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>
36 * Attempt to steal a page from a pipe buffer. This should perhaps go into
37 * a vm helper function, it's already simplified quite a bit by the
38 * addition of remove_mapping(). If success is returned, the caller may
39 * attempt to reuse this page for another destination.
41 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
42 struct pipe_buffer *buf)
44 struct page *page = buf->page;
45 struct address_space *mapping;
49 mapping = page_mapping(page);
51 WARN_ON(!PageUptodate(page));
54 * At least for ext2 with nobh option, we need to wait on
55 * writeback completing on this page, since we'll remove it
56 * from the pagecache. Otherwise truncate wont wait on the
57 * page, allowing the disk blocks to be reused by someone else
58 * before we actually wrote our data to them. fs corruption
61 wait_on_page_writeback(page);
63 if (page_has_private(page) &&
64 !try_to_release_page(page, GFP_KERNEL))
68 * If we succeeded in removing the mapping, set LRU flag
71 if (remove_mapping(mapping, page)) {
72 buf->flags |= PIPE_BUF_FLAG_LRU;
78 * Raced with truncate or failed to remove page from current
79 * address space, unlock and return failure.
86 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
87 struct pipe_buffer *buf)
89 page_cache_release(buf->page);
90 buf->flags &= ~PIPE_BUF_FLAG_LRU;
94 * Check whether the contents of buf is OK to access. Since the content
95 * is a page cache page, IO may be in flight.
97 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
98 struct pipe_buffer *buf)
100 struct page *page = buf->page;
103 if (!PageUptodate(page)) {
107 * Page got truncated/unhashed. This will cause a 0-byte
108 * splice, if this is the first page.
110 if (!page->mapping) {
116 * Uh oh, read-error from disk.
118 if (!PageUptodate(page)) {
124 * Page is ok afterall, we are done.
135 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
137 .map = generic_pipe_buf_map,
138 .unmap = generic_pipe_buf_unmap,
139 .confirm = page_cache_pipe_buf_confirm,
140 .release = page_cache_pipe_buf_release,
141 .steal = page_cache_pipe_buf_steal,
142 .get = generic_pipe_buf_get,
145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
146 struct pipe_buffer *buf)
148 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
151 buf->flags |= PIPE_BUF_FLAG_LRU;
152 return generic_pipe_buf_steal(pipe, buf);
155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
157 .map = generic_pipe_buf_map,
158 .unmap = generic_pipe_buf_unmap,
159 .confirm = generic_pipe_buf_confirm,
160 .release = page_cache_pipe_buf_release,
161 .steal = user_page_pipe_buf_steal,
162 .get = generic_pipe_buf_get,
166 * splice_to_pipe - fill passed data into a pipe
167 * @pipe: pipe to fill
171 * @spd contains a map of pages and len/offset tuples, along with
172 * the struct pipe_buf_operations associated with these pages. This
173 * function will link that data to the pipe.
176 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
177 struct splice_pipe_desc *spd)
179 unsigned int spd_pages = spd->nr_pages;
180 int ret, do_wakeup, page_nr;
189 if (!pipe->readers) {
190 send_sig(SIGPIPE, current, 0);
196 if (pipe->nrbufs < pipe->buffers) {
197 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
198 struct pipe_buffer *buf = pipe->bufs + newbuf;
200 buf->page = spd->pages[page_nr];
201 buf->offset = spd->partial[page_nr].offset;
202 buf->len = spd->partial[page_nr].len;
203 buf->private = spd->partial[page_nr].private;
205 if (spd->flags & SPLICE_F_GIFT)
206 buf->flags |= PIPE_BUF_FLAG_GIFT;
215 if (!--spd->nr_pages)
217 if (pipe->nrbufs < pipe->buffers)
223 if (spd->flags & SPLICE_F_NONBLOCK) {
229 if (signal_pending(current)) {
237 if (waitqueue_active(&pipe->wait))
238 wake_up_interruptible_sync(&pipe->wait);
239 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
243 pipe->waiting_writers++;
245 pipe->waiting_writers--;
252 if (waitqueue_active(&pipe->wait))
253 wake_up_interruptible(&pipe->wait);
254 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
257 while (page_nr < spd_pages)
258 spd->spd_release(spd, page_nr++);
263 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
265 page_cache_release(spd->pages[i]);
269 * Check if we need to grow the arrays holding pages and partial page
272 int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
274 if (pipe->buffers <= PIPE_DEF_BUFFERS)
277 spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
278 spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
280 if (spd->pages && spd->partial)
288 void splice_shrink_spd(struct pipe_inode_info *pipe,
289 struct splice_pipe_desc *spd)
291 if (pipe->buffers <= PIPE_DEF_BUFFERS)
299 __generic_file_splice_read(struct file *in, loff_t *ppos,
300 struct pipe_inode_info *pipe, size_t len,
303 struct address_space *mapping = in->f_mapping;
304 unsigned int loff, nr_pages, req_pages;
305 struct page *pages[PIPE_DEF_BUFFERS];
306 struct partial_page partial[PIPE_DEF_BUFFERS];
308 pgoff_t index, end_index;
311 struct splice_pipe_desc spd = {
315 .ops = &page_cache_pipe_buf_ops,
316 .spd_release = spd_release_page,
319 if (splice_grow_spd(pipe, &spd))
322 index = *ppos >> PAGE_CACHE_SHIFT;
323 loff = *ppos & ~PAGE_CACHE_MASK;
324 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
325 nr_pages = min(req_pages, pipe->buffers);
328 * Lookup the (hopefully) full range of pages we need.
330 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
331 index += spd.nr_pages;
334 * If find_get_pages_contig() returned fewer pages than we needed,
335 * readahead/allocate the rest and fill in the holes.
337 if (spd.nr_pages < nr_pages)
338 page_cache_sync_readahead(mapping, &in->f_ra, in,
339 index, req_pages - spd.nr_pages);
342 while (spd.nr_pages < nr_pages) {
344 * Page could be there, find_get_pages_contig() breaks on
347 page = find_get_page(mapping, index);
350 * page didn't exist, allocate one.
352 page = page_cache_alloc_cold(mapping);
356 error = add_to_page_cache_lru(page, mapping, index,
358 if (unlikely(error)) {
359 page_cache_release(page);
360 if (error == -EEXIST)
365 * add_to_page_cache() locks the page, unlock it
366 * to avoid convoluting the logic below even more.
371 spd.pages[spd.nr_pages++] = page;
376 * Now loop over the map and see if we need to start IO on any
377 * pages, fill in the partial map, etc.
379 index = *ppos >> PAGE_CACHE_SHIFT;
380 nr_pages = spd.nr_pages;
382 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
383 unsigned int this_len;
389 * this_len is the max we'll use from this page
391 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
392 page = spd.pages[page_nr];
394 if (PageReadahead(page))
395 page_cache_async_readahead(mapping, &in->f_ra, in,
396 page, index, req_pages - page_nr);
399 * If the page isn't uptodate, we may need to start io on it
401 if (!PageUptodate(page)) {
403 * If in nonblock mode then dont block on waiting
404 * for an in-flight io page
406 if (flags & SPLICE_F_NONBLOCK) {
407 if (!trylock_page(page)) {
415 * Page was truncated, or invalidated by the
416 * filesystem. Redo the find/create, but this time the
417 * page is kept locked, so there's no chance of another
418 * race with truncate/invalidate.
420 if (!page->mapping) {
422 page = find_or_create_page(mapping, index,
423 mapping_gfp_mask(mapping));
429 page_cache_release(spd.pages[page_nr]);
430 spd.pages[page_nr] = page;
433 * page was already under io and is now done, great
435 if (PageUptodate(page)) {
441 * need to read in the page
443 error = mapping->a_ops->readpage(in, page);
444 if (unlikely(error)) {
446 * We really should re-lookup the page here,
447 * but it complicates things a lot. Instead
448 * lets just do what we already stored, and
449 * we'll get it the next time we are called.
451 if (error == AOP_TRUNCATED_PAGE)
459 * i_size must be checked after PageUptodate.
461 isize = i_size_read(mapping->host);
462 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
463 if (unlikely(!isize || index > end_index))
467 * if this is the last page, see if we need to shrink
468 * the length and stop
470 if (end_index == index) {
474 * max good bytes in this page
476 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
481 * force quit after adding this page
483 this_len = min(this_len, plen - loff);
487 spd.partial[page_nr].offset = loff;
488 spd.partial[page_nr].len = this_len;
496 * Release any pages at the end, if we quit early. 'page_nr' is how far
497 * we got, 'nr_pages' is how many pages are in the map.
499 while (page_nr < nr_pages)
500 page_cache_release(spd.pages[page_nr++]);
501 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
504 error = splice_to_pipe(pipe, &spd);
506 splice_shrink_spd(pipe, &spd);
511 * generic_file_splice_read - splice data from file to a pipe
512 * @in: file to splice from
513 * @ppos: position in @in
514 * @pipe: pipe to splice to
515 * @len: number of bytes to splice
516 * @flags: splice modifier flags
519 * Will read pages from given file and fill them into a pipe. Can be
520 * used as long as the address_space operations for the source implements
524 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
525 struct pipe_inode_info *pipe, size_t len,
531 isize = i_size_read(in->f_mapping->host);
532 if (unlikely(*ppos >= isize))
535 left = isize - *ppos;
536 if (unlikely(left < len))
539 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
547 EXPORT_SYMBOL(generic_file_splice_read);
549 static const struct pipe_buf_operations default_pipe_buf_ops = {
551 .map = generic_pipe_buf_map,
552 .unmap = generic_pipe_buf_unmap,
553 .confirm = generic_pipe_buf_confirm,
554 .release = generic_pipe_buf_release,
555 .steal = generic_pipe_buf_steal,
556 .get = generic_pipe_buf_get,
559 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
560 unsigned long vlen, loff_t offset)
568 /* The cast to a user pointer is valid due to the set_fs() */
569 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
575 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
583 /* The cast to a user pointer is valid due to the set_fs() */
584 res = vfs_write(file, (const char __user *)buf, count, &pos);
590 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
591 struct pipe_inode_info *pipe, size_t len,
594 unsigned int nr_pages;
595 unsigned int nr_freed;
597 struct page *pages[PIPE_DEF_BUFFERS];
598 struct partial_page partial[PIPE_DEF_BUFFERS];
599 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
604 struct splice_pipe_desc spd = {
608 .ops = &default_pipe_buf_ops,
609 .spd_release = spd_release_page,
612 if (splice_grow_spd(pipe, &spd))
617 if (pipe->buffers > PIPE_DEF_BUFFERS) {
618 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
623 offset = *ppos & ~PAGE_CACHE_MASK;
624 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
626 for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
629 page = alloc_page(GFP_USER);
634 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
635 vec[i].iov_base = (void __user *) page_address(page);
636 vec[i].iov_len = this_len;
643 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
654 for (i = 0; i < spd.nr_pages; i++) {
655 this_len = min_t(size_t, vec[i].iov_len, res);
656 spd.partial[i].offset = 0;
657 spd.partial[i].len = this_len;
659 __free_page(spd.pages[i]);
665 spd.nr_pages -= nr_freed;
667 res = splice_to_pipe(pipe, &spd);
674 splice_shrink_spd(pipe, &spd);
678 for (i = 0; i < spd.nr_pages; i++)
679 __free_page(spd.pages[i]);
684 EXPORT_SYMBOL(default_file_splice_read);
687 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
688 * using sendpage(). Return the number of bytes sent.
690 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
691 struct pipe_buffer *buf, struct splice_desc *sd)
693 struct file *file = sd->u.file;
694 loff_t pos = sd->pos;
697 ret = buf->ops->confirm(pipe, buf);
699 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
700 if (file->f_op && file->f_op->sendpage)
701 ret = file->f_op->sendpage(file, buf->page, buf->offset,
702 sd->len, &pos, more);
711 * This is a little more tricky than the file -> pipe splicing. There are
712 * basically three cases:
714 * - Destination page already exists in the address space and there
715 * are users of it. For that case we have no other option that
716 * copying the data. Tough luck.
717 * - Destination page already exists in the address space, but there
718 * are no users of it. Make sure it's uptodate, then drop it. Fall
719 * through to last case.
720 * - Destination page does not exist, we can add the pipe page to
721 * the page cache and avoid the copy.
723 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
724 * sd->flags), we attempt to migrate pages from the pipe to the output
725 * file address space page cache. This is possible if no one else has
726 * the pipe page referenced outside of the pipe and page cache. If
727 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
728 * a new page in the output file page cache and fill/dirty that.
730 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
731 struct splice_desc *sd)
733 struct file *file = sd->u.file;
734 struct address_space *mapping = file->f_mapping;
735 unsigned int offset, this_len;
741 * make sure the data in this buffer is uptodate
743 ret = buf->ops->confirm(pipe, buf);
747 offset = sd->pos & ~PAGE_CACHE_MASK;
750 if (this_len + offset > PAGE_CACHE_SIZE)
751 this_len = PAGE_CACHE_SIZE - offset;
753 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
754 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
758 if (buf->page != page) {
760 * Careful, ->map() uses KM_USER0!
762 char *src = buf->ops->map(pipe, buf, 1);
763 char *dst = kmap_atomic(page, KM_USER1);
765 memcpy(dst + offset, src + buf->offset, this_len);
766 flush_dcache_page(page);
767 kunmap_atomic(dst, KM_USER1);
768 buf->ops->unmap(pipe, buf, src);
770 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
775 EXPORT_SYMBOL(pipe_to_file);
777 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
780 if (waitqueue_active(&pipe->wait))
781 wake_up_interruptible(&pipe->wait);
782 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
786 * splice_from_pipe_feed - feed available data from a pipe to a file
787 * @pipe: pipe to splice from
788 * @sd: information to @actor
789 * @actor: handler that splices the data
792 * This function loops over the pipe and calls @actor to do the
793 * actual moving of a single struct pipe_buffer to the desired
794 * destination. It returns when there's no more buffers left in
795 * the pipe or if the requested number of bytes (@sd->total_len)
796 * have been copied. It returns a positive number (one) if the
797 * pipe needs to be filled with more data, zero if the required
798 * number of bytes have been copied and -errno on error.
800 * This, together with splice_from_pipe_{begin,end,next}, may be
801 * used to implement the functionality of __splice_from_pipe() when
802 * locking is required around copying the pipe buffers to the
805 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
810 while (pipe->nrbufs) {
811 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
812 const struct pipe_buf_operations *ops = buf->ops;
815 if (sd->len > sd->total_len)
816 sd->len = sd->total_len;
818 ret = actor(pipe, buf, sd);
827 sd->num_spliced += ret;
830 sd->total_len -= ret;
834 ops->release(pipe, buf);
835 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
838 sd->need_wakeup = true;
847 EXPORT_SYMBOL(splice_from_pipe_feed);
850 * splice_from_pipe_next - wait for some data to splice from
851 * @pipe: pipe to splice from
852 * @sd: information about the splice operation
855 * This function will wait for some data and return a positive
856 * value (one) if pipe buffers are available. It will return zero
857 * or -errno if no more data needs to be spliced.
859 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
861 while (!pipe->nrbufs) {
865 if (!pipe->waiting_writers && sd->num_spliced)
868 if (sd->flags & SPLICE_F_NONBLOCK)
871 if (signal_pending(current))
874 if (sd->need_wakeup) {
875 wakeup_pipe_writers(pipe);
876 sd->need_wakeup = false;
884 EXPORT_SYMBOL(splice_from_pipe_next);
887 * splice_from_pipe_begin - start splicing from pipe
888 * @sd: information about the splice operation
891 * This function should be called before a loop containing
892 * splice_from_pipe_next() and splice_from_pipe_feed() to
893 * initialize the necessary fields of @sd.
895 void splice_from_pipe_begin(struct splice_desc *sd)
898 sd->need_wakeup = false;
900 EXPORT_SYMBOL(splice_from_pipe_begin);
903 * splice_from_pipe_end - finish splicing from pipe
904 * @pipe: pipe to splice from
905 * @sd: information about the splice operation
908 * This function will wake up pipe writers if necessary. It should
909 * be called after a loop containing splice_from_pipe_next() and
910 * splice_from_pipe_feed().
912 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
915 wakeup_pipe_writers(pipe);
917 EXPORT_SYMBOL(splice_from_pipe_end);
920 * __splice_from_pipe - splice data from a pipe to given actor
921 * @pipe: pipe to splice from
922 * @sd: information to @actor
923 * @actor: handler that splices the data
926 * This function does little more than loop over the pipe and call
927 * @actor to do the actual moving of a single struct pipe_buffer to
928 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
932 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
937 splice_from_pipe_begin(sd);
939 ret = splice_from_pipe_next(pipe, sd);
941 ret = splice_from_pipe_feed(pipe, sd, actor);
943 splice_from_pipe_end(pipe, sd);
945 return sd->num_spliced ? sd->num_spliced : ret;
947 EXPORT_SYMBOL(__splice_from_pipe);
950 * splice_from_pipe - splice data from a pipe to a file
951 * @pipe: pipe to splice from
952 * @out: file to splice to
953 * @ppos: position in @out
954 * @len: how many bytes to splice
955 * @flags: splice modifier flags
956 * @actor: handler that splices the data
959 * See __splice_from_pipe. This function locks the pipe inode,
960 * otherwise it's identical to __splice_from_pipe().
963 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
964 loff_t *ppos, size_t len, unsigned int flags,
968 struct splice_desc sd = {
976 ret = __splice_from_pipe(pipe, &sd, actor);
983 * generic_file_splice_write - splice data from a pipe to a file
985 * @out: file to write to
986 * @ppos: position in @out
987 * @len: number of bytes to splice
988 * @flags: splice modifier flags
991 * Will either move or copy pages (determined by @flags options) from
992 * the given pipe inode to the given file.
996 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
997 loff_t *ppos, size_t len, unsigned int flags)
999 struct address_space *mapping = out->f_mapping;
1000 struct inode *inode = mapping->host;
1001 struct splice_desc sd = {
1011 splice_from_pipe_begin(&sd);
1013 ret = splice_from_pipe_next(pipe, &sd);
1017 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1018 ret = file_remove_suid(out);
1020 file_update_time(out);
1021 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1023 mutex_unlock(&inode->i_mutex);
1025 splice_from_pipe_end(pipe, &sd);
1030 ret = sd.num_spliced;
1033 unsigned long nr_pages;
1036 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1038 err = generic_write_sync(out, *ppos, ret);
1043 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1049 EXPORT_SYMBOL(generic_file_splice_write);
1051 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1052 struct splice_desc *sd)
1057 ret = buf->ops->confirm(pipe, buf);
1061 data = buf->ops->map(pipe, buf, 0);
1062 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1063 buf->ops->unmap(pipe, buf, data);
1068 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1069 struct file *out, loff_t *ppos,
1070 size_t len, unsigned int flags)
1074 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1082 * generic_splice_sendpage - splice data from a pipe to a socket
1083 * @pipe: pipe to splice from
1084 * @out: socket to write to
1085 * @ppos: position in @out
1086 * @len: number of bytes to splice
1087 * @flags: splice modifier flags
1090 * Will send @len bytes from the pipe to a network socket. No data copying
1094 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1095 loff_t *ppos, size_t len, unsigned int flags)
1097 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1100 EXPORT_SYMBOL(generic_splice_sendpage);
1103 * Attempt to initiate a splice from pipe to file.
1105 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1106 loff_t *ppos, size_t len, unsigned int flags)
1108 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1109 loff_t *, size_t, unsigned int);
1112 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1115 if (unlikely(out->f_flags & O_APPEND))
1118 ret = rw_verify_area(WRITE, out, ppos, len);
1119 if (unlikely(ret < 0))
1122 if (out->f_op && out->f_op->splice_write)
1123 splice_write = out->f_op->splice_write;
1125 splice_write = default_file_splice_write;
1127 return splice_write(pipe, out, ppos, len, flags);
1131 * Attempt to initiate a splice from a file to a pipe.
1133 static long do_splice_to(struct file *in, loff_t *ppos,
1134 struct pipe_inode_info *pipe, size_t len,
1137 ssize_t (*splice_read)(struct file *, loff_t *,
1138 struct pipe_inode_info *, size_t, unsigned int);
1141 if (unlikely(!(in->f_mode & FMODE_READ)))
1144 ret = rw_verify_area(READ, in, ppos, len);
1145 if (unlikely(ret < 0))
1148 if (in->f_op && in->f_op->splice_read)
1149 splice_read = in->f_op->splice_read;
1151 splice_read = default_file_splice_read;
1153 return splice_read(in, ppos, pipe, len, flags);
1157 * splice_direct_to_actor - splices data directly between two non-pipes
1158 * @in: file to splice from
1159 * @sd: actor information on where to splice to
1160 * @actor: handles the data splicing
1163 * This is a special case helper to splice directly between two
1164 * points, without requiring an explicit pipe. Internally an allocated
1165 * pipe is cached in the process, and reused during the lifetime of
1169 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1170 splice_direct_actor *actor)
1172 struct pipe_inode_info *pipe;
1179 * We require the input being a regular file, as we don't want to
1180 * randomly drop data for eg socket -> socket splicing. Use the
1181 * piped splicing for that!
1183 i_mode = in->f_path.dentry->d_inode->i_mode;
1184 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1188 * neither in nor out is a pipe, setup an internal pipe attached to
1189 * 'out' and transfer the wanted data from 'in' to 'out' through that
1191 pipe = current->splice_pipe;
1192 if (unlikely(!pipe)) {
1193 pipe = alloc_pipe_info(NULL);
1198 * We don't have an immediate reader, but we'll read the stuff
1199 * out of the pipe right after the splice_to_pipe(). So set
1200 * PIPE_READERS appropriately.
1204 current->splice_pipe = pipe;
1212 len = sd->total_len;
1216 * Don't block on output, we have to drain the direct pipe.
1218 sd->flags &= ~SPLICE_F_NONBLOCK;
1222 loff_t pos = sd->pos, prev_pos = pos;
1224 ret = do_splice_to(in, &pos, pipe, len, flags);
1225 if (unlikely(ret <= 0))
1229 sd->total_len = read_len;
1232 * NOTE: nonblocking mode only applies to the input. We
1233 * must not do the output in nonblocking mode as then we
1234 * could get stuck data in the internal pipe:
1236 ret = actor(pipe, sd);
1237 if (unlikely(ret <= 0)) {
1246 if (ret < read_len) {
1247 sd->pos = prev_pos + ret;
1253 pipe->nrbufs = pipe->curbuf = 0;
1259 * If we did an incomplete transfer we must release
1260 * the pipe buffers in question:
1262 for (i = 0; i < pipe->buffers; i++) {
1263 struct pipe_buffer *buf = pipe->bufs + i;
1266 buf->ops->release(pipe, buf);
1276 EXPORT_SYMBOL(splice_direct_to_actor);
1278 static int direct_splice_actor(struct pipe_inode_info *pipe,
1279 struct splice_desc *sd)
1281 struct file *file = sd->u.file;
1283 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1288 * do_splice_direct - splices data directly between two files
1289 * @in: file to splice from
1290 * @ppos: input file offset
1291 * @out: file to splice to
1292 * @len: number of bytes to splice
1293 * @flags: splice modifier flags
1296 * For use by do_sendfile(). splice can easily emulate sendfile, but
1297 * doing it in the application would incur an extra system call
1298 * (splice in + splice out, as compared to just sendfile()). So this helper
1299 * can splice directly through a process-private pipe.
1302 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1303 size_t len, unsigned int flags)
1305 struct splice_desc sd = {
1314 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1321 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1322 struct pipe_inode_info *opipe,
1323 size_t len, unsigned int flags);
1325 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1326 * location, so checking ->i_pipe is not enough to verify that this is a
1329 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1331 if (S_ISFIFO(inode->i_mode))
1332 return inode->i_pipe;
1338 * Determine where to splice to/from.
1340 static long do_splice(struct file *in, loff_t __user *off_in,
1341 struct file *out, loff_t __user *off_out,
1342 size_t len, unsigned int flags)
1344 struct pipe_inode_info *ipipe;
1345 struct pipe_inode_info *opipe;
1346 loff_t offset, *off;
1349 ipipe = pipe_info(in->f_path.dentry->d_inode);
1350 opipe = pipe_info(out->f_path.dentry->d_inode);
1352 if (ipipe && opipe) {
1353 if (off_in || off_out)
1356 if (!(in->f_mode & FMODE_READ))
1359 if (!(out->f_mode & FMODE_WRITE))
1362 /* Splicing to self would be fun, but... */
1366 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1373 if (!(out->f_mode & FMODE_PWRITE))
1375 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1381 ret = do_splice_from(ipipe, out, off, len, flags);
1383 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1393 if (!(in->f_mode & FMODE_PREAD))
1395 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1401 ret = do_splice_to(in, off, opipe, len, flags);
1403 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1413 * Map an iov into an array of pages and offset/length tupples. With the
1414 * partial_page structure, we can map several non-contiguous ranges into
1415 * our ones pages[] map instead of splitting that operation into pieces.
1416 * Could easily be exported as a generic helper for other users, in which
1417 * case one would probably want to add a 'max_nr_pages' parameter as well.
1419 static int get_iovec_page_array(const struct iovec __user *iov,
1420 unsigned int nr_vecs, struct page **pages,
1421 struct partial_page *partial, int aligned,
1422 unsigned int pipe_buffers)
1424 int buffers = 0, error = 0;
1427 unsigned long off, npages;
1434 if (copy_from_user(&entry, iov, sizeof(entry)))
1437 base = entry.iov_base;
1438 len = entry.iov_len;
1441 * Sanity check this iovec. 0 read succeeds.
1447 if (!access_ok(VERIFY_READ, base, len))
1451 * Get this base offset and number of pages, then map
1452 * in the user pages.
1454 off = (unsigned long) base & ~PAGE_MASK;
1457 * If asked for alignment, the offset must be zero and the
1458 * length a multiple of the PAGE_SIZE.
1461 if (aligned && (off || len & ~PAGE_MASK))
1464 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1465 if (npages > pipe_buffers - buffers)
1466 npages = pipe_buffers - buffers;
1468 error = get_user_pages_fast((unsigned long)base, npages,
1469 0, &pages[buffers]);
1471 if (unlikely(error <= 0))
1475 * Fill this contiguous range into the partial page map.
1477 for (i = 0; i < error; i++) {
1478 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1480 partial[buffers].offset = off;
1481 partial[buffers].len = plen;
1489 * We didn't complete this iov, stop here since it probably
1490 * means we have to move some of this into a pipe to
1491 * be able to continue.
1497 * Don't continue if we mapped fewer pages than we asked for,
1498 * or if we mapped the max number of pages that we have
1501 if (error < npages || buffers == pipe_buffers)
1514 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1515 struct splice_desc *sd)
1520 ret = buf->ops->confirm(pipe, buf);
1525 * See if we can use the atomic maps, by prefaulting in the
1526 * pages and doing an atomic copy
1528 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1529 src = buf->ops->map(pipe, buf, 1);
1530 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1532 buf->ops->unmap(pipe, buf, src);
1540 * No dice, use slow non-atomic map and copy
1542 src = buf->ops->map(pipe, buf, 0);
1545 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1548 buf->ops->unmap(pipe, buf, src);
1551 sd->u.userptr += ret;
1556 * For lack of a better implementation, implement vmsplice() to userspace
1557 * as a simple copy of the pipes pages to the user iov.
1559 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1560 unsigned long nr_segs, unsigned int flags)
1562 struct pipe_inode_info *pipe;
1563 struct splice_desc sd;
1568 pipe = pipe_info(file->f_path.dentry->d_inode);
1580 * Get user address base and length for this iovec.
1582 error = get_user(base, &iov->iov_base);
1583 if (unlikely(error))
1585 error = get_user(len, &iov->iov_len);
1586 if (unlikely(error))
1590 * Sanity check this iovec. 0 read succeeds.
1594 if (unlikely(!base)) {
1599 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1607 sd.u.userptr = base;
1610 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1636 * vmsplice splices a user address range into a pipe. It can be thought of
1637 * as splice-from-memory, where the regular splice is splice-from-file (or
1638 * to file). In both cases the output is a pipe, naturally.
1640 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1641 unsigned long nr_segs, unsigned int flags)
1643 struct pipe_inode_info *pipe;
1644 struct page *pages[PIPE_DEF_BUFFERS];
1645 struct partial_page partial[PIPE_DEF_BUFFERS];
1646 struct splice_pipe_desc spd = {
1650 .ops = &user_page_pipe_buf_ops,
1651 .spd_release = spd_release_page,
1655 pipe = pipe_info(file->f_path.dentry->d_inode);
1659 if (splice_grow_spd(pipe, &spd))
1662 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1663 spd.partial, flags & SPLICE_F_GIFT,
1665 if (spd.nr_pages <= 0)
1668 ret = splice_to_pipe(pipe, &spd);
1670 splice_shrink_spd(pipe, &spd);
1675 * Note that vmsplice only really supports true splicing _from_ user memory
1676 * to a pipe, not the other way around. Splicing from user memory is a simple
1677 * operation that can be supported without any funky alignment restrictions
1678 * or nasty vm tricks. We simply map in the user memory and fill them into
1679 * a pipe. The reverse isn't quite as easy, though. There are two possible
1680 * solutions for that:
1682 * - memcpy() the data internally, at which point we might as well just
1683 * do a regular read() on the buffer anyway.
1684 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1685 * has restriction limitations on both ends of the pipe).
1687 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1690 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1691 unsigned long, nr_segs, unsigned int, flags)
1697 if (unlikely(nr_segs > UIO_MAXIOV))
1699 else if (unlikely(!nr_segs))
1703 file = fget_light(fd, &fput);
1705 if (file->f_mode & FMODE_WRITE)
1706 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1707 else if (file->f_mode & FMODE_READ)
1708 error = vmsplice_to_user(file, iov, nr_segs, flags);
1710 fput_light(file, fput);
1716 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1717 int, fd_out, loff_t __user *, off_out,
1718 size_t, len, unsigned int, flags)
1721 struct file *in, *out;
1722 int fput_in, fput_out;
1728 in = fget_light(fd_in, &fput_in);
1730 if (in->f_mode & FMODE_READ) {
1731 out = fget_light(fd_out, &fput_out);
1733 if (out->f_mode & FMODE_WRITE)
1734 error = do_splice(in, off_in,
1737 fput_light(out, fput_out);
1741 fput_light(in, fput_in);
1748 * Make sure there's data to read. Wait for input if we can, otherwise
1749 * return an appropriate error.
1751 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1756 * Check ->nrbufs without the inode lock first. This function
1757 * is speculative anyways, so missing one is ok.
1765 while (!pipe->nrbufs) {
1766 if (signal_pending(current)) {
1772 if (!pipe->waiting_writers) {
1773 if (flags & SPLICE_F_NONBLOCK) {
1786 * Make sure there's writeable room. Wait for room if we can, otherwise
1787 * return an appropriate error.
1789 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1794 * Check ->nrbufs without the inode lock first. This function
1795 * is speculative anyways, so missing one is ok.
1797 if (pipe->nrbufs < pipe->buffers)
1803 while (pipe->nrbufs >= pipe->buffers) {
1804 if (!pipe->readers) {
1805 send_sig(SIGPIPE, current, 0);
1809 if (flags & SPLICE_F_NONBLOCK) {
1813 if (signal_pending(current)) {
1817 pipe->waiting_writers++;
1819 pipe->waiting_writers--;
1827 * Splice contents of ipipe to opipe.
1829 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1830 struct pipe_inode_info *opipe,
1831 size_t len, unsigned int flags)
1833 struct pipe_buffer *ibuf, *obuf;
1835 bool input_wakeup = false;
1839 ret = ipipe_prep(ipipe, flags);
1843 ret = opipe_prep(opipe, flags);
1848 * Potential ABBA deadlock, work around it by ordering lock
1849 * grabbing by pipe info address. Otherwise two different processes
1850 * could deadlock (one doing tee from A -> B, the other from B -> A).
1852 pipe_double_lock(ipipe, opipe);
1855 if (!opipe->readers) {
1856 send_sig(SIGPIPE, current, 0);
1862 if (!ipipe->nrbufs && !ipipe->writers)
1866 * Cannot make any progress, because either the input
1867 * pipe is empty or the output pipe is full.
1869 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1870 /* Already processed some buffers, break */
1874 if (flags & SPLICE_F_NONBLOCK) {
1880 * We raced with another reader/writer and haven't
1881 * managed to process any buffers. A zero return
1882 * value means EOF, so retry instead.
1889 ibuf = ipipe->bufs + ipipe->curbuf;
1890 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1891 obuf = opipe->bufs + nbuf;
1893 if (len >= ibuf->len) {
1895 * Simply move the whole buffer from ipipe to opipe
1900 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1902 input_wakeup = true;
1905 * Get a reference to this pipe buffer,
1906 * so we can copy the contents over.
1908 ibuf->ops->get(ipipe, ibuf);
1912 * Don't inherit the gift flag, we need to
1913 * prevent multiple steals of this page.
1915 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1919 ibuf->offset += obuf->len;
1920 ibuf->len -= obuf->len;
1930 * If we put data in the output pipe, wakeup any potential readers.
1934 if (waitqueue_active(&opipe->wait))
1935 wake_up_interruptible(&opipe->wait);
1936 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1939 wakeup_pipe_writers(ipipe);
1945 * Link contents of ipipe to opipe.
1947 static int link_pipe(struct pipe_inode_info *ipipe,
1948 struct pipe_inode_info *opipe,
1949 size_t len, unsigned int flags)
1951 struct pipe_buffer *ibuf, *obuf;
1952 int ret = 0, i = 0, nbuf;
1955 * Potential ABBA deadlock, work around it by ordering lock
1956 * grabbing by pipe info address. Otherwise two different processes
1957 * could deadlock (one doing tee from A -> B, the other from B -> A).
1959 pipe_double_lock(ipipe, opipe);
1962 if (!opipe->readers) {
1963 send_sig(SIGPIPE, current, 0);
1970 * If we have iterated all input buffers or ran out of
1971 * output room, break.
1973 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1976 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1977 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1980 * Get a reference to this pipe buffer,
1981 * so we can copy the contents over.
1983 ibuf->ops->get(ipipe, ibuf);
1985 obuf = opipe->bufs + nbuf;
1989 * Don't inherit the gift flag, we need to
1990 * prevent multiple steals of this page.
1992 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1994 if (obuf->len > len)
2004 * return EAGAIN if we have the potential of some data in the
2005 * future, otherwise just return 0
2007 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2014 * If we put data in the output pipe, wakeup any potential readers.
2018 if (waitqueue_active(&opipe->wait))
2019 wake_up_interruptible(&opipe->wait);
2020 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2027 * This is a tee(1) implementation that works on pipes. It doesn't copy
2028 * any data, it simply references the 'in' pages on the 'out' pipe.
2029 * The 'flags' used are the SPLICE_F_* variants, currently the only
2030 * applicable one is SPLICE_F_NONBLOCK.
2032 static long do_tee(struct file *in, struct file *out, size_t len,
2035 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2036 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2040 * Duplicate the contents of ipipe to opipe without actually
2043 if (ipipe && opipe && ipipe != opipe) {
2045 * Keep going, unless we encounter an error. The ipipe/opipe
2046 * ordering doesn't really matter.
2048 ret = ipipe_prep(ipipe, flags);
2050 ret = opipe_prep(opipe, flags);
2052 ret = link_pipe(ipipe, opipe, len, flags);
2059 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2068 in = fget_light(fdin, &fput_in);
2070 if (in->f_mode & FMODE_READ) {
2072 struct file *out = fget_light(fdout, &fput_out);
2075 if (out->f_mode & FMODE_WRITE)
2076 error = do_tee(in, out, len, flags);
2077 fput_light(out, fput_out);
2080 fput_light(in, fput_in);