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 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);
1155 EXPORT_SYMBOL(do_splice_from);
1158 * Attempt to initiate a splice from a file to a pipe.
1160 long do_splice_to(struct file *in, loff_t *ppos,
1161 struct pipe_inode_info *pipe, size_t len,
1164 ssize_t (*splice_read)(struct file *, loff_t *,
1165 struct pipe_inode_info *, size_t, unsigned int);
1168 if (unlikely(!(in->f_mode & FMODE_READ)))
1171 ret = rw_verify_area(READ, in, ppos, len);
1172 if (unlikely(ret < 0))
1175 if (in->f_op && in->f_op->splice_read)
1176 splice_read = in->f_op->splice_read;
1178 splice_read = default_file_splice_read;
1180 return splice_read(in, ppos, pipe, len, flags);
1182 EXPORT_SYMBOL(do_splice_to);
1185 * splice_direct_to_actor - splices data directly between two non-pipes
1186 * @in: file to splice from
1187 * @sd: actor information on where to splice to
1188 * @actor: handles the data splicing
1191 * This is a special case helper to splice directly between two
1192 * points, without requiring an explicit pipe. Internally an allocated
1193 * pipe is cached in the process, and reused during the lifetime of
1197 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1198 splice_direct_actor *actor)
1200 struct pipe_inode_info *pipe;
1207 * We require the input being a regular file, as we don't want to
1208 * randomly drop data for eg socket -> socket splicing. Use the
1209 * piped splicing for that!
1211 i_mode = in->f_path.dentry->d_inode->i_mode;
1212 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1216 * neither in nor out is a pipe, setup an internal pipe attached to
1217 * 'out' and transfer the wanted data from 'in' to 'out' through that
1219 pipe = current->splice_pipe;
1220 if (unlikely(!pipe)) {
1221 pipe = alloc_pipe_info(NULL);
1226 * We don't have an immediate reader, but we'll read the stuff
1227 * out of the pipe right after the splice_to_pipe(). So set
1228 * PIPE_READERS appropriately.
1232 current->splice_pipe = pipe;
1240 len = sd->total_len;
1244 * Don't block on output, we have to drain the direct pipe.
1246 sd->flags &= ~SPLICE_F_NONBLOCK;
1247 more = sd->flags & SPLICE_F_MORE;
1251 loff_t pos = sd->pos, prev_pos = pos;
1253 ret = do_splice_to(in, &pos, pipe, len, flags);
1254 if (unlikely(ret <= 0))
1258 sd->total_len = read_len;
1261 * If more data is pending, set SPLICE_F_MORE
1262 * If this is the last data and SPLICE_F_MORE was not set
1263 * initially, clears it.
1266 sd->flags |= SPLICE_F_MORE;
1268 sd->flags &= ~SPLICE_F_MORE;
1270 * NOTE: nonblocking mode only applies to the input. We
1271 * must not do the output in nonblocking mode as then we
1272 * could get stuck data in the internal pipe:
1274 ret = actor(pipe, sd);
1275 if (unlikely(ret <= 0)) {
1284 if (ret < read_len) {
1285 sd->pos = prev_pos + ret;
1291 pipe->nrbufs = pipe->curbuf = 0;
1297 * If we did an incomplete transfer we must release
1298 * the pipe buffers in question:
1300 for (i = 0; i < pipe->buffers; i++) {
1301 struct pipe_buffer *buf = pipe->bufs + i;
1304 buf->ops->release(pipe, buf);
1314 EXPORT_SYMBOL(splice_direct_to_actor);
1316 static int direct_splice_actor(struct pipe_inode_info *pipe,
1317 struct splice_desc *sd)
1319 struct file *file = sd->u.file;
1321 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1326 * do_splice_direct - splices data directly between two files
1327 * @in: file to splice from
1328 * @ppos: input file offset
1329 * @out: file to splice to
1330 * @len: number of bytes to splice
1331 * @flags: splice modifier flags
1334 * For use by do_sendfile(). splice can easily emulate sendfile, but
1335 * doing it in the application would incur an extra system call
1336 * (splice in + splice out, as compared to just sendfile()). So this helper
1337 * can splice directly through a process-private pipe.
1340 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1341 size_t len, unsigned int flags)
1343 struct splice_desc sd = {
1352 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1359 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1360 struct pipe_inode_info *opipe,
1361 size_t len, unsigned int flags);
1364 * Determine where to splice to/from.
1366 static long do_splice(struct file *in, loff_t __user *off_in,
1367 struct file *out, loff_t __user *off_out,
1368 size_t len, unsigned int flags)
1370 struct pipe_inode_info *ipipe;
1371 struct pipe_inode_info *opipe;
1372 loff_t offset, *off;
1375 ipipe = get_pipe_info(in);
1376 opipe = get_pipe_info(out);
1378 if (ipipe && opipe) {
1379 if (off_in || off_out)
1382 if (!(in->f_mode & FMODE_READ))
1385 if (!(out->f_mode & FMODE_WRITE))
1388 /* Splicing to self would be fun, but... */
1392 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1399 if (!(out->f_mode & FMODE_PWRITE))
1401 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1407 ret = do_splice_from(ipipe, out, off, len, flags);
1409 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1419 if (!(in->f_mode & FMODE_PREAD))
1421 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1427 ret = do_splice_to(in, off, opipe, len, flags);
1429 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1439 * Map an iov into an array of pages and offset/length tupples. With the
1440 * partial_page structure, we can map several non-contiguous ranges into
1441 * our ones pages[] map instead of splitting that operation into pieces.
1442 * Could easily be exported as a generic helper for other users, in which
1443 * case one would probably want to add a 'max_nr_pages' parameter as well.
1445 static int get_iovec_page_array(const struct iovec __user *iov,
1446 unsigned int nr_vecs, struct page **pages,
1447 struct partial_page *partial, int aligned,
1448 unsigned int pipe_buffers)
1450 int buffers = 0, error = 0;
1453 unsigned long off, npages;
1460 if (copy_from_user(&entry, iov, sizeof(entry)))
1463 base = entry.iov_base;
1464 len = entry.iov_len;
1467 * Sanity check this iovec. 0 read succeeds.
1473 if (!access_ok(VERIFY_READ, base, len))
1477 * Get this base offset and number of pages, then map
1478 * in the user pages.
1480 off = (unsigned long) base & ~PAGE_MASK;
1483 * If asked for alignment, the offset must be zero and the
1484 * length a multiple of the PAGE_SIZE.
1487 if (aligned && (off || len & ~PAGE_MASK))
1490 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1491 if (npages > pipe_buffers - buffers)
1492 npages = pipe_buffers - buffers;
1494 error = get_user_pages_fast((unsigned long)base, npages,
1495 0, &pages[buffers]);
1497 if (unlikely(error <= 0))
1501 * Fill this contiguous range into the partial page map.
1503 for (i = 0; i < error; i++) {
1504 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1506 partial[buffers].offset = off;
1507 partial[buffers].len = plen;
1515 * We didn't complete this iov, stop here since it probably
1516 * means we have to move some of this into a pipe to
1517 * be able to continue.
1523 * Don't continue if we mapped fewer pages than we asked for,
1524 * or if we mapped the max number of pages that we have
1527 if (error < npages || buffers == pipe_buffers)
1540 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1541 struct splice_desc *sd)
1547 * See if we can use the atomic maps, by prefaulting in the
1548 * pages and doing an atomic copy
1550 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1551 src = buf->ops->map(pipe, buf, 1);
1552 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1554 buf->ops->unmap(pipe, buf, src);
1562 * No dice, use slow non-atomic map and copy
1564 src = buf->ops->map(pipe, buf, 0);
1567 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1570 buf->ops->unmap(pipe, buf, src);
1573 sd->u.userptr += ret;
1578 * For lack of a better implementation, implement vmsplice() to userspace
1579 * as a simple copy of the pipes pages to the user iov.
1581 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1582 unsigned long nr_segs, unsigned int flags)
1584 struct pipe_inode_info *pipe;
1585 struct splice_desc sd;
1590 pipe = get_pipe_info(file);
1602 * Get user address base and length for this iovec.
1604 error = get_user(base, &iov->iov_base);
1605 if (unlikely(error))
1607 error = get_user(len, &iov->iov_len);
1608 if (unlikely(error))
1612 * Sanity check this iovec. 0 read succeeds.
1616 if (unlikely(!base)) {
1621 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1629 sd.u.userptr = base;
1632 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1658 * vmsplice splices a user address range into a pipe. It can be thought of
1659 * as splice-from-memory, where the regular splice is splice-from-file (or
1660 * to file). In both cases the output is a pipe, naturally.
1662 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1663 unsigned long nr_segs, unsigned int flags)
1665 struct pipe_inode_info *pipe;
1666 struct page *pages[PIPE_DEF_BUFFERS];
1667 struct partial_page partial[PIPE_DEF_BUFFERS];
1668 struct splice_pipe_desc spd = {
1671 .nr_pages_max = PIPE_DEF_BUFFERS,
1673 .ops = &user_page_pipe_buf_ops,
1674 .spd_release = spd_release_page,
1678 pipe = get_pipe_info(file);
1682 if (splice_grow_spd(pipe, &spd))
1685 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1686 spd.partial, flags & SPLICE_F_GIFT,
1688 if (spd.nr_pages <= 0)
1691 ret = splice_to_pipe(pipe, &spd);
1693 splice_shrink_spd(&spd);
1698 * Note that vmsplice only really supports true splicing _from_ user memory
1699 * to a pipe, not the other way around. Splicing from user memory is a simple
1700 * operation that can be supported without any funky alignment restrictions
1701 * or nasty vm tricks. We simply map in the user memory and fill them into
1702 * a pipe. The reverse isn't quite as easy, though. There are two possible
1703 * solutions for that:
1705 * - memcpy() the data internally, at which point we might as well just
1706 * do a regular read() on the buffer anyway.
1707 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1708 * has restriction limitations on both ends of the pipe).
1710 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1713 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1714 unsigned long, nr_segs, unsigned int, flags)
1720 if (unlikely(nr_segs > UIO_MAXIOV))
1722 else if (unlikely(!nr_segs))
1726 file = fget_light(fd, &fput);
1728 if (file->f_mode & FMODE_WRITE)
1729 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1730 else if (file->f_mode & FMODE_READ)
1731 error = vmsplice_to_user(file, iov, nr_segs, flags);
1733 fput_light(file, fput);
1739 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1740 int, fd_out, loff_t __user *, off_out,
1741 size_t, len, unsigned int, flags)
1744 struct file *in, *out;
1745 int fput_in, fput_out;
1751 in = fget_light(fd_in, &fput_in);
1753 if (in->f_mode & FMODE_READ) {
1754 out = fget_light(fd_out, &fput_out);
1756 if (out->f_mode & FMODE_WRITE)
1757 error = do_splice(in, off_in,
1760 fput_light(out, fput_out);
1764 fput_light(in, fput_in);
1771 * Make sure there's data to read. Wait for input if we can, otherwise
1772 * return an appropriate error.
1774 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1779 * Check ->nrbufs without the inode lock first. This function
1780 * is speculative anyways, so missing one is ok.
1788 while (!pipe->nrbufs) {
1789 if (signal_pending(current)) {
1795 if (!pipe->waiting_writers) {
1796 if (flags & SPLICE_F_NONBLOCK) {
1809 * Make sure there's writeable room. Wait for room if we can, otherwise
1810 * return an appropriate error.
1812 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1817 * Check ->nrbufs without the inode lock first. This function
1818 * is speculative anyways, so missing one is ok.
1820 if (pipe->nrbufs < pipe->buffers)
1826 while (pipe->nrbufs >= pipe->buffers) {
1827 if (!pipe->readers) {
1828 send_sig(SIGPIPE, current, 0);
1832 if (flags & SPLICE_F_NONBLOCK) {
1836 if (signal_pending(current)) {
1840 pipe->waiting_writers++;
1842 pipe->waiting_writers--;
1850 * Splice contents of ipipe to opipe.
1852 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1853 struct pipe_inode_info *opipe,
1854 size_t len, unsigned int flags)
1856 struct pipe_buffer *ibuf, *obuf;
1858 bool input_wakeup = false;
1862 ret = ipipe_prep(ipipe, flags);
1866 ret = opipe_prep(opipe, flags);
1871 * Potential ABBA deadlock, work around it by ordering lock
1872 * grabbing by pipe info address. Otherwise two different processes
1873 * could deadlock (one doing tee from A -> B, the other from B -> A).
1875 pipe_double_lock(ipipe, opipe);
1878 if (!opipe->readers) {
1879 send_sig(SIGPIPE, current, 0);
1885 if (!ipipe->nrbufs && !ipipe->writers)
1889 * Cannot make any progress, because either the input
1890 * pipe is empty or the output pipe is full.
1892 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1893 /* Already processed some buffers, break */
1897 if (flags & SPLICE_F_NONBLOCK) {
1903 * We raced with another reader/writer and haven't
1904 * managed to process any buffers. A zero return
1905 * value means EOF, so retry instead.
1912 ibuf = ipipe->bufs + ipipe->curbuf;
1913 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1914 obuf = opipe->bufs + nbuf;
1916 if (len >= ibuf->len) {
1918 * Simply move the whole buffer from ipipe to opipe
1923 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1925 input_wakeup = true;
1928 * Get a reference to this pipe buffer,
1929 * so we can copy the contents over.
1931 ibuf->ops->get(ipipe, ibuf);
1935 * Don't inherit the gift flag, we need to
1936 * prevent multiple steals of this page.
1938 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1942 ibuf->offset += obuf->len;
1943 ibuf->len -= obuf->len;
1953 * If we put data in the output pipe, wakeup any potential readers.
1956 wakeup_pipe_readers(opipe);
1959 wakeup_pipe_writers(ipipe);
1965 * Link contents of ipipe to opipe.
1967 static int link_pipe(struct pipe_inode_info *ipipe,
1968 struct pipe_inode_info *opipe,
1969 size_t len, unsigned int flags)
1971 struct pipe_buffer *ibuf, *obuf;
1972 int ret = 0, i = 0, nbuf;
1975 * Potential ABBA deadlock, work around it by ordering lock
1976 * grabbing by pipe info address. Otherwise two different processes
1977 * could deadlock (one doing tee from A -> B, the other from B -> A).
1979 pipe_double_lock(ipipe, opipe);
1982 if (!opipe->readers) {
1983 send_sig(SIGPIPE, current, 0);
1990 * If we have iterated all input buffers or ran out of
1991 * output room, break.
1993 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1996 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1997 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2000 * Get a reference to this pipe buffer,
2001 * so we can copy the contents over.
2003 ibuf->ops->get(ipipe, ibuf);
2005 obuf = opipe->bufs + nbuf;
2009 * Don't inherit the gift flag, we need to
2010 * prevent multiple steals of this page.
2012 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2014 if (obuf->len > len)
2024 * return EAGAIN if we have the potential of some data in the
2025 * future, otherwise just return 0
2027 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2034 * If we put data in the output pipe, wakeup any potential readers.
2037 wakeup_pipe_readers(opipe);
2043 * This is a tee(1) implementation that works on pipes. It doesn't copy
2044 * any data, it simply references the 'in' pages on the 'out' pipe.
2045 * The 'flags' used are the SPLICE_F_* variants, currently the only
2046 * applicable one is SPLICE_F_NONBLOCK.
2048 static long do_tee(struct file *in, struct file *out, size_t len,
2051 struct pipe_inode_info *ipipe = get_pipe_info(in);
2052 struct pipe_inode_info *opipe = get_pipe_info(out);
2056 * Duplicate the contents of ipipe to opipe without actually
2059 if (ipipe && opipe && ipipe != opipe) {
2061 * Keep going, unless we encounter an error. The ipipe/opipe
2062 * ordering doesn't really matter.
2064 ret = ipipe_prep(ipipe, flags);
2066 ret = opipe_prep(opipe, flags);
2068 ret = link_pipe(ipipe, opipe, len, flags);
2075 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2084 in = fget_light(fdin, &fput_in);
2086 if (in->f_mode & FMODE_READ) {
2088 struct file *out = fget_light(fdout, &fput_out);
2091 if (out->f_mode & FMODE_WRITE)
2092 error = do_tee(in, out, len, flags);
2093 fput_light(out, fput_out);
2096 fput_light(in, fput_in);