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/export.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32 #include <linux/gfp.h>
33 #include <linux/socket.h>
34 #include <linux/compat.h>
38 * Attempt to steal a page from a pipe buffer. This should perhaps go into
39 * a vm helper function, it's already simplified quite a bit by the
40 * addition of remove_mapping(). If success is returned, the caller may
41 * attempt to reuse this page for another destination.
43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
44 struct pipe_buffer *buf)
46 struct page *page = buf->page;
47 struct address_space *mapping;
51 mapping = page_mapping(page);
53 WARN_ON(!PageUptodate(page));
56 * At least for ext2 with nobh option, we need to wait on
57 * writeback completing on this page, since we'll remove it
58 * from the pagecache. Otherwise truncate wont wait on the
59 * page, allowing the disk blocks to be reused by someone else
60 * before we actually wrote our data to them. fs corruption
63 wait_on_page_writeback(page);
65 if (page_has_private(page) &&
66 !try_to_release_page(page, GFP_KERNEL))
70 * If we succeeded in removing the mapping, set LRU flag
73 if (remove_mapping(mapping, page)) {
74 buf->flags |= PIPE_BUF_FLAG_LRU;
80 * Raced with truncate or failed to remove page from current
81 * address space, unlock and return failure.
88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
89 struct pipe_buffer *buf)
91 page_cache_release(buf->page);
92 buf->flags &= ~PIPE_BUF_FLAG_LRU;
96 * Check whether the contents of buf is OK to access. Since the content
97 * is a page cache page, IO may be in flight.
99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
100 struct pipe_buffer *buf)
102 struct page *page = buf->page;
105 if (!PageUptodate(page)) {
109 * Page got truncated/unhashed. This will cause a 0-byte
110 * splice, if this is the first page.
112 if (!page->mapping) {
118 * Uh oh, read-error from disk.
120 if (!PageUptodate(page)) {
126 * Page is ok afterall, we are done.
137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
139 .map = generic_pipe_buf_map,
140 .unmap = generic_pipe_buf_unmap,
141 .confirm = page_cache_pipe_buf_confirm,
142 .release = page_cache_pipe_buf_release,
143 .steal = page_cache_pipe_buf_steal,
144 .get = generic_pipe_buf_get,
147 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
148 struct pipe_buffer *buf)
150 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
153 buf->flags |= PIPE_BUF_FLAG_LRU;
154 return generic_pipe_buf_steal(pipe, buf);
157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
159 .map = generic_pipe_buf_map,
160 .unmap = generic_pipe_buf_unmap,
161 .confirm = generic_pipe_buf_confirm,
162 .release = page_cache_pipe_buf_release,
163 .steal = user_page_pipe_buf_steal,
164 .get = generic_pipe_buf_get,
167 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
170 if (waitqueue_active(&pipe->wait))
171 wake_up_interruptible(&pipe->wait);
172 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
176 * splice_to_pipe - fill passed data into a pipe
177 * @pipe: pipe to fill
181 * @spd contains a map of pages and len/offset tuples, along with
182 * the struct pipe_buf_operations associated with these pages. This
183 * function will link that data to the pipe.
186 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
187 struct splice_pipe_desc *spd)
189 unsigned int spd_pages = spd->nr_pages;
190 int ret, do_wakeup, page_nr;
199 if (!pipe->readers) {
200 send_sig(SIGPIPE, current, 0);
206 if (pipe->nrbufs < pipe->buffers) {
207 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
208 struct pipe_buffer *buf = pipe->bufs + newbuf;
210 buf->page = spd->pages[page_nr];
211 buf->offset = spd->partial[page_nr].offset;
212 buf->len = spd->partial[page_nr].len;
213 buf->private = spd->partial[page_nr].private;
215 if (spd->flags & SPLICE_F_GIFT)
216 buf->flags |= PIPE_BUF_FLAG_GIFT;
225 if (!--spd->nr_pages)
227 if (pipe->nrbufs < pipe->buffers)
233 if (spd->flags & SPLICE_F_NONBLOCK) {
239 if (signal_pending(current)) {
247 if (waitqueue_active(&pipe->wait))
248 wake_up_interruptible_sync(&pipe->wait);
249 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
253 pipe->waiting_writers++;
255 pipe->waiting_writers--;
261 wakeup_pipe_readers(pipe);
263 while (page_nr < spd_pages)
264 spd->spd_release(spd, page_nr++);
269 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
271 page_cache_release(spd->pages[i]);
275 * Check if we need to grow the arrays holding pages and partial page
278 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
280 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
282 spd->nr_pages_max = buffers;
283 if (buffers <= PIPE_DEF_BUFFERS)
286 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
287 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
289 if (spd->pages && spd->partial)
297 void splice_shrink_spd(struct splice_pipe_desc *spd)
299 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
307 __generic_file_splice_read(struct file *in, loff_t *ppos,
308 struct pipe_inode_info *pipe, size_t len,
311 struct address_space *mapping = in->f_mapping;
312 unsigned int loff, nr_pages, req_pages;
313 struct page *pages[PIPE_DEF_BUFFERS];
314 struct partial_page partial[PIPE_DEF_BUFFERS];
316 pgoff_t index, end_index;
319 struct splice_pipe_desc spd = {
322 .nr_pages_max = PIPE_DEF_BUFFERS,
324 .ops = &page_cache_pipe_buf_ops,
325 .spd_release = spd_release_page,
328 if (splice_grow_spd(pipe, &spd))
331 index = *ppos >> PAGE_CACHE_SHIFT;
332 loff = *ppos & ~PAGE_CACHE_MASK;
333 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
334 nr_pages = min(req_pages, spd.nr_pages_max);
337 * Lookup the (hopefully) full range of pages we need.
339 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
340 index += spd.nr_pages;
343 * If find_get_pages_contig() returned fewer pages than we needed,
344 * readahead/allocate the rest and fill in the holes.
346 if (spd.nr_pages < nr_pages)
347 page_cache_sync_readahead(mapping, &in->f_ra, in,
348 index, req_pages - spd.nr_pages);
351 while (spd.nr_pages < nr_pages) {
353 * Page could be there, find_get_pages_contig() breaks on
356 page = find_get_page(mapping, index);
359 * page didn't exist, allocate one.
361 page = page_cache_alloc_cold(mapping);
365 error = add_to_page_cache_lru(page, mapping, index,
367 if (unlikely(error)) {
368 page_cache_release(page);
369 if (error == -EEXIST)
374 * add_to_page_cache() locks the page, unlock it
375 * to avoid convoluting the logic below even more.
380 spd.pages[spd.nr_pages++] = page;
385 * Now loop over the map and see if we need to start IO on any
386 * pages, fill in the partial map, etc.
388 index = *ppos >> PAGE_CACHE_SHIFT;
389 nr_pages = spd.nr_pages;
391 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
392 unsigned int this_len;
398 * this_len is the max we'll use from this page
400 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
401 page = spd.pages[page_nr];
403 if (PageReadahead(page))
404 page_cache_async_readahead(mapping, &in->f_ra, in,
405 page, index, req_pages - page_nr);
408 * If the page isn't uptodate, we may need to start io on it
410 if (!PageUptodate(page)) {
414 * Page was truncated, or invalidated by the
415 * filesystem. Redo the find/create, but this time the
416 * page is kept locked, so there's no chance of another
417 * race with truncate/invalidate.
419 if (!page->mapping) {
421 page = find_or_create_page(mapping, index,
422 mapping_gfp_mask(mapping));
428 page_cache_release(spd.pages[page_nr]);
429 spd.pages[page_nr] = page;
432 * page was already under io and is now done, great
434 if (PageUptodate(page)) {
440 * need to read in the page
442 error = mapping->a_ops->readpage(in, page);
443 if (unlikely(error)) {
445 * We really should re-lookup the page here,
446 * but it complicates things a lot. Instead
447 * lets just do what we already stored, and
448 * we'll get it the next time we are called.
450 if (error == AOP_TRUNCATED_PAGE)
458 * i_size must be checked after PageUptodate.
460 isize = i_size_read(mapping->host);
461 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
462 if (unlikely(!isize || index > end_index))
466 * if this is the last page, see if we need to shrink
467 * the length and stop
469 if (end_index == index) {
473 * max good bytes in this page
475 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
480 * force quit after adding this page
482 this_len = min(this_len, plen - loff);
486 spd.partial[page_nr].offset = loff;
487 spd.partial[page_nr].len = this_len;
495 * Release any pages at the end, if we quit early. 'page_nr' is how far
496 * we got, 'nr_pages' is how many pages are in the map.
498 while (page_nr < nr_pages)
499 page_cache_release(spd.pages[page_nr++]);
500 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
503 error = splice_to_pipe(pipe, &spd);
505 splice_shrink_spd(&spd);
510 * generic_file_splice_read - splice data from file to a pipe
511 * @in: file to splice from
512 * @ppos: position in @in
513 * @pipe: pipe to splice to
514 * @len: number of bytes to splice
515 * @flags: splice modifier flags
518 * Will read pages from given file and fill them into a pipe. Can be
519 * used as long as the address_space operations for the source implements
523 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
524 struct pipe_inode_info *pipe, size_t len,
530 isize = i_size_read(in->f_mapping->host);
531 if (unlikely(*ppos >= isize))
534 left = isize - *ppos;
535 if (unlikely(left < len))
538 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
546 EXPORT_SYMBOL(generic_file_splice_read);
548 static const struct pipe_buf_operations default_pipe_buf_ops = {
550 .map = generic_pipe_buf_map,
551 .unmap = generic_pipe_buf_unmap,
552 .confirm = generic_pipe_buf_confirm,
553 .release = generic_pipe_buf_release,
554 .steal = generic_pipe_buf_steal,
555 .get = generic_pipe_buf_get,
558 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
559 struct pipe_buffer *buf)
564 /* Pipe buffer operations for a socket and similar. */
565 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
567 .map = generic_pipe_buf_map,
568 .unmap = generic_pipe_buf_unmap,
569 .confirm = generic_pipe_buf_confirm,
570 .release = generic_pipe_buf_release,
571 .steal = generic_pipe_buf_nosteal,
572 .get = generic_pipe_buf_get,
574 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
576 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
577 unsigned long vlen, loff_t offset)
585 /* The cast to a user pointer is valid due to the set_fs() */
586 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
592 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
600 /* The cast to a user pointer is valid due to the set_fs() */
601 res = vfs_write(file, (__force const char __user *)buf, count, &pos);
606 EXPORT_SYMBOL(kernel_write);
608 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
609 struct pipe_inode_info *pipe, size_t len,
612 unsigned int nr_pages;
613 unsigned int nr_freed;
615 struct page *pages[PIPE_DEF_BUFFERS];
616 struct partial_page partial[PIPE_DEF_BUFFERS];
617 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
622 struct splice_pipe_desc spd = {
625 .nr_pages_max = PIPE_DEF_BUFFERS,
627 .ops = &default_pipe_buf_ops,
628 .spd_release = spd_release_page,
631 if (splice_grow_spd(pipe, &spd))
636 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
637 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
642 offset = *ppos & ~PAGE_CACHE_MASK;
643 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
645 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
648 page = alloc_page(GFP_USER);
653 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
654 vec[i].iov_base = (void __user *) page_address(page);
655 vec[i].iov_len = this_len;
662 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
673 for (i = 0; i < spd.nr_pages; i++) {
674 this_len = min_t(size_t, vec[i].iov_len, res);
675 spd.partial[i].offset = 0;
676 spd.partial[i].len = this_len;
678 __free_page(spd.pages[i]);
684 spd.nr_pages -= nr_freed;
686 res = splice_to_pipe(pipe, &spd);
693 splice_shrink_spd(&spd);
697 for (i = 0; i < spd.nr_pages; i++)
698 __free_page(spd.pages[i]);
703 EXPORT_SYMBOL(default_file_splice_read);
706 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
707 * using sendpage(). Return the number of bytes sent.
709 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
710 struct pipe_buffer *buf, struct splice_desc *sd)
712 struct file *file = sd->u.file;
713 loff_t pos = sd->pos;
716 if (!likely(file->f_op && file->f_op->sendpage))
719 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
721 if (sd->len < sd->total_len && pipe->nrbufs > 1)
722 more |= MSG_SENDPAGE_NOTLAST;
724 return file->f_op->sendpage(file, buf->page, buf->offset,
725 sd->len, &pos, more);
729 * This is a little more tricky than the file -> pipe splicing. There are
730 * basically three cases:
732 * - Destination page already exists in the address space and there
733 * are users of it. For that case we have no other option that
734 * copying the data. Tough luck.
735 * - Destination page already exists in the address space, but there
736 * are no users of it. Make sure it's uptodate, then drop it. Fall
737 * through to last case.
738 * - Destination page does not exist, we can add the pipe page to
739 * the page cache and avoid the copy.
741 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
742 * sd->flags), we attempt to migrate pages from the pipe to the output
743 * file address space page cache. This is possible if no one else has
744 * the pipe page referenced outside of the pipe and page cache. If
745 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
746 * a new page in the output file page cache and fill/dirty that.
748 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
749 struct splice_desc *sd)
751 struct file *file = sd->u.file;
752 struct address_space *mapping = file->f_mapping;
753 unsigned int offset, this_len;
758 offset = sd->pos & ~PAGE_CACHE_MASK;
761 if (this_len + offset > PAGE_CACHE_SIZE)
762 this_len = PAGE_CACHE_SIZE - offset;
764 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
765 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
769 if (buf->page != page) {
770 char *src = buf->ops->map(pipe, buf, 1);
771 char *dst = kmap_atomic(page);
773 memcpy(dst + offset, src + buf->offset, this_len);
774 flush_dcache_page(page);
776 buf->ops->unmap(pipe, buf, src);
778 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
783 EXPORT_SYMBOL(pipe_to_file);
785 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
788 if (waitqueue_active(&pipe->wait))
789 wake_up_interruptible(&pipe->wait);
790 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
794 * splice_from_pipe_feed - feed available data from a pipe to a file
795 * @pipe: pipe to splice from
796 * @sd: information to @actor
797 * @actor: handler that splices the data
800 * This function loops over the pipe and calls @actor to do the
801 * actual moving of a single struct pipe_buffer to the desired
802 * destination. It returns when there's no more buffers left in
803 * the pipe or if the requested number of bytes (@sd->total_len)
804 * have been copied. It returns a positive number (one) if the
805 * pipe needs to be filled with more data, zero if the required
806 * number of bytes have been copied and -errno on error.
808 * This, together with splice_from_pipe_{begin,end,next}, may be
809 * used to implement the functionality of __splice_from_pipe() when
810 * locking is required around copying the pipe buffers to the
813 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
818 while (pipe->nrbufs) {
819 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
820 const struct pipe_buf_operations *ops = buf->ops;
823 if (sd->len > sd->total_len)
824 sd->len = sd->total_len;
826 ret = buf->ops->confirm(pipe, buf);
833 ret = actor(pipe, buf, sd);
840 sd->num_spliced += ret;
843 sd->total_len -= ret;
847 ops->release(pipe, buf);
848 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
851 sd->need_wakeup = true;
860 EXPORT_SYMBOL(splice_from_pipe_feed);
863 * splice_from_pipe_next - wait for some data to splice from
864 * @pipe: pipe to splice from
865 * @sd: information about the splice operation
868 * This function will wait for some data and return a positive
869 * value (one) if pipe buffers are available. It will return zero
870 * or -errno if no more data needs to be spliced.
872 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
874 while (!pipe->nrbufs) {
878 if (!pipe->waiting_writers && sd->num_spliced)
881 if (sd->flags & SPLICE_F_NONBLOCK)
884 if (signal_pending(current))
887 if (sd->need_wakeup) {
888 wakeup_pipe_writers(pipe);
889 sd->need_wakeup = false;
897 EXPORT_SYMBOL(splice_from_pipe_next);
900 * splice_from_pipe_begin - start splicing from pipe
901 * @sd: information about the splice operation
904 * This function should be called before a loop containing
905 * splice_from_pipe_next() and splice_from_pipe_feed() to
906 * initialize the necessary fields of @sd.
908 void splice_from_pipe_begin(struct splice_desc *sd)
911 sd->need_wakeup = false;
913 EXPORT_SYMBOL(splice_from_pipe_begin);
916 * splice_from_pipe_end - finish splicing from pipe
917 * @pipe: pipe to splice from
918 * @sd: information about the splice operation
921 * This function will wake up pipe writers if necessary. It should
922 * be called after a loop containing splice_from_pipe_next() and
923 * splice_from_pipe_feed().
925 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
928 wakeup_pipe_writers(pipe);
930 EXPORT_SYMBOL(splice_from_pipe_end);
933 * __splice_from_pipe - splice data from a pipe to given actor
934 * @pipe: pipe to splice from
935 * @sd: information to @actor
936 * @actor: handler that splices the data
939 * This function does little more than loop over the pipe and call
940 * @actor to do the actual moving of a single struct pipe_buffer to
941 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
945 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
950 splice_from_pipe_begin(sd);
952 ret = splice_from_pipe_next(pipe, sd);
954 ret = splice_from_pipe_feed(pipe, sd, actor);
956 splice_from_pipe_end(pipe, sd);
958 return sd->num_spliced ? sd->num_spliced : ret;
960 EXPORT_SYMBOL(__splice_from_pipe);
963 * splice_from_pipe - splice data from a pipe to a file
964 * @pipe: pipe to splice from
965 * @out: file to splice to
966 * @ppos: position in @out
967 * @len: how many bytes to splice
968 * @flags: splice modifier flags
969 * @actor: handler that splices the data
972 * See __splice_from_pipe. This function locks the pipe inode,
973 * otherwise it's identical to __splice_from_pipe().
976 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
977 loff_t *ppos, size_t len, unsigned int flags,
981 struct splice_desc sd = {
989 ret = __splice_from_pipe(pipe, &sd, actor);
996 * generic_file_splice_write - splice data from a pipe to a file
998 * @out: file to write to
999 * @ppos: position in @out
1000 * @len: number of bytes to splice
1001 * @flags: splice modifier flags
1004 * Will either move or copy pages (determined by @flags options) from
1005 * the given pipe inode to the given file.
1009 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1010 loff_t *ppos, size_t len, unsigned int flags)
1012 struct address_space *mapping = out->f_mapping;
1013 struct inode *inode = mapping->host;
1014 struct splice_desc sd = {
1020 ret = generic_write_checks(out, ppos, &len, S_ISBLK(inode->i_mode));
1028 splice_from_pipe_begin(&sd);
1030 ret = splice_from_pipe_next(pipe, &sd);
1034 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1035 ret = file_remove_suid(out);
1037 ret = file_update_time(out);
1039 ret = splice_from_pipe_feed(pipe, &sd,
1042 mutex_unlock(&inode->i_mutex);
1044 splice_from_pipe_end(pipe, &sd);
1049 ret = sd.num_spliced;
1054 err = generic_write_sync(out, *ppos, ret);
1059 balance_dirty_pages_ratelimited(mapping);
1065 EXPORT_SYMBOL(generic_file_splice_write);
1067 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1068 struct splice_desc *sd)
1072 loff_t tmp = sd->pos;
1074 data = buf->ops->map(pipe, buf, 0);
1075 ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1076 buf->ops->unmap(pipe, buf, data);
1081 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1082 struct file *out, loff_t *ppos,
1083 size_t len, unsigned int flags)
1087 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1095 * generic_splice_sendpage - splice data from a pipe to a socket
1096 * @pipe: pipe to splice from
1097 * @out: socket to write to
1098 * @ppos: position in @out
1099 * @len: number of bytes to splice
1100 * @flags: splice modifier flags
1103 * Will send @len bytes from the pipe to a network socket. No data copying
1107 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1108 loff_t *ppos, size_t len, unsigned int flags)
1110 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1113 EXPORT_SYMBOL(generic_splice_sendpage);
1116 * Attempt to initiate a splice from pipe to file.
1118 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1119 loff_t *ppos, size_t len, unsigned int flags)
1121 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1122 loff_t *, size_t, unsigned int);
1125 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1128 if (unlikely(out->f_flags & O_APPEND))
1131 ret = rw_verify_area(WRITE, out, ppos, len);
1132 if (unlikely(ret < 0))
1135 if (out->f_op && out->f_op->splice_write)
1136 splice_write = out->f_op->splice_write;
1138 splice_write = default_file_splice_write;
1140 file_start_write(out);
1141 ret = splice_write(pipe, out, ppos, len, flags);
1142 file_end_write(out);
1147 * Attempt to initiate a splice from a file to a pipe.
1149 static long do_splice_to(struct file *in, loff_t *ppos,
1150 struct pipe_inode_info *pipe, size_t len,
1153 ssize_t (*splice_read)(struct file *, loff_t *,
1154 struct pipe_inode_info *, size_t, unsigned int);
1157 if (unlikely(!(in->f_mode & FMODE_READ)))
1160 ret = rw_verify_area(READ, in, ppos, len);
1161 if (unlikely(ret < 0))
1164 if (in->f_op && in->f_op->splice_read)
1165 splice_read = in->f_op->splice_read;
1167 splice_read = default_file_splice_read;
1169 return splice_read(in, ppos, pipe, len, flags);
1173 * splice_direct_to_actor - splices data directly between two non-pipes
1174 * @in: file to splice from
1175 * @sd: actor information on where to splice to
1176 * @actor: handles the data splicing
1179 * This is a special case helper to splice directly between two
1180 * points, without requiring an explicit pipe. Internally an allocated
1181 * pipe is cached in the process, and reused during the lifetime of
1185 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1186 splice_direct_actor *actor)
1188 struct pipe_inode_info *pipe;
1195 * We require the input being a regular file, as we don't want to
1196 * randomly drop data for eg socket -> socket splicing. Use the
1197 * piped splicing for that!
1199 i_mode = file_inode(in)->i_mode;
1200 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1204 * neither in nor out is a pipe, setup an internal pipe attached to
1205 * 'out' and transfer the wanted data from 'in' to 'out' through that
1207 pipe = current->splice_pipe;
1208 if (unlikely(!pipe)) {
1209 pipe = alloc_pipe_info();
1214 * We don't have an immediate reader, but we'll read the stuff
1215 * out of the pipe right after the splice_to_pipe(). So set
1216 * PIPE_READERS appropriately.
1220 current->splice_pipe = pipe;
1228 len = sd->total_len;
1232 * Don't block on output, we have to drain the direct pipe.
1234 sd->flags &= ~SPLICE_F_NONBLOCK;
1238 loff_t pos = sd->pos, prev_pos = pos;
1240 ret = do_splice_to(in, &pos, pipe, len, flags);
1241 if (unlikely(ret <= 0))
1245 sd->total_len = read_len;
1248 * NOTE: nonblocking mode only applies to the input. We
1249 * must not do the output in nonblocking mode as then we
1250 * could get stuck data in the internal pipe:
1252 ret = actor(pipe, sd);
1253 if (unlikely(ret <= 0)) {
1262 if (ret < read_len) {
1263 sd->pos = prev_pos + ret;
1269 pipe->nrbufs = pipe->curbuf = 0;
1275 * If we did an incomplete transfer we must release
1276 * the pipe buffers in question:
1278 for (i = 0; i < pipe->buffers; i++) {
1279 struct pipe_buffer *buf = pipe->bufs + i;
1282 buf->ops->release(pipe, buf);
1292 EXPORT_SYMBOL(splice_direct_to_actor);
1294 static int direct_splice_actor(struct pipe_inode_info *pipe,
1295 struct splice_desc *sd)
1297 struct file *file = sd->u.file;
1299 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1304 * do_splice_direct - splices data directly between two files
1305 * @in: file to splice from
1306 * @ppos: input file offset
1307 * @out: file to splice to
1308 * @opos: output file offset
1309 * @len: number of bytes to splice
1310 * @flags: splice modifier flags
1313 * For use by do_sendfile(). splice can easily emulate sendfile, but
1314 * doing it in the application would incur an extra system call
1315 * (splice in + splice out, as compared to just sendfile()). So this helper
1316 * can splice directly through a process-private pipe.
1319 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1320 loff_t *opos, size_t len, unsigned int flags)
1322 struct splice_desc sd = {
1332 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1339 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1340 struct pipe_inode_info *opipe,
1341 size_t len, unsigned int flags);
1344 * Determine where to splice to/from.
1346 static long do_splice(struct file *in, loff_t __user *off_in,
1347 struct file *out, loff_t __user *off_out,
1348 size_t len, unsigned int flags)
1350 struct pipe_inode_info *ipipe;
1351 struct pipe_inode_info *opipe;
1355 ipipe = get_pipe_info(in);
1356 opipe = get_pipe_info(out);
1358 if (ipipe && opipe) {
1359 if (off_in || off_out)
1362 if (!(in->f_mode & FMODE_READ))
1365 if (!(out->f_mode & FMODE_WRITE))
1368 /* Splicing to self would be fun, but... */
1372 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1379 if (!(out->f_mode & FMODE_PWRITE))
1381 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1384 offset = out->f_pos;
1387 ret = do_splice_from(ipipe, out, &offset, len, flags);
1390 out->f_pos = offset;
1391 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1401 if (!(in->f_mode & FMODE_PREAD))
1403 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1409 ret = do_splice_to(in, &offset, opipe, len, flags);
1413 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1423 * Map an iov into an array of pages and offset/length tupples. With the
1424 * partial_page structure, we can map several non-contiguous ranges into
1425 * our ones pages[] map instead of splitting that operation into pieces.
1426 * Could easily be exported as a generic helper for other users, in which
1427 * case one would probably want to add a 'max_nr_pages' parameter as well.
1429 static int get_iovec_page_array(const struct iovec __user *iov,
1430 unsigned int nr_vecs, struct page **pages,
1431 struct partial_page *partial, bool aligned,
1432 unsigned int pipe_buffers)
1434 int buffers = 0, error = 0;
1437 unsigned long off, npages;
1444 if (copy_from_user(&entry, iov, sizeof(entry)))
1447 base = entry.iov_base;
1448 len = entry.iov_len;
1451 * Sanity check this iovec. 0 read succeeds.
1457 if (!access_ok(VERIFY_READ, base, len))
1461 * Get this base offset and number of pages, then map
1462 * in the user pages.
1464 off = (unsigned long) base & ~PAGE_MASK;
1467 * If asked for alignment, the offset must be zero and the
1468 * length a multiple of the PAGE_SIZE.
1471 if (aligned && (off || len & ~PAGE_MASK))
1474 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1475 if (npages > pipe_buffers - buffers)
1476 npages = pipe_buffers - buffers;
1478 error = get_user_pages_fast((unsigned long)base, npages,
1479 0, &pages[buffers]);
1481 if (unlikely(error <= 0))
1485 * Fill this contiguous range into the partial page map.
1487 for (i = 0; i < error; i++) {
1488 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1490 partial[buffers].offset = off;
1491 partial[buffers].len = plen;
1499 * We didn't complete this iov, stop here since it probably
1500 * means we have to move some of this into a pipe to
1501 * be able to continue.
1507 * Don't continue if we mapped fewer pages than we asked for,
1508 * or if we mapped the max number of pages that we have
1511 if (error < npages || buffers == pipe_buffers)
1524 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1525 struct splice_desc *sd)
1531 * See if we can use the atomic maps, by prefaulting in the
1532 * pages and doing an atomic copy
1534 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1535 src = buf->ops->map(pipe, buf, 1);
1536 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1538 buf->ops->unmap(pipe, buf, src);
1546 * No dice, use slow non-atomic map and copy
1548 src = buf->ops->map(pipe, buf, 0);
1551 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1554 buf->ops->unmap(pipe, buf, src);
1557 sd->u.userptr += ret;
1562 * For lack of a better implementation, implement vmsplice() to userspace
1563 * as a simple copy of the pipes pages to the user iov.
1565 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1566 unsigned long nr_segs, unsigned int flags)
1568 struct pipe_inode_info *pipe;
1569 struct splice_desc sd;
1574 pipe = get_pipe_info(file);
1586 * Get user address base and length for this iovec.
1588 error = get_user(base, &iov->iov_base);
1589 if (unlikely(error))
1591 error = get_user(len, &iov->iov_len);
1592 if (unlikely(error))
1596 * Sanity check this iovec. 0 read succeeds.
1600 if (unlikely(!base)) {
1605 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1613 sd.u.userptr = base;
1616 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1642 * vmsplice splices a user address range into a pipe. It can be thought of
1643 * as splice-from-memory, where the regular splice is splice-from-file (or
1644 * to file). In both cases the output is a pipe, naturally.
1646 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1647 unsigned long nr_segs, unsigned int flags)
1649 struct pipe_inode_info *pipe;
1650 struct page *pages[PIPE_DEF_BUFFERS];
1651 struct partial_page partial[PIPE_DEF_BUFFERS];
1652 struct splice_pipe_desc spd = {
1655 .nr_pages_max = PIPE_DEF_BUFFERS,
1657 .ops = &user_page_pipe_buf_ops,
1658 .spd_release = spd_release_page,
1662 pipe = get_pipe_info(file);
1666 if (splice_grow_spd(pipe, &spd))
1669 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1672 if (spd.nr_pages <= 0)
1675 ret = splice_to_pipe(pipe, &spd);
1677 splice_shrink_spd(&spd);
1682 * Note that vmsplice only really supports true splicing _from_ user memory
1683 * to a pipe, not the other way around. Splicing from user memory is a simple
1684 * operation that can be supported without any funky alignment restrictions
1685 * or nasty vm tricks. We simply map in the user memory and fill them into
1686 * a pipe. The reverse isn't quite as easy, though. There are two possible
1687 * solutions for that:
1689 * - memcpy() the data internally, at which point we might as well just
1690 * do a regular read() on the buffer anyway.
1691 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1692 * has restriction limitations on both ends of the pipe).
1694 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1697 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1698 unsigned long, nr_segs, unsigned int, flags)
1703 if (unlikely(nr_segs > UIO_MAXIOV))
1705 else if (unlikely(!nr_segs))
1711 if (f.file->f_mode & FMODE_WRITE)
1712 error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1713 else if (f.file->f_mode & FMODE_READ)
1714 error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1722 #ifdef CONFIG_COMPAT
1723 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1724 unsigned int, nr_segs, unsigned int, flags)
1727 struct iovec __user *iov;
1728 if (nr_segs > UIO_MAXIOV)
1730 iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1731 for (i = 0; i < nr_segs; i++) {
1732 struct compat_iovec v;
1733 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1734 get_user(v.iov_len, &iov32[i].iov_len) ||
1735 put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1736 put_user(v.iov_len, &iov[i].iov_len))
1739 return sys_vmsplice(fd, iov, nr_segs, flags);
1743 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1744 int, fd_out, loff_t __user *, off_out,
1745 size_t, len, unsigned int, flags)
1756 if (in.file->f_mode & FMODE_READ) {
1757 out = fdget(fd_out);
1759 if (out.file->f_mode & FMODE_WRITE)
1760 error = do_splice(in.file, off_in,
1772 * Make sure there's data to read. Wait for input if we can, otherwise
1773 * return an appropriate error.
1775 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1780 * Check ->nrbufs without the inode lock first. This function
1781 * is speculative anyways, so missing one is ok.
1789 while (!pipe->nrbufs) {
1790 if (signal_pending(current)) {
1796 if (!pipe->waiting_writers) {
1797 if (flags & SPLICE_F_NONBLOCK) {
1810 * Make sure there's writeable room. Wait for room if we can, otherwise
1811 * return an appropriate error.
1813 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1818 * Check ->nrbufs without the inode lock first. This function
1819 * is speculative anyways, so missing one is ok.
1821 if (pipe->nrbufs < pipe->buffers)
1827 while (pipe->nrbufs >= pipe->buffers) {
1828 if (!pipe->readers) {
1829 send_sig(SIGPIPE, current, 0);
1833 if (flags & SPLICE_F_NONBLOCK) {
1837 if (signal_pending(current)) {
1841 pipe->waiting_writers++;
1843 pipe->waiting_writers--;
1851 * Splice contents of ipipe to opipe.
1853 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1854 struct pipe_inode_info *opipe,
1855 size_t len, unsigned int flags)
1857 struct pipe_buffer *ibuf, *obuf;
1859 bool input_wakeup = false;
1863 ret = ipipe_prep(ipipe, flags);
1867 ret = opipe_prep(opipe, flags);
1872 * Potential ABBA deadlock, work around it by ordering lock
1873 * grabbing by pipe info address. Otherwise two different processes
1874 * could deadlock (one doing tee from A -> B, the other from B -> A).
1876 pipe_double_lock(ipipe, opipe);
1879 if (!opipe->readers) {
1880 send_sig(SIGPIPE, current, 0);
1886 if (!ipipe->nrbufs && !ipipe->writers)
1890 * Cannot make any progress, because either the input
1891 * pipe is empty or the output pipe is full.
1893 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1894 /* Already processed some buffers, break */
1898 if (flags & SPLICE_F_NONBLOCK) {
1904 * We raced with another reader/writer and haven't
1905 * managed to process any buffers. A zero return
1906 * value means EOF, so retry instead.
1913 ibuf = ipipe->bufs + ipipe->curbuf;
1914 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1915 obuf = opipe->bufs + nbuf;
1917 if (len >= ibuf->len) {
1919 * Simply move the whole buffer from ipipe to opipe
1924 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1926 input_wakeup = true;
1929 * Get a reference to this pipe buffer,
1930 * so we can copy the contents over.
1932 ibuf->ops->get(ipipe, ibuf);
1936 * Don't inherit the gift flag, we need to
1937 * prevent multiple steals of this page.
1939 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1943 ibuf->offset += obuf->len;
1944 ibuf->len -= obuf->len;
1954 * If we put data in the output pipe, wakeup any potential readers.
1957 wakeup_pipe_readers(opipe);
1960 wakeup_pipe_writers(ipipe);
1966 * Link contents of ipipe to opipe.
1968 static int link_pipe(struct pipe_inode_info *ipipe,
1969 struct pipe_inode_info *opipe,
1970 size_t len, unsigned int flags)
1972 struct pipe_buffer *ibuf, *obuf;
1973 int ret = 0, i = 0, nbuf;
1976 * Potential ABBA deadlock, work around it by ordering lock
1977 * grabbing by pipe info address. Otherwise two different processes
1978 * could deadlock (one doing tee from A -> B, the other from B -> A).
1980 pipe_double_lock(ipipe, opipe);
1983 if (!opipe->readers) {
1984 send_sig(SIGPIPE, current, 0);
1991 * If we have iterated all input buffers or ran out of
1992 * output room, break.
1994 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1997 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1998 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2001 * Get a reference to this pipe buffer,
2002 * so we can copy the contents over.
2004 ibuf->ops->get(ipipe, ibuf);
2006 obuf = opipe->bufs + nbuf;
2010 * Don't inherit the gift flag, we need to
2011 * prevent multiple steals of this page.
2013 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2015 if (obuf->len > len)
2025 * return EAGAIN if we have the potential of some data in the
2026 * future, otherwise just return 0
2028 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2035 * If we put data in the output pipe, wakeup any potential readers.
2038 wakeup_pipe_readers(opipe);
2044 * This is a tee(1) implementation that works on pipes. It doesn't copy
2045 * any data, it simply references the 'in' pages on the 'out' pipe.
2046 * The 'flags' used are the SPLICE_F_* variants, currently the only
2047 * applicable one is SPLICE_F_NONBLOCK.
2049 static long do_tee(struct file *in, struct file *out, size_t len,
2052 struct pipe_inode_info *ipipe = get_pipe_info(in);
2053 struct pipe_inode_info *opipe = get_pipe_info(out);
2057 * Duplicate the contents of ipipe to opipe without actually
2060 if (ipipe && opipe && ipipe != opipe) {
2062 * Keep going, unless we encounter an error. The ipipe/opipe
2063 * ordering doesn't really matter.
2065 ret = ipipe_prep(ipipe, flags);
2067 ret = opipe_prep(opipe, flags);
2069 ret = link_pipe(ipipe, opipe, len, flags);
2076 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2087 if (in.file->f_mode & FMODE_READ) {
2088 struct fd out = fdget(fdout);
2090 if (out.file->f_mode & FMODE_WRITE)
2091 error = do_tee(in.file, out.file,