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 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,
603 EXPORT_SYMBOL(pipe_to_file);
605 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
608 if (waitqueue_active(&pipe->wait))
609 wake_up_interruptible(&pipe->wait);
610 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
614 * splice_from_pipe_feed - feed available data from a pipe to a file
615 * @pipe: pipe to splice from
616 * @sd: information to @actor
617 * @actor: handler that splices the data
621 * This function loops over the pipe and calls @actor to do the
622 * actual moving of a single struct pipe_buffer to the desired
623 * destination. It returns when there's no more buffers left in
624 * the pipe or if the requested number of bytes (@sd->total_len)
625 * have been copied. It returns a positive number (one) if the
626 * pipe needs to be filled with more data, zero if the required
627 * number of bytes have been copied and -errno on error.
629 * This, together with splice_from_pipe_{begin,end,next}, may be
630 * used to implement the functionality of __splice_from_pipe() when
631 * locking is required around copying the pipe buffers to the
634 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
639 while (pipe->nrbufs) {
640 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
641 const struct pipe_buf_operations *ops = buf->ops;
644 if (sd->len > sd->total_len)
645 sd->len = sd->total_len;
647 ret = actor(pipe, buf, sd);
656 sd->num_spliced += ret;
659 sd->total_len -= ret;
663 ops->release(pipe, buf);
664 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
667 sd->need_wakeup = true;
676 EXPORT_SYMBOL(splice_from_pipe_feed);
679 * splice_from_pipe_next - wait for some data to splice from
680 * @pipe: pipe to splice from
681 * @sd: information about the splice operation
684 * This function will wait for some data and return a positive
685 * value (one) if pipe buffers are available. It will return zero
686 * or -errno if no more data needs to be spliced.
688 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
690 while (!pipe->nrbufs) {
694 if (!pipe->waiting_writers && sd->num_spliced)
697 if (sd->flags & SPLICE_F_NONBLOCK)
700 if (signal_pending(current))
703 if (sd->need_wakeup) {
704 wakeup_pipe_writers(pipe);
705 sd->need_wakeup = false;
713 EXPORT_SYMBOL(splice_from_pipe_next);
716 * splice_from_pipe_begin - start splicing from pipe
717 * @pipe: pipe to splice from
720 * This function should be called before a loop containing
721 * splice_from_pipe_next() and splice_from_pipe_feed() to
722 * initialize the necessary fields of @sd.
724 void splice_from_pipe_begin(struct splice_desc *sd)
727 sd->need_wakeup = false;
729 EXPORT_SYMBOL(splice_from_pipe_begin);
732 * splice_from_pipe_end - finish splicing from pipe
733 * @pipe: pipe to splice from
734 * @sd: information about the splice operation
737 * This function will wake up pipe writers if necessary. It should
738 * be called after a loop containing splice_from_pipe_next() and
739 * splice_from_pipe_feed().
741 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
744 wakeup_pipe_writers(pipe);
746 EXPORT_SYMBOL(splice_from_pipe_end);
749 * __splice_from_pipe - splice data from a pipe to given actor
750 * @pipe: pipe to splice from
751 * @sd: information to @actor
752 * @actor: handler that splices the data
755 * This function does little more than loop over the pipe and call
756 * @actor to do the actual moving of a single struct pipe_buffer to
757 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
761 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
766 splice_from_pipe_begin(sd);
768 ret = splice_from_pipe_next(pipe, sd);
770 ret = splice_from_pipe_feed(pipe, sd, actor);
772 splice_from_pipe_end(pipe, sd);
774 return sd->num_spliced ? sd->num_spliced : ret;
776 EXPORT_SYMBOL(__splice_from_pipe);
779 * splice_from_pipe - splice data from a pipe to a file
780 * @pipe: pipe to splice from
781 * @out: file to splice to
782 * @ppos: position in @out
783 * @len: how many bytes to splice
784 * @flags: splice modifier flags
785 * @actor: handler that splices the data
788 * See __splice_from_pipe. This function locks the pipe inode,
789 * otherwise it's identical to __splice_from_pipe().
792 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
793 loff_t *ppos, size_t len, unsigned int flags,
797 struct splice_desc sd = {
805 mutex_lock(&pipe->inode->i_mutex);
806 ret = __splice_from_pipe(pipe, &sd, actor);
808 mutex_unlock(&pipe->inode->i_mutex);
814 * generic_file_splice_write - splice data from a pipe to a file
816 * @out: file to write to
817 * @ppos: position in @out
818 * @len: number of bytes to splice
819 * @flags: splice modifier flags
822 * Will either move or copy pages (determined by @flags options) from
823 * the given pipe inode to the given file.
827 generic_file_splice_write(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 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
843 splice_from_pipe_begin(&sd);
845 ret = splice_from_pipe_next(pipe, &sd);
849 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
850 ret = file_remove_suid(out);
852 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
853 mutex_unlock(&inode->i_mutex);
855 splice_from_pipe_end(pipe, &sd);
858 mutex_unlock(&pipe->inode->i_mutex);
861 ret = sd.num_spliced;
864 unsigned long nr_pages;
867 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
870 * If file or inode is SYNC and we actually wrote some data,
873 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
876 mutex_lock(&inode->i_mutex);
877 err = generic_osync_inode(inode, mapping,
878 OSYNC_METADATA|OSYNC_DATA);
879 mutex_unlock(&inode->i_mutex);
884 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
890 EXPORT_SYMBOL(generic_file_splice_write);
893 * generic_splice_sendpage - splice data from a pipe to a socket
894 * @pipe: pipe to splice from
895 * @out: socket to write to
896 * @ppos: position in @out
897 * @len: number of bytes to splice
898 * @flags: splice modifier flags
901 * Will send @len bytes from the pipe to a network socket. No data copying
905 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
906 loff_t *ppos, size_t len, unsigned int flags)
908 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
911 EXPORT_SYMBOL(generic_splice_sendpage);
914 * Attempt to initiate a splice from pipe to file.
916 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
917 loff_t *ppos, size_t len, unsigned int flags)
921 if (unlikely(!out->f_op || !out->f_op->splice_write))
924 if (unlikely(!(out->f_mode & FMODE_WRITE)))
927 if (unlikely(out->f_flags & O_APPEND))
930 ret = rw_verify_area(WRITE, out, ppos, len);
931 if (unlikely(ret < 0))
934 return out->f_op->splice_write(pipe, out, ppos, len, flags);
938 * Attempt to initiate a splice from a file to a pipe.
940 static long do_splice_to(struct file *in, loff_t *ppos,
941 struct pipe_inode_info *pipe, size_t len,
946 if (unlikely(!in->f_op || !in->f_op->splice_read))
949 if (unlikely(!(in->f_mode & FMODE_READ)))
952 ret = rw_verify_area(READ, in, ppos, len);
953 if (unlikely(ret < 0))
956 return in->f_op->splice_read(in, ppos, pipe, len, flags);
960 * splice_direct_to_actor - splices data directly between two non-pipes
961 * @in: file to splice from
962 * @sd: actor information on where to splice to
963 * @actor: handles the data splicing
966 * This is a special case helper to splice directly between two
967 * points, without requiring an explicit pipe. Internally an allocated
968 * pipe is cached in the process, and reused during the lifetime of
972 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
973 splice_direct_actor *actor)
975 struct pipe_inode_info *pipe;
982 * We require the input being a regular file, as we don't want to
983 * randomly drop data for eg socket -> socket splicing. Use the
984 * piped splicing for that!
986 i_mode = in->f_path.dentry->d_inode->i_mode;
987 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
991 * neither in nor out is a pipe, setup an internal pipe attached to
992 * 'out' and transfer the wanted data from 'in' to 'out' through that
994 pipe = current->splice_pipe;
995 if (unlikely(!pipe)) {
996 pipe = alloc_pipe_info(NULL);
1001 * We don't have an immediate reader, but we'll read the stuff
1002 * out of the pipe right after the splice_to_pipe(). So set
1003 * PIPE_READERS appropriately.
1007 current->splice_pipe = pipe;
1015 len = sd->total_len;
1019 * Don't block on output, we have to drain the direct pipe.
1021 sd->flags &= ~SPLICE_F_NONBLOCK;
1025 loff_t pos = sd->pos, prev_pos = pos;
1027 ret = do_splice_to(in, &pos, pipe, len, flags);
1028 if (unlikely(ret <= 0))
1032 sd->total_len = read_len;
1035 * NOTE: nonblocking mode only applies to the input. We
1036 * must not do the output in nonblocking mode as then we
1037 * could get stuck data in the internal pipe:
1039 ret = actor(pipe, sd);
1040 if (unlikely(ret <= 0)) {
1049 if (ret < read_len) {
1050 sd->pos = prev_pos + ret;
1056 pipe->nrbufs = pipe->curbuf = 0;
1062 * If we did an incomplete transfer we must release
1063 * the pipe buffers in question:
1065 for (i = 0; i < PIPE_BUFFERS; i++) {
1066 struct pipe_buffer *buf = pipe->bufs + i;
1069 buf->ops->release(pipe, buf);
1079 EXPORT_SYMBOL(splice_direct_to_actor);
1081 static int direct_splice_actor(struct pipe_inode_info *pipe,
1082 struct splice_desc *sd)
1084 struct file *file = sd->u.file;
1086 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1090 * do_splice_direct - splices data directly between two files
1091 * @in: file to splice from
1092 * @ppos: input file offset
1093 * @out: file to splice to
1094 * @len: number of bytes to splice
1095 * @flags: splice modifier flags
1098 * For use by do_sendfile(). splice can easily emulate sendfile, but
1099 * doing it in the application would incur an extra system call
1100 * (splice in + splice out, as compared to just sendfile()). So this helper
1101 * can splice directly through a process-private pipe.
1104 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1105 size_t len, unsigned int flags)
1107 struct splice_desc sd = {
1116 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1124 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1125 * location, so checking ->i_pipe is not enough to verify that this is a
1128 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1130 if (S_ISFIFO(inode->i_mode))
1131 return inode->i_pipe;
1137 * Determine where to splice to/from.
1139 static long do_splice(struct file *in, loff_t __user *off_in,
1140 struct file *out, loff_t __user *off_out,
1141 size_t len, unsigned int flags)
1143 struct pipe_inode_info *pipe;
1144 loff_t offset, *off;
1147 pipe = pipe_info(in->f_path.dentry->d_inode);
1152 if (out->f_op->llseek == no_llseek)
1154 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1160 ret = do_splice_from(pipe, out, off, len, flags);
1162 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1168 pipe = pipe_info(out->f_path.dentry->d_inode);
1173 if (in->f_op->llseek == no_llseek)
1175 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1181 ret = do_splice_to(in, off, pipe, len, flags);
1183 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1193 * Map an iov into an array of pages and offset/length tupples. With the
1194 * partial_page structure, we can map several non-contiguous ranges into
1195 * our ones pages[] map instead of splitting that operation into pieces.
1196 * Could easily be exported as a generic helper for other users, in which
1197 * case one would probably want to add a 'max_nr_pages' parameter as well.
1199 static int get_iovec_page_array(const struct iovec __user *iov,
1200 unsigned int nr_vecs, struct page **pages,
1201 struct partial_page *partial, int aligned)
1203 int buffers = 0, error = 0;
1206 unsigned long off, npages;
1213 if (copy_from_user(&entry, iov, sizeof(entry)))
1216 base = entry.iov_base;
1217 len = entry.iov_len;
1220 * Sanity check this iovec. 0 read succeeds.
1226 if (!access_ok(VERIFY_READ, base, len))
1230 * Get this base offset and number of pages, then map
1231 * in the user pages.
1233 off = (unsigned long) base & ~PAGE_MASK;
1236 * If asked for alignment, the offset must be zero and the
1237 * length a multiple of the PAGE_SIZE.
1240 if (aligned && (off || len & ~PAGE_MASK))
1243 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1244 if (npages > PIPE_BUFFERS - buffers)
1245 npages = PIPE_BUFFERS - buffers;
1247 error = get_user_pages_fast((unsigned long)base, npages,
1248 0, &pages[buffers]);
1250 if (unlikely(error <= 0))
1254 * Fill this contiguous range into the partial page map.
1256 for (i = 0; i < error; i++) {
1257 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1259 partial[buffers].offset = off;
1260 partial[buffers].len = plen;
1268 * We didn't complete this iov, stop here since it probably
1269 * means we have to move some of this into a pipe to
1270 * be able to continue.
1276 * Don't continue if we mapped fewer pages than we asked for,
1277 * or if we mapped the max number of pages that we have
1280 if (error < npages || buffers == PIPE_BUFFERS)
1293 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1294 struct splice_desc *sd)
1299 ret = buf->ops->confirm(pipe, buf);
1304 * See if we can use the atomic maps, by prefaulting in the
1305 * pages and doing an atomic copy
1307 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1308 src = buf->ops->map(pipe, buf, 1);
1309 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1311 buf->ops->unmap(pipe, buf, src);
1319 * No dice, use slow non-atomic map and copy
1321 src = buf->ops->map(pipe, buf, 0);
1324 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1327 buf->ops->unmap(pipe, buf, src);
1330 sd->u.userptr += ret;
1335 * For lack of a better implementation, implement vmsplice() to userspace
1336 * as a simple copy of the pipes pages to the user iov.
1338 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1339 unsigned long nr_segs, unsigned int flags)
1341 struct pipe_inode_info *pipe;
1342 struct splice_desc sd;
1347 pipe = pipe_info(file->f_path.dentry->d_inode);
1352 mutex_lock(&pipe->inode->i_mutex);
1360 * Get user address base and length for this iovec.
1362 error = get_user(base, &iov->iov_base);
1363 if (unlikely(error))
1365 error = get_user(len, &iov->iov_len);
1366 if (unlikely(error))
1370 * Sanity check this iovec. 0 read succeeds.
1374 if (unlikely(!base)) {
1379 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1387 sd.u.userptr = base;
1390 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1408 mutex_unlock(&pipe->inode->i_mutex);
1417 * vmsplice splices a user address range into a pipe. It can be thought of
1418 * as splice-from-memory, where the regular splice is splice-from-file (or
1419 * to file). In both cases the output is a pipe, naturally.
1421 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1422 unsigned long nr_segs, unsigned int flags)
1424 struct pipe_inode_info *pipe;
1425 struct page *pages[PIPE_BUFFERS];
1426 struct partial_page partial[PIPE_BUFFERS];
1427 struct splice_pipe_desc spd = {
1431 .ops = &user_page_pipe_buf_ops,
1432 .spd_release = spd_release_page,
1435 pipe = pipe_info(file->f_path.dentry->d_inode);
1439 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1440 flags & SPLICE_F_GIFT);
1441 if (spd.nr_pages <= 0)
1442 return spd.nr_pages;
1444 return splice_to_pipe(pipe, &spd);
1448 * Note that vmsplice only really supports true splicing _from_ user memory
1449 * to a pipe, not the other way around. Splicing from user memory is a simple
1450 * operation that can be supported without any funky alignment restrictions
1451 * or nasty vm tricks. We simply map in the user memory and fill them into
1452 * a pipe. The reverse isn't quite as easy, though. There are two possible
1453 * solutions for that:
1455 * - memcpy() the data internally, at which point we might as well just
1456 * do a regular read() on the buffer anyway.
1457 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1458 * has restriction limitations on both ends of the pipe).
1460 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1463 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1464 unsigned long, nr_segs, unsigned int, flags)
1470 if (unlikely(nr_segs > UIO_MAXIOV))
1472 else if (unlikely(!nr_segs))
1476 file = fget_light(fd, &fput);
1478 if (file->f_mode & FMODE_WRITE)
1479 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1480 else if (file->f_mode & FMODE_READ)
1481 error = vmsplice_to_user(file, iov, nr_segs, flags);
1483 fput_light(file, fput);
1489 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1490 int, fd_out, loff_t __user *, off_out,
1491 size_t, len, unsigned int, flags)
1494 struct file *in, *out;
1495 int fput_in, fput_out;
1501 in = fget_light(fd_in, &fput_in);
1503 if (in->f_mode & FMODE_READ) {
1504 out = fget_light(fd_out, &fput_out);
1506 if (out->f_mode & FMODE_WRITE)
1507 error = do_splice(in, off_in,
1510 fput_light(out, fput_out);
1514 fput_light(in, fput_in);
1521 * Make sure there's data to read. Wait for input if we can, otherwise
1522 * return an appropriate error.
1524 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1529 * Check ->nrbufs without the inode lock first. This function
1530 * is speculative anyways, so missing one is ok.
1536 mutex_lock(&pipe->inode->i_mutex);
1538 while (!pipe->nrbufs) {
1539 if (signal_pending(current)) {
1545 if (!pipe->waiting_writers) {
1546 if (flags & SPLICE_F_NONBLOCK) {
1554 mutex_unlock(&pipe->inode->i_mutex);
1559 * Make sure there's writeable room. Wait for room if we can, otherwise
1560 * return an appropriate error.
1562 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1567 * Check ->nrbufs without the inode lock first. This function
1568 * is speculative anyways, so missing one is ok.
1570 if (pipe->nrbufs < PIPE_BUFFERS)
1574 mutex_lock(&pipe->inode->i_mutex);
1576 while (pipe->nrbufs >= PIPE_BUFFERS) {
1577 if (!pipe->readers) {
1578 send_sig(SIGPIPE, current, 0);
1582 if (flags & SPLICE_F_NONBLOCK) {
1586 if (signal_pending(current)) {
1590 pipe->waiting_writers++;
1592 pipe->waiting_writers--;
1595 mutex_unlock(&pipe->inode->i_mutex);
1600 * Link contents of ipipe to opipe.
1602 static int link_pipe(struct pipe_inode_info *ipipe,
1603 struct pipe_inode_info *opipe,
1604 size_t len, unsigned int flags)
1606 struct pipe_buffer *ibuf, *obuf;
1607 int ret = 0, i = 0, nbuf;
1610 * Potential ABBA deadlock, work around it by ordering lock
1611 * grabbing by inode address. Otherwise two different processes
1612 * could deadlock (one doing tee from A -> B, the other from B -> A).
1614 inode_double_lock(ipipe->inode, opipe->inode);
1617 if (!opipe->readers) {
1618 send_sig(SIGPIPE, current, 0);
1625 * If we have iterated all input buffers or ran out of
1626 * output room, break.
1628 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1631 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1632 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1635 * Get a reference to this pipe buffer,
1636 * so we can copy the contents over.
1638 ibuf->ops->get(ipipe, ibuf);
1640 obuf = opipe->bufs + nbuf;
1644 * Don't inherit the gift flag, we need to
1645 * prevent multiple steals of this page.
1647 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1649 if (obuf->len > len)
1659 * return EAGAIN if we have the potential of some data in the
1660 * future, otherwise just return 0
1662 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1665 inode_double_unlock(ipipe->inode, opipe->inode);
1668 * If we put data in the output pipe, wakeup any potential readers.
1672 if (waitqueue_active(&opipe->wait))
1673 wake_up_interruptible(&opipe->wait);
1674 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1681 * This is a tee(1) implementation that works on pipes. It doesn't copy
1682 * any data, it simply references the 'in' pages on the 'out' pipe.
1683 * The 'flags' used are the SPLICE_F_* variants, currently the only
1684 * applicable one is SPLICE_F_NONBLOCK.
1686 static long do_tee(struct file *in, struct file *out, size_t len,
1689 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1690 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1694 * Duplicate the contents of ipipe to opipe without actually
1697 if (ipipe && opipe && ipipe != opipe) {
1699 * Keep going, unless we encounter an error. The ipipe/opipe
1700 * ordering doesn't really matter.
1702 ret = link_ipipe_prep(ipipe, flags);
1704 ret = link_opipe_prep(opipe, flags);
1706 ret = link_pipe(ipipe, opipe, len, flags);
1713 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1722 in = fget_light(fdin, &fput_in);
1724 if (in->f_mode & FMODE_READ) {
1726 struct file *out = fget_light(fdout, &fput_out);
1729 if (out->f_mode & FMODE_WRITE)
1730 error = do_tee(in, out, len, flags);
1731 fput_light(out, fput_out);
1734 fput_light(in, fput_in);