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;
186 mutex_lock(&pipe->inode->i_mutex);
189 if (!pipe->readers) {
190 send_sig(SIGPIPE, current, 0);
196 if (pipe->nrbufs < PIPE_BUFFERS) {
197 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
198 struct pipe_buffer *buf = pipe->bufs + newbuf;
200 buf->page = spd->pages[page_nr];
201 buf->offset = spd->partial[page_nr].offset;
202 buf->len = spd->partial[page_nr].len;
203 buf->private = spd->partial[page_nr].private;
205 if (spd->flags & SPLICE_F_GIFT)
206 buf->flags |= PIPE_BUF_FLAG_GIFT;
215 if (!--spd->nr_pages)
217 if (pipe->nrbufs < PIPE_BUFFERS)
223 if (spd->flags & SPLICE_F_NONBLOCK) {
229 if (signal_pending(current)) {
237 if (waitqueue_active(&pipe->wait))
238 wake_up_interruptible_sync(&pipe->wait);
239 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
243 pipe->waiting_writers++;
245 pipe->waiting_writers--;
249 mutex_unlock(&pipe->inode->i_mutex);
253 if (waitqueue_active(&pipe->wait))
254 wake_up_interruptible(&pipe->wait);
255 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
259 while (page_nr < spd_pages)
260 spd->spd_release(spd, page_nr++);
265 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
267 page_cache_release(spd->pages[i]);
271 __generic_file_splice_read(struct file *in, loff_t *ppos,
272 struct pipe_inode_info *pipe, size_t len,
275 struct address_space *mapping = in->f_mapping;
276 unsigned int loff, nr_pages, req_pages;
277 struct page *pages[PIPE_BUFFERS];
278 struct partial_page partial[PIPE_BUFFERS];
280 pgoff_t index, end_index;
283 struct splice_pipe_desc spd = {
287 .ops = &page_cache_pipe_buf_ops,
288 .spd_release = spd_release_page,
291 index = *ppos >> PAGE_CACHE_SHIFT;
292 loff = *ppos & ~PAGE_CACHE_MASK;
293 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
294 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
297 * Lookup the (hopefully) full range of pages we need.
299 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
300 index += spd.nr_pages;
303 * If find_get_pages_contig() returned fewer pages than we needed,
304 * readahead/allocate the rest and fill in the holes.
306 if (spd.nr_pages < nr_pages)
307 page_cache_sync_readahead(mapping, &in->f_ra, in,
308 index, req_pages - spd.nr_pages);
311 while (spd.nr_pages < nr_pages) {
313 * Page could be there, find_get_pages_contig() breaks on
316 page = find_get_page(mapping, index);
319 * page didn't exist, allocate one.
321 page = page_cache_alloc_cold(mapping);
325 error = add_to_page_cache_lru(page, mapping, index,
326 mapping_gfp_mask(mapping));
327 if (unlikely(error)) {
328 page_cache_release(page);
329 if (error == -EEXIST)
334 * add_to_page_cache() locks the page, unlock it
335 * to avoid convoluting the logic below even more.
340 pages[spd.nr_pages++] = page;
345 * Now loop over the map and see if we need to start IO on any
346 * pages, fill in the partial map, etc.
348 index = *ppos >> PAGE_CACHE_SHIFT;
349 nr_pages = spd.nr_pages;
351 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
352 unsigned int this_len;
358 * this_len is the max we'll use from this page
360 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
361 page = pages[page_nr];
363 if (PageReadahead(page))
364 page_cache_async_readahead(mapping, &in->f_ra, in,
365 page, index, req_pages - page_nr);
368 * If the page isn't uptodate, we may need to start io on it
370 if (!PageUptodate(page)) {
372 * If in nonblock mode then dont block on waiting
373 * for an in-flight io page
375 if (flags & SPLICE_F_NONBLOCK) {
376 if (!trylock_page(page)) {
384 * Page was truncated, or invalidated by the
385 * filesystem. Redo the find/create, but this time the
386 * page is kept locked, so there's no chance of another
387 * race with truncate/invalidate.
389 if (!page->mapping) {
391 page = find_or_create_page(mapping, index,
392 mapping_gfp_mask(mapping));
398 page_cache_release(pages[page_nr]);
399 pages[page_nr] = page;
402 * page was already under io and is now done, great
404 if (PageUptodate(page)) {
410 * need to read in the page
412 error = mapping->a_ops->readpage(in, page);
413 if (unlikely(error)) {
415 * We really should re-lookup the page here,
416 * but it complicates things a lot. Instead
417 * lets just do what we already stored, and
418 * we'll get it the next time we are called.
420 if (error == AOP_TRUNCATED_PAGE)
428 * i_size must be checked after PageUptodate.
430 isize = i_size_read(mapping->host);
431 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
432 if (unlikely(!isize || index > end_index))
436 * if this is the last page, see if we need to shrink
437 * the length and stop
439 if (end_index == index) {
443 * max good bytes in this page
445 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
450 * force quit after adding this page
452 this_len = min(this_len, plen - loff);
456 partial[page_nr].offset = loff;
457 partial[page_nr].len = this_len;
465 * Release any pages at the end, if we quit early. 'page_nr' is how far
466 * we got, 'nr_pages' is how many pages are in the map.
468 while (page_nr < nr_pages)
469 page_cache_release(pages[page_nr++]);
470 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
473 return splice_to_pipe(pipe, &spd);
479 * generic_file_splice_read - splice data from file to a pipe
480 * @in: file to splice from
481 * @ppos: position in @in
482 * @pipe: pipe to splice to
483 * @len: number of bytes to splice
484 * @flags: splice modifier flags
487 * Will read pages from given file and fill them into a pipe. Can be
488 * used as long as the address_space operations for the source implements
492 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
493 struct pipe_inode_info *pipe, size_t len,
499 isize = i_size_read(in->f_mapping->host);
500 if (unlikely(*ppos >= isize))
503 left = isize - *ppos;
504 if (unlikely(left < len))
507 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
514 EXPORT_SYMBOL(generic_file_splice_read);
517 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
518 * using sendpage(). Return the number of bytes sent.
520 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
521 struct pipe_buffer *buf, struct splice_desc *sd)
523 struct file *file = sd->u.file;
524 loff_t pos = sd->pos;
527 ret = buf->ops->confirm(pipe, buf);
529 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
531 ret = file->f_op->sendpage(file, buf->page, buf->offset,
532 sd->len, &pos, more);
539 * This is a little more tricky than the file -> pipe splicing. There are
540 * basically three cases:
542 * - Destination page already exists in the address space and there
543 * are users of it. For that case we have no other option that
544 * copying the data. Tough luck.
545 * - Destination page already exists in the address space, but there
546 * are no users of it. Make sure it's uptodate, then drop it. Fall
547 * through to last case.
548 * - Destination page does not exist, we can add the pipe page to
549 * the page cache and avoid the copy.
551 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
552 * sd->flags), we attempt to migrate pages from the pipe to the output
553 * file address space page cache. This is possible if no one else has
554 * the pipe page referenced outside of the pipe and page cache. If
555 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
556 * a new page in the output file page cache and fill/dirty that.
558 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
559 struct splice_desc *sd)
561 struct file *file = sd->u.file;
562 struct address_space *mapping = file->f_mapping;
563 unsigned int offset, this_len;
569 * make sure the data in this buffer is uptodate
571 ret = buf->ops->confirm(pipe, buf);
575 offset = sd->pos & ~PAGE_CACHE_MASK;
578 if (this_len + offset > PAGE_CACHE_SIZE)
579 this_len = PAGE_CACHE_SIZE - offset;
581 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
582 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
586 if (buf->page != page) {
588 * Careful, ->map() uses KM_USER0!
590 char *src = buf->ops->map(pipe, buf, 1);
591 char *dst = kmap_atomic(page, KM_USER1);
593 memcpy(dst + offset, src + buf->offset, this_len);
594 flush_dcache_page(page);
595 kunmap_atomic(dst, KM_USER1);
596 buf->ops->unmap(pipe, buf, src);
598 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
604 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
607 if (waitqueue_active(&pipe->wait))
608 wake_up_interruptible(&pipe->wait);
609 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
613 * splice_from_pipe_feed - feed available data from a pipe to a file
614 * @pipe: pipe to splice from
615 * @sd: information to @actor
616 * @actor: handler that splices the data
620 * This function loops over the pipe and calls @actor to do the
621 * actual moving of a single struct pipe_buffer to the desired
622 * destination. It returns when there's no more buffers left in
623 * the pipe or if the requested number of bytes (@sd->total_len)
624 * have been copied. It returns a positive number (one) if the
625 * pipe needs to be filled with more data, zero if the required
626 * number of bytes have been copied and -errno on error.
628 * This, together with splice_from_pipe_{begin,end,next}, may be
629 * used to implement the functionality of __splice_from_pipe() when
630 * locking is required around copying the pipe buffers to the
633 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
638 while (pipe->nrbufs) {
639 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
640 const struct pipe_buf_operations *ops = buf->ops;
643 if (sd->len > sd->total_len)
644 sd->len = sd->total_len;
646 ret = actor(pipe, buf, sd);
655 sd->num_spliced += ret;
658 sd->total_len -= ret;
662 ops->release(pipe, buf);
663 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
666 sd->need_wakeup = true;
675 EXPORT_SYMBOL(splice_from_pipe_feed);
678 * splice_from_pipe_next - wait for some data to splice from
679 * @pipe: pipe to splice from
680 * @sd: information about the splice operation
683 * This function will wait for some data and return a positive
684 * value (one) if pipe buffers are available. It will return zero
685 * or -errno if no more data needs to be spliced.
687 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
689 while (!pipe->nrbufs) {
693 if (!pipe->waiting_writers && sd->num_spliced)
696 if (sd->flags & SPLICE_F_NONBLOCK)
699 if (signal_pending(current))
702 if (sd->need_wakeup) {
703 wakeup_pipe_writers(pipe);
704 sd->need_wakeup = false;
712 EXPORT_SYMBOL(splice_from_pipe_next);
715 * splice_from_pipe_begin - start splicing from pipe
716 * @pipe: pipe to splice from
719 * This function should be called before a loop containing
720 * splice_from_pipe_next() and splice_from_pipe_feed() to
721 * initialize the necessary fields of @sd.
723 void splice_from_pipe_begin(struct splice_desc *sd)
726 sd->need_wakeup = false;
728 EXPORT_SYMBOL(splice_from_pipe_begin);
731 * splice_from_pipe_end - finish splicing from pipe
732 * @pipe: pipe to splice from
733 * @sd: information about the splice operation
736 * This function will wake up pipe writers if necessary. It should
737 * be called after a loop containing splice_from_pipe_next() and
738 * splice_from_pipe_feed().
740 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
743 wakeup_pipe_writers(pipe);
745 EXPORT_SYMBOL(splice_from_pipe_end);
748 * __splice_from_pipe - splice data from a pipe to given actor
749 * @pipe: pipe to splice from
750 * @sd: information to @actor
751 * @actor: handler that splices the data
754 * This function does little more than loop over the pipe and call
755 * @actor to do the actual moving of a single struct pipe_buffer to
756 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
760 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
765 splice_from_pipe_begin(sd);
767 ret = splice_from_pipe_next(pipe, sd);
769 ret = splice_from_pipe_feed(pipe, sd, actor);
771 splice_from_pipe_end(pipe, sd);
773 return sd->num_spliced ? sd->num_spliced : ret;
775 EXPORT_SYMBOL(__splice_from_pipe);
778 * splice_from_pipe - splice data from a pipe to a file
779 * @pipe: pipe to splice from
780 * @out: file to splice to
781 * @ppos: position in @out
782 * @len: how many bytes to splice
783 * @flags: splice modifier flags
784 * @actor: handler that splices the data
787 * See __splice_from_pipe. This function locks the pipe inode,
788 * otherwise it's identical to __splice_from_pipe().
791 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
792 loff_t *ppos, size_t len, unsigned int flags,
796 struct splice_desc sd = {
804 mutex_lock(&pipe->inode->i_mutex);
805 ret = __splice_from_pipe(pipe, &sd, actor);
807 mutex_unlock(&pipe->inode->i_mutex);
813 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
815 * @out: file to write to
816 * @ppos: position in @out
817 * @len: number of bytes to splice
818 * @flags: splice modifier flags
821 * Will either move or copy pages (determined by @flags options) from
822 * the given pipe inode to the given file. The caller is responsible
823 * for acquiring i_mutex on both inodes.
827 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
828 loff_t *ppos, size_t len, unsigned int flags)
830 struct address_space *mapping = out->f_mapping;
831 struct inode *inode = mapping->host;
832 struct splice_desc sd = {
841 err = file_remove_suid(out);
845 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
847 unsigned long nr_pages;
850 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
853 * If file or inode is SYNC and we actually wrote some data,
856 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
857 err = generic_osync_inode(inode, mapping,
858 OSYNC_METADATA|OSYNC_DATA);
863 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
869 EXPORT_SYMBOL(generic_file_splice_write_nolock);
872 * generic_file_splice_write - splice data from a pipe to a file
874 * @out: file to write to
875 * @ppos: position in @out
876 * @len: number of bytes to splice
877 * @flags: splice modifier flags
880 * Will either move or copy pages (determined by @flags options) from
881 * the given pipe inode to the given file.
885 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
886 loff_t *ppos, size_t len, unsigned int flags)
888 struct address_space *mapping = out->f_mapping;
889 struct inode *inode = mapping->host;
890 struct splice_desc sd = {
898 WARN_ON(S_ISFIFO(inode->i_mode));
899 mutex_lock_nested(&inode->i_mutex, I_MUTEX_PARENT);
900 ret = file_remove_suid(out);
903 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_CHILD);
904 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
906 mutex_unlock(&pipe->inode->i_mutex);
908 mutex_unlock(&inode->i_mutex);
910 unsigned long nr_pages;
913 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
916 * If file or inode is SYNC and we actually wrote some data,
919 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
922 mutex_lock(&inode->i_mutex);
923 err = generic_osync_inode(inode, mapping,
924 OSYNC_METADATA|OSYNC_DATA);
925 mutex_unlock(&inode->i_mutex);
930 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
936 EXPORT_SYMBOL(generic_file_splice_write);
939 * generic_splice_sendpage - splice data from a pipe to a socket
940 * @pipe: pipe to splice from
941 * @out: socket to write to
942 * @ppos: position in @out
943 * @len: number of bytes to splice
944 * @flags: splice modifier flags
947 * Will send @len bytes from the pipe to a network socket. No data copying
951 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
952 loff_t *ppos, size_t len, unsigned int flags)
954 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
957 EXPORT_SYMBOL(generic_splice_sendpage);
960 * Attempt to initiate a splice from pipe to file.
962 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
963 loff_t *ppos, size_t len, unsigned int flags)
967 if (unlikely(!out->f_op || !out->f_op->splice_write))
970 if (unlikely(!(out->f_mode & FMODE_WRITE)))
973 if (unlikely(out->f_flags & O_APPEND))
976 ret = rw_verify_area(WRITE, out, ppos, len);
977 if (unlikely(ret < 0))
980 return out->f_op->splice_write(pipe, out, ppos, len, flags);
984 * Attempt to initiate a splice from a file to a pipe.
986 static long do_splice_to(struct file *in, loff_t *ppos,
987 struct pipe_inode_info *pipe, size_t len,
992 if (unlikely(!in->f_op || !in->f_op->splice_read))
995 if (unlikely(!(in->f_mode & FMODE_READ)))
998 ret = rw_verify_area(READ, in, ppos, len);
999 if (unlikely(ret < 0))
1002 return in->f_op->splice_read(in, ppos, pipe, len, flags);
1006 * splice_direct_to_actor - splices data directly between two non-pipes
1007 * @in: file to splice from
1008 * @sd: actor information on where to splice to
1009 * @actor: handles the data splicing
1012 * This is a special case helper to splice directly between two
1013 * points, without requiring an explicit pipe. Internally an allocated
1014 * pipe is cached in the process, and reused during the lifetime of
1018 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1019 splice_direct_actor *actor)
1021 struct pipe_inode_info *pipe;
1028 * We require the input being a regular file, as we don't want to
1029 * randomly drop data for eg socket -> socket splicing. Use the
1030 * piped splicing for that!
1032 i_mode = in->f_path.dentry->d_inode->i_mode;
1033 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1037 * neither in nor out is a pipe, setup an internal pipe attached to
1038 * 'out' and transfer the wanted data from 'in' to 'out' through that
1040 pipe = current->splice_pipe;
1041 if (unlikely(!pipe)) {
1042 pipe = alloc_pipe_info(NULL);
1047 * We don't have an immediate reader, but we'll read the stuff
1048 * out of the pipe right after the splice_to_pipe(). So set
1049 * PIPE_READERS appropriately.
1053 current->splice_pipe = pipe;
1061 len = sd->total_len;
1065 * Don't block on output, we have to drain the direct pipe.
1067 sd->flags &= ~SPLICE_F_NONBLOCK;
1071 loff_t pos = sd->pos, prev_pos = pos;
1073 ret = do_splice_to(in, &pos, pipe, len, flags);
1074 if (unlikely(ret <= 0))
1078 sd->total_len = read_len;
1081 * NOTE: nonblocking mode only applies to the input. We
1082 * must not do the output in nonblocking mode as then we
1083 * could get stuck data in the internal pipe:
1085 ret = actor(pipe, sd);
1086 if (unlikely(ret <= 0)) {
1095 if (ret < read_len) {
1096 sd->pos = prev_pos + ret;
1102 pipe->nrbufs = pipe->curbuf = 0;
1108 * If we did an incomplete transfer we must release
1109 * the pipe buffers in question:
1111 for (i = 0; i < PIPE_BUFFERS; i++) {
1112 struct pipe_buffer *buf = pipe->bufs + i;
1115 buf->ops->release(pipe, buf);
1125 EXPORT_SYMBOL(splice_direct_to_actor);
1127 static int direct_splice_actor(struct pipe_inode_info *pipe,
1128 struct splice_desc *sd)
1130 struct file *file = sd->u.file;
1132 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1136 * do_splice_direct - splices data directly between two files
1137 * @in: file to splice from
1138 * @ppos: input file offset
1139 * @out: file to splice to
1140 * @len: number of bytes to splice
1141 * @flags: splice modifier flags
1144 * For use by do_sendfile(). splice can easily emulate sendfile, but
1145 * doing it in the application would incur an extra system call
1146 * (splice in + splice out, as compared to just sendfile()). So this helper
1147 * can splice directly through a process-private pipe.
1150 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1151 size_t len, unsigned int flags)
1153 struct splice_desc sd = {
1162 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1170 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1171 * location, so checking ->i_pipe is not enough to verify that this is a
1174 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1176 if (S_ISFIFO(inode->i_mode))
1177 return inode->i_pipe;
1183 * Determine where to splice to/from.
1185 static long do_splice(struct file *in, loff_t __user *off_in,
1186 struct file *out, loff_t __user *off_out,
1187 size_t len, unsigned int flags)
1189 struct pipe_inode_info *pipe;
1190 loff_t offset, *off;
1193 pipe = pipe_info(in->f_path.dentry->d_inode);
1198 if (out->f_op->llseek == no_llseek)
1200 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1206 ret = do_splice_from(pipe, out, off, len, flags);
1208 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1214 pipe = pipe_info(out->f_path.dentry->d_inode);
1219 if (in->f_op->llseek == no_llseek)
1221 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1227 ret = do_splice_to(in, off, pipe, len, flags);
1229 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1239 * Map an iov into an array of pages and offset/length tupples. With the
1240 * partial_page structure, we can map several non-contiguous ranges into
1241 * our ones pages[] map instead of splitting that operation into pieces.
1242 * Could easily be exported as a generic helper for other users, in which
1243 * case one would probably want to add a 'max_nr_pages' parameter as well.
1245 static int get_iovec_page_array(const struct iovec __user *iov,
1246 unsigned int nr_vecs, struct page **pages,
1247 struct partial_page *partial, int aligned)
1249 int buffers = 0, error = 0;
1252 unsigned long off, npages;
1259 if (copy_from_user(&entry, iov, sizeof(entry)))
1262 base = entry.iov_base;
1263 len = entry.iov_len;
1266 * Sanity check this iovec. 0 read succeeds.
1272 if (!access_ok(VERIFY_READ, base, len))
1276 * Get this base offset and number of pages, then map
1277 * in the user pages.
1279 off = (unsigned long) base & ~PAGE_MASK;
1282 * If asked for alignment, the offset must be zero and the
1283 * length a multiple of the PAGE_SIZE.
1286 if (aligned && (off || len & ~PAGE_MASK))
1289 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1290 if (npages > PIPE_BUFFERS - buffers)
1291 npages = PIPE_BUFFERS - buffers;
1293 error = get_user_pages_fast((unsigned long)base, npages,
1294 0, &pages[buffers]);
1296 if (unlikely(error <= 0))
1300 * Fill this contiguous range into the partial page map.
1302 for (i = 0; i < error; i++) {
1303 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1305 partial[buffers].offset = off;
1306 partial[buffers].len = plen;
1314 * We didn't complete this iov, stop here since it probably
1315 * means we have to move some of this into a pipe to
1316 * be able to continue.
1322 * Don't continue if we mapped fewer pages than we asked for,
1323 * or if we mapped the max number of pages that we have
1326 if (error < npages || buffers == PIPE_BUFFERS)
1339 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1340 struct splice_desc *sd)
1345 ret = buf->ops->confirm(pipe, buf);
1350 * See if we can use the atomic maps, by prefaulting in the
1351 * pages and doing an atomic copy
1353 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1354 src = buf->ops->map(pipe, buf, 1);
1355 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1357 buf->ops->unmap(pipe, buf, src);
1365 * No dice, use slow non-atomic map and copy
1367 src = buf->ops->map(pipe, buf, 0);
1370 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1373 buf->ops->unmap(pipe, buf, src);
1376 sd->u.userptr += ret;
1381 * For lack of a better implementation, implement vmsplice() to userspace
1382 * as a simple copy of the pipes pages to the user iov.
1384 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1385 unsigned long nr_segs, unsigned int flags)
1387 struct pipe_inode_info *pipe;
1388 struct splice_desc sd;
1393 pipe = pipe_info(file->f_path.dentry->d_inode);
1398 mutex_lock(&pipe->inode->i_mutex);
1406 * Get user address base and length for this iovec.
1408 error = get_user(base, &iov->iov_base);
1409 if (unlikely(error))
1411 error = get_user(len, &iov->iov_len);
1412 if (unlikely(error))
1416 * Sanity check this iovec. 0 read succeeds.
1420 if (unlikely(!base)) {
1425 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1433 sd.u.userptr = base;
1436 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1454 mutex_unlock(&pipe->inode->i_mutex);
1463 * vmsplice splices a user address range into a pipe. It can be thought of
1464 * as splice-from-memory, where the regular splice is splice-from-file (or
1465 * to file). In both cases the output is a pipe, naturally.
1467 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1468 unsigned long nr_segs, unsigned int flags)
1470 struct pipe_inode_info *pipe;
1471 struct page *pages[PIPE_BUFFERS];
1472 struct partial_page partial[PIPE_BUFFERS];
1473 struct splice_pipe_desc spd = {
1477 .ops = &user_page_pipe_buf_ops,
1478 .spd_release = spd_release_page,
1481 pipe = pipe_info(file->f_path.dentry->d_inode);
1485 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1486 flags & SPLICE_F_GIFT);
1487 if (spd.nr_pages <= 0)
1488 return spd.nr_pages;
1490 return splice_to_pipe(pipe, &spd);
1494 * Note that vmsplice only really supports true splicing _from_ user memory
1495 * to a pipe, not the other way around. Splicing from user memory is a simple
1496 * operation that can be supported without any funky alignment restrictions
1497 * or nasty vm tricks. We simply map in the user memory and fill them into
1498 * a pipe. The reverse isn't quite as easy, though. There are two possible
1499 * solutions for that:
1501 * - memcpy() the data internally, at which point we might as well just
1502 * do a regular read() on the buffer anyway.
1503 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1504 * has restriction limitations on both ends of the pipe).
1506 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1509 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1510 unsigned long, nr_segs, unsigned int, flags)
1516 if (unlikely(nr_segs > UIO_MAXIOV))
1518 else if (unlikely(!nr_segs))
1522 file = fget_light(fd, &fput);
1524 if (file->f_mode & FMODE_WRITE)
1525 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1526 else if (file->f_mode & FMODE_READ)
1527 error = vmsplice_to_user(file, iov, nr_segs, flags);
1529 fput_light(file, fput);
1535 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1536 int, fd_out, loff_t __user *, off_out,
1537 size_t, len, unsigned int, flags)
1540 struct file *in, *out;
1541 int fput_in, fput_out;
1547 in = fget_light(fd_in, &fput_in);
1549 if (in->f_mode & FMODE_READ) {
1550 out = fget_light(fd_out, &fput_out);
1552 if (out->f_mode & FMODE_WRITE)
1553 error = do_splice(in, off_in,
1556 fput_light(out, fput_out);
1560 fput_light(in, fput_in);
1567 * Make sure there's data to read. Wait for input if we can, otherwise
1568 * return an appropriate error.
1570 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1575 * Check ->nrbufs without the inode lock first. This function
1576 * is speculative anyways, so missing one is ok.
1582 mutex_lock(&pipe->inode->i_mutex);
1584 while (!pipe->nrbufs) {
1585 if (signal_pending(current)) {
1591 if (!pipe->waiting_writers) {
1592 if (flags & SPLICE_F_NONBLOCK) {
1600 mutex_unlock(&pipe->inode->i_mutex);
1605 * Make sure there's writeable room. Wait for room if we can, otherwise
1606 * return an appropriate error.
1608 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1613 * Check ->nrbufs without the inode lock first. This function
1614 * is speculative anyways, so missing one is ok.
1616 if (pipe->nrbufs < PIPE_BUFFERS)
1620 mutex_lock(&pipe->inode->i_mutex);
1622 while (pipe->nrbufs >= PIPE_BUFFERS) {
1623 if (!pipe->readers) {
1624 send_sig(SIGPIPE, current, 0);
1628 if (flags & SPLICE_F_NONBLOCK) {
1632 if (signal_pending(current)) {
1636 pipe->waiting_writers++;
1638 pipe->waiting_writers--;
1641 mutex_unlock(&pipe->inode->i_mutex);
1646 * Link contents of ipipe to opipe.
1648 static int link_pipe(struct pipe_inode_info *ipipe,
1649 struct pipe_inode_info *opipe,
1650 size_t len, unsigned int flags)
1652 struct pipe_buffer *ibuf, *obuf;
1653 int ret = 0, i = 0, nbuf;
1656 * Potential ABBA deadlock, work around it by ordering lock
1657 * grabbing by inode address. Otherwise two different processes
1658 * could deadlock (one doing tee from A -> B, the other from B -> A).
1660 inode_double_lock(ipipe->inode, opipe->inode);
1663 if (!opipe->readers) {
1664 send_sig(SIGPIPE, current, 0);
1671 * If we have iterated all input buffers or ran out of
1672 * output room, break.
1674 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1677 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1678 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1681 * Get a reference to this pipe buffer,
1682 * so we can copy the contents over.
1684 ibuf->ops->get(ipipe, ibuf);
1686 obuf = opipe->bufs + nbuf;
1690 * Don't inherit the gift flag, we need to
1691 * prevent multiple steals of this page.
1693 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1695 if (obuf->len > len)
1705 * return EAGAIN if we have the potential of some data in the
1706 * future, otherwise just return 0
1708 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1711 inode_double_unlock(ipipe->inode, opipe->inode);
1714 * If we put data in the output pipe, wakeup any potential readers.
1718 if (waitqueue_active(&opipe->wait))
1719 wake_up_interruptible(&opipe->wait);
1720 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1727 * This is a tee(1) implementation that works on pipes. It doesn't copy
1728 * any data, it simply references the 'in' pages on the 'out' pipe.
1729 * The 'flags' used are the SPLICE_F_* variants, currently the only
1730 * applicable one is SPLICE_F_NONBLOCK.
1732 static long do_tee(struct file *in, struct file *out, size_t len,
1735 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1736 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1740 * Duplicate the contents of ipipe to opipe without actually
1743 if (ipipe && opipe && ipipe != opipe) {
1745 * Keep going, unless we encounter an error. The ipipe/opipe
1746 * ordering doesn't really matter.
1748 ret = link_ipipe_prep(ipipe, flags);
1750 ret = link_opipe_prep(opipe, flags);
1752 ret = link_pipe(ipipe, opipe, len, flags);
1759 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1768 in = fget_light(fdin, &fput_in);
1770 if (in->f_mode & FMODE_READ) {
1772 struct file *out = fget_light(fdout, &fput_out);
1775 if (out->f_mode & FMODE_WRITE)
1776 error = do_tee(in, out, len, flags);
1777 fput_light(out, fput_out);
1780 fput_light(in, fput_in);