes
1984 * could deadlock (one doing tee from A -> B, the other from B -> A).
1986 pipe_double_lock(ipipe, opipe);
1989 if (!opipe->readers) {
1990 send_sig(SIGPIPE, current, 0);
1997 * If we have iterated all input buffers or ran out of
1998 * output room, break.
2000 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
2003 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
2004 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
2007 * Get a reference to this pipe buffer,
2008 * so we can copy the contents over.
2010 ibuf->ops->get(ipipe, ibuf);
2012 obuf = opipe->bufs + nbuf;
2016 * Don't inherit the gift flag, we need to
2017 * prevent multiple steals of this page.
2019 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2021 if (obuf->len > len)
2031 * return EAGAIN if we have the potential of some data in the
2032 * future, otherwise just return 0
2034 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2041 * If we put data in the output pipe, wakeup any potential readers.
2045 if (waitqueue_active(&opipe->wait))
2046 wake_up_interruptible(&opipe->wait);
2047 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2054 * This is a tee(1) implementation that works on pipes. It doesn't copy
2055 * any data, it simply references the 'in' pages on the 'out' pipe.
2056 * The 'flags' used are the SPLICE_F_* variants, currently the only
2057 * applicable one is SPLICE_F_NONBLOCK.
2059 static long do_tee(struct file *in, struct file *out, size_t len,
2062 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2063 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2067 * Duplicate the contents of ipipe to opipe without actually
2070 if (ipipe && opipe && ipipe != opipe) {
2072 * Keep going, unless we encounter an error. The ipipe/opipe
2073 * ordering doesn't really matter.
2075 ret = ipipe_prep(ipipe, flags);
2077 ret = opipe_prep(opipe, flags);
2079 ret = link_pipe(ipipe, opipe, len, flags);
2086 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2095 in = fget_light(fdin, &fput_in);
2097 if (in->f_mode & FMODE_READ) {
2099 struct file *out = fget_light(fdout, &fput_out);
2102 if (out->f_mode & FMODE_WRITE)
2103 error = do_tee(in, out, len, flags);
2104 fput_light(out, fput_out);
2107 fput_light(in, fput_in);