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 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
138 .map = generic_pipe_buf_map,
139 .unmap = generic_pipe_buf_unmap,
140 .confirm = page_cache_pipe_buf_confirm,
141 .release = page_cache_pipe_buf_release,
142 .steal = page_cache_pipe_buf_steal,
143 .get = generic_pipe_buf_get,
146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
147 struct pipe_buffer *buf)
149 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
152 buf->flags |= PIPE_BUF_FLAG_LRU;
153 return generic_pipe_buf_steal(pipe, buf);
156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158 .map = generic_pipe_buf_map,
159 .unmap = generic_pipe_buf_unmap,
160 .confirm = generic_pipe_buf_confirm,
161 .release = page_cache_pipe_buf_release,
162 .steal = user_page_pipe_buf_steal,
163 .get = generic_pipe_buf_get,
166 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
169 if (waitqueue_active(&pipe->wait))
170 wake_up_interruptible(&pipe->wait);
171 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
175 * splice_to_pipe - fill passed data into a pipe
176 * @pipe: pipe to fill
180 * @spd contains a map of pages and len/offset tuples, along with
181 * the struct pipe_buf_operations associated with these pages. This
182 * function will link that data to the pipe.
185 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
186 struct splice_pipe_desc *spd)
188 unsigned int spd_pages = spd->nr_pages;
189 int ret, do_wakeup, page_nr;
198 if (!pipe->readers) {
199 send_sig(SIGPIPE, current, 0);
205 if (pipe->nrbufs < pipe->buffers) {
206 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
207 struct pipe_buffer *buf = pipe->bufs + newbuf;
209 buf->page = spd->pages[page_nr];
210 buf->offset = spd->partial[page_nr].offset;
211 buf->len = spd->partial[page_nr].len;
212 buf->private = spd->partial[page_nr].private;
214 if (spd->flags & SPLICE_F_GIFT)
215 buf->flags |= PIPE_BUF_FLAG_GIFT;
224 if (!--spd->nr_pages)
226 if (pipe->nrbufs < pipe->buffers)
232 if (spd->flags & SPLICE_F_NONBLOCK) {
238 if (signal_pending(current)) {
246 if (waitqueue_active(&pipe->wait))
247 wake_up_interruptible_sync(&pipe->wait);
248 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
252 pipe->waiting_writers++;
254 pipe->waiting_writers--;
260 wakeup_pipe_readers(pipe);
262 while (page_nr < spd_pages)
263 spd->spd_release(spd, page_nr++);
268 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
270 page_cache_release(spd->pages[i]);
274 * Check if we need to grow the arrays holding pages and partial page
277 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
279 if (pipe->buffers <= PIPE_DEF_BUFFERS)
282 spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
283 spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
285 if (spd->pages && spd->partial)
293 void splice_shrink_spd(struct pipe_inode_info *pipe,
294 struct splice_pipe_desc *spd)
296 if (pipe->buffers <= PIPE_DEF_BUFFERS)
304 __generic_file_splice_read(struct file *in, loff_t *ppos,
305 struct pipe_inode_info *pipe, size_t len,
308 struct address_space *mapping = in->f_mapping;
309 unsigned int loff, nr_pages, req_pages;
310 struct page *pages[PIPE_DEF_BUFFERS];
311 struct partial_page partial[PIPE_DEF_BUFFERS];
313 pgoff_t index, end_index;
316 struct splice_pipe_desc spd = {
320 .ops = &page_cache_pipe_buf_ops,
321 .spd_release = spd_release_page,
324 if (splice_grow_spd(pipe, &spd))
327 index = *ppos >> PAGE_CACHE_SHIFT;
328 loff = *ppos & ~PAGE_CACHE_MASK;
329 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
330 nr_pages = min(req_pages, pipe->buffers);
333 * Lookup the (hopefully) full range of pages we need.
335 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
336 index += spd.nr_pages;
339 * If find_get_pages_contig() returned fewer pages than we needed,
340 * readahead/allocate the rest and fill in the holes.
342 if (spd.nr_pages < nr_pages)
343 page_cache_sync_readahead(mapping, &in->f_ra, in,
344 index, req_pages - spd.nr_pages);
347 while (spd.nr_pages < nr_pages) {
349 * Page could be there, find_get_pages_contig() breaks on
352 page = find_get_page(mapping, index);
355 * page didn't exist, allocate one.
357 page = page_cache_alloc_cold(mapping);
361 error = add_to_page_cache_lru(page, mapping, index,
363 if (unlikely(error)) {
364 page_cache_release(page);
365 if (error == -EEXIST)
370 * add_to_page_cache() locks the page, unlock it
371 * to avoid convoluting the logic below even more.
376 spd.pages[spd.nr_pages++] = page;
381 * Now loop over the map and see if we need to start IO on any
382 * pages, fill in the partial map, etc.
384 index = *ppos >> PAGE_CACHE_SHIFT;
385 nr_pages = spd.nr_pages;
387 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
388 unsigned int this_len;
394 * this_len is the max we'll use from this page
396 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
397 page = spd.pages[page_nr];
399 if (PageReadahead(page))
400 page_cache_async_readahead(mapping, &in->f_ra, in,
401 page, index, req_pages - page_nr);
404 * If the page isn't uptodate, we may need to start io on it
406 if (!PageUptodate(page)) {
410 * Page was truncated, or invalidated by the
411 * filesystem. Redo the find/create, but this time the
412 * page is kept locked, so there's no chance of another
413 * race with truncate/invalidate.
415 if (!page->mapping) {
417 page = find_or_create_page(mapping, index,
418 mapping_gfp_mask(mapping));
424 page_cache_release(spd.pages[page_nr]);
425 spd.pages[page_nr] = page;
428 * page was already under io and is now done, great
430 if (PageUptodate(page)) {
436 * need to read in the page
438 error = mapping->a_ops->readpage(in, page);
439 if (unlikely(error)) {
441 * We really should re-lookup the page here,
442 * but it complicates things a lot. Instead
443 * lets just do what we already stored, and
444 * we'll get it the next time we are called.
446 if (error == AOP_TRUNCATED_PAGE)
454 * i_size must be checked after PageUptodate.
456 isize = i_size_read(mapping->host);
457 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
458 if (unlikely(!isize || index > end_index))
462 * if this is the last page, see if we need to shrink
463 * the length and stop
465 if (end_index == index) {
469 * max good bytes in this page
471 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
476 * force quit after adding this page
478 this_len = min(this_len, plen - loff);
482 spd.partial[page_nr].offset = loff;
483 spd.partial[page_nr].len = this_len;
491 * Release any pages at the end, if we quit early. 'page_nr' is how far
492 * we got, 'nr_pages' is how many pages are in the map.
494 while (page_nr < nr_pages)
495 page_cache_release(spd.pages[page_nr++]);
496 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
499 error = splice_to_pipe(pipe, &spd);
501 splice_shrink_spd(pipe, &spd);
506 * generic_file_splice_read - splice data from file to a pipe
507 * @in: file to splice from
508 * @ppos: position in @in
509 * @pipe: pipe to splice to
510 * @len: number of bytes to splice
511 * @flags: splice modifier flags
514 * Will read pages from given file and fill them into a pipe. Can be
515 * used as long as the address_space operations for the source implements
519 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
520 struct pipe_inode_info *pipe, size_t len,
526 isize = i_size_read(in->f_mapping->host);
527 if (unlikely(*ppos >= isize))
530 left = isize - *ppos;
531 if (unlikely(left < len))
534 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
542 EXPORT_SYMBOL(generic_file_splice_read);
544 static const struct pipe_buf_operations default_pipe_buf_ops = {
546 .map = generic_pipe_buf_map,
547 .unmap = generic_pipe_buf_unmap,
548 .confirm = generic_pipe_buf_confirm,
549 .release = generic_pipe_buf_release,
550 .steal = generic_pipe_buf_steal,
551 .get = generic_pipe_buf_get,
554 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
555 unsigned long vlen, loff_t offset)
563 /* The cast to a user pointer is valid due to the set_fs() */
564 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
570 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
578 /* The cast to a user pointer is valid due to the set_fs() */
579 res = vfs_write(file, (const char __user *)buf, count, &pos);
585 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
586 struct pipe_inode_info *pipe, size_t len,
589 unsigned int nr_pages;
590 unsigned int nr_freed;
592 struct page *pages[PIPE_DEF_BUFFERS];
593 struct partial_page partial[PIPE_DEF_BUFFERS];
594 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
599 struct splice_pipe_desc spd = {
603 .ops = &default_pipe_buf_ops,
604 .spd_release = spd_release_page,
607 if (splice_grow_spd(pipe, &spd))
612 if (pipe->buffers > PIPE_DEF_BUFFERS) {
613 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
618 offset = *ppos & ~PAGE_CACHE_MASK;
619 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
621 for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
624 page = alloc_page(GFP_USER);
629 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
630 vec[i].iov_base = (void __user *) page_address(page);
631 vec[i].iov_len = this_len;
638 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
649 for (i = 0; i < spd.nr_pages; i++) {
650 this_len = min_t(size_t, vec[i].iov_len, res);
651 spd.partial[i].offset = 0;
652 spd.partial[i].len = this_len;
654 __free_page(spd.pages[i]);
660 spd.nr_pages -= nr_freed;
662 res = splice_to_pipe(pipe, &spd);
669 splice_shrink_spd(pipe, &spd);
673 for (i = 0; i < spd.nr_pages; i++)
674 __free_page(spd.pages[i]);
679 EXPORT_SYMBOL(default_file_splice_read);
682 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
683 * using sendpage(). Return the number of bytes sent.
685 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
686 struct pipe_buffer *buf, struct splice_desc *sd)
688 struct file *file = sd->u.file;
689 loff_t pos = sd->pos;
692 if (!likely(file->f_op && file->f_op->sendpage))
695 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
697 if (sd->len < sd->total_len && pipe->nrbufs > 1)
698 more |= MSG_SENDPAGE_NOTLAST;
700 return file->f_op->sendpage(file, buf->page, buf->offset,
701 sd->len, &pos, more);
705 * This is a little more tricky than the file -> pipe splicing. There are
706 * basically three cases:
708 * - Destination page already exists in the address space and there
709 * are users of it. For that case we have no other option that
710 * copying the data. Tough luck.
711 * - Destination page already exists in the address space, but there
712 * are no users of it. Make sure it's uptodate, then drop it. Fall
713 * through to last case.
714 * - Destination page does not exist, we can add the pipe page to
715 * the page cache and avoid the copy.
717 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
718 * sd->flags), we attempt to migrate pages from the pipe to the output
719 * file address space page cache. This is possible if no one else has
720 * the pipe page referenced outside of the pipe and page cache. If
721 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
722 * a new page in the output file page cache and fill/dirty that.
724 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
725 struct splice_desc *sd)
727 struct file *file = sd->u.file;
728 struct address_space *mapping = file->f_mapping;
729 unsigned int offset, this_len;
734 offset = sd->pos & ~PAGE_CACHE_MASK;
737 if (this_len + offset > PAGE_CACHE_SIZE)
738 this_len = PAGE_CACHE_SIZE - offset;
740 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
741 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
745 if (buf->page != page) {
747 * Careful, ->map() uses KM_USER0!
749 char *src = buf->ops->map(pipe, buf, 1);
750 char *dst = kmap_atomic(page, KM_USER1);
752 memcpy(dst + offset, src + buf->offset, this_len);
753 flush_dcache_page(page);
754 kunmap_atomic(dst, KM_USER1);
755 buf->ops->unmap(pipe, buf, src);
757 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
762 EXPORT_SYMBOL(pipe_to_file);
764 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
767 if (waitqueue_active(&pipe->wait))
768 wake_up_interruptible(&pipe->wait);
769 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
773 * splice_from_pipe_feed - feed available data from a pipe to a file
774 * @pipe: pipe to splice from
775 * @sd: information to @actor
776 * @actor: handler that splices the data
779 * This function loops over the pipe and calls @actor to do the
780 * actual moving of a single struct pipe_buffer to the desired
781 * destination. It returns when there's no more buffers left in
782 * the pipe or if the requested number of bytes (@sd->total_len)
783 * have been copied. It returns a positive number (one) if the
784 * pipe needs to be filled with more data, zero if the required
785 * number of bytes have been copied and -errno on error.
787 * This, together with splice_from_pipe_{begin,end,next}, may be
788 * used to implement the functionality of __splice_from_pipe() when
789 * locking is required around copying the pipe buffers to the
792 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
797 while (pipe->nrbufs) {
798 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
799 const struct pipe_buf_operations *ops = buf->ops;
802 if (sd->len > sd->total_len)
803 sd->len = sd->total_len;
805 ret = buf->ops->confirm(pipe, buf);
812 ret = actor(pipe, buf, sd);
819 sd->num_spliced += ret;
822 sd->total_len -= ret;
826 ops->release(pipe, buf);
827 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
830 sd->need_wakeup = true;
839 EXPORT_SYMBOL(splice_from_pipe_feed);
842 * splice_from_pipe_next - wait for some data to splice from
843 * @pipe: pipe to splice from
844 * @sd: information about the splice operation
847 * This function will wait for some data and return a positive
848 * value (one) if pipe buffers are available. It will return zero
849 * or -errno if no more data needs to be spliced.
851 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
853 while (!pipe->nrbufs) {
857 if (!pipe->waiting_writers && sd->num_spliced)
860 if (sd->flags & SPLICE_F_NONBLOCK)
863 if (signal_pending(current))
866 if (sd->need_wakeup) {
867 wakeup_pipe_writers(pipe);
868 sd->need_wakeup = false;
876 EXPORT_SYMBOL(splice_from_pipe_next);
879 * splice_from_pipe_begin - start splicing from pipe
880 * @sd: information about the splice operation
883 * This function should be called before a loop containing
884 * splice_from_pipe_next() and splice_from_pipe_feed() to
885 * initialize the necessary fields of @sd.
887 void splice_from_pipe_begin(struct splice_desc *sd)
890 sd->need_wakeup = false;
892 EXPORT_SYMBOL(splice_from_pipe_begin);
895 * splice_from_pipe_end - finish splicing from pipe
896 * @pipe: pipe to splice from
897 * @sd: information about the splice operation
900 * This function will wake up pipe writers if necessary. It should
901 * be called after a loop containing splice_from_pipe_next() and
902 * splice_from_pipe_feed().
904 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
907 wakeup_pipe_writers(pipe);
909 EXPORT_SYMBOL(splice_from_pipe_end);
912 * __splice_from_pipe - splice data from a pipe to given actor
913 * @pipe: pipe to splice from
914 * @sd: information to @actor
915 * @actor: handler that splices the data
918 * This function does little more than loop over the pipe and call
919 * @actor to do the actual moving of a single struct pipe_buffer to
920 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
924 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
929 splice_from_pipe_begin(sd);
931 ret = splice_from_pipe_next(pipe, sd);
933 ret = splice_from_pipe_feed(pipe, sd, actor);
935 splice_from_pipe_end(pipe, sd);
937 return sd->num_spliced ? sd->num_spliced : ret;
939 EXPORT_SYMBOL(__splice_from_pipe);
942 * splice_from_pipe - splice data from a pipe to a file
943 * @pipe: pipe to splice from
944 * @out: file to splice to
945 * @ppos: position in @out
946 * @len: how many bytes to splice
947 * @flags: splice modifier flags
948 * @actor: handler that splices the data
951 * See __splice_from_pipe. This function locks the pipe inode,
952 * otherwise it's identical to __splice_from_pipe().
955 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
956 loff_t *ppos, size_t len, unsigned int flags,
960 struct splice_desc sd = {
968 ret = __splice_from_pipe(pipe, &sd, actor);
975 * generic_file_splice_write - splice data from a pipe to a file
977 * @out: file to write to
978 * @ppos: position in @out
979 * @len: number of bytes to splice
980 * @flags: splice modifier flags
983 * Will either move or copy pages (determined by @flags options) from
984 * the given pipe inode to the given file.
988 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
989 loff_t *ppos, size_t len, unsigned int flags)
991 struct address_space *mapping = out->f_mapping;
992 struct inode *inode = mapping->host;
993 struct splice_desc sd = {
1003 splice_from_pipe_begin(&sd);
1005 ret = splice_from_pipe_next(pipe, &sd);
1009 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1010 ret = file_remove_suid(out);
1012 file_update_time(out);
1013 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1015 mutex_unlock(&inode->i_mutex);
1017 splice_from_pipe_end(pipe, &sd);
1022 ret = sd.num_spliced;
1025 unsigned long nr_pages;
1028 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1030 err = generic_write_sync(out, *ppos, ret);
1035 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1041 EXPORT_SYMBOL(generic_file_splice_write);
1043 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1044 struct splice_desc *sd)
1049 data = buf->ops->map(pipe, buf, 0);
1050 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1051 buf->ops->unmap(pipe, buf, data);
1056 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1057 struct file *out, loff_t *ppos,
1058 size_t len, unsigned int flags)
1062 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1070 * generic_splice_sendpage - splice data from a pipe to a socket
1071 * @pipe: pipe to splice from
1072 * @out: socket to write to
1073 * @ppos: position in @out
1074 * @len: number of bytes to splice
1075 * @flags: splice modifier flags
1078 * Will send @len bytes from the pipe to a network socket. No data copying
1082 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1083 loff_t *ppos, size_t len, unsigned int flags)
1085 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1088 EXPORT_SYMBOL(generic_splice_sendpage);
1091 * Attempt to initiate a splice from pipe to file.
1093 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1094 loff_t *ppos, size_t len, unsigned int flags)
1096 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1097 loff_t *, size_t, unsigned int);
1100 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1103 if (unlikely(out->f_flags & O_APPEND))
1106 ret = rw_verify_area(WRITE, out, ppos, len);
1107 if (unlikely(ret < 0))
1110 if (out->f_op && out->f_op->splice_write)
1111 splice_write = out->f_op->splice_write;
1113 splice_write = default_file_splice_write;
1115 return splice_write(pipe, out, ppos, len, flags);
1119 * Attempt to initiate a splice from a file to a pipe.
1121 static long do_splice_to(struct file *in, loff_t *ppos,
1122 struct pipe_inode_info *pipe, size_t len,
1125 ssize_t (*splice_read)(struct file *, loff_t *,
1126 struct pipe_inode_info *, size_t, unsigned int);
1129 if (unlikely(!(in->f_mode & FMODE_READ)))
1132 ret = rw_verify_area(READ, in, ppos, len);
1133 if (unlikely(ret < 0))
1136 if (in->f_op && in->f_op->splice_read)
1137 splice_read = in->f_op->splice_read;
1139 splice_read = default_file_splice_read;
1141 return splice_read(in, ppos, pipe, len, flags);
1145 * splice_direct_to_actor - splices data directly between two non-pipes
1146 * @in: file to splice from
1147 * @sd: actor information on where to splice to
1148 * @actor: handles the data splicing
1151 * This is a special case helper to splice directly between two
1152 * points, without requiring an explicit pipe. Internally an allocated
1153 * pipe is cached in the process, and reused during the lifetime of
1157 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1158 splice_direct_actor *actor)
1160 struct pipe_inode_info *pipe;
1167 * We require the input being a regular file, as we don't want to
1168 * randomly drop data for eg socket -> socket splicing. Use the
1169 * piped splicing for that!
1171 i_mode = in->f_path.dentry->d_inode->i_mode;
1172 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1176 * neither in nor out is a pipe, setup an internal pipe attached to
1177 * 'out' and transfer the wanted data from 'in' to 'out' through that
1179 pipe = current->splice_pipe;
1180 if (unlikely(!pipe)) {
1181 pipe = alloc_pipe_info(NULL);
1186 * We don't have an immediate reader, but we'll read the stuff
1187 * out of the pipe right after the splice_to_pipe(). So set
1188 * PIPE_READERS appropriately.
1192 current->splice_pipe = pipe;
1200 len = sd->total_len;
1204 * Don't block on output, we have to drain the direct pipe.
1206 sd->flags &= ~SPLICE_F_NONBLOCK;
1210 loff_t pos = sd->pos, prev_pos = pos;
1212 ret = do_splice_to(in, &pos, pipe, len, flags);
1213 if (unlikely(ret <= 0))
1217 sd->total_len = read_len;
1220 * NOTE: nonblocking mode only applies to the input. We
1221 * must not do the output in nonblocking mode as then we
1222 * could get stuck data in the internal pipe:
1224 ret = actor(pipe, sd);
1225 if (unlikely(ret <= 0)) {
1234 if (ret < read_len) {
1235 sd->pos = prev_pos + ret;
1241 pipe->nrbufs = pipe->curbuf = 0;
1247 * If we did an incomplete transfer we must release
1248 * the pipe buffers in question:
1250 for (i = 0; i < pipe->buffers; i++) {
1251 struct pipe_buffer *buf = pipe->bufs + i;
1254 buf->ops->release(pipe, buf);
1264 EXPORT_SYMBOL(splice_direct_to_actor);
1266 static int direct_splice_actor(struct pipe_inode_info *pipe,
1267 struct splice_desc *sd)
1269 struct file *file = sd->u.file;
1271 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1276 * do_splice_direct - splices data directly between two files
1277 * @in: file to splice from
1278 * @ppos: input file offset
1279 * @out: file to splice to
1280 * @len: number of bytes to splice
1281 * @flags: splice modifier flags
1284 * For use by do_sendfile(). splice can easily emulate sendfile, but
1285 * doing it in the application would incur an extra system call
1286 * (splice in + splice out, as compared to just sendfile()). So this helper
1287 * can splice directly through a process-private pipe.
1290 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1291 size_t len, unsigned int flags)
1293 struct splice_desc sd = {
1302 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1309 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1310 struct pipe_inode_info *opipe,
1311 size_t len, unsigned int flags);
1314 * Determine where to splice to/from.
1316 static long do_splice(struct file *in, loff_t __user *off_in,
1317 struct file *out, loff_t __user *off_out,
1318 size_t len, unsigned int flags)
1320 struct pipe_inode_info *ipipe;
1321 struct pipe_inode_info *opipe;
1322 loff_t offset, *off;
1325 ipipe = get_pipe_info(in);
1326 opipe = get_pipe_info(out);
1328 if (ipipe && opipe) {
1329 if (off_in || off_out)
1332 if (!(in->f_mode & FMODE_READ))
1335 if (!(out->f_mode & FMODE_WRITE))
1338 /* Splicing to self would be fun, but... */
1342 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1349 if (!(out->f_mode & FMODE_PWRITE))
1351 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1357 ret = do_splice_from(ipipe, out, off, len, flags);
1359 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1369 if (!(in->f_mode & FMODE_PREAD))
1371 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1377 ret = do_splice_to(in, off, opipe, len, flags);
1379 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1389 * Map an iov into an array of pages and offset/length tupples. With the
1390 * partial_page structure, we can map several non-contiguous ranges into
1391 * our ones pages[] map instead of splitting that operation into pieces.
1392 * Could easily be exported as a generic helper for other users, in which
1393 * case one would probably want to add a 'max_nr_pages' parameter as well.
1395 static int get_iovec_page_array(const struct iovec __user *iov,
1396 unsigned int nr_vecs, struct page **pages,
1397 struct partial_page *partial, int aligned,
1398 unsigned int pipe_buffers)
1400 int buffers = 0, error = 0;
1403 unsigned long off, npages;
1410 if (copy_from_user(&entry, iov, sizeof(entry)))
1413 base = entry.iov_base;
1414 len = entry.iov_len;
1417 * Sanity check this iovec. 0 read succeeds.
1423 if (!access_ok(VERIFY_READ, base, len))
1427 * Get this base offset and number of pages, then map
1428 * in the user pages.
1430 off = (unsigned long) base & ~PAGE_MASK;
1433 * If asked for alignment, the offset must be zero and the
1434 * length a multiple of the PAGE_SIZE.
1437 if (aligned && (off || len & ~PAGE_MASK))
1440 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1441 if (npages > pipe_buffers - buffers)
1442 npages = pipe_buffers - buffers;
1444 error = get_user_pages_fast((unsigned long)base, npages,
1445 0, &pages[buffers]);
1447 if (unlikely(error <= 0))
1451 * Fill this contiguous range into the partial page map.
1453 for (i = 0; i < error; i++) {
1454 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1456 partial[buffers].offset = off;
1457 partial[buffers].len = plen;
1465 * We didn't complete this iov, stop here since it probably
1466 * means we have to move some of this into a pipe to
1467 * be able to continue.
1473 * Don't continue if we mapped fewer pages than we asked for,
1474 * or if we mapped the max number of pages that we have
1477 if (error < npages || buffers == pipe_buffers)
1490 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1491 struct splice_desc *sd)
1497 * See if we can use the atomic maps, by prefaulting in the
1498 * pages and doing an atomic copy
1500 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1501 src = buf->ops->map(pipe, buf, 1);
1502 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1504 buf->ops->unmap(pipe, buf, src);
1512 * No dice, use slow non-atomic map and copy
1514 src = buf->ops->map(pipe, buf, 0);
1517 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1520 buf->ops->unmap(pipe, buf, src);
1523 sd->u.userptr += ret;
1528 * For lack of a better implementation, implement vmsplice() to userspace
1529 * as a simple copy of the pipes pages to the user iov.
1531 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1532 unsigned long nr_segs, unsigned int flags)
1534 struct pipe_inode_info *pipe;
1535 struct splice_desc sd;
1540 pipe = get_pipe_info(file);
1552 * Get user address base and length for this iovec.
1554 error = get_user(base, &iov->iov_base);
1555 if (unlikely(error))
1557 error = get_user(len, &iov->iov_len);
1558 if (unlikely(error))
1562 * Sanity check this iovec. 0 read succeeds.
1566 if (unlikely(!base)) {
1571 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1579 sd.u.userptr = base;
1582 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1608 * vmsplice splices a user address range into a pipe. It can be thought of
1609 * as splice-from-memory, where the regular splice is splice-from-file (or
1610 * to file). In both cases the output is a pipe, naturally.
1612 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1613 unsigned long nr_segs, unsigned int flags)
1615 struct pipe_inode_info *pipe;
1616 struct page *pages[PIPE_DEF_BUFFERS];
1617 struct partial_page partial[PIPE_DEF_BUFFERS];
1618 struct splice_pipe_desc spd = {
1622 .ops = &user_page_pipe_buf_ops,
1623 .spd_release = spd_release_page,
1627 pipe = get_pipe_info(file);
1631 if (splice_grow_spd(pipe, &spd))
1634 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1635 spd.partial, flags & SPLICE_F_GIFT,
1637 if (spd.nr_pages <= 0)
1640 ret = splice_to_pipe(pipe, &spd);
1642 splice_shrink_spd(pipe, &spd);
1647 * Note that vmsplice only really supports true splicing _from_ user memory
1648 * to a pipe, not the other way around. Splicing from user memory is a simple
1649 * operation that can be supported without any funky alignment restrictions
1650 * or nasty vm tricks. We simply map in the user memory and fill them into
1651 * a pipe. The reverse isn't quite as easy, though. There are two possible
1652 * solutions for that:
1654 * - memcpy() the data internally, at which point we might as well just
1655 * do a regular read() on the buffer anyway.
1656 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1657 * has restriction limitations on both ends of the pipe).
1659 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1662 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1663 unsigned long, nr_segs, unsigned int, flags)
1669 if (unlikely(nr_segs > UIO_MAXIOV))
1671 else if (unlikely(!nr_segs))
1675 file = fget_light(fd, &fput);
1677 if (file->f_mode & FMODE_WRITE)
1678 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1679 else if (file->f_mode & FMODE_READ)
1680 error = vmsplice_to_user(file, iov, nr_segs, flags);
1682 fput_light(file, fput);
1688 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1689 int, fd_out, loff_t __user *, off_out,
1690 size_t, len, unsigned int, flags)
1693 struct file *in, *out;
1694 int fput_in, fput_out;
1700 in = fget_light(fd_in, &fput_in);
1702 if (in->f_mode & FMODE_READ) {
1703 out = fget_light(fd_out, &fput_out);
1705 if (out->f_mode & FMODE_WRITE)
1706 error = do_splice(in, off_in,
1709 fput_light(out, fput_out);
1713 fput_light(in, fput_in);
1720 * Make sure there's data to read. Wait for input if we can, otherwise
1721 * return an appropriate error.
1723 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1728 * Check ->nrbufs without the inode lock first. This function
1729 * is speculative anyways, so missing one is ok.
1737 while (!pipe->nrbufs) {
1738 if (signal_pending(current)) {
1744 if (!pipe->waiting_writers) {
1745 if (flags & SPLICE_F_NONBLOCK) {
1758 * Make sure there's writeable room. Wait for room if we can, otherwise
1759 * return an appropriate error.
1761 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1766 * Check ->nrbufs without the inode lock first. This function
1767 * is speculative anyways, so missing one is ok.
1769 if (pipe->nrbufs < pipe->buffers)
1775 while (pipe->nrbufs >= pipe->buffers) {
1776 if (!pipe->readers) {
1777 send_sig(SIGPIPE, current, 0);
1781 if (flags & SPLICE_F_NONBLOCK) {
1785 if (signal_pending(current)) {
1789 pipe->waiting_writers++;
1791 pipe->waiting_writers--;
1799 * Splice contents of ipipe to opipe.
1801 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1802 struct pipe_inode_info *opipe,
1803 size_t len, unsigned int flags)
1805 struct pipe_buffer *ibuf, *obuf;
1807 bool input_wakeup = false;
1811 ret = ipipe_prep(ipipe, flags);
1815 ret = opipe_prep(opipe, flags);
1820 * Potential ABBA deadlock, work around it by ordering lock
1821 * grabbing by pipe info address. Otherwise two different processes
1822 * could deadlock (one doing tee from A -> B, the other from B -> A).
1824 pipe_double_lock(ipipe, opipe);
1827 if (!opipe->readers) {
1828 send_sig(SIGPIPE, current, 0);
1834 if (!ipipe->nrbufs && !ipipe->writers)
1838 * Cannot make any progress, because either the input
1839 * pipe is empty or the output pipe is full.
1841 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1842 /* Already processed some buffers, break */
1846 if (flags & SPLICE_F_NONBLOCK) {
1852 * We raced with another reader/writer and haven't
1853 * managed to process any buffers. A zero return
1854 * value means EOF, so retry instead.
1861 ibuf = ipipe->bufs + ipipe->curbuf;
1862 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1863 obuf = opipe->bufs + nbuf;
1865 if (len >= ibuf->len) {
1867 * Simply move the whole buffer from ipipe to opipe
1872 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1874 input_wakeup = true;
1877 * Get a reference to this pipe buffer,
1878 * so we can copy the contents over.
1880 ibuf->ops->get(ipipe, ibuf);
1884 * Don't inherit the gift flag, we need to
1885 * prevent multiple steals of this page.
1887 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1891 ibuf->offset += obuf->len;
1892 ibuf->len -= obuf->len;
1902 * If we put data in the output pipe, wakeup any potential readers.
1905 wakeup_pipe_readers(opipe);
1908 wakeup_pipe_writers(ipipe);
1914 * Link contents of ipipe to opipe.
1916 static int link_pipe(struct pipe_inode_info *ipipe,
1917 struct pipe_inode_info *opipe,
1918 size_t len, unsigned int flags)
1920 struct pipe_buffer *ibuf, *obuf;
1921 int ret = 0, i = 0, nbuf;
1924 * Potential ABBA deadlock, work around it by ordering lock
1925 * grabbing by pipe info address. Otherwise two different processes
1926 * could deadlock (one doing tee from A -> B, the other from B -> A).
1928 pipe_double_lock(ipipe, opipe);
1931 if (!opipe->readers) {
1932 send_sig(SIGPIPE, current, 0);
1939 * If we have iterated all input buffers or ran out of
1940 * output room, break.
1942 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1945 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1946 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1949 * Get a reference to this pipe buffer,
1950 * so we can copy the contents over.
1952 ibuf->ops->get(ipipe, ibuf);
1954 obuf = opipe->bufs + nbuf;
1958 * Don't inherit the gift flag, we need to
1959 * prevent multiple steals of this page.
1961 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1963 if (obuf->len > len)
1973 * return EAGAIN if we have the potential of some data in the
1974 * future, otherwise just return 0
1976 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1983 * If we put data in the output pipe, wakeup any potential readers.
1986 wakeup_pipe_readers(opipe);
1992 * This is a tee(1) implementation that works on pipes. It doesn't copy
1993 * any data, it simply references the 'in' pages on the 'out' pipe.
1994 * The 'flags' used are the SPLICE_F_* variants, currently the only
1995 * applicable one is SPLICE_F_NONBLOCK.
1997 static long do_tee(struct file *in, struct file *out, size_t len,
2000 struct pipe_inode_info *ipipe = get_pipe_info(in);
2001 struct pipe_inode_info *opipe = get_pipe_info(out);
2005 * Duplicate the contents of ipipe to opipe without actually
2008 if (ipipe && opipe && ipipe != opipe) {
2010 * Keep going, unless we encounter an error. The ipipe/opipe
2011 * ordering doesn't really matter.
2013 ret = ipipe_prep(ipipe, flags);
2015 ret = opipe_prep(opipe, flags);
2017 ret = link_pipe(ipipe, opipe, len, flags);
2024 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2033 in = fget_light(fdin, &fput_in);
2035 if (in->f_mode & FMODE_READ) {
2037 struct file *out = fget_light(fdout, &fput_out);
2040 if (out->f_mode & FMODE_WRITE)
2041 error = do_tee(in, out, len, flags);
2042 fput_light(out, fput_out);
2045 fput_light(in, fput_in);