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)) {
405 * Page was truncated, or invalidated by the
406 * filesystem. Redo the find/create, but this time the
407 * page is kept locked, so there's no chance of another
408 * race with truncate/invalidate.
410 if (!page->mapping) {
412 page = find_or_create_page(mapping, index,
413 mapping_gfp_mask(mapping));
419 page_cache_release(spd.pages[page_nr]);
420 spd.pages[page_nr] = page;
423 * page was already under io and is now done, great
425 if (PageUptodate(page)) {
431 * need to read in the page
433 error = mapping->a_ops->readpage(in, page);
434 if (unlikely(error)) {
436 * We really should re-lookup the page here,
437 * but it complicates things a lot. Instead
438 * lets just do what we already stored, and
439 * we'll get it the next time we are called.
441 if (error == AOP_TRUNCATED_PAGE)
449 * i_size must be checked after PageUptodate.
451 isize = i_size_read(mapping->host);
452 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
453 if (unlikely(!isize || index > end_index))
457 * if this is the last page, see if we need to shrink
458 * the length and stop
460 if (end_index == index) {
464 * max good bytes in this page
466 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
471 * force quit after adding this page
473 this_len = min(this_len, plen - loff);
477 spd.partial[page_nr].offset = loff;
478 spd.partial[page_nr].len = this_len;
486 * Release any pages at the end, if we quit early. 'page_nr' is how far
487 * we got, 'nr_pages' is how many pages are in the map.
489 while (page_nr < nr_pages)
490 page_cache_release(spd.pages[page_nr++]);
491 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
494 error = splice_to_pipe(pipe, &spd);
496 splice_shrink_spd(pipe, &spd);
501 * generic_file_splice_read - splice data from file to a pipe
502 * @in: file to splice from
503 * @ppos: position in @in
504 * @pipe: pipe to splice to
505 * @len: number of bytes to splice
506 * @flags: splice modifier flags
509 * Will read pages from given file and fill them into a pipe. Can be
510 * used as long as the address_space operations for the source implements
514 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
515 struct pipe_inode_info *pipe, size_t len,
521 isize = i_size_read(in->f_mapping->host);
522 if (unlikely(*ppos >= isize))
525 left = isize - *ppos;
526 if (unlikely(left < len))
529 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
537 EXPORT_SYMBOL(generic_file_splice_read);
539 static const struct pipe_buf_operations default_pipe_buf_ops = {
541 .map = generic_pipe_buf_map,
542 .unmap = generic_pipe_buf_unmap,
543 .confirm = generic_pipe_buf_confirm,
544 .release = generic_pipe_buf_release,
545 .steal = generic_pipe_buf_steal,
546 .get = generic_pipe_buf_get,
549 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
550 unsigned long vlen, loff_t offset)
558 /* The cast to a user pointer is valid due to the set_fs() */
559 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
565 static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
573 /* The cast to a user pointer is valid due to the set_fs() */
574 res = vfs_write(file, (const char __user *)buf, count, &pos);
580 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
581 struct pipe_inode_info *pipe, size_t len,
584 unsigned int nr_pages;
585 unsigned int nr_freed;
587 struct page *pages[PIPE_DEF_BUFFERS];
588 struct partial_page partial[PIPE_DEF_BUFFERS];
589 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
594 struct splice_pipe_desc spd = {
598 .ops = &default_pipe_buf_ops,
599 .spd_release = spd_release_page,
602 if (splice_grow_spd(pipe, &spd))
607 if (pipe->buffers > PIPE_DEF_BUFFERS) {
608 vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
613 offset = *ppos & ~PAGE_CACHE_MASK;
614 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
616 for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
619 page = alloc_page(GFP_USER);
624 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
625 vec[i].iov_base = (void __user *) page_address(page);
626 vec[i].iov_len = this_len;
633 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
644 for (i = 0; i < spd.nr_pages; i++) {
645 this_len = min_t(size_t, vec[i].iov_len, res);
646 spd.partial[i].offset = 0;
647 spd.partial[i].len = this_len;
649 __free_page(spd.pages[i]);
655 spd.nr_pages -= nr_freed;
657 res = splice_to_pipe(pipe, &spd);
664 splice_shrink_spd(pipe, &spd);
668 for (i = 0; i < spd.nr_pages; i++)
669 __free_page(spd.pages[i]);
674 EXPORT_SYMBOL(default_file_splice_read);
677 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
678 * using sendpage(). Return the number of bytes sent.
680 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
681 struct pipe_buffer *buf, struct splice_desc *sd)
683 struct file *file = sd->u.file;
684 loff_t pos = sd->pos;
687 if (!likely(file->f_op && file->f_op->sendpage))
690 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
691 return file->f_op->sendpage(file, buf->page, buf->offset,
692 sd->len, &pos, more);
696 * This is a little more tricky than the file -> pipe splicing. There are
697 * basically three cases:
699 * - Destination page already exists in the address space and there
700 * are users of it. For that case we have no other option that
701 * copying the data. Tough luck.
702 * - Destination page already exists in the address space, but there
703 * are no users of it. Make sure it's uptodate, then drop it. Fall
704 * through to last case.
705 * - Destination page does not exist, we can add the pipe page to
706 * the page cache and avoid the copy.
708 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
709 * sd->flags), we attempt to migrate pages from the pipe to the output
710 * file address space page cache. This is possible if no one else has
711 * the pipe page referenced outside of the pipe and page cache. If
712 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
713 * a new page in the output file page cache and fill/dirty that.
715 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
716 struct splice_desc *sd)
718 struct file *file = sd->u.file;
719 struct address_space *mapping = file->f_mapping;
720 unsigned int offset, this_len;
725 offset = sd->pos & ~PAGE_CACHE_MASK;
728 if (this_len + offset > PAGE_CACHE_SIZE)
729 this_len = PAGE_CACHE_SIZE - offset;
731 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
732 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
736 if (buf->page != page) {
738 * Careful, ->map() uses KM_USER0!
740 char *src = buf->ops->map(pipe, buf, 1);
741 char *dst = kmap_atomic(page, KM_USER1);
743 memcpy(dst + offset, src + buf->offset, this_len);
744 flush_dcache_page(page);
745 kunmap_atomic(dst, KM_USER1);
746 buf->ops->unmap(pipe, buf, src);
748 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
753 EXPORT_SYMBOL(pipe_to_file);
755 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
758 if (waitqueue_active(&pipe->wait))
759 wake_up_interruptible(&pipe->wait);
760 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
764 * splice_from_pipe_feed - feed available data from a pipe to a file
765 * @pipe: pipe to splice from
766 * @sd: information to @actor
767 * @actor: handler that splices the data
770 * This function loops over the pipe and calls @actor to do the
771 * actual moving of a single struct pipe_buffer to the desired
772 * destination. It returns when there's no more buffers left in
773 * the pipe or if the requested number of bytes (@sd->total_len)
774 * have been copied. It returns a positive number (one) if the
775 * pipe needs to be filled with more data, zero if the required
776 * number of bytes have been copied and -errno on error.
778 * This, together with splice_from_pipe_{begin,end,next}, may be
779 * used to implement the functionality of __splice_from_pipe() when
780 * locking is required around copying the pipe buffers to the
783 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
788 while (pipe->nrbufs) {
789 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
790 const struct pipe_buf_operations *ops = buf->ops;
793 if (sd->len > sd->total_len)
794 sd->len = sd->total_len;
796 ret = buf->ops->confirm(pipe, buf);
803 ret = actor(pipe, buf, sd);
810 sd->num_spliced += ret;
813 sd->total_len -= ret;
817 ops->release(pipe, buf);
818 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
821 sd->need_wakeup = true;
830 EXPORT_SYMBOL(splice_from_pipe_feed);
833 * splice_from_pipe_next - wait for some data to splice from
834 * @pipe: pipe to splice from
835 * @sd: information about the splice operation
838 * This function will wait for some data and return a positive
839 * value (one) if pipe buffers are available. It will return zero
840 * or -errno if no more data needs to be spliced.
842 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
844 while (!pipe->nrbufs) {
848 if (!pipe->waiting_writers && sd->num_spliced)
851 if (sd->flags & SPLICE_F_NONBLOCK)
854 if (signal_pending(current))
857 if (sd->need_wakeup) {
858 wakeup_pipe_writers(pipe);
859 sd->need_wakeup = false;
867 EXPORT_SYMBOL(splice_from_pipe_next);
870 * splice_from_pipe_begin - start splicing from pipe
871 * @sd: information about the splice operation
874 * This function should be called before a loop containing
875 * splice_from_pipe_next() and splice_from_pipe_feed() to
876 * initialize the necessary fields of @sd.
878 void splice_from_pipe_begin(struct splice_desc *sd)
881 sd->need_wakeup = false;
883 EXPORT_SYMBOL(splice_from_pipe_begin);
886 * splice_from_pipe_end - finish splicing from pipe
887 * @pipe: pipe to splice from
888 * @sd: information about the splice operation
891 * This function will wake up pipe writers if necessary. It should
892 * be called after a loop containing splice_from_pipe_next() and
893 * splice_from_pipe_feed().
895 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
898 wakeup_pipe_writers(pipe);
900 EXPORT_SYMBOL(splice_from_pipe_end);
903 * __splice_from_pipe - splice data from a pipe to given actor
904 * @pipe: pipe to splice from
905 * @sd: information to @actor
906 * @actor: handler that splices the data
909 * This function does little more than loop over the pipe and call
910 * @actor to do the actual moving of a single struct pipe_buffer to
911 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
915 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
920 splice_from_pipe_begin(sd);
922 ret = splice_from_pipe_next(pipe, sd);
924 ret = splice_from_pipe_feed(pipe, sd, actor);
926 splice_from_pipe_end(pipe, sd);
928 return sd->num_spliced ? sd->num_spliced : ret;
930 EXPORT_SYMBOL(__splice_from_pipe);
933 * splice_from_pipe - splice data from a pipe to a file
934 * @pipe: pipe to splice from
935 * @out: file to splice to
936 * @ppos: position in @out
937 * @len: how many bytes to splice
938 * @flags: splice modifier flags
939 * @actor: handler that splices the data
942 * See __splice_from_pipe. This function locks the pipe inode,
943 * otherwise it's identical to __splice_from_pipe().
946 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
947 loff_t *ppos, size_t len, unsigned int flags,
951 struct splice_desc sd = {
959 ret = __splice_from_pipe(pipe, &sd, actor);
966 * generic_file_splice_write - splice data from a pipe to a file
968 * @out: file to write to
969 * @ppos: position in @out
970 * @len: number of bytes to splice
971 * @flags: splice modifier flags
974 * Will either move or copy pages (determined by @flags options) from
975 * the given pipe inode to the given file.
979 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
980 loff_t *ppos, size_t len, unsigned int flags)
982 struct address_space *mapping = out->f_mapping;
983 struct inode *inode = mapping->host;
984 struct splice_desc sd = {
994 splice_from_pipe_begin(&sd);
996 ret = splice_from_pipe_next(pipe, &sd);
1000 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1001 ret = file_remove_suid(out);
1003 file_update_time(out);
1004 ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1006 mutex_unlock(&inode->i_mutex);
1008 splice_from_pipe_end(pipe, &sd);
1013 ret = sd.num_spliced;
1016 unsigned long nr_pages;
1019 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1021 err = generic_write_sync(out, *ppos, ret);
1026 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1032 EXPORT_SYMBOL(generic_file_splice_write);
1034 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1035 struct splice_desc *sd)
1040 data = buf->ops->map(pipe, buf, 0);
1041 ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1042 buf->ops->unmap(pipe, buf, data);
1047 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1048 struct file *out, loff_t *ppos,
1049 size_t len, unsigned int flags)
1053 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1061 * generic_splice_sendpage - splice data from a pipe to a socket
1062 * @pipe: pipe to splice from
1063 * @out: socket to write to
1064 * @ppos: position in @out
1065 * @len: number of bytes to splice
1066 * @flags: splice modifier flags
1069 * Will send @len bytes from the pipe to a network socket. No data copying
1073 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1074 loff_t *ppos, size_t len, unsigned int flags)
1076 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1079 EXPORT_SYMBOL(generic_splice_sendpage);
1082 * Attempt to initiate a splice from pipe to file.
1084 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1085 loff_t *ppos, size_t len, unsigned int flags)
1087 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1088 loff_t *, size_t, unsigned int);
1091 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1094 if (unlikely(out->f_flags & O_APPEND))
1097 ret = rw_verify_area(WRITE, out, ppos, len);
1098 if (unlikely(ret < 0))
1101 if (out->f_op && out->f_op->splice_write)
1102 splice_write = out->f_op->splice_write;
1104 splice_write = default_file_splice_write;
1106 return splice_write(pipe, out, ppos, len, flags);
1110 * Attempt to initiate a splice from a file to a pipe.
1112 static long do_splice_to(struct file *in, loff_t *ppos,
1113 struct pipe_inode_info *pipe, size_t len,
1116 ssize_t (*splice_read)(struct file *, loff_t *,
1117 struct pipe_inode_info *, size_t, unsigned int);
1120 if (unlikely(!(in->f_mode & FMODE_READ)))
1123 ret = rw_verify_area(READ, in, ppos, len);
1124 if (unlikely(ret < 0))
1127 if (in->f_op && in->f_op->splice_read)
1128 splice_read = in->f_op->splice_read;
1130 splice_read = default_file_splice_read;
1132 return splice_read(in, ppos, pipe, len, flags);
1136 * splice_direct_to_actor - splices data directly between two non-pipes
1137 * @in: file to splice from
1138 * @sd: actor information on where to splice to
1139 * @actor: handles the data splicing
1142 * This is a special case helper to splice directly between two
1143 * points, without requiring an explicit pipe. Internally an allocated
1144 * pipe is cached in the process, and reused during the lifetime of
1148 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1149 splice_direct_actor *actor)
1151 struct pipe_inode_info *pipe;
1158 * We require the input being a regular file, as we don't want to
1159 * randomly drop data for eg socket -> socket splicing. Use the
1160 * piped splicing for that!
1162 i_mode = in->f_path.dentry->d_inode->i_mode;
1163 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1167 * neither in nor out is a pipe, setup an internal pipe attached to
1168 * 'out' and transfer the wanted data from 'in' to 'out' through that
1170 pipe = current->splice_pipe;
1171 if (unlikely(!pipe)) {
1172 pipe = alloc_pipe_info(NULL);
1177 * We don't have an immediate reader, but we'll read the stuff
1178 * out of the pipe right after the splice_to_pipe(). So set
1179 * PIPE_READERS appropriately.
1183 current->splice_pipe = pipe;
1191 len = sd->total_len;
1195 * Don't block on output, we have to drain the direct pipe.
1197 sd->flags &= ~SPLICE_F_NONBLOCK;
1201 loff_t pos = sd->pos, prev_pos = pos;
1203 ret = do_splice_to(in, &pos, pipe, len, flags);
1204 if (unlikely(ret <= 0))
1208 sd->total_len = read_len;
1211 * NOTE: nonblocking mode only applies to the input. We
1212 * must not do the output in nonblocking mode as then we
1213 * could get stuck data in the internal pipe:
1215 ret = actor(pipe, sd);
1216 if (unlikely(ret <= 0)) {
1225 if (ret < read_len) {
1226 sd->pos = prev_pos + ret;
1232 pipe->nrbufs = pipe->curbuf = 0;
1238 * If we did an incomplete transfer we must release
1239 * the pipe buffers in question:
1241 for (i = 0; i < pipe->buffers; i++) {
1242 struct pipe_buffer *buf = pipe->bufs + i;
1245 buf->ops->release(pipe, buf);
1255 EXPORT_SYMBOL(splice_direct_to_actor);
1257 static int direct_splice_actor(struct pipe_inode_info *pipe,
1258 struct splice_desc *sd)
1260 struct file *file = sd->u.file;
1262 return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1267 * do_splice_direct - splices data directly between two files
1268 * @in: file to splice from
1269 * @ppos: input file offset
1270 * @out: file to splice to
1271 * @len: number of bytes to splice
1272 * @flags: splice modifier flags
1275 * For use by do_sendfile(). splice can easily emulate sendfile, but
1276 * doing it in the application would incur an extra system call
1277 * (splice in + splice out, as compared to just sendfile()). So this helper
1278 * can splice directly through a process-private pipe.
1281 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1282 size_t len, unsigned int flags)
1284 struct splice_desc sd = {
1293 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1300 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1301 struct pipe_inode_info *opipe,
1302 size_t len, unsigned int flags);
1304 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1305 * location, so checking ->i_pipe is not enough to verify that this is a
1308 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1310 if (S_ISFIFO(inode->i_mode))
1311 return inode->i_pipe;
1317 * Determine where to splice to/from.
1319 static long do_splice(struct file *in, loff_t __user *off_in,
1320 struct file *out, loff_t __user *off_out,
1321 size_t len, unsigned int flags)
1323 struct pipe_inode_info *ipipe;
1324 struct pipe_inode_info *opipe;
1325 loff_t offset, *off;
1328 ipipe = pipe_info(in->f_path.dentry->d_inode);
1329 opipe = pipe_info(out->f_path.dentry->d_inode);
1331 if (ipipe && opipe) {
1332 if (off_in || off_out)
1335 if (!(in->f_mode & FMODE_READ))
1338 if (!(out->f_mode & FMODE_WRITE))
1341 /* Splicing to self would be fun, but... */
1345 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1352 if (!(out->f_mode & FMODE_PWRITE))
1354 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1360 ret = do_splice_from(ipipe, out, off, len, flags);
1362 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1372 if (!(in->f_mode & FMODE_PREAD))
1374 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1380 ret = do_splice_to(in, off, opipe, len, flags);
1382 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1392 * Map an iov into an array of pages and offset/length tupples. With the
1393 * partial_page structure, we can map several non-contiguous ranges into
1394 * our ones pages[] map instead of splitting that operation into pieces.
1395 * Could easily be exported as a generic helper for other users, in which
1396 * case one would probably want to add a 'max_nr_pages' parameter as well.
1398 static int get_iovec_page_array(const struct iovec __user *iov,
1399 unsigned int nr_vecs, struct page **pages,
1400 struct partial_page *partial, int aligned,
1401 unsigned int pipe_buffers)
1403 int buffers = 0, error = 0;
1406 unsigned long off, npages;
1413 if (copy_from_user(&entry, iov, sizeof(entry)))
1416 base = entry.iov_base;
1417 len = entry.iov_len;
1420 * Sanity check this iovec. 0 read succeeds.
1426 if (!access_ok(VERIFY_READ, base, len))
1430 * Get this base offset and number of pages, then map
1431 * in the user pages.
1433 off = (unsigned long) base & ~PAGE_MASK;
1436 * If asked for alignment, the offset must be zero and the
1437 * length a multiple of the PAGE_SIZE.
1440 if (aligned && (off || len & ~PAGE_MASK))
1443 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1444 if (npages > pipe_buffers - buffers)
1445 npages = pipe_buffers - buffers;
1447 error = get_user_pages_fast((unsigned long)base, npages,
1448 0, &pages[buffers]);
1450 if (unlikely(error <= 0))
1454 * Fill this contiguous range into the partial page map.
1456 for (i = 0; i < error; i++) {
1457 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1459 partial[buffers].offset = off;
1460 partial[buffers].len = plen;
1468 * We didn't complete this iov, stop here since it probably
1469 * means we have to move some of this into a pipe to
1470 * be able to continue.
1476 * Don't continue if we mapped fewer pages than we asked for,
1477 * or if we mapped the max number of pages that we have
1480 if (error < npages || buffers == pipe_buffers)
1493 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1494 struct splice_desc *sd)
1500 * See if we can use the atomic maps, by prefaulting in the
1501 * pages and doing an atomic copy
1503 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1504 src = buf->ops->map(pipe, buf, 1);
1505 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1507 buf->ops->unmap(pipe, buf, src);
1515 * No dice, use slow non-atomic map and copy
1517 src = buf->ops->map(pipe, buf, 0);
1520 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1523 buf->ops->unmap(pipe, buf, src);
1526 sd->u.userptr += ret;
1531 * For lack of a better implementation, implement vmsplice() to userspace
1532 * as a simple copy of the pipes pages to the user iov.
1534 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1535 unsigned long nr_segs, unsigned int flags)
1537 struct pipe_inode_info *pipe;
1538 struct splice_desc sd;
1543 pipe = pipe_info(file->f_path.dentry->d_inode);
1555 * Get user address base and length for this iovec.
1557 error = get_user(base, &iov->iov_base);
1558 if (unlikely(error))
1560 error = get_user(len, &iov->iov_len);
1561 if (unlikely(error))
1565 * Sanity check this iovec. 0 read succeeds.
1569 if (unlikely(!base)) {
1574 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1582 sd.u.userptr = base;
1585 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1611 * vmsplice splices a user address range into a pipe. It can be thought of
1612 * as splice-from-memory, where the regular splice is splice-from-file (or
1613 * to file). In both cases the output is a pipe, naturally.
1615 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1616 unsigned long nr_segs, unsigned int flags)
1618 struct pipe_inode_info *pipe;
1619 struct page *pages[PIPE_DEF_BUFFERS];
1620 struct partial_page partial[PIPE_DEF_BUFFERS];
1621 struct splice_pipe_desc spd = {
1625 .ops = &user_page_pipe_buf_ops,
1626 .spd_release = spd_release_page,
1630 pipe = pipe_info(file->f_path.dentry->d_inode);
1634 if (splice_grow_spd(pipe, &spd))
1637 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1638 spd.partial, flags & SPLICE_F_GIFT,
1640 if (spd.nr_pages <= 0)
1643 ret = splice_to_pipe(pipe, &spd);
1645 splice_shrink_spd(pipe, &spd);
1650 * Note that vmsplice only really supports true splicing _from_ user memory
1651 * to a pipe, not the other way around. Splicing from user memory is a simple
1652 * operation that can be supported without any funky alignment restrictions
1653 * or nasty vm tricks. We simply map in the user memory and fill them into
1654 * a pipe. The reverse isn't quite as easy, though. There are two possible
1655 * solutions for that:
1657 * - memcpy() the data internally, at which point we might as well just
1658 * do a regular read() on the buffer anyway.
1659 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1660 * has restriction limitations on both ends of the pipe).
1662 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1665 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1666 unsigned long, nr_segs, unsigned int, flags)
1672 if (unlikely(nr_segs > UIO_MAXIOV))
1674 else if (unlikely(!nr_segs))
1678 file = fget_light(fd, &fput);
1680 if (file->f_mode & FMODE_WRITE)
1681 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1682 else if (file->f_mode & FMODE_READ)
1683 error = vmsplice_to_user(file, iov, nr_segs, flags);
1685 fput_light(file, fput);
1691 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1692 int, fd_out, loff_t __user *, off_out,
1693 size_t, len, unsigned int, flags)
1696 struct file *in, *out;
1697 int fput_in, fput_out;
1703 in = fget_light(fd_in, &fput_in);
1705 if (in->f_mode & FMODE_READ) {
1706 out = fget_light(fd_out, &fput_out);
1708 if (out->f_mode & FMODE_WRITE)
1709 error = do_splice(in, off_in,
1712 fput_light(out, fput_out);
1716 fput_light(in, fput_in);
1723 * Make sure there's data to read. Wait for input if we can, otherwise
1724 * return an appropriate error.
1726 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1731 * Check ->nrbufs without the inode lock first. This function
1732 * is speculative anyways, so missing one is ok.
1740 while (!pipe->nrbufs) {
1741 if (signal_pending(current)) {
1747 if (!pipe->waiting_writers) {
1748 if (flags & SPLICE_F_NONBLOCK) {
1761 * Make sure there's writeable room. Wait for room if we can, otherwise
1762 * return an appropriate error.
1764 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1769 * Check ->nrbufs without the inode lock first. This function
1770 * is speculative anyways, so missing one is ok.
1772 if (pipe->nrbufs < pipe->buffers)
1778 while (pipe->nrbufs >= pipe->buffers) {
1779 if (!pipe->readers) {
1780 send_sig(SIGPIPE, current, 0);
1784 if (flags & SPLICE_F_NONBLOCK) {
1788 if (signal_pending(current)) {
1792 pipe->waiting_writers++;
1794 pipe->waiting_writers--;
1802 * Splice contents of ipipe to opipe.
1804 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1805 struct pipe_inode_info *opipe,
1806 size_t len, unsigned int flags)
1808 struct pipe_buffer *ibuf, *obuf;
1810 bool input_wakeup = false;
1814 ret = ipipe_prep(ipipe, flags);
1818 ret = opipe_prep(opipe, flags);
1823 * Potential ABBA deadlock, work around it by ordering lock
1824 * grabbing by pipe info address. Otherwise two different processes
1825 * could deadlock (one doing tee from A -> B, the other from B -> A).
1827 pipe_double_lock(ipipe, opipe);
1830 if (!opipe->readers) {
1831 send_sig(SIGPIPE, current, 0);
1837 if (!ipipe->nrbufs && !ipipe->writers)
1841 * Cannot make any progress, because either the input
1842 * pipe is empty or the output pipe is full.
1844 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1845 /* Already processed some buffers, break */
1849 if (flags & SPLICE_F_NONBLOCK) {
1855 * We raced with another reader/writer and haven't
1856 * managed to process any buffers. A zero return
1857 * value means EOF, so retry instead.
1864 ibuf = ipipe->bufs + ipipe->curbuf;
1865 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1866 obuf = opipe->bufs + nbuf;
1868 if (len >= ibuf->len) {
1870 * Simply move the whole buffer from ipipe to opipe
1875 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1877 input_wakeup = true;
1880 * Get a reference to this pipe buffer,
1881 * so we can copy the contents over.
1883 ibuf->ops->get(ipipe, ibuf);
1887 * Don't inherit the gift flag, we need to
1888 * prevent multiple steals of this page.
1890 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1894 ibuf->offset += obuf->len;
1895 ibuf->len -= obuf->len;
1905 * If we put data in the output pipe, wakeup any potential readers.
1909 if (waitqueue_active(&opipe->wait))
1910 wake_up_interruptible(&opipe->wait);
1911 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1914 wakeup_pipe_writers(ipipe);
1920 * Link contents of ipipe to opipe.
1922 static int link_pipe(struct pipe_inode_info *ipipe,
1923 struct pipe_inode_info *opipe,
1924 size_t len, unsigned int flags)
1926 struct pipe_buffer *ibuf, *obuf;
1927 int ret = 0, i = 0, nbuf;
1930 * Potential ABBA deadlock, work around it by ordering lock
1931 * grabbing by pipe info address. Otherwise two different processes
1932 * could deadlock (one doing tee from A -> B, the other from B -> A).
1934 pipe_double_lock(ipipe, opipe);
1937 if (!opipe->readers) {
1938 send_sig(SIGPIPE, current, 0);
1945 * If we have iterated all input buffers or ran out of
1946 * output room, break.
1948 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1951 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1952 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1955 * Get a reference to this pipe buffer,
1956 * so we can copy the contents over.
1958 ibuf->ops->get(ipipe, ibuf);
1960 obuf = opipe->bufs + nbuf;
1964 * Don't inherit the gift flag, we need to
1965 * prevent multiple steals of this page.
1967 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1969 if (obuf->len > len)
1979 * return EAGAIN if we have the potential of some data in the
1980 * future, otherwise just return 0
1982 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1989 * If we put data in the output pipe, wakeup any potential readers.
1993 if (waitqueue_active(&opipe->wait))
1994 wake_up_interruptible(&opipe->wait);
1995 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
2002 * This is a tee(1) implementation that works on pipes. It doesn't copy
2003 * any data, it simply references the 'in' pages on the 'out' pipe.
2004 * The 'flags' used are the SPLICE_F_* variants, currently the only
2005 * applicable one is SPLICE_F_NONBLOCK.
2007 static long do_tee(struct file *in, struct file *out, size_t len,
2010 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
2011 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
2015 * Duplicate the contents of ipipe to opipe without actually
2018 if (ipipe && opipe && ipipe != opipe) {
2020 * Keep going, unless we encounter an error. The ipipe/opipe
2021 * ordering doesn't really matter.
2023 ret = ipipe_prep(ipipe, flags);
2025 ret = opipe_prep(opipe, flags);
2027 ret = link_pipe(ipipe, opipe, len, flags);
2034 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2043 in = fget_light(fdin, &fput_in);
2045 if (in->f_mode & FMODE_READ) {
2047 struct file *out = fget_light(fdout, &fput_out);
2050 if (out->f_mode & FMODE_WRITE)
2051 error = do_tee(in, out, len, flags);
2052 fput_light(out, fput_out);
2055 fput_light(in, fput_in);