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>
33 #include <linux/gfp.h>
34 #include <linux/socket.h>
37 * Attempt to steal a page from a pipe buffer. This should perhaps go into
38 * a vm helper function, it's already simplified quite a bit by the
39 * addition of remove_mapping(). If success is returned, the caller may
40 * attempt to reuse this page for another destination.
42 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
43 struct pipe_buffer *buf)
45 struct page *page = buf->page;
46 struct address_space *mapping;
50 mapping = page_mapping(page);
52 WARN_ON(!PageUptodate(page));
55 * At least for ext2 with nobh option, we need to wait on
56 * writeback completing on this page, since we'll remove it
57 * from the pagecache. Otherwise truncate wont wait on the
58 * page, allowing the disk blocks to be reused by someone else
59 * before we actually wrote our data to them. fs corruption
62 wait_on_page_writeback(page);
64 if (page_has_private(page) &&
65 !try_to_release_page(page, GFP_KERNEL))
69 * If we succeeded in removing the mapping, set LRU flag
72 if (remove_mapping(mapping, page)) {
73 buf->flags |= PIPE_BUF_FLAG_LRU;
79 * Raced with truncate or failed to remove page from current
80 * address space, unlock and return failure.
87 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
88 struct pipe_buffer *buf)
90 page_cache_release(buf->page);
91 buf->flags &= ~PIPE_BUF_FLAG_LRU;
95 * Check whether the contents of buf is OK to access. Since the content
96 * is a page cache page, IO may be in flight.
98 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
99 struct pipe_buffer *buf)
101 struct page *page = buf->page;
104 if (!PageUptodate(page)) {
108 * Page got truncated/unhashed. This will cause a 0-byte
109 * splice, if this is the first page.
111 if (!page->mapping) {
117 * Uh oh, read-error from disk.
119 if (!PageUptodate(page)) {
125 * Page is ok afterall, we are done.
136 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
138 .map = generic_pipe_buf_map,
139 .unmap = generic_pipe_buf_unmap,
140 .confirm = page_cache_pipe_buf_confirm,
141 .release = page_cache_pipe_buf_release,
142 .steal = page_cache_pipe_buf_steal,
143 .get = generic_pipe_buf_get,
146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
147 struct pipe_buffer *buf)
149 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
152 buf->flags |= PIPE_BUF_FLAG_LRU;
153 return generic_pipe_buf_steal(pipe, buf);
156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158 .map = generic_pipe_buf_map,
159 .unmap = generic_pipe_buf_unmap,
160 .confirm = generic_pipe_buf_confirm,
161 .release = page_cache_pipe_buf_release,
162 .steal = user_page_pipe_buf_steal,
163 .get = generic_pipe_buf_get,
166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
169 if (waitqueue_active(&pipe->wait))
170 wake_up_interruptible(&pipe->wait);
171 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
175 * splice_to_pipe - fill passed data into a pipe
176 * @pipe: pipe to fill
180 * @spd contains a map of pages and len/offset tuples, along with
181 * the struct pipe_buf_operations associated with these pages. This
182 * function will link that data to the pipe.
185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
186 struct splice_pipe_desc *spd)
188 unsigned int spd_pages = spd->nr_pages;
189 int ret, do_wakeup, page_nr;
198 if (!pipe->readers) {
199 send_sig(SIGPIPE, current, 0);
205 if (pipe->nrbufs < pipe->buffers) {
206 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
207 struct pipe_buffer *buf = pipe->bufs + newbuf;
209 buf->page = spd->pages[page_nr];
210 buf->offset = spd->partial[page_nr].offset;
211 buf->len = spd->partial[page_nr].len;
212 buf->private = spd->partial[page_nr].private;
214 if (spd->flags & SPLICE_F_GIFT)
215 buf->flags |= PIPE_BUF_FLAG_GIFT;
224 if (!--spd->nr_pages)
226 if (pipe->nrbufs < pipe->buffers)
232 if (spd->flags & SPLICE_F_NONBLOCK) {
238 if (signal_pending(current)) {
246 if (waitqueue_active(&pipe->wait))
247 wake_up_interruptible_sync(&pipe->wait);
248 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
252 pipe->waiting_writers++;
254 pipe->waiting_writers--;
260 wakeup_pipe_readers(pipe);
262 while (page_nr < spd_pages)
263 spd->spd_release(spd, page_nr++);
268 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
270 page_cache_release(spd->pages[i]);
274 * Check if we need to grow the arrays holding pages and partial page
277 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
279 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
281 spd->nr_pages_max = buffers;
282 if (buffers <= PIPE_DEF_BUFFERS)
285 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
286 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
288 if (spd->pages && spd->partial)
296 void splice_shrink_spd(struct splice_pipe_desc *spd)
298 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
306 __generic_file_splice_read(struct file *in, loff_t *ppos,
307 struct pipe_inode_info *pipe, size_t len,
310 struct address_space *mapping = in->f_mapping;
311 unsigned int loff, nr_pages, req_pages;
312 struct page *pages[PIPE_DEF_BUFFERS];
313 struct partial_page partial[PIPE_DEF_BUFFERS];
315 pgoff_t index, end_index;
318 struct splice_pipe_desc spd = {
321 .nr_pages_max = PIPE_DEF_BUFFERS,
323 .ops = &page_cache_pipe_buf_ops,
324 .spd_release = spd_release_page,
327 if (splice_grow_spd(pipe, &spd))
330 index = *ppos >> PAGE_CACHE_SHIFT;
331 loff = *ppos & ~PAGE_CACHE_MASK;
332 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
333 nr_pages = min(req_pages, spd.nr_pages_max);
336 * Lookup the (hopefully) full range of pages we need.
338 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
339 index += spd.nr_pages;
342 * If find_get_pages_contig() returned fewer pages than we needed,
343 * readahead/allocate the rest and fill in the holes.
345 if (spd.nr_pages < nr_pages)
346 page_cache_sync_readahead(mapping, &in->f_ra, in,
347 index, req_pages - spd.nr_pages);
350 while (spd.nr_pages < nr_pages) {
352 * Page could be there, find_get_pages_contig() breaks on
355 page = find_get_page(mapping, index);
358 * page didn't exist, allocate one.
360 page = page_cache_alloc_cold(mapping);
364 error = add_to_page_cache_lru(page, mapping, index,
366 if (unlikely(error)) {
367 page_cache_release(page);
368 if (error == -EEXIST)
373 * add_to_page_cache() locks the page, unlock it
374 * to avoid convoluting the logic below even more.
379 spd.pages[spd.nr_pages++] = page;
384 * Now loop over the map and see if we need to start IO on any
385 * pages, fill in the partial map, etc.
387 index = *ppos >> PAGE_CACHE_SHIFT;
388 nr_pages = spd.nr_pages;
390 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
391 unsigned int this_len;
397 * this_len is the max we'll use from this page
399 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
400 page = spd.pages[page_nr];
402 if (PageReadahead(page))
403 page_cache_async_readahead(mapping, &in->f_ra, in,
404 page, index, req_pages - page_nr);
407 * If the page isn't uptodate, we may need to start io on it
409 if (!PageUptodate(page)) {
413 * Page was truncated, or invalidated by the
414 * filesystem. Redo the find/create, but this time the
415 * page is kept locked, so there's no chance of another
416 * race with truncate/invalidate.
418 if (!page->mapping) {
420 page = find_or_create_page(mapping, index,
421 mapping_gfp_mask(mapping));
427 page_cache_release(spd.pages[page_nr]);
428 spd.pages[page_nr] = page;
431 * page was already under io and is now done, great
433 if (PageUptodate(page)) {
439 * need to read in the page
441 error = mapping->a_ops->readpage(in, page);
442 if (unlikely(error)) {
444 * We really should re-lookup the page here,
445 * but it complicates things a lot. Instead
446 * lets just do what we already stored, and
447 * we'll get it the next time we are called.
449 if (error == AOP_TRUNCATED_PAGE)
457 * i_size must be checked after PageUptodate.
459 isize = i_size_read(mapping->host);
460 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
461 if (unlikely(!isize || index > end_index))
465 * if this is the last page, see if we need to shrink
466 * the length and stop
468 if (end_index == index) {
472 * max good bytes in this page
474 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
479 * force quit after adding this page
481 this_len = min(this_len, plen - loff);
485 spd.partial[page_nr].offset = loff;
486 spd.partial[page_nr].len = this_len;
494 * Release any pages at the end, if we quit early. 'page_nr' is how far
495 * we got, 'nr_pages' is how many pages are in the map.
497 while (page_nr < nr_pages)
498 page_cache_release(spd.pages[page_nr++]);
499 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
502 error = splice_to_pipe(pipe, &spd);
504 splice_shrink_spd(&spd);
509 * generic_file_splice_read - splice data from file to a pipe
510 * @in: file to splice from
511 * @ppos: position in @in
512 * @pipe: pipe to splice to
513 * @len: number of bytes to splice
514 * @flags: splice modifier flags
517 * Will read pages from given file and fill them into a pipe. Can be
518 * used as long as the address_space operations for the source implements
522 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
523 struct pipe_inode_info *pipe, size_t len,
529 isize = i_size_read(in->f_mapping->host);
530 if (unlikely(*ppos >= isize))
533 left = isize - *ppos;
534 if (unlikely(left < len))
537 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
545 EXPORT_SYMBOL(generic_file_splice_read);
547 static const struct pipe_buf_operations default_pipe_buf_ops = {
549 .map = generic_pipe_buf_map,
550 .unmap = generic_pipe_buf_unmap,
551 .confirm = generic_pipe_buf_confirm,
552 .release = generic_pipe_buf_release,
553 .steal = generic_pipe_buf_steal,
554 .get = generic_pipe_buf_get,
557 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
558 struct pipe_buffer *buf)
563 /* Pipe buffer operations for a socket and similar. */
564 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
566 .map = generic_pipe_buf_map,
567 .unmap = generic_pipe_buf_unmap,
568 .confirm = generic_pipe_buf_confirm,
569 .release = generic_pipe_buf_release,
570 .steal = generic_pipe_buf_nosteal,
571 .get = generic_pipe_buf_get,
573 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
575 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
576 unsigned long vlen, loff_t offset)
584 /* The cast to a user pointer is valid due to the set_fs() */
585 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
591 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
599 /* The cast to a user pointer is valid due to the set_fs() */
600 res = vfs_write(file, (const char __user *)buf, count, &pos);
606 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
607 struct pipe_inode_info *pipe, size_t len,
610 unsigned int nr_pages;
611 unsigned int nr_freed;
613 struct page *pages[PIPE_DEF_BUFFERS];
614 struct partial_page partial[PIPE_DEF_BUFFERS];
615 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
620 struct splice_pipe_desc spd = {
623 .nr_pages_max = PIPE_DEF_BUFFERS,
625 .ops = &default_pipe_buf_ops,
626 .spd_release = spd_release_page,
629 if (splice_grow_spd(pipe, &spd))
634 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
635 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
640 offset = *ppos & ~PAGE_CACHE_MASK;
641 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
643 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
646 page = alloc_page(GFP_USER);
651 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
652 vec[i].iov_base = (void __user *) page_address(page);
653 vec[i].iov_len = this_len;
660 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
671 for (i = 0; i < spd.nr_pages; i++) {
672 this_len = min_t(size_t, vec[i].iov_len, res);
673 spd.partial[i].offset = 0;
674 spd.partial[i].len = this_len;
676 __free_page(spd.pages[i]);
682 spd.nr_pages -= nr_freed;
684 res = splice_to_pipe(pipe, &spd);
691 splice_shrink_spd(&spd);
695 for (i = 0; i < spd.nr_pages; i++)
696 __free_page(spd.pages[i]);
701 EXPORT_SYMBOL(default_file_splice_read);
704 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
705 * using sendpage(). Return the number of bytes sent.
707 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
708 struct pipe_buffer *buf, struct splice_desc *sd)
710 struct file *file = sd->u.file;
711 loff_t pos = sd->pos;
714 if (!likely(file->f_op && file->f_op->sendpage))
717 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
719 if (sd->len < sd->total_len && pipe->nrbufs > 1)
720 more |= MSG_SENDPAGE_NOTLAST;
722 return file->f_op->sendpage(file, buf->page, buf->offset,
723 sd->len, &pos, more);
727 * This is a little more tricky than the file -> pipe splicing. There are
728 * basically three cases:
730 * - Destination page already exists in the address space and there
731 * are users of it. For that case we have no other option that
732 * copying the data. Tough luck.
733 * - Destination page already exists in the address space, but there
734 * are no users of it. Make sure it's uptodate, then drop it. Fall
735 * through to last case.
736 * - Destination page does not exist, we can add the pipe page to
737 * the page cache and avoid the copy.
739 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
740 * sd->flags), we attempt to migrate pages from the pipe to the output
741 * file address space page cache. This is possible if no one else has
742 * the pipe page referenced outside of the pipe and page cache. If
743 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
744 * a new page in the output file page cache and fill/dirty that.
746 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
747 struct splice_desc *sd)
749 struct file *file = sd->u.file;
750 struct address_space *mapping = file->f_mapping;
751 unsigned int offset, this_len;
756 offset = sd->pos & ~PAGE_CACHE_MASK;
759 if (this_len + offset > PAGE_CACHE_SIZE)
760 this_len = PAGE_CACHE_SIZE - offset;
762 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
763 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
767 if (buf->page != page) {
769 * Careful, ->map() uses KM_USER0!
771 char *src = buf->ops->map(pipe, buf, 1);
772 char *dst = kmap_atomic(page, KM_USER1);
774 memcpy(dst + offset, src + buf->offset, this_len);
775 flush_dcache_page(page);
776 kunmap_atomic(dst, KM_USER1);
777 buf->ops->unmap(pipe, buf, src);
779 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
784 EXPORT_SYMBOL(pipe_to_file);
786 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
789 if (waitqueue_active(&pipe->wait))
790 wake_up_interruptible(&pipe->wait);
791 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
795 * splice_from_pipe_feed - feed available data from a pipe to a file
796 * @pipe: pipe to splice from
797 * @sd: information to @actor
798 * @actor: handler that splices the data
801 * This function loops over the pipe and calls @actor to do the
802 * actual moving of a single struct pipe_buffer to the desired
803 * destination. It returns when there's no more buffers left in
804 * the pipe or if the requested number of bytes (@sd->total_len)
805 * have been copied. It returns a positive number (one) if the
806 * pipe needs to be filled with more data, zero if the required
807 * number of bytes have been copied and -errno on error.
809 * This, together with splice_from_pipe_{begin,end,next}, may be
810 * used to implement the functionality of __splice_from_pipe() when
811 * locking is required around copying the pipe buffers to the
814 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
819 while (pipe->nrbufs) {
820 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
821 const struct pipe_buf_operations *ops = buf->ops;
824 if (sd->len > sd->total_len)
825 sd->len = sd->total_len;
827 ret = buf->ops->confirm(pipe, buf);
834 ret = actor(pipe, buf, sd);
841 sd->num_spliced += ret;
844 sd->total_len -= ret;
848 ops->release(pipe, buf);
849 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
852 sd->need_wakeup = true;
861 EXPORT_SYMBOL(splice_from_pipe_feed);
864 * splice_from_pipe_next - wait for some data to splice from
865 * @pipe: pipe to splice from
866 * @sd: information about the splice operation
869 * This function will wait for some data and return a positive
870 * value (one) if pipe buffers are available. It will return zero
871 * or -errno if no more data needs to be spliced.
873 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
875 while (!pipe->nrbufs) {
879 if (!pipe->waiting_writers && sd->num_spliced)
882 if (sd->flags & SPLICE_F_NONBLOCK)
885 if (signal_pending(current))
888 if (sd->need_wakeup) {
889 wakeup_pipe_writers(pipe);
890 sd->need_wakeup = false;
898 EXPORT_SYMBOL(splice_from_pipe_next);
901 * splice_from_pipe_begin - start splicing from pipe
902 * @sd: information about the splice operation
905 * This function should be called before a loop containing
906 * splice_from_pipe_next() and splice_from_pipe_feed() to
907 * initialize the necessary fields of @sd.
909 void splice_from_pipe_begin(struct splice_desc *sd)
912 sd->need_wakeup = false;
914 EXPORT_SYMBOL(splice_from_pipe_begin);
917 * splice_from_pipe_end - finish splicing from pipe
918 * @pipe: pipe to splice from
919 * @sd: information about the splice operation
922 * This function will wake up pipe writers if necessary. It should
923 * be called after a loop containing splice_from_pipe_next() and
924 * splice_from_pipe_feed().
926 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
929 wakeup_pipe_writers(pipe);
931 EXPORT_SYMBOL(splice_from_pipe_end);
934 * __splice_from_pipe - splice data from a pipe to given actor
935 * @pipe: pipe to splice from
936 * @sd: information to @actor
937 * @actor: handler that splices the data
940 * This function does little more than loop over the pipe and call
941 * @actor to do the actual moving of a single struct pipe_buffer to
942 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
946 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
951 splice_from_pipe_begin(sd);
953 ret = splice_from_pipe_next(pipe, sd);
955 ret = splice_from_pipe_feed(pipe, sd, actor);
957 splice_from_pipe_end(pipe, sd);
959 return sd->num_spliced ? sd->num_spliced : ret;
961 EXPORT_SYMBOL(__splice_from_pipe);
964 * splice_from_pipe - splice data from a pipe to a file
965 * @pipe: pipe to splice from
966 * @out: file to splice to
967 * @ppos: position in @out
968 * @len: how many bytes to splice
969 * @flags: splice modifier flags
970 * @actor: handler that splices the data
973 * See __splice_from_pipe. This function locks the pipe inode,
974 * otherwise it's identical to __splice_from_pipe().
977 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
978 loff_t *ppos, size_t len, unsigned int flags,
982 struct splice_desc sd = {
990 ret = __splice_from_pipe(pipe, &sd, actor);
997 * generic_file_splice_write - splice data from a pipe to a file
999 * @out: file to write to
1000 * @ppos: position in @out
1001 * @len: number of bytes to splice
1002 * @flags: splice modifier flags
1005 * Will either move or copy pages (determined by @flags options) from
1006 * the given pipe inode to the given file.
1010 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1011 loff_t *ppos, size_t len, unsigned int flags)
1013 struct address_space *mapping = out->f_mapping;
1014 struct inode *inode = mapping->host;
1015 struct splice_desc sd = {
1025 splice_from_pipe_begin(&sd);
1027 ret = splice_from_pipe_next(pipe, &sd);
1031 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1032 ret = file_remove_suid(out);
1034 file_update_time(out);
1035 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1037 mutex_unlock(&inode->i_mutex);
1039 splice_from_pipe_end(pipe, &sd);
1044 ret = sd.num_spliced;
1047 unsigned long nr_pages;
1050 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1052 err = generic_write_sync(out, *ppos, ret);
1057 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1063 EXPORT_SYMBOL(generic_file_splice_write);
1065 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1066 struct splice_desc *sd)
1071 data = buf->ops->map(pipe, buf, 0);
1072 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1073 buf->ops->unmap(pipe, buf, data);
1078 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1079 struct file *out, loff_t *ppos,
1080 size_t len, unsigned int flags)
1084 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1092 * generic_splice_sendpage - splice data from a pipe to a socket
1093 * @pipe: pipe to splice from
1094 * @out: socket to write to
1095 * @ppos: position in @out
1096 * @len: number of bytes to splice
1097 * @flags: splice modifier flags
1100 * Will send @len bytes from the pipe to a network socket. No data copying
1104 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1105 loff_t *ppos, size_t len, unsigned int flags)
1107 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1110 EXPORT_SYMBOL(generic_splice_sendpage);
1113 * Attempt to initiate a splice from pipe to file.
1115 long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1116 loff_t *ppos, size_t len, unsigned int flags)
1118 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1119 loff_t *, size_t, unsigned int);
1122 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1125 if (unlikely(out->f_flags & O_APPEND))
1128 ret = rw_verify_area(WRITE, out, ppos, len);
1129 if (unlikely(ret < 0))
1132 if (out->f_op && out->f_op->splice_write)
1133 splice_write = out->f_op->splice_write;
1135 splice_write = default_file_splice_write;
1137 return splice_write(pipe, out, ppos, len, flags);
1139 EXPORT_SYMBOL(do_splice_from);
1142 * Attempt to initiate a splice from a file to a pipe.
1144 long do_splice_to(struct file *in, loff_t *ppos,
1145 struct pipe_inode_info *pipe, size_t len,
1148 ssize_t (*splice_read)(struct file *, loff_t *,
1149 struct pipe_inode_info *, size_t, unsigned int);
1152 if (unlikely(!(in->f_mode & FMODE_READ)))
1155 ret = rw_verify_area(READ, in, ppos, len);
1156 if (unlikely(ret < 0))
1159 if (in->f_op && in->f_op->splice_read)
1160 splice_read = in->f_op->splice_read;
1162 splice_read = default_file_splice_read;
1164 return splice_read(in, ppos, pipe, len, flags);
1166 EXPORT_SYMBOL(do_splice_to);
1169 * splice_direct_to_actor - splices data directly between two non-pipes
1170 * @in: file to splice from
1171 * @sd: actor information on where to splice to
1172 * @actor: handles the data splicing
1175 * This is a special case helper to splice directly between two
1176 * points, without requiring an explicit pipe. Internally an allocated
1177 * pipe is cached in the process, and reused during the lifetime of
1181 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1182 splice_direct_actor *actor)
1184 struct pipe_inode_info *pipe;
1191 * We require the input being a regular file, as we don't want to
1192 * randomly drop data for eg socket -> socket splicing. Use the
1193 * piped splicing for that!
1195 i_mode = in->f_path.dentry->d_inode->i_mode;
1196 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1200 * neither in nor out is a pipe, setup an internal pipe attached to
1201 * 'out' and transfer the wanted data from 'in' to 'out' through that
1203 pipe = current->splice_pipe;
1204 if (unlikely(!pipe)) {
1205 pipe = alloc_pipe_info(NULL);
1210 * We don't have an immediate reader, but we'll read the stuff
1211 * out of the pipe right after the splice_to_pipe(). So set
1212 * PIPE_READERS appropriately.
1216 current->splice_pipe = pipe;
1224 len = sd->total_len;
1228 * Don't block on output, we have to drain the direct pipe.
1230 sd->flags &= ~SPLICE_F_NONBLOCK;
1234 loff_t pos = sd->pos, prev_pos = pos;
1236 ret = do_splice_to(in, &pos, pipe, len, flags);
1237 if (unlikely(ret <= 0))
1241 sd->total_len = read_len;
1244 * NOTE: nonblocking mode only applies to the input. We
1245 * must not do the output in nonblocking mode as then we
1246 * could get stuck data in the internal pipe:
1248 ret = actor(pipe, sd);
1249 if (unlikely(ret <= 0)) {
1258 if (ret < read_len) {
1259 sd->pos = prev_pos + ret;
1265 pipe->nrbufs = pipe->curbuf = 0;
1271 * If we did an incomplete transfer we must release
1272 * the pipe buffers in question:
1274 for (i = 0; i < pipe->buffers; i++) {
1275 struct pipe_buffer *buf = pipe->bufs + i;
1278 buf->ops->release(pipe, buf);
1288 EXPORT_SYMBOL(splice_direct_to_actor);
1290 static int direct_splice_actor(struct pipe_inode_info *pipe,
1291 struct splice_desc *sd)
1293 struct file *file = sd->u.file;
1295 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1300 * do_splice_direct - splices data directly between two files
1301 * @in: file to splice from
1302 * @ppos: input file offset
1303 * @out: file to splice to
1304 * @len: number of bytes to splice
1305 * @flags: splice modifier flags
1308 * For use by do_sendfile(). splice can easily emulate sendfile, but
1309 * doing it in the application would incur an extra system call
1310 * (splice in + splice out, as compared to just sendfile()). So this helper
1311 * can splice directly through a process-private pipe.
1314 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1315 size_t len, unsigned int flags)
1317 struct splice_desc sd = {
1326 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1333 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1334 struct pipe_inode_info *opipe,
1335 size_t len, unsigned int flags);
1338 * Determine where to splice to/from.
1340 static long do_splice(struct file *in, loff_t __user *off_in,
1341 struct file *out, loff_t __user *off_out,
1342 size_t len, unsigned int flags)
1344 struct pipe_inode_info *ipipe;
1345 struct pipe_inode_info *opipe;
1346 loff_t offset, *off;
1349 ipipe = get_pipe_info(in);
1350 opipe = get_pipe_info(out);
1352 if (ipipe && opipe) {
1353 if (off_in || off_out)
1356 if (!(in->f_mode & FMODE_READ))
1359 if (!(out->f_mode & FMODE_WRITE))
1362 /* Splicing to self would be fun, but... */
1366 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1373 if (!(out->f_mode & FMODE_PWRITE))
1375 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1381 ret = do_splice_from(ipipe, out, off, len, flags);
1383 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1393 if (!(in->f_mode & FMODE_PREAD))
1395 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1401 ret = do_splice_to(in, off, opipe, len, flags);
1403 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1413 * Map an iov into an array of pages and offset/length tupples. With the
1414 * partial_page structure, we can map several non-contiguous ranges into
1415 * our ones pages[] map instead of splitting that operation into pieces.
1416 * Could easily be exported as a generic helper for other users, in which
1417 * case one would probably want to add a 'max_nr_pages' parameter as well.
1419 static int get_iovec_page_array(const struct iovec __user *iov,
1420 unsigned int nr_vecs, struct page **pages,
1421 struct partial_page *partial, int aligned,
1422 unsigned int pipe_buffers)
1424 int buffers = 0, error = 0;
1427 unsigned long off, npages;
1434 if (copy_from_user(&entry, iov, sizeof(entry)))
1437 base = entry.iov_base;
1438 len = entry.iov_len;
1441 * Sanity check this iovec. 0 read succeeds.
1447 if (!access_ok(VERIFY_READ, base, len))
1451 * Get this base offset and number of pages, then map
1452 * in the user pages.
1454 off = (unsigned long) base & ~PAGE_MASK;
1457 * If asked for alignment, the offset must be zero and the
1458 * length a multiple of the PAGE_SIZE.
1461 if (aligned && (off || len & ~PAGE_MASK))
1464 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1465 if (npages > pipe_buffers - buffers)
1466 npages = pipe_buffers - buffers;
1468 error = get_user_pages_fast((unsigned long)base, npages,
1469 0, &pages[buffers]);
1471 if (unlikely(error <= 0))
1475 * Fill this contiguous range into the partial page map.
1477 for (i = 0; i < error; i++) {
1478 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1480 partial[buffers].offset = off;
1481 partial[buffers].len = plen;
1489 * We didn't complete this iov, stop here since it probably
1490 * means we have to move some of this into a pipe to
1491 * be able to continue.
1497 * Don't continue if we mapped fewer pages than we asked for,
1498 * or if we mapped the max number of pages that we have
1501 if (error < npages || buffers == pipe_buffers)
1514 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1515 struct splice_desc *sd)
1521 * See if we can use the atomic maps, by prefaulting in the
1522 * pages and doing an atomic copy
1524 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1525 src = buf->ops->map(pipe, buf, 1);
1526 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1528 buf->ops->unmap(pipe, buf, src);
1536 * No dice, use slow non-atomic map and copy
1538 src = buf->ops->map(pipe, buf, 0);
1541 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1544 buf->ops->unmap(pipe, buf, src);
1547 sd->u.userptr += ret;
1552 * For lack of a better implementation, implement vmsplice() to userspace
1553 * as a simple copy of the pipes pages to the user iov.
1555 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1556 unsigned long nr_segs, unsigned int flags)
1558 struct pipe_inode_info *pipe;
1559 struct splice_desc sd;
1564 pipe = get_pipe_info(file);
1576 * Get user address base and length for this iovec.
1578 error = get_user(base, &iov->iov_base);
1579 if (unlikely(error))
1581 error = get_user(len, &iov->iov_len);
1582 if (unlikely(error))
1586 * Sanity check this iovec. 0 read succeeds.
1590 if (unlikely(!base)) {
1595 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1603 sd.u.userptr = base;
1606 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1632 * vmsplice splices a user address range into a pipe. It can be thought of
1633 * as splice-from-memory, where the regular splice is splice-from-file (or
1634 * to file). In both cases the output is a pipe, naturally.
1636 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1637 unsigned long nr_segs, unsigned int flags)
1639 struct pipe_inode_info *pipe;
1640 struct page *pages[PIPE_DEF_BUFFERS];
1641 struct partial_page partial[PIPE_DEF_BUFFERS];
1642 struct splice_pipe_desc spd = {
1645 .nr_pages_max = PIPE_DEF_BUFFERS,
1647 .ops = &user_page_pipe_buf_ops,
1648 .spd_release = spd_release_page,
1652 pipe = get_pipe_info(file);
1656 if (splice_grow_spd(pipe, &spd))
1659 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1660 spd.partial, flags & SPLICE_F_GIFT,
1662 if (spd.nr_pages <= 0)
1665 ret = splice_to_pipe(pipe, &spd);
1667 splice_shrink_spd(&spd);
1672 * Note that vmsplice only really supports true splicing _from_ user memory
1673 * to a pipe, not the other way around. Splicing from user memory is a simple
1674 * operation that can be supported without any funky alignment restrictions
1675 * or nasty vm tricks. We simply map in the user memory and fill them into
1676 * a pipe. The reverse isn't quite as easy, though. There are two possible
1677 * solutions for that:
1679 * - memcpy() the data internally, at which point we might as well just
1680 * do a regular read() on the buffer anyway.
1681 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1682 * has restriction limitations on both ends of the pipe).
1684 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1687 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1688 unsigned long, nr_segs, unsigned int, flags)
1694 if (unlikely(nr_segs > UIO_MAXIOV))
1696 else if (unlikely(!nr_segs))
1700 file = fget_light(fd, &fput);
1702 if (file->f_mode & FMODE_WRITE)
1703 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1704 else if (file->f_mode & FMODE_READ)
1705 error = vmsplice_to_user(file, iov, nr_segs, flags);
1707 fput_light(file, fput);
1713 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1714 int, fd_out, loff_t __user *, off_out,
1715 size_t, len, unsigned int, flags)
1718 struct file *in, *out;
1719 int fput_in, fput_out;
1725 in = fget_light(fd_in, &fput_in);
1727 if (in->f_mode & FMODE_READ) {
1728 out = fget_light(fd_out, &fput_out);
1730 if (out->f_mode & FMODE_WRITE)
1731 error = do_splice(in, off_in,
1734 fput_light(out, fput_out);
1738 fput_light(in, fput_in);
1745 * Make sure there's data to read. Wait for input if we can, otherwise
1746 * return an appropriate error.
1748 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1753 * Check ->nrbufs without the inode lock first. This function
1754 * is speculative anyways, so missing one is ok.
1762 while (!pipe->nrbufs) {
1763 if (signal_pending(current)) {
1769 if (!pipe->waiting_writers) {
1770 if (flags & SPLICE_F_NONBLOCK) {
1783 * Make sure there's writeable room. Wait for room if we can, otherwise
1784 * return an appropriate error.
1786 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1791 * Check ->nrbufs without the inode lock first. This function
1792 * is speculative anyways, so missing one is ok.
1794 if (pipe->nrbufs < pipe->buffers)
1800 while (pipe->nrbufs >= pipe->buffers) {
1801 if (!pipe->readers) {
1802 send_sig(SIGPIPE, current, 0);
1806 if (flags & SPLICE_F_NONBLOCK) {
1810 if (signal_pending(current)) {
1814 pipe->waiting_writers++;
1816 pipe->waiting_writers--;
1824 * Splice contents of ipipe to opipe.
1826 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1827 struct pipe_inode_info *opipe,
1828 size_t len, unsigned int flags)
1830 struct pipe_buffer *ibuf, *obuf;
1832 bool input_wakeup = false;
1836 ret = ipipe_prep(ipipe, flags);
1840 ret = opipe_prep(opipe, flags);
1845 * Potential ABBA deadlock, work around it by ordering lock
1846 * grabbing by pipe info address. Otherwise two different processes
1847 * could deadlock (one doing tee from A -> B, the other from B -> A).
1849 pipe_double_lock(ipipe, opipe);
1852 if (!opipe->readers) {
1853 send_sig(SIGPIPE, current, 0);
1859 if (!ipipe->nrbufs && !ipipe->writers)
1863 * Cannot make any progress, because either the input
1864 * pipe is empty or the output pipe is full.
1866 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1867 /* Already processed some buffers, break */
1871 if (flags & SPLICE_F_NONBLOCK) {
1877 * We raced with another reader/writer and haven't
1878 * managed to process any buffers. A zero return
1879 * value means EOF, so retry instead.
1886 ibuf = ipipe->bufs + ipipe->curbuf;
1887 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1888 obuf = opipe->bufs + nbuf;
1890 if (len >= ibuf->len) {
1892 * Simply move the whole buffer from ipipe to opipe
1897 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1899 input_wakeup = true;
1902 * Get a reference to this pipe buffer,
1903 * so we can copy the contents over.
1905 ibuf->ops->get(ipipe, ibuf);
1909 * Don't inherit the gift flag, we need to
1910 * prevent multiple steals of this page.
1912 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1916 ibuf->offset += obuf->len;
1917 ibuf->len -= obuf->len;
1927 * If we put data in the output pipe, wakeup any potential readers.
1930 wakeup_pipe_readers(opipe);
1933 wakeup_pipe_writers(ipipe);
1939 * Link contents of ipipe to opipe.
1941 static int link_pipe(struct pipe_inode_info *ipipe,
1942 struct pipe_inode_info *opipe,
1943 size_t len, unsigned int flags)
1945 struct pipe_buffer *ibuf, *obuf;
1946 int ret = 0, i = 0, nbuf;
1949 * Potential ABBA deadlock, work around it by ordering lock
1950 * grabbing by pipe info address. Otherwise two different processes
1951 * could deadlock (one doing tee from A -> B, the other from B -> A).
1953 pipe_double_lock(ipipe, opipe);
1956 if (!opipe->readers) {
1957 send_sig(SIGPIPE, current, 0);
1964 * If we have iterated all input buffers or ran out of
1965 * output room, break.
1967 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1970 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1971 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1974 * Get a reference to this pipe buffer,
1975 * so we can copy the contents over.
1977 ibuf->ops->get(ipipe, ibuf);
1979 obuf = opipe->bufs + nbuf;
1983 * Don't inherit the gift flag, we need to
1984 * prevent multiple steals of this page.
1986 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1988 if (obuf->len > len)
1998 * return EAGAIN if we have the potential of some data in the
1999 * future, otherwise just return 0
2001 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2008 * If we put data in the output pipe, wakeup any potential readers.
2011 wakeup_pipe_readers(opipe);
2017 * This is a tee(1) implementation that works on pipes. It doesn't copy
2018 * any data, it simply references the 'in' pages on the 'out' pipe.
2019 * The 'flags' used are the SPLICE_F_* variants, currently the only
2020 * applicable one is SPLICE_F_NONBLOCK.
2022 static long do_tee(struct file *in, struct file *out, size_t len,
2025 struct pipe_inode_info *ipipe = get_pipe_info(in);
2026 struct pipe_inode_info *opipe = get_pipe_info(out);
2030 * Duplicate the contents of ipipe to opipe without actually
2033 if (ipipe && opipe && ipipe != opipe) {
2035 * Keep going, unless we encounter an error. The ipipe/opipe
2036 * ordering doesn't really matter.
2038 ret = ipipe_prep(ipipe, flags);
2040 ret = opipe_prep(opipe, flags);
2042 ret = link_pipe(ipipe, opipe, len, flags);
2049 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2058 in = fget_light(fdin, &fput_in);
2060 if (in->f_mode & FMODE_READ) {
2062 struct file *out = fget_light(fdout, &fput_out);
2065 if (out->f_mode & FMODE_WRITE)
2066 error = do_tee(in, out, len, flags);
2067 fput_light(out, fput_out);
2070 fput_light(in, fput_in);