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;
217 if (spd->flags & SPLICE_F_GIFT)
218 buf->flags |= PIPE_BUF_FLAG_GIFT;
227 if (!--spd->nr_pages)
229 if (pipe->nrbufs < pipe->buffers)
235 if (spd->flags & SPLICE_F_NONBLOCK) {
241 if (signal_pending(current)) {
249 if (waitqueue_active(&pipe->wait))
250 wake_up_interruptible_sync(&pipe->wait);
251 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
255 pipe->waiting_writers++;
257 pipe->waiting_writers--;
263 wakeup_pipe_readers(pipe);
265 while (page_nr < spd_pages)
266 spd->spd_release(spd, page_nr++);
271 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
273 page_cache_release(spd->pages[i]);
277 * Check if we need to grow the arrays holding pages and partial page
280 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
282 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
284 spd->nr_pages_max = buffers;
285 if (buffers <= PIPE_DEF_BUFFERS)
288 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
289 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
291 if (spd->pages && spd->partial)
299 void splice_shrink_spd(struct splice_pipe_desc *spd)
301 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
309 __generic_file_splice_read(struct file *in, loff_t *ppos,
310 struct pipe_inode_info *pipe, size_t len,
313 struct address_space *mapping = in->f_mapping;
314 unsigned int loff, nr_pages, req_pages;
315 struct page *pages[PIPE_DEF_BUFFERS];
316 struct partial_page partial[PIPE_DEF_BUFFERS];
318 pgoff_t index, end_index;
321 struct splice_pipe_desc spd = {
324 .nr_pages_max = PIPE_DEF_BUFFERS,
326 .ops = &page_cache_pipe_buf_ops,
327 .spd_release = spd_release_page,
330 if (splice_grow_spd(pipe, &spd))
333 index = *ppos >> PAGE_CACHE_SHIFT;
334 loff = *ppos & ~PAGE_CACHE_MASK;
335 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
336 nr_pages = min(req_pages, spd.nr_pages_max);
339 * Lookup the (hopefully) full range of pages we need.
341 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
342 index += spd.nr_pages;
345 * If find_get_pages_contig() returned fewer pages than we needed,
346 * readahead/allocate the rest and fill in the holes.
348 if (spd.nr_pages < nr_pages)
349 page_cache_sync_readahead(mapping, &in->f_ra, in,
350 index, req_pages - spd.nr_pages);
353 while (spd.nr_pages < nr_pages) {
355 * Page could be there, find_get_pages_contig() breaks on
358 page = find_get_page(mapping, index);
361 * page didn't exist, allocate one.
363 page = page_cache_alloc_cold(mapping);
367 error = add_to_page_cache_lru(page, mapping, index,
369 if (unlikely(error)) {
370 page_cache_release(page);
371 if (error == -EEXIST)
376 * add_to_page_cache() locks the page, unlock it
377 * to avoid convoluting the logic below even more.
382 spd.pages[spd.nr_pages++] = page;
387 * Now loop over the map and see if we need to start IO on any
388 * pages, fill in the partial map, etc.
390 index = *ppos >> PAGE_CACHE_SHIFT;
391 nr_pages = spd.nr_pages;
393 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
394 unsigned int this_len;
400 * this_len is the max we'll use from this page
402 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
403 page = spd.pages[page_nr];
405 if (PageReadahead(page))
406 page_cache_async_readahead(mapping, &in->f_ra, in,
407 page, index, req_pages - page_nr);
410 * If the page isn't uptodate, we may need to start io on it
412 if (!PageUptodate(page)) {
416 * Page was truncated, or invalidated by the
417 * filesystem. Redo the find/create, but this time the
418 * page is kept locked, so there's no chance of another
419 * race with truncate/invalidate.
421 if (!page->mapping) {
423 page = find_or_create_page(mapping, index,
424 mapping_gfp_mask(mapping));
430 page_cache_release(spd.pages[page_nr]);
431 spd.pages[page_nr] = page;
434 * page was already under io and is now done, great
436 if (PageUptodate(page)) {
442 * need to read in the page
444 error = mapping->a_ops->readpage(in, page);
445 if (unlikely(error)) {
447 * We really should re-lookup the page here,
448 * but it complicates things a lot. Instead
449 * lets just do what we already stored, and
450 * we'll get it the next time we are called.
452 if (error == AOP_TRUNCATED_PAGE)
460 * i_size must be checked after PageUptodate.
462 isize = i_size_read(mapping->host);
463 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
464 if (unlikely(!isize || index > end_index))
468 * if this is the last page, see if we need to shrink
469 * the length and stop
471 if (end_index == index) {
475 * max good bytes in this page
477 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
482 * force quit after adding this page
484 this_len = min(this_len, plen - loff);
488 spd.partial[page_nr].offset = loff;
489 spd.partial[page_nr].len = this_len;
497 * Release any pages at the end, if we quit early. 'page_nr' is how far
498 * we got, 'nr_pages' is how many pages are in the map.
500 while (page_nr < nr_pages)
501 page_cache_release(spd.pages[page_nr++]);
502 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
505 error = splice_to_pipe(pipe, &spd);
507 splice_shrink_spd(&spd);
512 * generic_file_splice_read - splice data from file to a pipe
513 * @in: file to splice from
514 * @ppos: position in @in
515 * @pipe: pipe to splice to
516 * @len: number of bytes to splice
517 * @flags: splice modifier flags
520 * Will read pages from given file and fill them into a pipe. Can be
521 * used as long as the address_space operations for the source implements
525 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
526 struct pipe_inode_info *pipe, size_t len,
532 isize = i_size_read(in->f_mapping->host);
533 if (unlikely(*ppos >= isize))
536 left = isize - *ppos;
537 if (unlikely(left < len))
540 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
548 EXPORT_SYMBOL(generic_file_splice_read);
550 static const struct pipe_buf_operations default_pipe_buf_ops = {
552 .map = generic_pipe_buf_map,
553 .unmap = generic_pipe_buf_unmap,
554 .confirm = generic_pipe_buf_confirm,
555 .release = generic_pipe_buf_release,
556 .steal = generic_pipe_buf_steal,
557 .get = generic_pipe_buf_get,
560 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
561 struct pipe_buffer *buf)
566 /* Pipe buffer operations for a socket and similar. */
567 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
569 .map = generic_pipe_buf_map,
570 .unmap = generic_pipe_buf_unmap,
571 .confirm = generic_pipe_buf_confirm,
572 .release = generic_pipe_buf_release,
573 .steal = generic_pipe_buf_nosteal,
574 .get = generic_pipe_buf_get,
576 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
578 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
579 unsigned long vlen, loff_t offset)
587 /* The cast to a user pointer is valid due to the set_fs() */
588 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
594 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
602 /* The cast to a user pointer is valid due to the set_fs() */
603 res = vfs_write(file, (const char __user *)buf, count, &pos);
609 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
610 struct pipe_inode_info *pipe, size_t len,
613 unsigned int nr_pages;
614 unsigned int nr_freed;
616 struct page *pages[PIPE_DEF_BUFFERS];
617 struct partial_page partial[PIPE_DEF_BUFFERS];
618 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
623 struct splice_pipe_desc spd = {
626 .nr_pages_max = PIPE_DEF_BUFFERS,
628 .ops = &default_pipe_buf_ops,
629 .spd_release = spd_release_page,
632 if (splice_grow_spd(pipe, &spd))
637 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
638 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
643 offset = *ppos & ~PAGE_CACHE_MASK;
644 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
646 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
649 page = alloc_page(GFP_USER);
654 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
655 vec[i].iov_base = (void __user *) page_address(page);
656 vec[i].iov_len = this_len;
663 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
674 for (i = 0; i < spd.nr_pages; i++) {
675 this_len = min_t(size_t, vec[i].iov_len, res);
676 spd.partial[i].offset = 0;
677 spd.partial[i].len = this_len;
679 __free_page(spd.pages[i]);
685 spd.nr_pages -= nr_freed;
687 res = splice_to_pipe(pipe, &spd);
694 splice_shrink_spd(&spd);
698 for (i = 0; i < spd.nr_pages; i++)
699 __free_page(spd.pages[i]);
704 EXPORT_SYMBOL(default_file_splice_read);
707 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
708 * using sendpage(). Return the number of bytes sent.
710 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
711 struct pipe_buffer *buf, struct splice_desc *sd)
713 struct file *file = sd->u.file;
714 loff_t pos = sd->pos;
717 if (!likely(file->f_op && file->f_op->sendpage))
720 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
722 if (sd->len < sd->total_len && pipe->nrbufs > 1)
723 more |= MSG_SENDPAGE_NOTLAST;
725 return file->f_op->sendpage(file, buf->page, buf->offset,
726 sd->len, &pos, more);
730 * This is a little more tricky than the file -> pipe splicing. There are
731 * basically three cases:
733 * - Destination page already exists in the address space and there
734 * are users of it. For that case we have no other option that
735 * copying the data. Tough luck.
736 * - Destination page already exists in the address space, but there
737 * are no users of it. Make sure it's uptodate, then drop it. Fall
738 * through to last case.
739 * - Destination page does not exist, we can add the pipe page to
740 * the page cache and avoid the copy.
742 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
743 * sd->flags), we attempt to migrate pages from the pipe to the output
744 * file address space page cache. This is possible if no one else has
745 * the pipe page referenced outside of the pipe and page cache. If
746 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
747 * a new page in the output file page cache and fill/dirty that.
749 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
750 struct splice_desc *sd)
752 struct file *file = sd->u.file;
753 struct address_space *mapping = file->f_mapping;
754 unsigned int offset, this_len;
759 offset = sd->pos & ~PAGE_CACHE_MASK;
762 if (this_len + offset > PAGE_CACHE_SIZE)
763 this_len = PAGE_CACHE_SIZE - offset;
765 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
766 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
770 if (buf->page != page) {
772 * Careful, ->map() uses KM_USER0!
774 char *src = buf->ops->map(pipe, buf, 1);
775 char *dst = kmap_atomic(page, KM_USER1);
777 memcpy(dst + offset, src + buf->offset, this_len);
778 flush_dcache_page(page);
779 kunmap_atomic(dst, KM_USER1);
780 buf->ops->unmap(pipe, buf, src);
782 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
787 EXPORT_SYMBOL(pipe_to_file);
789 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
792 if (waitqueue_active(&pipe->wait))
793 wake_up_interruptible(&pipe->wait);
794 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
798 * splice_from_pipe_feed - feed available data from a pipe to a file
799 * @pipe: pipe to splice from
800 * @sd: information to @actor
801 * @actor: handler that splices the data
804 * This function loops over the pipe and calls @actor to do the
805 * actual moving of a single struct pipe_buffer to the desired
806 * destination. It returns when there's no more buffers left in
807 * the pipe or if the requested number of bytes (@sd->total_len)
808 * have been copied. It returns a positive number (one) if the
809 * pipe needs to be filled with more data, zero if the required
810 * number of bytes have been copied and -errno on error.
812 * This, together with splice_from_pipe_{begin,end,next}, may be
813 * used to implement the functionality of __splice_from_pipe() when
814 * locking is required around copying the pipe buffers to the
817 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
822 while (pipe->nrbufs) {
823 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
824 const struct pipe_buf_operations *ops = buf->ops;
827 if (sd->len > sd->total_len)
828 sd->len = sd->total_len;
830 ret = buf->ops->confirm(pipe, buf);
837 ret = actor(pipe, buf, sd);
844 sd->num_spliced += ret;
847 sd->total_len -= ret;
851 ops->release(pipe, buf);
852 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
855 sd->need_wakeup = true;
864 EXPORT_SYMBOL(splice_from_pipe_feed);
867 * splice_from_pipe_next - wait for some data to splice from
868 * @pipe: pipe to splice from
869 * @sd: information about the splice operation
872 * This function will wait for some data and return a positive
873 * value (one) if pipe buffers are available. It will return zero
874 * or -errno if no more data needs to be spliced.
876 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
879 * Check for signal early to make process killable when there are
880 * always buffers available
882 if (signal_pending(current))
885 while (!pipe->nrbufs) {
889 if (!pipe->waiting_writers && sd->num_spliced)
892 if (sd->flags & SPLICE_F_NONBLOCK)
895 if (signal_pending(current))
898 if (sd->need_wakeup) {
899 wakeup_pipe_writers(pipe);
900 sd->need_wakeup = false;
908 EXPORT_SYMBOL(splice_from_pipe_next);
911 * splice_from_pipe_begin - start splicing from pipe
912 * @sd: information about the splice operation
915 * This function should be called before a loop containing
916 * splice_from_pipe_next() and splice_from_pipe_feed() to
917 * initialize the necessary fields of @sd.
919 void splice_from_pipe_begin(struct splice_desc *sd)
922 sd->need_wakeup = false;
924 EXPORT_SYMBOL(splice_from_pipe_begin);
927 * splice_from_pipe_end - finish splicing from pipe
928 * @pipe: pipe to splice from
929 * @sd: information about the splice operation
932 * This function will wake up pipe writers if necessary. It should
933 * be called after a loop containing splice_from_pipe_next() and
934 * splice_from_pipe_feed().
936 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
939 wakeup_pipe_writers(pipe);
941 EXPORT_SYMBOL(splice_from_pipe_end);
944 * __splice_from_pipe - splice data from a pipe to given actor
945 * @pipe: pipe to splice from
946 * @sd: information to @actor
947 * @actor: handler that splices the data
950 * This function does little more than loop over the pipe and call
951 * @actor to do the actual moving of a single struct pipe_buffer to
952 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
956 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
961 splice_from_pipe_begin(sd);
964 ret = splice_from_pipe_next(pipe, sd);
966 ret = splice_from_pipe_feed(pipe, sd, actor);
968 splice_from_pipe_end(pipe, sd);
970 return sd->num_spliced ? sd->num_spliced : ret;
972 EXPORT_SYMBOL(__splice_from_pipe);
975 * splice_from_pipe - splice data from a pipe to a file
976 * @pipe: pipe to splice from
977 * @out: file to splice to
978 * @ppos: position in @out
979 * @len: how many bytes to splice
980 * @flags: splice modifier flags
981 * @actor: handler that splices the data
984 * See __splice_from_pipe. This function locks the pipe inode,
985 * otherwise it's identical to __splice_from_pipe().
988 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
989 loff_t *ppos, size_t len, unsigned int flags,
993 struct splice_desc sd = {
1001 ret = __splice_from_pipe(pipe, &sd, actor);
1008 * generic_file_splice_write - splice data from a pipe to a file
1010 * @out: file to write to
1011 * @ppos: position in @out
1012 * @len: number of bytes to splice
1013 * @flags: splice modifier flags
1016 * Will either move or copy pages (determined by @flags options) from
1017 * the given pipe inode to the given file.
1021 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1022 loff_t *ppos, size_t len, unsigned int flags)
1024 struct address_space *mapping = out->f_mapping;
1025 struct inode *inode = mapping->host;
1026 struct splice_desc sd = {
1032 ret = generic_write_checks(out, ppos, &len, S_ISBLK(inode->i_mode));
1040 splice_from_pipe_begin(&sd);
1042 ret = splice_from_pipe_next(pipe, &sd);
1046 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1047 ret = file_remove_suid(out);
1049 file_update_time(out);
1050 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1052 mutex_unlock(&inode->i_mutex);
1054 splice_from_pipe_end(pipe, &sd);
1059 ret = sd.num_spliced;
1062 unsigned long nr_pages;
1065 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1067 err = generic_write_sync(out, *ppos, ret);
1072 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1078 EXPORT_SYMBOL(generic_file_splice_write);
1080 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1081 struct splice_desc *sd)
1086 data = buf->ops->map(pipe, buf, 0);
1087 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1088 buf->ops->unmap(pipe, buf, data);
1093 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1094 struct file *out, loff_t *ppos,
1095 size_t len, unsigned int flags)
1099 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1107 * generic_splice_sendpage - splice data from a pipe to a socket
1108 * @pipe: pipe to splice from
1109 * @out: socket to write to
1110 * @ppos: position in @out
1111 * @len: number of bytes to splice
1112 * @flags: splice modifier flags
1115 * Will send @len bytes from the pipe to a network socket. No data copying
1119 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1120 loff_t *ppos, size_t len, unsigned int flags)
1122 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1125 EXPORT_SYMBOL(generic_splice_sendpage);
1128 * Attempt to initiate a splice from pipe to file.
1130 long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1131 loff_t *ppos, size_t len, unsigned int flags)
1133 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1134 loff_t *, size_t, unsigned int);
1137 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1140 if (unlikely(out->f_flags & O_APPEND))
1143 ret = rw_verify_area(WRITE, out, ppos, len);
1144 if (unlikely(ret < 0))
1147 if (out->f_op && out->f_op->splice_write)
1148 splice_write = out->f_op->splice_write;
1150 splice_write = default_file_splice_write;
1152 return splice_write(pipe, out, ppos, len, flags);
1154 EXPORT_SYMBOL(do_splice_from);
1157 * Attempt to initiate a splice from a file to a pipe.
1159 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);
1181 EXPORT_SYMBOL(do_splice_to);
1184 * splice_direct_to_actor - splices data directly between two non-pipes
1185 * @in: file to splice from
1186 * @sd: actor information on where to splice to
1187 * @actor: handles the data splicing
1190 * This is a special case helper to splice directly between two
1191 * points, without requiring an explicit pipe. Internally an allocated
1192 * pipe is cached in the process, and reused during the lifetime of
1196 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1197 splice_direct_actor *actor)
1199 struct pipe_inode_info *pipe;
1206 * We require the input being a regular file, as we don't want to
1207 * randomly drop data for eg socket -> socket splicing. Use the
1208 * piped splicing for that!
1210 i_mode = in->f_path.dentry->d_inode->i_mode;
1211 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1215 * neither in nor out is a pipe, setup an internal pipe attached to
1216 * 'out' and transfer the wanted data from 'in' to 'out' through that
1218 pipe = current->splice_pipe;
1219 if (unlikely(!pipe)) {
1220 pipe = alloc_pipe_info(NULL);
1225 * We don't have an immediate reader, but we'll read the stuff
1226 * out of the pipe right after the splice_to_pipe(). So set
1227 * PIPE_READERS appropriately.
1231 current->splice_pipe = pipe;
1239 len = sd->total_len;
1243 * Don't block on output, we have to drain the direct pipe.
1245 sd->flags &= ~SPLICE_F_NONBLOCK;
1246 more = sd->flags & SPLICE_F_MORE;
1250 loff_t pos = sd->pos, prev_pos = pos;
1252 ret = do_splice_to(in, &pos, pipe, len, flags);
1253 if (unlikely(ret <= 0))
1257 sd->total_len = read_len;
1260 * If more data is pending, set SPLICE_F_MORE
1261 * If this is the last data and SPLICE_F_MORE was not set
1262 * initially, clears it.
1265 sd->flags |= SPLICE_F_MORE;
1267 sd->flags &= ~SPLICE_F_MORE;
1269 * NOTE: nonblocking mode only applies to the input. We
1270 * must not do the output in nonblocking mode as then we
1271 * could get stuck data in the internal pipe:
1273 ret = actor(pipe, sd);
1274 if (unlikely(ret <= 0)) {
1283 if (ret < read_len) {
1284 sd->pos = prev_pos + ret;
1290 pipe->nrbufs = pipe->curbuf = 0;
1296 * If we did an incomplete transfer we must release
1297 * the pipe buffers in question:
1299 for (i = 0; i < pipe->buffers; i++) {
1300 struct pipe_buffer *buf = pipe->bufs + i;
1303 buf->ops->release(pipe, buf);
1313 EXPORT_SYMBOL(splice_direct_to_actor);
1315 static int direct_splice_actor(struct pipe_inode_info *pipe,
1316 struct splice_desc *sd)
1318 struct file *file = sd->u.file;
1320 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1325 * do_splice_direct - splices data directly between two files
1326 * @in: file to splice from
1327 * @ppos: input file offset
1328 * @out: file to splice to
1329 * @len: number of bytes to splice
1330 * @flags: splice modifier flags
1333 * For use by do_sendfile(). splice can easily emulate sendfile, but
1334 * doing it in the application would incur an extra system call
1335 * (splice in + splice out, as compared to just sendfile()). So this helper
1336 * can splice directly through a process-private pipe.
1339 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1340 size_t len, unsigned int flags)
1342 struct splice_desc sd = {
1351 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1358 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1359 struct pipe_inode_info *opipe,
1360 size_t len, unsigned int flags);
1363 * Determine where to splice to/from.
1365 static long do_splice(struct file *in, loff_t __user *off_in,
1366 struct file *out, loff_t __user *off_out,
1367 size_t len, unsigned int flags)
1369 struct pipe_inode_info *ipipe;
1370 struct pipe_inode_info *opipe;
1371 loff_t offset, *off;
1374 ipipe = get_pipe_info(in);
1375 opipe = get_pipe_info(out);
1377 if (ipipe && opipe) {
1378 if (off_in || off_out)
1381 if (!(in->f_mode & FMODE_READ))
1384 if (!(out->f_mode & FMODE_WRITE))
1387 /* Splicing to self would be fun, but... */
1391 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1398 if (!(out->f_mode & FMODE_PWRITE))
1400 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1406 ret = do_splice_from(ipipe, out, off, len, flags);
1408 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1418 if (!(in->f_mode & FMODE_PREAD))
1420 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1426 ret = do_splice_to(in, off, opipe, len, flags);
1428 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1438 * Map an iov into an array of pages and offset/length tupples. With the
1439 * partial_page structure, we can map several non-contiguous ranges into
1440 * our ones pages[] map instead of splitting that operation into pieces.
1441 * Could easily be exported as a generic helper for other users, in which
1442 * case one would probably want to add a 'max_nr_pages' parameter as well.
1444 static int get_iovec_page_array(const struct iovec __user *iov,
1445 unsigned int nr_vecs, struct page **pages,
1446 struct partial_page *partial, int aligned,
1447 unsigned int pipe_buffers)
1449 int buffers = 0, error = 0;
1452 unsigned long off, npages;
1459 if (copy_from_user(&entry, iov, sizeof(entry)))
1462 base = entry.iov_base;
1463 len = entry.iov_len;
1466 * Sanity check this iovec. 0 read succeeds.
1472 if (!access_ok(VERIFY_READ, base, len))
1476 * Get this base offset and number of pages, then map
1477 * in the user pages.
1479 off = (unsigned long) base & ~PAGE_MASK;
1482 * If asked for alignment, the offset must be zero and the
1483 * length a multiple of the PAGE_SIZE.
1486 if (aligned && (off || len & ~PAGE_MASK))
1489 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1490 if (npages > pipe_buffers - buffers)
1491 npages = pipe_buffers - buffers;
1493 error = get_user_pages_fast((unsigned long)base, npages,
1494 0, &pages[buffers]);
1496 if (unlikely(error <= 0))
1500 * Fill this contiguous range into the partial page map.
1502 for (i = 0; i < error; i++) {
1503 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1505 partial[buffers].offset = off;
1506 partial[buffers].len = plen;
1514 * We didn't complete this iov, stop here since it probably
1515 * means we have to move some of this into a pipe to
1516 * be able to continue.
1522 * Don't continue if we mapped fewer pages than we asked for,
1523 * or if we mapped the max number of pages that we have
1526 if (error < npages || buffers == pipe_buffers)
1539 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1540 struct splice_desc *sd)
1546 * See if we can use the atomic maps, by prefaulting in the
1547 * pages and doing an atomic copy
1549 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1550 src = buf->ops->map(pipe, buf, 1);
1551 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1553 buf->ops->unmap(pipe, buf, src);
1561 * No dice, use slow non-atomic map and copy
1563 src = buf->ops->map(pipe, buf, 0);
1566 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1569 buf->ops->unmap(pipe, buf, src);
1572 sd->u.userptr += ret;
1577 * For lack of a better implementation, implement vmsplice() to userspace
1578 * as a simple copy of the pipes pages to the user iov.
1580 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1581 unsigned long nr_segs, unsigned int flags)
1583 struct pipe_inode_info *pipe;
1584 struct splice_desc sd;
1589 pipe = get_pipe_info(file);
1601 * Get user address base and length for this iovec.
1603 error = get_user(base, &iov->iov_base);
1604 if (unlikely(error))
1606 error = get_user(len, &iov->iov_len);
1607 if (unlikely(error))
1611 * Sanity check this iovec. 0 read succeeds.
1615 if (unlikely(!base)) {
1620 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1628 sd.u.userptr = base;
1631 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1657 * vmsplice splices a user address range into a pipe. It can be thought of
1658 * as splice-from-memory, where the regular splice is splice-from-file (or
1659 * to file). In both cases the output is a pipe, naturally.
1661 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1662 unsigned long nr_segs, unsigned int flags)
1664 struct pipe_inode_info *pipe;
1665 struct page *pages[PIPE_DEF_BUFFERS];
1666 struct partial_page partial[PIPE_DEF_BUFFERS];
1667 struct splice_pipe_desc spd = {
1670 .nr_pages_max = PIPE_DEF_BUFFERS,
1672 .ops = &user_page_pipe_buf_ops,
1673 .spd_release = spd_release_page,
1677 pipe = get_pipe_info(file);
1681 if (splice_grow_spd(pipe, &spd))
1684 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1685 spd.partial, flags & SPLICE_F_GIFT,
1687 if (spd.nr_pages <= 0)
1690 ret = splice_to_pipe(pipe, &spd);
1692 splice_shrink_spd(&spd);
1697 * Note that vmsplice only really supports true splicing _from_ user memory
1698 * to a pipe, not the other way around. Splicing from user memory is a simple
1699 * operation that can be supported without any funky alignment restrictions
1700 * or nasty vm tricks. We simply map in the user memory and fill them into
1701 * a pipe. The reverse isn't quite as easy, though. There are two possible
1702 * solutions for that:
1704 * - memcpy() the data internally, at which point we might as well just
1705 * do a regular read() on the buffer anyway.
1706 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1707 * has restriction limitations on both ends of the pipe).
1709 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1712 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1713 unsigned long, nr_segs, unsigned int, flags)
1719 if (unlikely(nr_segs > UIO_MAXIOV))
1721 else if (unlikely(!nr_segs))
1725 file = fget_light(fd, &fput);
1727 if (file->f_mode & FMODE_WRITE)
1728 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1729 else if (file->f_mode & FMODE_READ)
1730 error = vmsplice_to_user(file, iov, nr_segs, flags);
1732 fput_light(file, fput);
1738 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1739 int, fd_out, loff_t __user *, off_out,
1740 size_t, len, unsigned int, flags)
1743 struct file *in, *out;
1744 int fput_in, fput_out;
1750 in = fget_light(fd_in, &fput_in);
1752 if (in->f_mode & FMODE_READ) {
1753 out = fget_light(fd_out, &fput_out);
1755 if (out->f_mode & FMODE_WRITE)
1756 error = do_splice(in, off_in,
1759 fput_light(out, fput_out);
1763 fput_light(in, fput_in);
1770 * Make sure there's data to read. Wait for input if we can, otherwise
1771 * return an appropriate error.
1773 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1778 * Check ->nrbufs without the inode lock first. This function
1779 * is speculative anyways, so missing one is ok.
1787 while (!pipe->nrbufs) {
1788 if (signal_pending(current)) {
1794 if (!pipe->waiting_writers) {
1795 if (flags & SPLICE_F_NONBLOCK) {
1808 * Make sure there's writeable room. Wait for room if we can, otherwise
1809 * return an appropriate error.
1811 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1816 * Check ->nrbufs without the inode lock first. This function
1817 * is speculative anyways, so missing one is ok.
1819 if (pipe->nrbufs < pipe->buffers)
1825 while (pipe->nrbufs >= pipe->buffers) {
1826 if (!pipe->readers) {
1827 send_sig(SIGPIPE, current, 0);
1831 if (flags & SPLICE_F_NONBLOCK) {
1835 if (signal_pending(current)) {
1839 pipe->waiting_writers++;
1841 pipe->waiting_writers--;
1849 * Splice contents of ipipe to opipe.
1851 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1852 struct pipe_inode_info *opipe,
1853 size_t len, unsigned int flags)
1855 struct pipe_buffer *ibuf, *obuf;
1857 bool input_wakeup = false;
1861 ret = ipipe_prep(ipipe, flags);
1865 ret = opipe_prep(opipe, flags);
1870 * Potential ABBA deadlock, work around it by ordering lock
1871 * grabbing by pipe info address. Otherwise two different processes
1872 * could deadlock (one doing tee from A -> B, the other from B -> A).
1874 pipe_double_lock(ipipe, opipe);
1877 if (!opipe->readers) {
1878 send_sig(SIGPIPE, current, 0);
1884 if (!ipipe->nrbufs && !ipipe->writers)
1888 * Cannot make any progress, because either the input
1889 * pipe is empty or the output pipe is full.
1891 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1892 /* Already processed some buffers, break */
1896 if (flags & SPLICE_F_NONBLOCK) {
1902 * We raced with another reader/writer and haven't
1903 * managed to process any buffers. A zero return
1904 * value means EOF, so retry instead.
1911 ibuf = ipipe->bufs + ipipe->curbuf;
1912 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1913 obuf = opipe->bufs + nbuf;
1915 if (len >= ibuf->len) {
1917 * Simply move the whole buffer from ipipe to opipe
1922 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1924 input_wakeup = true;
1927 * Get a reference to this pipe buffer,
1928 * so we can copy the contents over.
1930 ibuf->ops->get(ipipe, ibuf);
1934 * Don't inherit the gift flag, we need to
1935 * prevent multiple steals of this page.
1937 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1941 ibuf->offset += obuf->len;
1942 ibuf->len -= obuf->len;
1952 * If we put data in the output pipe, wakeup any potential readers.
1955 wakeup_pipe_readers(opipe);
1958 wakeup_pipe_writers(ipipe);
1964 * Link contents of ipipe to opipe.
1966 static int link_pipe(struct pipe_inode_info *ipipe,
1967 struct pipe_inode_info *opipe,
1968 size_t len, unsigned int flags)
1970 struct pipe_buffer *ibuf, *obuf;
1971 int ret = 0, i = 0, nbuf;
1974 * Potential ABBA deadlock, work around it by ordering lock
1975 * grabbing by pipe info address. Otherwise two different processes
1976 * could deadlock (one doing tee from A -> B, the other from B -> A).
1978 pipe_double_lock(ipipe, opipe);
1981 if (!opipe->readers) {
1982 send_sig(SIGPIPE, current, 0);
1989 * If we have iterated all input buffers or ran out of
1990 * output room, break.
1992 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1995 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1996 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1999 * Get a reference to this pipe buffer,
2000 * so we can copy the contents over.
2002 ibuf->ops->get(ipipe, ibuf);
2004 obuf = opipe->bufs + nbuf;
2008 * Don't inherit the gift flag, we need to
2009 * prevent multiple steals of this page.
2011 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2013 if (obuf->len > len)
2023 * return EAGAIN if we have the potential of some data in the
2024 * future, otherwise just return 0
2026 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2033 * If we put data in the output pipe, wakeup any potential readers.
2036 wakeup_pipe_readers(opipe);
2042 * This is a tee(1) implementation that works on pipes. It doesn't copy
2043 * any data, it simply references the 'in' pages on the 'out' pipe.
2044 * The 'flags' used are the SPLICE_F_* variants, currently the only
2045 * applicable one is SPLICE_F_NONBLOCK.
2047 static long do_tee(struct file *in, struct file *out, size_t len,
2050 struct pipe_inode_info *ipipe = get_pipe_info(in);
2051 struct pipe_inode_info *opipe = get_pipe_info(out);
2055 * Duplicate the contents of ipipe to opipe without actually
2058 if (ipipe && opipe && ipipe != opipe) {
2060 * Keep going, unless we encounter an error. The ipipe/opipe
2061 * ordering doesn't really matter.
2063 ret = ipipe_prep(ipipe, flags);
2065 ret = opipe_prep(opipe, flags);
2067 ret = link_pipe(ipipe, opipe, len, flags);
2074 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2083 in = fget_light(fdin, &fput_in);
2085 if (in->f_mode & FMODE_READ) {
2087 struct file *out = fget_light(fdout, &fput_out);
2090 if (out->f_mode & FMODE_WRITE)
2091 error = do_tee(in, out, len, flags);
2092 fput_light(out, fput_out);
2095 fput_light(in, fput_in);