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;
201 if (!pipe->readers) {
202 send_sig(SIGPIPE, current, 0);
208 if (pipe->nrbufs < pipe->buffers) {
209 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
210 struct pipe_buffer *buf = pipe->bufs + newbuf;
212 buf->page = spd->pages[page_nr];
213 buf->offset = spd->partial[page_nr].offset;
214 buf->len = spd->partial[page_nr].len;
215 buf->private = spd->partial[page_nr].private;
218 if (spd->flags & SPLICE_F_GIFT)
219 buf->flags |= PIPE_BUF_FLAG_GIFT;
228 if (!--spd->nr_pages)
230 if (pipe->nrbufs < pipe->buffers)
236 if (spd->flags & SPLICE_F_NONBLOCK) {
242 if (signal_pending(current)) {
250 if (waitqueue_active(&pipe->wait))
251 wake_up_interruptible_sync(&pipe->wait);
252 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
256 pipe->waiting_writers++;
258 pipe->waiting_writers--;
264 wakeup_pipe_readers(pipe);
266 while (page_nr < spd_pages)
267 spd->spd_release(spd, page_nr++);
272 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
274 page_cache_release(spd->pages[i]);
278 * Check if we need to grow the arrays holding pages and partial page
281 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
283 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
285 spd->nr_pages_max = buffers;
286 if (buffers <= PIPE_DEF_BUFFERS)
289 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
290 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
292 if (spd->pages && spd->partial)
300 void splice_shrink_spd(struct splice_pipe_desc *spd)
302 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
310 __generic_file_splice_read(struct file *in, loff_t *ppos,
311 struct pipe_inode_info *pipe, size_t len,
314 struct address_space *mapping = in->f_mapping;
315 unsigned int loff, nr_pages, req_pages;
316 struct page *pages[PIPE_DEF_BUFFERS];
317 struct partial_page partial[PIPE_DEF_BUFFERS];
319 pgoff_t index, end_index;
322 struct splice_pipe_desc spd = {
325 .nr_pages_max = PIPE_DEF_BUFFERS,
327 .ops = &page_cache_pipe_buf_ops,
328 .spd_release = spd_release_page,
331 if (splice_grow_spd(pipe, &spd))
334 index = *ppos >> PAGE_CACHE_SHIFT;
335 loff = *ppos & ~PAGE_CACHE_MASK;
336 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
337 nr_pages = min(req_pages, spd.nr_pages_max);
340 * Lookup the (hopefully) full range of pages we need.
342 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
343 index += spd.nr_pages;
346 * If find_get_pages_contig() returned fewer pages than we needed,
347 * readahead/allocate the rest and fill in the holes.
349 if (spd.nr_pages < nr_pages)
350 page_cache_sync_readahead(mapping, &in->f_ra, in,
351 index, req_pages - spd.nr_pages);
354 while (spd.nr_pages < nr_pages) {
356 * Page could be there, find_get_pages_contig() breaks on
359 page = find_get_page(mapping, index);
362 * page didn't exist, allocate one.
364 page = page_cache_alloc_cold(mapping);
368 error = add_to_page_cache_lru(page, mapping, index,
370 if (unlikely(error)) {
371 page_cache_release(page);
372 if (error == -EEXIST)
377 * add_to_page_cache() locks the page, unlock it
378 * to avoid convoluting the logic below even more.
383 spd.pages[spd.nr_pages++] = page;
388 * Now loop over the map and see if we need to start IO on any
389 * pages, fill in the partial map, etc.
391 index = *ppos >> PAGE_CACHE_SHIFT;
392 nr_pages = spd.nr_pages;
394 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
395 unsigned int this_len;
401 * this_len is the max we'll use from this page
403 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
404 page = spd.pages[page_nr];
406 if (PageReadahead(page))
407 page_cache_async_readahead(mapping, &in->f_ra, in,
408 page, index, req_pages - page_nr);
411 * If the page isn't uptodate, we may need to start io on it
413 if (!PageUptodate(page)) {
417 * Page was truncated, or invalidated by the
418 * filesystem. Redo the find/create, but this time the
419 * page is kept locked, so there's no chance of another
420 * race with truncate/invalidate.
422 if (!page->mapping) {
424 page = find_or_create_page(mapping, index,
425 mapping_gfp_mask(mapping));
431 page_cache_release(spd.pages[page_nr]);
432 spd.pages[page_nr] = page;
435 * page was already under io and is now done, great
437 if (PageUptodate(page)) {
443 * need to read in the page
445 error = mapping->a_ops->readpage(in, page);
446 if (unlikely(error)) {
448 * We really should re-lookup the page here,
449 * but it complicates things a lot. Instead
450 * lets just do what we already stored, and
451 * we'll get it the next time we are called.
453 if (error == AOP_TRUNCATED_PAGE)
461 * i_size must be checked after PageUptodate.
463 isize = i_size_read(mapping->host);
464 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
465 if (unlikely(!isize || index > end_index))
469 * if this is the last page, see if we need to shrink
470 * the length and stop
472 if (end_index == index) {
476 * max good bytes in this page
478 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
483 * force quit after adding this page
485 this_len = min(this_len, plen - loff);
489 spd.partial[page_nr].offset = loff;
490 spd.partial[page_nr].len = this_len;
498 * Release any pages at the end, if we quit early. 'page_nr' is how far
499 * we got, 'nr_pages' is how many pages are in the map.
501 while (page_nr < nr_pages)
502 page_cache_release(spd.pages[page_nr++]);
503 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
506 error = splice_to_pipe(pipe, &spd);
508 splice_shrink_spd(&spd);
513 * generic_file_splice_read - splice data from file to a pipe
514 * @in: file to splice from
515 * @ppos: position in @in
516 * @pipe: pipe to splice to
517 * @len: number of bytes to splice
518 * @flags: splice modifier flags
521 * Will read pages from given file and fill them into a pipe. Can be
522 * used as long as the address_space operations for the source implements
526 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
527 struct pipe_inode_info *pipe, size_t len,
533 isize = i_size_read(in->f_mapping->host);
534 if (unlikely(*ppos >= isize))
537 left = isize - *ppos;
538 if (unlikely(left < len))
541 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
549 EXPORT_SYMBOL(generic_file_splice_read);
551 static const struct pipe_buf_operations default_pipe_buf_ops = {
553 .map = generic_pipe_buf_map,
554 .unmap = generic_pipe_buf_unmap,
555 .confirm = generic_pipe_buf_confirm,
556 .release = generic_pipe_buf_release,
557 .steal = generic_pipe_buf_steal,
558 .get = generic_pipe_buf_get,
561 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
562 struct pipe_buffer *buf)
567 /* Pipe buffer operations for a socket and similar. */
568 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
570 .map = generic_pipe_buf_map,
571 .unmap = generic_pipe_buf_unmap,
572 .confirm = generic_pipe_buf_confirm,
573 .release = generic_pipe_buf_release,
574 .steal = generic_pipe_buf_nosteal,
575 .get = generic_pipe_buf_get,
577 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
579 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
580 unsigned long vlen, loff_t offset)
588 /* The cast to a user pointer is valid due to the set_fs() */
589 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
595 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
603 /* The cast to a user pointer is valid due to the set_fs() */
604 res = vfs_write(file, (const char __user *)buf, count, &pos);
610 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
611 struct pipe_inode_info *pipe, size_t len,
614 unsigned int nr_pages;
615 unsigned int nr_freed;
617 struct page *pages[PIPE_DEF_BUFFERS];
618 struct partial_page partial[PIPE_DEF_BUFFERS];
619 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
624 struct splice_pipe_desc spd = {
627 .nr_pages_max = PIPE_DEF_BUFFERS,
629 .ops = &default_pipe_buf_ops,
630 .spd_release = spd_release_page,
633 if (splice_grow_spd(pipe, &spd))
638 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
639 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
644 offset = *ppos & ~PAGE_CACHE_MASK;
645 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
647 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
650 page = alloc_page(GFP_USER);
655 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
656 vec[i].iov_base = (void __user *) page_address(page);
657 vec[i].iov_len = this_len;
664 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
675 for (i = 0; i < spd.nr_pages; i++) {
676 this_len = min_t(size_t, vec[i].iov_len, res);
677 spd.partial[i].offset = 0;
678 spd.partial[i].len = this_len;
680 __free_page(spd.pages[i]);
686 spd.nr_pages -= nr_freed;
688 res = splice_to_pipe(pipe, &spd);
695 splice_shrink_spd(&spd);
699 for (i = 0; i < spd.nr_pages; i++)
700 __free_page(spd.pages[i]);
705 EXPORT_SYMBOL(default_file_splice_read);
708 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
709 * using sendpage(). Return the number of bytes sent.
711 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
712 struct pipe_buffer *buf, struct splice_desc *sd)
714 struct file *file = sd->u.file;
715 loff_t pos = sd->pos;
718 if (!likely(file->f_op && file->f_op->sendpage))
721 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
723 if (sd->len < sd->total_len && pipe->nrbufs > 1)
724 more |= MSG_SENDPAGE_NOTLAST;
726 return file->f_op->sendpage(file, buf->page, buf->offset,
727 sd->len, &pos, more);
731 * This is a little more tricky than the file -> pipe splicing. There are
732 * basically three cases:
734 * - Destination page already exists in the address space and there
735 * are users of it. For that case we have no other option that
736 * copying the data. Tough luck.
737 * - Destination page already exists in the address space, but there
738 * are no users of it. Make sure it's uptodate, then drop it. Fall
739 * through to last case.
740 * - Destination page does not exist, we can add the pipe page to
741 * the page cache and avoid the copy.
743 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
744 * sd->flags), we attempt to migrate pages from the pipe to the output
745 * file address space page cache. This is possible if no one else has
746 * the pipe page referenced outside of the pipe and page cache. If
747 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
748 * a new page in the output file page cache and fill/dirty that.
750 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
751 struct splice_desc *sd)
753 struct file *file = sd->u.file;
754 struct address_space *mapping = file->f_mapping;
755 unsigned int offset, this_len;
760 offset = sd->pos & ~PAGE_CACHE_MASK;
763 if (this_len + offset > PAGE_CACHE_SIZE)
764 this_len = PAGE_CACHE_SIZE - offset;
766 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
767 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
771 if (buf->page != page) {
773 * Careful, ->map() uses KM_USER0!
775 char *src = buf->ops->map(pipe, buf, 1);
776 char *dst = kmap_atomic(page, KM_USER1);
778 memcpy(dst + offset, src + buf->offset, this_len);
779 flush_dcache_page(page);
780 kunmap_atomic(dst, KM_USER1);
781 buf->ops->unmap(pipe, buf, src);
783 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
788 EXPORT_SYMBOL(pipe_to_file);
790 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
793 if (waitqueue_active(&pipe->wait))
794 wake_up_interruptible(&pipe->wait);
795 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
799 * splice_from_pipe_feed - feed available data from a pipe to a file
800 * @pipe: pipe to splice from
801 * @sd: information to @actor
802 * @actor: handler that splices the data
805 * This function loops over the pipe and calls @actor to do the
806 * actual moving of a single struct pipe_buffer to the desired
807 * destination. It returns when there's no more buffers left in
808 * the pipe or if the requested number of bytes (@sd->total_len)
809 * have been copied. It returns a positive number (one) if the
810 * pipe needs to be filled with more data, zero if the required
811 * number of bytes have been copied and -errno on error.
813 * This, together with splice_from_pipe_{begin,end,next}, may be
814 * used to implement the functionality of __splice_from_pipe() when
815 * locking is required around copying the pipe buffers to the
818 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
823 while (pipe->nrbufs) {
824 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
825 const struct pipe_buf_operations *ops = buf->ops;
828 if (sd->len > sd->total_len)
829 sd->len = sd->total_len;
831 ret = buf->ops->confirm(pipe, buf);
838 ret = actor(pipe, buf, sd);
845 sd->num_spliced += ret;
848 sd->total_len -= ret;
852 ops->release(pipe, buf);
853 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
856 sd->need_wakeup = true;
865 EXPORT_SYMBOL(splice_from_pipe_feed);
868 * splice_from_pipe_next - wait for some data to splice from
869 * @pipe: pipe to splice from
870 * @sd: information about the splice operation
873 * This function will wait for some data and return a positive
874 * value (one) if pipe buffers are available. It will return zero
875 * or -errno if no more data needs to be spliced.
877 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
880 * Check for signal early to make process killable when there are
881 * always buffers available
883 if (signal_pending(current))
886 while (!pipe->nrbufs) {
890 if (!pipe->waiting_writers && sd->num_spliced)
893 if (sd->flags & SPLICE_F_NONBLOCK)
896 if (signal_pending(current))
899 if (sd->need_wakeup) {
900 wakeup_pipe_writers(pipe);
901 sd->need_wakeup = false;
909 EXPORT_SYMBOL(splice_from_pipe_next);
912 * splice_from_pipe_begin - start splicing from pipe
913 * @sd: information about the splice operation
916 * This function should be called before a loop containing
917 * splice_from_pipe_next() and splice_from_pipe_feed() to
918 * initialize the necessary fields of @sd.
920 void splice_from_pipe_begin(struct splice_desc *sd)
923 sd->need_wakeup = false;
925 EXPORT_SYMBOL(splice_from_pipe_begin);
928 * splice_from_pipe_end - finish splicing from pipe
929 * @pipe: pipe to splice from
930 * @sd: information about the splice operation
933 * This function will wake up pipe writers if necessary. It should
934 * be called after a loop containing splice_from_pipe_next() and
935 * splice_from_pipe_feed().
937 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
940 wakeup_pipe_writers(pipe);
942 EXPORT_SYMBOL(splice_from_pipe_end);
945 * __splice_from_pipe - splice data from a pipe to given actor
946 * @pipe: pipe to splice from
947 * @sd: information to @actor
948 * @actor: handler that splices the data
951 * This function does little more than loop over the pipe and call
952 * @actor to do the actual moving of a single struct pipe_buffer to
953 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
957 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
962 splice_from_pipe_begin(sd);
965 ret = splice_from_pipe_next(pipe, sd);
967 ret = splice_from_pipe_feed(pipe, sd, actor);
969 splice_from_pipe_end(pipe, sd);
971 return sd->num_spliced ? sd->num_spliced : ret;
973 EXPORT_SYMBOL(__splice_from_pipe);
976 * splice_from_pipe - splice data from a pipe to a file
977 * @pipe: pipe to splice from
978 * @out: file to splice to
979 * @ppos: position in @out
980 * @len: how many bytes to splice
981 * @flags: splice modifier flags
982 * @actor: handler that splices the data
985 * See __splice_from_pipe. This function locks the pipe inode,
986 * otherwise it's identical to __splice_from_pipe().
989 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
990 loff_t *ppos, size_t len, unsigned int flags,
994 struct splice_desc sd = {
1002 ret = __splice_from_pipe(pipe, &sd, actor);
1009 * generic_file_splice_write - splice data from a pipe to a file
1011 * @out: file to write to
1012 * @ppos: position in @out
1013 * @len: number of bytes to splice
1014 * @flags: splice modifier flags
1017 * Will either move or copy pages (determined by @flags options) from
1018 * the given pipe inode to the given file.
1022 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1023 loff_t *ppos, size_t len, unsigned int flags)
1025 struct address_space *mapping = out->f_mapping;
1026 struct inode *inode = mapping->host;
1027 struct splice_desc sd = {
1033 ret = generic_write_checks(out, ppos, &len, S_ISBLK(inode->i_mode));
1041 splice_from_pipe_begin(&sd);
1043 ret = splice_from_pipe_next(pipe, &sd);
1047 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1048 ret = file_remove_suid(out);
1050 file_update_time(out);
1051 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1053 mutex_unlock(&inode->i_mutex);
1055 splice_from_pipe_end(pipe, &sd);
1060 ret = sd.num_spliced;
1063 unsigned long nr_pages;
1066 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1068 err = generic_write_sync(out, *ppos, ret);
1073 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1079 EXPORT_SYMBOL(generic_file_splice_write);
1081 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1082 struct splice_desc *sd)
1087 data = buf->ops->map(pipe, buf, 0);
1088 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1089 buf->ops->unmap(pipe, buf, data);
1094 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1095 struct file *out, loff_t *ppos,
1096 size_t len, unsigned int flags)
1100 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1108 * generic_splice_sendpage - splice data from a pipe to a socket
1109 * @pipe: pipe to splice from
1110 * @out: socket to write to
1111 * @ppos: position in @out
1112 * @len: number of bytes to splice
1113 * @flags: splice modifier flags
1116 * Will send @len bytes from the pipe to a network socket. No data copying
1120 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1121 loff_t *ppos, size_t len, unsigned int flags)
1123 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1126 EXPORT_SYMBOL(generic_splice_sendpage);
1129 * Attempt to initiate a splice from pipe to file.
1131 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1132 loff_t *ppos, size_t len, unsigned int flags)
1134 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1135 loff_t *, size_t, unsigned int);
1138 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1141 if (unlikely(out->f_flags & O_APPEND))
1144 ret = rw_verify_area(WRITE, out, ppos, len);
1145 if (unlikely(ret < 0))
1148 if (out->f_op && out->f_op->splice_write)
1149 splice_write = out->f_op->splice_write;
1151 splice_write = default_file_splice_write;
1153 return splice_write(pipe, out, ppos, len, flags);
1157 * Attempt to initiate a splice from a file to a pipe.
1159 static long do_splice_to(struct file *in, loff_t *ppos,
1160 struct pipe_inode_info *pipe, size_t len,
1163 ssize_t (*splice_read)(struct file *, loff_t *,
1164 struct pipe_inode_info *, size_t, unsigned int);
1167 if (unlikely(!(in->f_mode & FMODE_READ)))
1170 ret = rw_verify_area(READ, in, ppos, len);
1171 if (unlikely(ret < 0))
1174 if (in->f_op && in->f_op->splice_read)
1175 splice_read = in->f_op->splice_read;
1177 splice_read = default_file_splice_read;
1179 return splice_read(in, ppos, pipe, len, flags);
1183 * splice_direct_to_actor - splices data directly between two non-pipes
1184 * @in: file to splice from
1185 * @sd: actor information on where to splice to
1186 * @actor: handles the data splicing
1189 * This is a special case helper to splice directly between two
1190 * points, without requiring an explicit pipe. Internally an allocated
1191 * pipe is cached in the process, and reused during the lifetime of
1195 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1196 splice_direct_actor *actor)
1198 struct pipe_inode_info *pipe;
1205 * We require the input being a regular file, as we don't want to
1206 * randomly drop data for eg socket -> socket splicing. Use the
1207 * piped splicing for that!
1209 i_mode = in->f_path.dentry->d_inode->i_mode;
1210 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1214 * neither in nor out is a pipe, setup an internal pipe attached to
1215 * 'out' and transfer the wanted data from 'in' to 'out' through that
1217 pipe = current->splice_pipe;
1218 if (unlikely(!pipe)) {
1219 pipe = alloc_pipe_info(NULL);
1224 * We don't have an immediate reader, but we'll read the stuff
1225 * out of the pipe right after the splice_to_pipe(). So set
1226 * PIPE_READERS appropriately.
1230 current->splice_pipe = pipe;
1238 len = sd->total_len;
1242 * Don't block on output, we have to drain the direct pipe.
1244 sd->flags &= ~SPLICE_F_NONBLOCK;
1245 more = sd->flags & SPLICE_F_MORE;
1249 loff_t pos = sd->pos, prev_pos = pos;
1251 ret = do_splice_to(in, &pos, pipe, len, flags);
1252 if (unlikely(ret <= 0))
1256 sd->total_len = read_len;
1259 * If more data is pending, set SPLICE_F_MORE
1260 * If this is the last data and SPLICE_F_MORE was not set
1261 * initially, clears it.
1264 sd->flags |= SPLICE_F_MORE;
1266 sd->flags &= ~SPLICE_F_MORE;
1268 * NOTE: nonblocking mode only applies to the input. We
1269 * must not do the output in nonblocking mode as then we
1270 * could get stuck data in the internal pipe:
1272 ret = actor(pipe, sd);
1273 if (unlikely(ret <= 0)) {
1282 if (ret < read_len) {
1283 sd->pos = prev_pos + ret;
1289 pipe->nrbufs = pipe->curbuf = 0;
1295 * If we did an incomplete transfer we must release
1296 * the pipe buffers in question:
1298 for (i = 0; i < pipe->buffers; i++) {
1299 struct pipe_buffer *buf = pipe->bufs + i;
1302 buf->ops->release(pipe, buf);
1312 EXPORT_SYMBOL(splice_direct_to_actor);
1314 static int direct_splice_actor(struct pipe_inode_info *pipe,
1315 struct splice_desc *sd)
1317 struct file *file = sd->u.file;
1319 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1324 * do_splice_direct - splices data directly between two files
1325 * @in: file to splice from
1326 * @ppos: input file offset
1327 * @out: file to splice to
1328 * @len: number of bytes to splice
1329 * @flags: splice modifier flags
1332 * For use by do_sendfile(). splice can easily emulate sendfile, but
1333 * doing it in the application would incur an extra system call
1334 * (splice in + splice out, as compared to just sendfile()). So this helper
1335 * can splice directly through a process-private pipe.
1338 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1339 size_t len, unsigned int flags)
1341 struct splice_desc sd = {
1350 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1357 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1358 struct pipe_inode_info *opipe,
1359 size_t len, unsigned int flags);
1362 * Determine where to splice to/from.
1364 static long do_splice(struct file *in, loff_t __user *off_in,
1365 struct file *out, loff_t __user *off_out,
1366 size_t len, unsigned int flags)
1368 struct pipe_inode_info *ipipe;
1369 struct pipe_inode_info *opipe;
1370 loff_t offset, *off;
1373 ipipe = get_pipe_info(in);
1374 opipe = get_pipe_info(out);
1376 if (ipipe && opipe) {
1377 if (off_in || off_out)
1380 if (!(in->f_mode & FMODE_READ))
1383 if (!(out->f_mode & FMODE_WRITE))
1386 /* Splicing to self would be fun, but... */
1390 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1397 if (!(out->f_mode & FMODE_PWRITE))
1399 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1405 ret = do_splice_from(ipipe, out, off, len, flags);
1407 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1417 if (!(in->f_mode & FMODE_PREAD))
1419 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1425 ret = do_splice_to(in, off, opipe, len, flags);
1427 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1437 * Map an iov into an array of pages and offset/length tupples. With the
1438 * partial_page structure, we can map several non-contiguous ranges into
1439 * our ones pages[] map instead of splitting that operation into pieces.
1440 * Could easily be exported as a generic helper for other users, in which
1441 * case one would probably want to add a 'max_nr_pages' parameter as well.
1443 static int get_iovec_page_array(const struct iovec __user *iov,
1444 unsigned int nr_vecs, struct page **pages,
1445 struct partial_page *partial, int aligned,
1446 unsigned int pipe_buffers)
1448 int buffers = 0, error = 0;
1451 unsigned long off, npages;
1458 if (copy_from_user(&entry, iov, sizeof(entry)))
1461 base = entry.iov_base;
1462 len = entry.iov_len;
1465 * Sanity check this iovec. 0 read succeeds.
1471 if (!access_ok(VERIFY_READ, base, len))
1475 * Get this base offset and number of pages, then map
1476 * in the user pages.
1478 off = (unsigned long) base & ~PAGE_MASK;
1481 * If asked for alignment, the offset must be zero and the
1482 * length a multiple of the PAGE_SIZE.
1485 if (aligned && (off || len & ~PAGE_MASK))
1488 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1489 if (npages > pipe_buffers - buffers)
1490 npages = pipe_buffers - buffers;
1492 error = get_user_pages_fast((unsigned long)base, npages,
1493 0, &pages[buffers]);
1495 if (unlikely(error <= 0))
1499 * Fill this contiguous range into the partial page map.
1501 for (i = 0; i < error; i++) {
1502 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1504 partial[buffers].offset = off;
1505 partial[buffers].len = plen;
1513 * We didn't complete this iov, stop here since it probably
1514 * means we have to move some of this into a pipe to
1515 * be able to continue.
1521 * Don't continue if we mapped fewer pages than we asked for,
1522 * or if we mapped the max number of pages that we have
1525 if (error < npages || buffers == pipe_buffers)
1538 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1539 struct splice_desc *sd)
1545 * See if we can use the atomic maps, by prefaulting in the
1546 * pages and doing an atomic copy
1548 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1549 src = buf->ops->map(pipe, buf, 1);
1550 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1552 buf->ops->unmap(pipe, buf, src);
1560 * No dice, use slow non-atomic map and copy
1562 src = buf->ops->map(pipe, buf, 0);
1565 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1568 buf->ops->unmap(pipe, buf, src);
1571 sd->u.userptr += ret;
1576 * For lack of a better implementation, implement vmsplice() to userspace
1577 * as a simple copy of the pipes pages to the user iov.
1579 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1580 unsigned long nr_segs, unsigned int flags)
1582 struct pipe_inode_info *pipe;
1583 struct splice_desc sd;
1588 pipe = get_pipe_info(file);
1600 * Get user address base and length for this iovec.
1602 error = get_user(base, &iov->iov_base);
1603 if (unlikely(error))
1605 error = get_user(len, &iov->iov_len);
1606 if (unlikely(error))
1610 * Sanity check this iovec. 0 read succeeds.
1614 if (unlikely(!base)) {
1619 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1627 sd.u.userptr = base;
1630 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1656 * vmsplice splices a user address range into a pipe. It can be thought of
1657 * as splice-from-memory, where the regular splice is splice-from-file (or
1658 * to file). In both cases the output is a pipe, naturally.
1660 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1661 unsigned long nr_segs, unsigned int flags)
1663 struct pipe_inode_info *pipe;
1664 struct page *pages[PIPE_DEF_BUFFERS];
1665 struct partial_page partial[PIPE_DEF_BUFFERS];
1666 struct splice_pipe_desc spd = {
1669 .nr_pages_max = PIPE_DEF_BUFFERS,
1671 .ops = &user_page_pipe_buf_ops,
1672 .spd_release = spd_release_page,
1676 pipe = get_pipe_info(file);
1680 if (splice_grow_spd(pipe, &spd))
1683 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1684 spd.partial, flags & SPLICE_F_GIFT,
1686 if (spd.nr_pages <= 0)
1689 ret = splice_to_pipe(pipe, &spd);
1691 splice_shrink_spd(&spd);
1696 * Note that vmsplice only really supports true splicing _from_ user memory
1697 * to a pipe, not the other way around. Splicing from user memory is a simple
1698 * operation that can be supported without any funky alignment restrictions
1699 * or nasty vm tricks. We simply map in the user memory and fill them into
1700 * a pipe. The reverse isn't quite as easy, though. There are two possible
1701 * solutions for that:
1703 * - memcpy() the data internally, at which point we might as well just
1704 * do a regular read() on the buffer anyway.
1705 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1706 * has restriction limitations on both ends of the pipe).
1708 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1711 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1712 unsigned long, nr_segs, unsigned int, flags)
1718 if (unlikely(nr_segs > UIO_MAXIOV))
1720 else if (unlikely(!nr_segs))
1724 file = fget_light(fd, &fput);
1726 if (file->f_mode & FMODE_WRITE)
1727 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1728 else if (file->f_mode & FMODE_READ)
1729 error = vmsplice_to_user(file, iov, nr_segs, flags);
1731 fput_light(file, fput);
1737 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1738 int, fd_out, loff_t __user *, off_out,
1739 size_t, len, unsigned int, flags)
1742 struct file *in, *out;
1743 int fput_in, fput_out;
1749 in = fget_light(fd_in, &fput_in);
1751 if (in->f_mode & FMODE_READ) {
1752 out = fget_light(fd_out, &fput_out);
1754 if (out->f_mode & FMODE_WRITE)
1755 error = do_splice(in, off_in,
1758 fput_light(out, fput_out);
1762 fput_light(in, fput_in);
1769 * Make sure there's data to read. Wait for input if we can, otherwise
1770 * return an appropriate error.
1772 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1777 * Check ->nrbufs without the inode lock first. This function
1778 * is speculative anyways, so missing one is ok.
1786 while (!pipe->nrbufs) {
1787 if (signal_pending(current)) {
1793 if (!pipe->waiting_writers) {
1794 if (flags & SPLICE_F_NONBLOCK) {
1807 * Make sure there's writeable room. Wait for room if we can, otherwise
1808 * return an appropriate error.
1810 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1815 * Check ->nrbufs without the inode lock first. This function
1816 * is speculative anyways, so missing one is ok.
1818 if (pipe->nrbufs < pipe->buffers)
1824 while (pipe->nrbufs >= pipe->buffers) {
1825 if (!pipe->readers) {
1826 send_sig(SIGPIPE, current, 0);
1830 if (flags & SPLICE_F_NONBLOCK) {
1834 if (signal_pending(current)) {
1838 pipe->waiting_writers++;
1840 pipe->waiting_writers--;
1848 * Splice contents of ipipe to opipe.
1850 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1851 struct pipe_inode_info *opipe,
1852 size_t len, unsigned int flags)
1854 struct pipe_buffer *ibuf, *obuf;
1856 bool input_wakeup = false;
1860 ret = ipipe_prep(ipipe, flags);
1864 ret = opipe_prep(opipe, flags);
1869 * Potential ABBA deadlock, work around it by ordering lock
1870 * grabbing by pipe info address. Otherwise two different processes
1871 * could deadlock (one doing tee from A -> B, the other from B -> A).
1873 pipe_double_lock(ipipe, opipe);
1876 if (!opipe->readers) {
1877 send_sig(SIGPIPE, current, 0);
1883 if (!ipipe->nrbufs && !ipipe->writers)
1887 * Cannot make any progress, because either the input
1888 * pipe is empty or the output pipe is full.
1890 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1891 /* Already processed some buffers, break */
1895 if (flags & SPLICE_F_NONBLOCK) {
1901 * We raced with another reader/writer and haven't
1902 * managed to process any buffers. A zero return
1903 * value means EOF, so retry instead.
1910 ibuf = ipipe->bufs + ipipe->curbuf;
1911 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1912 obuf = opipe->bufs + nbuf;
1914 if (len >= ibuf->len) {
1916 * Simply move the whole buffer from ipipe to opipe
1921 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1923 input_wakeup = true;
1926 * Get a reference to this pipe buffer,
1927 * so we can copy the contents over.
1929 ibuf->ops->get(ipipe, ibuf);
1933 * Don't inherit the gift flag, we need to
1934 * prevent multiple steals of this page.
1936 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1940 ibuf->offset += obuf->len;
1941 ibuf->len -= obuf->len;
1951 * If we put data in the output pipe, wakeup any potential readers.
1954 wakeup_pipe_readers(opipe);
1957 wakeup_pipe_writers(ipipe);
1963 * Link contents of ipipe to opipe.
1965 static int link_pipe(struct pipe_inode_info *ipipe,
1966 struct pipe_inode_info *opipe,
1967 size_t len, unsigned int flags)
1969 struct pipe_buffer *ibuf, *obuf;
1970 int ret = 0, i = 0, nbuf;
1973 * Potential ABBA deadlock, work around it by ordering lock
1974 * grabbing by pipe info address. Otherwise two different processes
1975 * could deadlock (one doing tee from A -> B, the other from B -> A).
1977 pipe_double_lock(ipipe, opipe);
1980 if (!opipe->readers) {
1981 send_sig(SIGPIPE, current, 0);
1988 * If we have iterated all input buffers or ran out of
1989 * output room, break.
1991 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1994 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1995 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1998 * Get a reference to this pipe buffer,
1999 * so we can copy the contents over.
2001 ibuf->ops->get(ipipe, ibuf);
2003 obuf = opipe->bufs + nbuf;
2007 * Don't inherit the gift flag, we need to
2008 * prevent multiple steals of this page.
2010 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2012 if (obuf->len > len)
2022 * return EAGAIN if we have the potential of some data in the
2023 * future, otherwise just return 0
2025 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2032 * If we put data in the output pipe, wakeup any potential readers.
2035 wakeup_pipe_readers(opipe);
2041 * This is a tee(1) implementation that works on pipes. It doesn't copy
2042 * any data, it simply references the 'in' pages on the 'out' pipe.
2043 * The 'flags' used are the SPLICE_F_* variants, currently the only
2044 * applicable one is SPLICE_F_NONBLOCK.
2046 static long do_tee(struct file *in, struct file *out, size_t len,
2049 struct pipe_inode_info *ipipe = get_pipe_info(in);
2050 struct pipe_inode_info *opipe = get_pipe_info(out);
2054 * Duplicate the contents of ipipe to opipe without actually
2057 if (ipipe && opipe && ipipe != opipe) {
2059 * Keep going, unless we encounter an error. The ipipe/opipe
2060 * ordering doesn't really matter.
2062 ret = ipipe_prep(ipipe, flags);
2064 ret = opipe_prep(opipe, flags);
2066 ret = link_pipe(ipipe, opipe, len, flags);
2073 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2082 in = fget_light(fdin, &fput_in);
2084 if (in->f_mode & FMODE_READ) {
2086 struct file *out = fget_light(fdout, &fput_out);
2089 if (out->f_mode & FMODE_WRITE)
2090 error = do_tee(in, out, len, flags);
2091 fput_light(out, fput_out);
2094 fput_light(in, fput_in);