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;
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 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(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
279 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
281 spd->nr_pages_max = buffers;
282 if (buffers <= PIPE_DEF_BUFFERS)
285 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
286 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
288 if (spd->pages && spd->partial)
296 void splice_shrink_spd(struct splice_pipe_desc *spd)
298 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
306 __generic_file_splice_read(struct file *in, loff_t *ppos,
307 struct pipe_inode_info *pipe, size_t len,
310 struct address_space *mapping = in->f_mapping;
311 unsigned int loff, nr_pages, req_pages;
312 struct page *pages[PIPE_DEF_BUFFERS];
313 struct partial_page partial[PIPE_DEF_BUFFERS];
315 pgoff_t index, end_index;
318 struct splice_pipe_desc spd = {
321 .nr_pages_max = PIPE_DEF_BUFFERS,
323 .ops = &page_cache_pipe_buf_ops,
324 .spd_release = spd_release_page,
327 if (splice_grow_spd(pipe, &spd))
330 index = *ppos >> PAGE_CACHE_SHIFT;
331 loff = *ppos & ~PAGE_CACHE_MASK;
332 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
333 nr_pages = min(req_pages, spd.nr_pages_max);
336 * Lookup the (hopefully) full range of pages we need.
338 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
339 index += spd.nr_pages;
342 * If find_get_pages_contig() returned fewer pages than we needed,
343 * readahead/allocate the rest and fill in the holes.
345 if (spd.nr_pages < nr_pages)
346 page_cache_sync_readahead(mapping, &in->f_ra, in,
347 index, req_pages - spd.nr_pages);
350 while (spd.nr_pages < nr_pages) {
352 * Page could be there, find_get_pages_contig() breaks on
355 page = find_get_page(mapping, index);
358 * page didn't exist, allocate one.
360 page = page_cache_alloc_cold(mapping);
364 error = add_to_page_cache_lru(page, mapping, index,
366 if (unlikely(error)) {
367 page_cache_release(page);
368 if (error == -EEXIST)
373 * add_to_page_cache() locks the page, unlock it
374 * to avoid convoluting the logic below even more.
379 spd.pages[spd.nr_pages++] = page;
384 * Now loop over the map and see if we need to start IO on any
385 * pages, fill in the partial map, etc.
387 index = *ppos >> PAGE_CACHE_SHIFT;
388 nr_pages = spd.nr_pages;
390 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
391 unsigned int this_len;
397 * this_len is the max we'll use from this page
399 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
400 page = spd.pages[page_nr];
402 if (PageReadahead(page))
403 page_cache_async_readahead(mapping, &in->f_ra, in,
404 page, index, req_pages - page_nr);
407 * If the page isn't uptodate, we may need to start io on it
409 if (!PageUptodate(page)) {
413 * Page was truncated, or invalidated by the
414 * filesystem. Redo the find/create, but this time the
415 * page is kept locked, so there's no chance of another
416 * race with truncate/invalidate.
418 if (!page->mapping) {
420 page = find_or_create_page(mapping, index,
421 mapping_gfp_mask(mapping));
427 page_cache_release(spd.pages[page_nr]);
428 spd.pages[page_nr] = page;
431 * page was already under io and is now done, great
433 if (PageUptodate(page)) {
439 * need to read in the page
441 error = mapping->a_ops->readpage(in, page);
442 if (unlikely(error)) {
444 * We really should re-lookup the page here,
445 * but it complicates things a lot. Instead
446 * lets just do what we already stored, and
447 * we'll get it the next time we are called.
449 if (error == AOP_TRUNCATED_PAGE)
457 * i_size must be checked after PageUptodate.
459 isize = i_size_read(mapping->host);
460 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
461 if (unlikely(!isize || index > end_index))
465 * if this is the last page, see if we need to shrink
466 * the length and stop
468 if (end_index == index) {
472 * max good bytes in this page
474 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
479 * force quit after adding this page
481 this_len = min(this_len, plen - loff);
485 spd.partial[page_nr].offset = loff;
486 spd.partial[page_nr].len = this_len;
494 * Release any pages at the end, if we quit early. 'page_nr' is how far
495 * we got, 'nr_pages' is how many pages are in the map.
497 while (page_nr < nr_pages)
498 page_cache_release(spd.pages[page_nr++]);
499 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
502 error = splice_to_pipe(pipe, &spd);
504 splice_shrink_spd(&spd);
509 * generic_file_splice_read - splice data from file to a pipe
510 * @in: file to splice from
511 * @ppos: position in @in
512 * @pipe: pipe to splice to
513 * @len: number of bytes to splice
514 * @flags: splice modifier flags
517 * Will read pages from given file and fill them into a pipe. Can be
518 * used as long as the address_space operations for the source implements
522 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
523 struct pipe_inode_info *pipe, size_t len,
529 isize = i_size_read(in->f_mapping->host);
530 if (unlikely(*ppos >= isize))
533 left = isize - *ppos;
534 if (unlikely(left < len))
537 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
545 EXPORT_SYMBOL(generic_file_splice_read);
547 static const struct pipe_buf_operations default_pipe_buf_ops = {
549 .map = generic_pipe_buf_map,
550 .unmap = generic_pipe_buf_unmap,
551 .confirm = generic_pipe_buf_confirm,
552 .release = generic_pipe_buf_release,
553 .steal = generic_pipe_buf_steal,
554 .get = generic_pipe_buf_get,
557 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
558 unsigned long vlen, loff_t offset)
566 /* The cast to a user pointer is valid due to the set_fs() */
567 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
573 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
581 /* The cast to a user pointer is valid due to the set_fs() */
582 res = vfs_write(file, (const char __user *)buf, count, &pos);
588 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
589 struct pipe_inode_info *pipe, size_t len,
592 unsigned int nr_pages;
593 unsigned int nr_freed;
595 struct page *pages[PIPE_DEF_BUFFERS];
596 struct partial_page partial[PIPE_DEF_BUFFERS];
597 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
602 struct splice_pipe_desc spd = {
605 .nr_pages_max = PIPE_DEF_BUFFERS,
607 .ops = &default_pipe_buf_ops,
608 .spd_release = spd_release_page,
611 if (splice_grow_spd(pipe, &spd))
616 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
617 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
622 offset = *ppos & ~PAGE_CACHE_MASK;
623 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
625 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
628 page = alloc_page(GFP_USER);
633 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
634 vec[i].iov_base = (void __user *) page_address(page);
635 vec[i].iov_len = this_len;
642 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
653 for (i = 0; i < spd.nr_pages; i++) {
654 this_len = min_t(size_t, vec[i].iov_len, res);
655 spd.partial[i].offset = 0;
656 spd.partial[i].len = this_len;
658 __free_page(spd.pages[i]);
664 spd.nr_pages -= nr_freed;
666 res = splice_to_pipe(pipe, &spd);
673 splice_shrink_spd(&spd);
677 for (i = 0; i < spd.nr_pages; i++)
678 __free_page(spd.pages[i]);
683 EXPORT_SYMBOL(default_file_splice_read);
686 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
687 * using sendpage(). Return the number of bytes sent.
689 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
690 struct pipe_buffer *buf, struct splice_desc *sd)
692 struct file *file = sd->u.file;
693 loff_t pos = sd->pos;
696 if (!likely(file->f_op && file->f_op->sendpage))
699 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
701 if (sd->len < sd->total_len && pipe->nrbufs > 1)
702 more |= MSG_SENDPAGE_NOTLAST;
704 return file->f_op->sendpage(file, buf->page, buf->offset,
705 sd->len, &pos, more);
709 * This is a little more tricky than the file -> pipe splicing. There are
710 * basically three cases:
712 * - Destination page already exists in the address space and there
713 * are users of it. For that case we have no other option that
714 * copying the data. Tough luck.
715 * - Destination page already exists in the address space, but there
716 * are no users of it. Make sure it's uptodate, then drop it. Fall
717 * through to last case.
718 * - Destination page does not exist, we can add the pipe page to
719 * the page cache and avoid the copy.
721 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
722 * sd->flags), we attempt to migrate pages from the pipe to the output
723 * file address space page cache. This is possible if no one else has
724 * the pipe page referenced outside of the pipe and page cache. If
725 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
726 * a new page in the output file page cache and fill/dirty that.
728 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
729 struct splice_desc *sd)
731 struct file *file = sd->u.file;
732 struct address_space *mapping = file->f_mapping;
733 unsigned int offset, this_len;
738 offset = sd->pos & ~PAGE_CACHE_MASK;
741 if (this_len + offset > PAGE_CACHE_SIZE)
742 this_len = PAGE_CACHE_SIZE - offset;
744 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
745 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
749 if (buf->page != page) {
751 * Careful, ->map() uses KM_USER0!
753 char *src = buf->ops->map(pipe, buf, 1);
754 char *dst = kmap_atomic(page, KM_USER1);
756 memcpy(dst + offset, src + buf->offset, this_len);
757 flush_dcache_page(page);
758 kunmap_atomic(dst, KM_USER1);
759 buf->ops->unmap(pipe, buf, src);
761 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
766 EXPORT_SYMBOL(pipe_to_file);
768 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
771 if (waitqueue_active(&pipe->wait))
772 wake_up_interruptible(&pipe->wait);
773 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
777 * splice_from_pipe_feed - feed available data from a pipe to a file
778 * @pipe: pipe to splice from
779 * @sd: information to @actor
780 * @actor: handler that splices the data
783 * This function loops over the pipe and calls @actor to do the
784 * actual moving of a single struct pipe_buffer to the desired
785 * destination. It returns when there's no more buffers left in
786 * the pipe or if the requested number of bytes (@sd->total_len)
787 * have been copied. It returns a positive number (one) if the
788 * pipe needs to be filled with more data, zero if the required
789 * number of bytes have been copied and -errno on error.
791 * This, together with splice_from_pipe_{begin,end,next}, may be
792 * used to implement the functionality of __splice_from_pipe() when
793 * locking is required around copying the pipe buffers to the
796 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
801 while (pipe->nrbufs) {
802 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
803 const struct pipe_buf_operations *ops = buf->ops;
806 if (sd->len > sd->total_len)
807 sd->len = sd->total_len;
809 ret = buf->ops->confirm(pipe, buf);
816 ret = actor(pipe, buf, sd);
823 sd->num_spliced += ret;
826 sd->total_len -= ret;
830 ops->release(pipe, buf);
831 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
834 sd->need_wakeup = true;
843 EXPORT_SYMBOL(splice_from_pipe_feed);
846 * splice_from_pipe_next - wait for some data to splice from
847 * @pipe: pipe to splice from
848 * @sd: information about the splice operation
851 * This function will wait for some data and return a positive
852 * value (one) if pipe buffers are available. It will return zero
853 * or -errno if no more data needs to be spliced.
855 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
857 while (!pipe->nrbufs) {
861 if (!pipe->waiting_writers && sd->num_spliced)
864 if (sd->flags & SPLICE_F_NONBLOCK)
867 if (signal_pending(current))
870 if (sd->need_wakeup) {
871 wakeup_pipe_writers(pipe);
872 sd->need_wakeup = false;
880 EXPORT_SYMBOL(splice_from_pipe_next);
883 * splice_from_pipe_begin - start splicing from pipe
884 * @sd: information about the splice operation
887 * This function should be called before a loop containing
888 * splice_from_pipe_next() and splice_from_pipe_feed() to
889 * initialize the necessary fields of @sd.
891 void splice_from_pipe_begin(struct splice_desc *sd)
894 sd->need_wakeup = false;
896 EXPORT_SYMBOL(splice_from_pipe_begin);
899 * splice_from_pipe_end - finish splicing from pipe
900 * @pipe: pipe to splice from
901 * @sd: information about the splice operation
904 * This function will wake up pipe writers if necessary. It should
905 * be called after a loop containing splice_from_pipe_next() and
906 * splice_from_pipe_feed().
908 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
911 wakeup_pipe_writers(pipe);
913 EXPORT_SYMBOL(splice_from_pipe_end);
916 * __splice_from_pipe - splice data from a pipe to given actor
917 * @pipe: pipe to splice from
918 * @sd: information to @actor
919 * @actor: handler that splices the data
922 * This function does little more than loop over the pipe and call
923 * @actor to do the actual moving of a single struct pipe_buffer to
924 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
928 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
933 splice_from_pipe_begin(sd);
935 ret = splice_from_pipe_next(pipe, sd);
937 ret = splice_from_pipe_feed(pipe, sd, actor);
939 splice_from_pipe_end(pipe, sd);
941 return sd->num_spliced ? sd->num_spliced : ret;
943 EXPORT_SYMBOL(__splice_from_pipe);
946 * splice_from_pipe - splice data from a pipe to a file
947 * @pipe: pipe to splice from
948 * @out: file to splice to
949 * @ppos: position in @out
950 * @len: how many bytes to splice
951 * @flags: splice modifier flags
952 * @actor: handler that splices the data
955 * See __splice_from_pipe. This function locks the pipe inode,
956 * otherwise it's identical to __splice_from_pipe().
959 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
960 loff_t *ppos, size_t len, unsigned int flags,
964 struct splice_desc sd = {
972 ret = __splice_from_pipe(pipe, &sd, actor);
979 * generic_file_splice_write - splice data from a pipe to a file
981 * @out: file to write to
982 * @ppos: position in @out
983 * @len: number of bytes to splice
984 * @flags: splice modifier flags
987 * Will either move or copy pages (determined by @flags options) from
988 * the given pipe inode to the given file.
992 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
993 loff_t *ppos, size_t len, unsigned int flags)
995 struct address_space *mapping = out->f_mapping;
996 struct inode *inode = mapping->host;
997 struct splice_desc sd = {
1007 splice_from_pipe_begin(&sd);
1009 ret = splice_from_pipe_next(pipe, &sd);
1013 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1014 ret = file_remove_suid(out);
1016 file_update_time(out);
1017 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1019 mutex_unlock(&inode->i_mutex);
1021 splice_from_pipe_end(pipe, &sd);
1026 ret = sd.num_spliced;
1029 unsigned long nr_pages;
1032 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1034 err = generic_write_sync(out, *ppos, ret);
1039 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1045 EXPORT_SYMBOL(generic_file_splice_write);
1047 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1048 struct splice_desc *sd)
1053 data = buf->ops->map(pipe, buf, 0);
1054 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1055 buf->ops->unmap(pipe, buf, data);
1060 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1061 struct file *out, loff_t *ppos,
1062 size_t len, unsigned int flags)
1066 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1074 * generic_splice_sendpage - splice data from a pipe to a socket
1075 * @pipe: pipe to splice from
1076 * @out: socket to write to
1077 * @ppos: position in @out
1078 * @len: number of bytes to splice
1079 * @flags: splice modifier flags
1082 * Will send @len bytes from the pipe to a network socket. No data copying
1086 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1087 loff_t *ppos, size_t len, unsigned int flags)
1089 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1092 EXPORT_SYMBOL(generic_splice_sendpage);
1095 * Attempt to initiate a splice from pipe to file.
1097 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1098 loff_t *ppos, size_t len, unsigned int flags)
1100 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1101 loff_t *, size_t, unsigned int);
1104 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1107 if (unlikely(out->f_flags & O_APPEND))
1110 ret = rw_verify_area(WRITE, out, ppos, len);
1111 if (unlikely(ret < 0))
1114 if (out->f_op && out->f_op->splice_write)
1115 splice_write = out->f_op->splice_write;
1117 splice_write = default_file_splice_write;
1119 return splice_write(pipe, out, ppos, len, flags);
1123 * Attempt to initiate a splice from a file to a pipe.
1125 static long do_splice_to(struct file *in, loff_t *ppos,
1126 struct pipe_inode_info *pipe, size_t len,
1129 ssize_t (*splice_read)(struct file *, loff_t *,
1130 struct pipe_inode_info *, size_t, unsigned int);
1133 if (unlikely(!(in->f_mode & FMODE_READ)))
1136 ret = rw_verify_area(READ, in, ppos, len);
1137 if (unlikely(ret < 0))
1140 if (in->f_op && in->f_op->splice_read)
1141 splice_read = in->f_op->splice_read;
1143 splice_read = default_file_splice_read;
1145 return splice_read(in, ppos, pipe, len, flags);
1149 * splice_direct_to_actor - splices data directly between two non-pipes
1150 * @in: file to splice from
1151 * @sd: actor information on where to splice to
1152 * @actor: handles the data splicing
1155 * This is a special case helper to splice directly between two
1156 * points, without requiring an explicit pipe. Internally an allocated
1157 * pipe is cached in the process, and reused during the lifetime of
1161 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1162 splice_direct_actor *actor)
1164 struct pipe_inode_info *pipe;
1171 * We require the input being a regular file, as we don't want to
1172 * randomly drop data for eg socket -> socket splicing. Use the
1173 * piped splicing for that!
1175 i_mode = in->f_path.dentry->d_inode->i_mode;
1176 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1180 * neither in nor out is a pipe, setup an internal pipe attached to
1181 * 'out' and transfer the wanted data from 'in' to 'out' through that
1183 pipe = current->splice_pipe;
1184 if (unlikely(!pipe)) {
1185 pipe = alloc_pipe_info(NULL);
1190 * We don't have an immediate reader, but we'll read the stuff
1191 * out of the pipe right after the splice_to_pipe(). So set
1192 * PIPE_READERS appropriately.
1196 current->splice_pipe = pipe;
1204 len = sd->total_len;
1208 * Don't block on output, we have to drain the direct pipe.
1210 sd->flags &= ~SPLICE_F_NONBLOCK;
1214 loff_t pos = sd->pos, prev_pos = pos;
1216 ret = do_splice_to(in, &pos, pipe, len, flags);
1217 if (unlikely(ret <= 0))
1221 sd->total_len = read_len;
1224 * NOTE: nonblocking mode only applies to the input. We
1225 * must not do the output in nonblocking mode as then we
1226 * could get stuck data in the internal pipe:
1228 ret = actor(pipe, sd);
1229 if (unlikely(ret <= 0)) {
1238 if (ret < read_len) {
1239 sd->pos = prev_pos + ret;
1245 pipe->nrbufs = pipe->curbuf = 0;
1251 * If we did an incomplete transfer we must release
1252 * the pipe buffers in question:
1254 for (i = 0; i < pipe->buffers; i++) {
1255 struct pipe_buffer *buf = pipe->bufs + i;
1258 buf->ops->release(pipe, buf);
1268 EXPORT_SYMBOL(splice_direct_to_actor);
1270 static int direct_splice_actor(struct pipe_inode_info *pipe,
1271 struct splice_desc *sd)
1273 struct file *file = sd->u.file;
1275 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1280 * do_splice_direct - splices data directly between two files
1281 * @in: file to splice from
1282 * @ppos: input file offset
1283 * @out: file to splice to
1284 * @len: number of bytes to splice
1285 * @flags: splice modifier flags
1288 * For use by do_sendfile(). splice can easily emulate sendfile, but
1289 * doing it in the application would incur an extra system call
1290 * (splice in + splice out, as compared to just sendfile()). So this helper
1291 * can splice directly through a process-private pipe.
1294 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1295 size_t len, unsigned int flags)
1297 struct splice_desc sd = {
1306 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1313 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1314 struct pipe_inode_info *opipe,
1315 size_t len, unsigned int flags);
1318 * Determine where to splice to/from.
1320 static long do_splice(struct file *in, loff_t __user *off_in,
1321 struct file *out, loff_t __user *off_out,
1322 size_t len, unsigned int flags)
1324 struct pipe_inode_info *ipipe;
1325 struct pipe_inode_info *opipe;
1326 loff_t offset, *off;
1329 ipipe = get_pipe_info(in);
1330 opipe = get_pipe_info(out);
1332 if (ipipe && opipe) {
1333 if (off_in || off_out)
1336 if (!(in->f_mode & FMODE_READ))
1339 if (!(out->f_mode & FMODE_WRITE))
1342 /* Splicing to self would be fun, but... */
1346 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1353 if (!(out->f_mode & FMODE_PWRITE))
1355 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1361 ret = do_splice_from(ipipe, out, off, len, flags);
1363 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1373 if (!(in->f_mode & FMODE_PREAD))
1375 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1381 ret = do_splice_to(in, off, opipe, len, flags);
1383 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1393 * Map an iov into an array of pages and offset/length tupples. With the
1394 * partial_page structure, we can map several non-contiguous ranges into
1395 * our ones pages[] map instead of splitting that operation into pieces.
1396 * Could easily be exported as a generic helper for other users, in which
1397 * case one would probably want to add a 'max_nr_pages' parameter as well.
1399 static int get_iovec_page_array(const struct iovec __user *iov,
1400 unsigned int nr_vecs, struct page **pages,
1401 struct partial_page *partial, int aligned,
1402 unsigned int pipe_buffers)
1404 int buffers = 0, error = 0;
1407 unsigned long off, npages;
1414 if (copy_from_user(&entry, iov, sizeof(entry)))
1417 base = entry.iov_base;
1418 len = entry.iov_len;
1421 * Sanity check this iovec. 0 read succeeds.
1427 if (!access_ok(VERIFY_READ, base, len))
1431 * Get this base offset and number of pages, then map
1432 * in the user pages.
1434 off = (unsigned long) base & ~PAGE_MASK;
1437 * If asked for alignment, the offset must be zero and the
1438 * length a multiple of the PAGE_SIZE.
1441 if (aligned && (off || len & ~PAGE_MASK))
1444 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1445 if (npages > pipe_buffers - buffers)
1446 npages = pipe_buffers - buffers;
1448 error = get_user_pages_fast((unsigned long)base, npages,
1449 0, &pages[buffers]);
1451 if (unlikely(error <= 0))
1455 * Fill this contiguous range into the partial page map.
1457 for (i = 0; i < error; i++) {
1458 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1460 partial[buffers].offset = off;
1461 partial[buffers].len = plen;
1469 * We didn't complete this iov, stop here since it probably
1470 * means we have to move some of this into a pipe to
1471 * be able to continue.
1477 * Don't continue if we mapped fewer pages than we asked for,
1478 * or if we mapped the max number of pages that we have
1481 if (error < npages || buffers == pipe_buffers)
1494 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1495 struct splice_desc *sd)
1501 * See if we can use the atomic maps, by prefaulting in the
1502 * pages and doing an atomic copy
1504 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1505 src = buf->ops->map(pipe, buf, 1);
1506 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1508 buf->ops->unmap(pipe, buf, src);
1516 * No dice, use slow non-atomic map and copy
1518 src = buf->ops->map(pipe, buf, 0);
1521 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1524 buf->ops->unmap(pipe, buf, src);
1527 sd->u.userptr += ret;
1532 * For lack of a better implementation, implement vmsplice() to userspace
1533 * as a simple copy of the pipes pages to the user iov.
1535 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1536 unsigned long nr_segs, unsigned int flags)
1538 struct pipe_inode_info *pipe;
1539 struct splice_desc sd;
1544 pipe = get_pipe_info(file);
1556 * Get user address base and length for this iovec.
1558 error = get_user(base, &iov->iov_base);
1559 if (unlikely(error))
1561 error = get_user(len, &iov->iov_len);
1562 if (unlikely(error))
1566 * Sanity check this iovec. 0 read succeeds.
1570 if (unlikely(!base)) {
1575 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1583 sd.u.userptr = base;
1586 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1612 * vmsplice splices a user address range into a pipe. It can be thought of
1613 * as splice-from-memory, where the regular splice is splice-from-file (or
1614 * to file). In both cases the output is a pipe, naturally.
1616 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1617 unsigned long nr_segs, unsigned int flags)
1619 struct pipe_inode_info *pipe;
1620 struct page *pages[PIPE_DEF_BUFFERS];
1621 struct partial_page partial[PIPE_DEF_BUFFERS];
1622 struct splice_pipe_desc spd = {
1625 .nr_pages_max = PIPE_DEF_BUFFERS,
1627 .ops = &user_page_pipe_buf_ops,
1628 .spd_release = spd_release_page,
1632 pipe = get_pipe_info(file);
1636 if (splice_grow_spd(pipe, &spd))
1639 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1640 spd.partial, flags & SPLICE_F_GIFT,
1642 if (spd.nr_pages <= 0)
1645 ret = splice_to_pipe(pipe, &spd);
1647 splice_shrink_spd(&spd);
1652 * Note that vmsplice only really supports true splicing _from_ user memory
1653 * to a pipe, not the other way around. Splicing from user memory is a simple
1654 * operation that can be supported without any funky alignment restrictions
1655 * or nasty vm tricks. We simply map in the user memory and fill them into
1656 * a pipe. The reverse isn't quite as easy, though. There are two possible
1657 * solutions for that:
1659 * - memcpy() the data internally, at which point we might as well just
1660 * do a regular read() on the buffer anyway.
1661 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1662 * has restriction limitations on both ends of the pipe).
1664 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1667 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1668 unsigned long, nr_segs, unsigned int, flags)
1674 if (unlikely(nr_segs > UIO_MAXIOV))
1676 else if (unlikely(!nr_segs))
1680 file = fget_light(fd, &fput);
1682 if (file->f_mode & FMODE_WRITE)
1683 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1684 else if (file->f_mode & FMODE_READ)
1685 error = vmsplice_to_user(file, iov, nr_segs, flags);
1687 fput_light(file, fput);
1693 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1694 int, fd_out, loff_t __user *, off_out,
1695 size_t, len, unsigned int, flags)
1698 struct file *in, *out;
1699 int fput_in, fput_out;
1705 in = fget_light(fd_in, &fput_in);
1707 if (in->f_mode & FMODE_READ) {
1708 out = fget_light(fd_out, &fput_out);
1710 if (out->f_mode & FMODE_WRITE)
1711 error = do_splice(in, off_in,
1714 fput_light(out, fput_out);
1718 fput_light(in, fput_in);
1725 * Make sure there's data to read. Wait for input if we can, otherwise
1726 * return an appropriate error.
1728 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1733 * Check ->nrbufs without the inode lock first. This function
1734 * is speculative anyways, so missing one is ok.
1742 while (!pipe->nrbufs) {
1743 if (signal_pending(current)) {
1749 if (!pipe->waiting_writers) {
1750 if (flags & SPLICE_F_NONBLOCK) {
1763 * Make sure there's writeable room. Wait for room if we can, otherwise
1764 * return an appropriate error.
1766 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1771 * Check ->nrbufs without the inode lock first. This function
1772 * is speculative anyways, so missing one is ok.
1774 if (pipe->nrbufs < pipe->buffers)
1780 while (pipe->nrbufs >= pipe->buffers) {
1781 if (!pipe->readers) {
1782 send_sig(SIGPIPE, current, 0);
1786 if (flags & SPLICE_F_NONBLOCK) {
1790 if (signal_pending(current)) {
1794 pipe->waiting_writers++;
1796 pipe->waiting_writers--;
1804 * Splice contents of ipipe to opipe.
1806 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1807 struct pipe_inode_info *opipe,
1808 size_t len, unsigned int flags)
1810 struct pipe_buffer *ibuf, *obuf;
1812 bool input_wakeup = false;
1816 ret = ipipe_prep(ipipe, flags);
1820 ret = opipe_prep(opipe, flags);
1825 * Potential ABBA deadlock, work around it by ordering lock
1826 * grabbing by pipe info address. Otherwise two different processes
1827 * could deadlock (one doing tee from A -> B, the other from B -> A).
1829 pipe_double_lock(ipipe, opipe);
1832 if (!opipe->readers) {
1833 send_sig(SIGPIPE, current, 0);
1839 if (!ipipe->nrbufs && !ipipe->writers)
1843 * Cannot make any progress, because either the input
1844 * pipe is empty or the output pipe is full.
1846 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1847 /* Already processed some buffers, break */
1851 if (flags & SPLICE_F_NONBLOCK) {
1857 * We raced with another reader/writer and haven't
1858 * managed to process any buffers. A zero return
1859 * value means EOF, so retry instead.
1866 ibuf = ipipe->bufs + ipipe->curbuf;
1867 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1868 obuf = opipe->bufs + nbuf;
1870 if (len >= ibuf->len) {
1872 * Simply move the whole buffer from ipipe to opipe
1877 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1879 input_wakeup = true;
1882 * Get a reference to this pipe buffer,
1883 * so we can copy the contents over.
1885 ibuf->ops->get(ipipe, ibuf);
1889 * Don't inherit the gift flag, we need to
1890 * prevent multiple steals of this page.
1892 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1896 ibuf->offset += obuf->len;
1897 ibuf->len -= obuf->len;
1907 * If we put data in the output pipe, wakeup any potential readers.
1910 wakeup_pipe_readers(opipe);
1913 wakeup_pipe_writers(ipipe);
1919 * Link contents of ipipe to opipe.
1921 static int link_pipe(struct pipe_inode_info *ipipe,
1922 struct pipe_inode_info *opipe,
1923 size_t len, unsigned int flags)
1925 struct pipe_buffer *ibuf, *obuf;
1926 int ret = 0, i = 0, nbuf;
1929 * Potential ABBA deadlock, work around it by ordering lock
1930 * grabbing by pipe info address. Otherwise two different processes
1931 * could deadlock (one doing tee from A -> B, the other from B -> A).
1933 pipe_double_lock(ipipe, opipe);
1936 if (!opipe->readers) {
1937 send_sig(SIGPIPE, current, 0);
1944 * If we have iterated all input buffers or ran out of
1945 * output room, break.
1947 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1950 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1951 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1954 * Get a reference to this pipe buffer,
1955 * so we can copy the contents over.
1957 ibuf->ops->get(ipipe, ibuf);
1959 obuf = opipe->bufs + nbuf;
1963 * Don't inherit the gift flag, we need to
1964 * prevent multiple steals of this page.
1966 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1968 if (obuf->len > len)
1978 * return EAGAIN if we have the potential of some data in the
1979 * future, otherwise just return 0
1981 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1988 * If we put data in the output pipe, wakeup any potential readers.
1991 wakeup_pipe_readers(opipe);
1997 * This is a tee(1) implementation that works on pipes. It doesn't copy
1998 * any data, it simply references the 'in' pages on the 'out' pipe.
1999 * The 'flags' used are the SPLICE_F_* variants, currently the only
2000 * applicable one is SPLICE_F_NONBLOCK.
2002 static long do_tee(struct file *in, struct file *out, size_t len,
2005 struct pipe_inode_info *ipipe = get_pipe_info(in);
2006 struct pipe_inode_info *opipe = get_pipe_info(out);
2010 * Duplicate the contents of ipipe to opipe without actually
2013 if (ipipe && opipe && ipipe != opipe) {
2015 * Keep going, unless we encounter an error. The ipipe/opipe
2016 * ordering doesn't really matter.
2018 ret = ipipe_prep(ipipe, flags);
2020 ret = opipe_prep(opipe, flags);
2022 ret = link_pipe(ipipe, opipe, len, flags);
2029 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2038 in = fget_light(fdin, &fput_in);
2040 if (in->f_mode & FMODE_READ) {
2042 struct file *out = fget_light(fdout, &fput_out);
2045 if (out->f_mode & FMODE_WRITE)
2046 error = do_tee(in, out, len, flags);
2047 fput_light(out, fput_out);
2050 fput_light(in, fput_in);