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/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
34 * Attempt to steal a page from a pipe buffer. This should perhaps go into
35 * a vm helper function, it's already simplified quite a bit by the
36 * addition of remove_mapping(). If success is returned, the caller may
37 * attempt to reuse this page for another destination.
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
40 struct pipe_buffer *buf)
42 struct page *page = buf->page;
43 struct address_space *mapping;
47 mapping = page_mapping(page);
49 WARN_ON(!PageUptodate(page));
52 * At least for ext2 with nobh option, we need to wait on
53 * writeback completing on this page, since we'll remove it
54 * from the pagecache. Otherwise truncate wont wait on the
55 * page, allowing the disk blocks to be reused by someone else
56 * before we actually wrote our data to them. fs corruption
59 wait_on_page_writeback(page);
61 if (PagePrivate(page))
62 try_to_release_page(page, GFP_KERNEL);
65 * If we succeeded in removing the mapping, set LRU flag
68 if (remove_mapping(mapping, page)) {
69 buf->flags |= PIPE_BUF_FLAG_LRU;
75 * Raced with truncate or failed to remove page from current
76 * address space, unlock and return failure.
82 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
83 struct pipe_buffer *buf)
85 page_cache_release(buf->page);
86 buf->flags &= ~PIPE_BUF_FLAG_LRU;
90 * Check whether the contents of buf is OK to access. Since the content
91 * is a page cache page, IO may be in flight.
93 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
94 struct pipe_buffer *buf)
96 struct page *page = buf->page;
99 if (!PageUptodate(page)) {
103 * Page got truncated/unhashed. This will cause a 0-byte
104 * splice, if this is the first page.
106 if (!page->mapping) {
112 * Uh oh, read-error from disk.
114 if (!PageUptodate(page)) {
120 * Page is ok afterall, we are done.
131 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
133 .map = generic_pipe_buf_map,
134 .unmap = generic_pipe_buf_unmap,
135 .confirm = page_cache_pipe_buf_confirm,
136 .release = page_cache_pipe_buf_release,
137 .steal = page_cache_pipe_buf_steal,
138 .get = generic_pipe_buf_get,
141 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
142 struct pipe_buffer *buf)
144 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
147 buf->flags |= PIPE_BUF_FLAG_LRU;
148 return generic_pipe_buf_steal(pipe, buf);
151 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
153 .map = generic_pipe_buf_map,
154 .unmap = generic_pipe_buf_unmap,
155 .confirm = generic_pipe_buf_confirm,
156 .release = page_cache_pipe_buf_release,
157 .steal = user_page_pipe_buf_steal,
158 .get = generic_pipe_buf_get,
162 * splice_to_pipe - fill passed data into a pipe
163 * @pipe: pipe to fill
167 * @spd contains a map of pages and len/offset tupples, a long with
168 * the struct pipe_buf_operations associated with these pages. This
169 * function will link that data to the pipe.
172 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
173 struct splice_pipe_desc *spd)
175 unsigned int spd_pages = spd->nr_pages;
176 int ret, do_wakeup, page_nr;
183 mutex_lock(&pipe->inode->i_mutex);
186 if (!pipe->readers) {
187 send_sig(SIGPIPE, current, 0);
193 if (pipe->nrbufs < PIPE_BUFFERS) {
194 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
195 struct pipe_buffer *buf = pipe->bufs + newbuf;
197 buf->page = spd->pages[page_nr];
198 buf->offset = spd->partial[page_nr].offset;
199 buf->len = spd->partial[page_nr].len;
200 buf->private = spd->partial[page_nr].private;
202 if (spd->flags & SPLICE_F_GIFT)
203 buf->flags |= PIPE_BUF_FLAG_GIFT;
212 if (!--spd->nr_pages)
214 if (pipe->nrbufs < PIPE_BUFFERS)
220 if (spd->flags & SPLICE_F_NONBLOCK) {
226 if (signal_pending(current)) {
234 if (waitqueue_active(&pipe->wait))
235 wake_up_interruptible_sync(&pipe->wait);
236 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
240 pipe->waiting_writers++;
242 pipe->waiting_writers--;
246 mutex_unlock(&pipe->inode->i_mutex);
250 if (waitqueue_active(&pipe->wait))
251 wake_up_interruptible(&pipe->wait);
252 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
256 while (page_nr < spd_pages)
257 page_cache_release(spd->pages[page_nr++]);
263 __generic_file_splice_read(struct file *in, loff_t *ppos,
264 struct pipe_inode_info *pipe, size_t len,
267 struct address_space *mapping = in->f_mapping;
268 unsigned int loff, nr_pages, req_pages;
269 struct page *pages[PIPE_BUFFERS];
270 struct partial_page partial[PIPE_BUFFERS];
272 pgoff_t index, end_index;
275 struct splice_pipe_desc spd = {
279 .ops = &page_cache_pipe_buf_ops,
282 index = *ppos >> PAGE_CACHE_SHIFT;
283 loff = *ppos & ~PAGE_CACHE_MASK;
284 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
285 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
288 * Lookup the (hopefully) full range of pages we need.
290 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
291 index += spd.nr_pages;
294 * If find_get_pages_contig() returned fewer pages than we needed,
295 * readahead/allocate the rest and fill in the holes.
297 if (spd.nr_pages < nr_pages)
298 page_cache_sync_readahead(mapping, &in->f_ra, in,
299 index, req_pages - spd.nr_pages);
302 while (spd.nr_pages < nr_pages) {
304 * Page could be there, find_get_pages_contig() breaks on
307 page = find_get_page(mapping, index);
310 * page didn't exist, allocate one.
312 page = page_cache_alloc_cold(mapping);
316 error = add_to_page_cache_lru(page, mapping, index,
318 if (unlikely(error)) {
319 page_cache_release(page);
320 if (error == -EEXIST)
325 * add_to_page_cache() locks the page, unlock it
326 * to avoid convoluting the logic below even more.
331 pages[spd.nr_pages++] = page;
336 * Now loop over the map and see if we need to start IO on any
337 * pages, fill in the partial map, etc.
339 index = *ppos >> PAGE_CACHE_SHIFT;
340 nr_pages = spd.nr_pages;
342 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
343 unsigned int this_len;
349 * this_len is the max we'll use from this page
351 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
352 page = pages[page_nr];
354 if (PageReadahead(page))
355 page_cache_async_readahead(mapping, &in->f_ra, in,
356 page, index, req_pages - page_nr);
359 * If the page isn't uptodate, we may need to start io on it
361 if (!PageUptodate(page)) {
363 * If in nonblock mode then dont block on waiting
364 * for an in-flight io page
366 if (flags & SPLICE_F_NONBLOCK) {
367 if (TestSetPageLocked(page))
373 * page was truncated, stop here. if this isn't the
374 * first page, we'll just complete what we already
377 if (!page->mapping) {
382 * page was already under io and is now done, great
384 if (PageUptodate(page)) {
390 * need to read in the page
392 error = mapping->a_ops->readpage(in, page);
393 if (unlikely(error)) {
395 * We really should re-lookup the page here,
396 * but it complicates things a lot. Instead
397 * lets just do what we already stored, and
398 * we'll get it the next time we are called.
400 if (error == AOP_TRUNCATED_PAGE)
408 * i_size must be checked after PageUptodate.
410 isize = i_size_read(mapping->host);
411 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
412 if (unlikely(!isize || index > end_index))
416 * if this is the last page, see if we need to shrink
417 * the length and stop
419 if (end_index == index) {
423 * max good bytes in this page
425 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
430 * force quit after adding this page
432 this_len = min(this_len, plen - loff);
436 partial[page_nr].offset = loff;
437 partial[page_nr].len = this_len;
445 * Release any pages at the end, if we quit early. 'page_nr' is how far
446 * we got, 'nr_pages' is how many pages are in the map.
448 while (page_nr < nr_pages)
449 page_cache_release(pages[page_nr++]);
450 in->f_ra.prev_index = index;
453 return splice_to_pipe(pipe, &spd);
459 * generic_file_splice_read - splice data from file to a pipe
460 * @in: file to splice from
461 * @ppos: position in @in
462 * @pipe: pipe to splice to
463 * @len: number of bytes to splice
464 * @flags: splice modifier flags
467 * Will read pages from given file and fill them into a pipe. Can be
468 * used as long as the address_space operations for the source implements
472 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
473 struct pipe_inode_info *pipe, size_t len,
480 isize = i_size_read(in->f_mapping->host);
481 if (unlikely(*ppos >= isize))
484 left = isize - *ppos;
485 if (unlikely(left < len))
490 while (len && !spliced) {
491 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
498 if (flags & SPLICE_F_NONBLOCK) {
515 EXPORT_SYMBOL(generic_file_splice_read);
518 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
519 * using sendpage(). Return the number of bytes sent.
521 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
522 struct pipe_buffer *buf, struct splice_desc *sd)
524 struct file *file = sd->u.file;
525 loff_t pos = sd->pos;
528 ret = buf->ops->confirm(pipe, buf);
530 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
532 ret = file->f_op->sendpage(file, buf->page, buf->offset,
533 sd->len, &pos, more);
540 * This is a little more tricky than the file -> pipe splicing. There are
541 * basically three cases:
543 * - Destination page already exists in the address space and there
544 * are users of it. For that case we have no other option that
545 * copying the data. Tough luck.
546 * - Destination page already exists in the address space, but there
547 * are no users of it. Make sure it's uptodate, then drop it. Fall
548 * through to last case.
549 * - Destination page does not exist, we can add the pipe page to
550 * the page cache and avoid the copy.
552 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
553 * sd->flags), we attempt to migrate pages from the pipe to the output
554 * file address space page cache. This is possible if no one else has
555 * the pipe page referenced outside of the pipe and page cache. If
556 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
557 * a new page in the output file page cache and fill/dirty that.
559 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
560 struct splice_desc *sd)
562 struct file *file = sd->u.file;
563 struct address_space *mapping = file->f_mapping;
564 unsigned int offset, this_len;
570 * make sure the data in this buffer is uptodate
572 ret = buf->ops->confirm(pipe, buf);
576 index = sd->pos >> PAGE_CACHE_SHIFT;
577 offset = sd->pos & ~PAGE_CACHE_MASK;
580 if (this_len + offset > PAGE_CACHE_SIZE)
581 this_len = PAGE_CACHE_SIZE - offset;
584 page = find_lock_page(mapping, index);
587 page = page_cache_alloc_cold(mapping);
592 * This will also lock the page
594 ret = add_to_page_cache_lru(page, mapping, index,
600 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
602 loff_t isize = i_size_read(mapping->host);
604 if (ret != AOP_TRUNCATED_PAGE)
606 page_cache_release(page);
607 if (ret == AOP_TRUNCATED_PAGE)
611 * prepare_write() may have instantiated a few blocks
612 * outside i_size. Trim these off again.
614 if (sd->pos + this_len > isize)
615 vmtruncate(mapping->host, isize);
620 if (buf->page != page) {
622 * Careful, ->map() uses KM_USER0!
624 char *src = buf->ops->map(pipe, buf, 1);
625 char *dst = kmap_atomic(page, KM_USER1);
627 memcpy(dst + offset, src + buf->offset, this_len);
628 flush_dcache_page(page);
629 kunmap_atomic(dst, KM_USER1);
630 buf->ops->unmap(pipe, buf, src);
633 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
635 if (ret == AOP_TRUNCATED_PAGE) {
636 page_cache_release(page);
642 * Partial write has happened, so 'ret' already initialized by
643 * number of bytes written, Where is nothing we have to do here.
648 * Return the number of bytes written and mark page as
649 * accessed, we are now done!
651 mark_page_accessed(page);
653 page_cache_release(page);
660 * __splice_from_pipe - splice data from a pipe to given actor
661 * @pipe: pipe to splice from
662 * @sd: information to @actor
663 * @actor: handler that splices the data
666 * This function does little more than loop over the pipe and call
667 * @actor to do the actual moving of a single struct pipe_buffer to
668 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
672 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
675 int ret, do_wakeup, err;
682 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
683 const struct pipe_buf_operations *ops = buf->ops;
686 if (sd->len > sd->total_len)
687 sd->len = sd->total_len;
689 err = actor(pipe, buf, sd);
691 if (!ret && err != -ENODATA)
703 sd->total_len -= err;
709 ops->release(pipe, buf);
710 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
724 if (!pipe->waiting_writers) {
729 if (sd->flags & SPLICE_F_NONBLOCK) {
735 if (signal_pending(current)) {
743 if (waitqueue_active(&pipe->wait))
744 wake_up_interruptible_sync(&pipe->wait);
745 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
754 if (waitqueue_active(&pipe->wait))
755 wake_up_interruptible(&pipe->wait);
756 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
761 EXPORT_SYMBOL(__splice_from_pipe);
764 * splice_from_pipe - splice data from a pipe to a file
765 * @pipe: pipe to splice from
766 * @out: file to splice to
767 * @ppos: position in @out
768 * @len: how many bytes to splice
769 * @flags: splice modifier flags
770 * @actor: handler that splices the data
773 * See __splice_from_pipe. This function locks the input and output inodes,
774 * otherwise it's identical to __splice_from_pipe().
777 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
778 loff_t *ppos, size_t len, unsigned int flags,
782 struct inode *inode = out->f_mapping->host;
783 struct splice_desc sd = {
791 * The actor worker might be calling ->prepare_write and
792 * ->commit_write. Most of the time, these expect i_mutex to
793 * be held. Since this may result in an ABBA deadlock with
794 * pipe->inode, we have to order lock acquiry here.
796 inode_double_lock(inode, pipe->inode);
797 ret = __splice_from_pipe(pipe, &sd, actor);
798 inode_double_unlock(inode, pipe->inode);
804 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
806 * @out: file to write to
807 * @ppos: position in @out
808 * @len: number of bytes to splice
809 * @flags: splice modifier flags
812 * Will either move or copy pages (determined by @flags options) from
813 * the given pipe inode to the given file. The caller is responsible
814 * for acquiring i_mutex on both inodes.
818 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
819 loff_t *ppos, size_t len, unsigned int flags)
821 struct address_space *mapping = out->f_mapping;
822 struct inode *inode = mapping->host;
823 struct splice_desc sd = {
832 err = remove_suid(out->f_path.dentry);
836 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
838 unsigned long nr_pages;
841 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
844 * If file or inode is SYNC and we actually wrote some data,
847 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
848 err = generic_osync_inode(inode, mapping,
849 OSYNC_METADATA|OSYNC_DATA);
854 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
860 EXPORT_SYMBOL(generic_file_splice_write_nolock);
863 * generic_file_splice_write - splice data from a pipe to a file
865 * @out: file to write to
866 * @ppos: position in @out
867 * @len: number of bytes to splice
868 * @flags: splice modifier flags
871 * Will either move or copy pages (determined by @flags options) from
872 * the given pipe inode to the given file.
876 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
877 loff_t *ppos, size_t len, unsigned int flags)
879 struct address_space *mapping = out->f_mapping;
880 struct inode *inode = mapping->host;
884 err = should_remove_suid(out->f_path.dentry);
886 mutex_lock(&inode->i_mutex);
887 err = __remove_suid(out->f_path.dentry, err);
888 mutex_unlock(&inode->i_mutex);
893 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
895 unsigned long nr_pages;
898 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
901 * If file or inode is SYNC and we actually wrote some data,
904 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
905 mutex_lock(&inode->i_mutex);
906 err = generic_osync_inode(inode, mapping,
907 OSYNC_METADATA|OSYNC_DATA);
908 mutex_unlock(&inode->i_mutex);
913 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
919 EXPORT_SYMBOL(generic_file_splice_write);
922 * generic_splice_sendpage - splice data from a pipe to a socket
923 * @pipe: pipe to splice from
924 * @out: socket to write to
925 * @ppos: position in @out
926 * @len: number of bytes to splice
927 * @flags: splice modifier flags
930 * Will send @len bytes from the pipe to a network socket. No data copying
934 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
935 loff_t *ppos, size_t len, unsigned int flags)
937 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
940 EXPORT_SYMBOL(generic_splice_sendpage);
943 * Attempt to initiate a splice from pipe to file.
945 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
946 loff_t *ppos, size_t len, unsigned int flags)
950 if (unlikely(!out->f_op || !out->f_op->splice_write))
953 if (unlikely(!(out->f_mode & FMODE_WRITE)))
956 ret = rw_verify_area(WRITE, out, ppos, len);
957 if (unlikely(ret < 0))
960 ret = security_file_permission(out, MAY_WRITE);
961 if (unlikely(ret < 0))
964 return out->f_op->splice_write(pipe, out, ppos, len, flags);
968 * Attempt to initiate a splice from a file to a pipe.
970 static long do_splice_to(struct file *in, loff_t *ppos,
971 struct pipe_inode_info *pipe, size_t len,
976 if (unlikely(!in->f_op || !in->f_op->splice_read))
979 if (unlikely(!(in->f_mode & FMODE_READ)))
982 ret = rw_verify_area(READ, in, ppos, len);
983 if (unlikely(ret < 0))
986 ret = security_file_permission(in, MAY_READ);
987 if (unlikely(ret < 0))
990 return in->f_op->splice_read(in, ppos, pipe, len, flags);
994 * splice_direct_to_actor - splices data directly between two non-pipes
995 * @in: file to splice from
996 * @sd: actor information on where to splice to
997 * @actor: handles the data splicing
1000 * This is a special case helper to splice directly between two
1001 * points, without requiring an explicit pipe. Internally an allocated
1002 * pipe is cached in the process, and reused during the life time of
1006 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1007 splice_direct_actor *actor)
1009 struct pipe_inode_info *pipe;
1016 * We require the input being a regular file, as we don't want to
1017 * randomly drop data for eg socket -> socket splicing. Use the
1018 * piped splicing for that!
1020 i_mode = in->f_path.dentry->d_inode->i_mode;
1021 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1025 * neither in nor out is a pipe, setup an internal pipe attached to
1026 * 'out' and transfer the wanted data from 'in' to 'out' through that
1028 pipe = current->splice_pipe;
1029 if (unlikely(!pipe)) {
1030 pipe = alloc_pipe_info(NULL);
1035 * We don't have an immediate reader, but we'll read the stuff
1036 * out of the pipe right after the splice_to_pipe(). So set
1037 * PIPE_READERS appropriately.
1041 current->splice_pipe = pipe;
1049 len = sd->total_len;
1053 * Don't block on output, we have to drain the direct pipe.
1055 sd->flags &= ~SPLICE_F_NONBLOCK;
1059 loff_t pos = sd->pos;
1061 ret = do_splice_to(in, &pos, pipe, len, flags);
1062 if (unlikely(ret <= 0))
1066 sd->total_len = read_len;
1069 * NOTE: nonblocking mode only applies to the input. We
1070 * must not do the output in nonblocking mode as then we
1071 * could get stuck data in the internal pipe:
1073 ret = actor(pipe, sd);
1074 if (unlikely(ret <= 0))
1085 pipe->nrbufs = pipe->curbuf = 0;
1090 * If we did an incomplete transfer we must release
1091 * the pipe buffers in question:
1093 for (i = 0; i < PIPE_BUFFERS; i++) {
1094 struct pipe_buffer *buf = pipe->bufs + i;
1097 buf->ops->release(pipe, buf);
1101 pipe->nrbufs = pipe->curbuf = 0;
1104 * If we transferred some data, return the number of bytes:
1112 EXPORT_SYMBOL(splice_direct_to_actor);
1114 static int direct_splice_actor(struct pipe_inode_info *pipe,
1115 struct splice_desc *sd)
1117 struct file *file = sd->u.file;
1119 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1123 * do_splice_direct - splices data directly between two files
1124 * @in: file to splice from
1125 * @ppos: input file offset
1126 * @out: file to splice to
1127 * @len: number of bytes to splice
1128 * @flags: splice modifier flags
1131 * For use by do_sendfile(). splice can easily emulate sendfile, but
1132 * doing it in the application would incur an extra system call
1133 * (splice in + splice out, as compared to just sendfile()). So this helper
1134 * can splice directly through a process-private pipe.
1137 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1138 size_t len, unsigned int flags)
1140 struct splice_desc sd = {
1149 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1157 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1158 * location, so checking ->i_pipe is not enough to verify that this is a
1161 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1163 if (S_ISFIFO(inode->i_mode))
1164 return inode->i_pipe;
1170 * Determine where to splice to/from.
1172 static long do_splice(struct file *in, loff_t __user *off_in,
1173 struct file *out, loff_t __user *off_out,
1174 size_t len, unsigned int flags)
1176 struct pipe_inode_info *pipe;
1177 loff_t offset, *off;
1180 pipe = pipe_info(in->f_path.dentry->d_inode);
1185 if (out->f_op->llseek == no_llseek)
1187 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1193 ret = do_splice_from(pipe, out, off, len, flags);
1195 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1201 pipe = pipe_info(out->f_path.dentry->d_inode);
1206 if (in->f_op->llseek == no_llseek)
1208 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1214 ret = do_splice_to(in, off, pipe, len, flags);
1216 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1226 * Map an iov into an array of pages and offset/length tupples. With the
1227 * partial_page structure, we can map several non-contiguous ranges into
1228 * our ones pages[] map instead of splitting that operation into pieces.
1229 * Could easily be exported as a generic helper for other users, in which
1230 * case one would probably want to add a 'max_nr_pages' parameter as well.
1232 static int get_iovec_page_array(const struct iovec __user *iov,
1233 unsigned int nr_vecs, struct page **pages,
1234 struct partial_page *partial, int aligned)
1236 int buffers = 0, error = 0;
1239 * It's ok to take the mmap_sem for reading, even
1240 * across a "get_user()".
1242 down_read(¤t->mm->mmap_sem);
1245 unsigned long off, npages;
1251 * Get user address base and length for this iovec.
1253 error = get_user(base, &iov->iov_base);
1254 if (unlikely(error))
1256 error = get_user(len, &iov->iov_len);
1257 if (unlikely(error))
1261 * Sanity check this iovec. 0 read succeeds.
1266 if (unlikely(!base))
1270 * Get this base offset and number of pages, then map
1271 * in the user pages.
1273 off = (unsigned long) base & ~PAGE_MASK;
1276 * If asked for alignment, the offset must be zero and the
1277 * length a multiple of the PAGE_SIZE.
1280 if (aligned && (off || len & ~PAGE_MASK))
1283 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1284 if (npages > PIPE_BUFFERS - buffers)
1285 npages = PIPE_BUFFERS - buffers;
1287 error = get_user_pages(current, current->mm,
1288 (unsigned long) base, npages, 0, 0,
1289 &pages[buffers], NULL);
1291 if (unlikely(error <= 0))
1295 * Fill this contiguous range into the partial page map.
1297 for (i = 0; i < error; i++) {
1298 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1300 partial[buffers].offset = off;
1301 partial[buffers].len = plen;
1309 * We didn't complete this iov, stop here since it probably
1310 * means we have to move some of this into a pipe to
1311 * be able to continue.
1317 * Don't continue if we mapped fewer pages than we asked for,
1318 * or if we mapped the max number of pages that we have
1321 if (error < npages || buffers == PIPE_BUFFERS)
1328 up_read(¤t->mm->mmap_sem);
1336 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1337 struct splice_desc *sd)
1342 ret = buf->ops->confirm(pipe, buf);
1347 * See if we can use the atomic maps, by prefaulting in the
1348 * pages and doing an atomic copy
1350 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1351 src = buf->ops->map(pipe, buf, 1);
1352 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1354 buf->ops->unmap(pipe, buf, src);
1362 * No dice, use slow non-atomic map and copy
1364 src = buf->ops->map(pipe, buf, 0);
1367 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1372 sd->u.userptr += ret;
1373 buf->ops->unmap(pipe, buf, src);
1378 * For lack of a better implementation, implement vmsplice() to userspace
1379 * as a simple copy of the pipes pages to the user iov.
1381 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1382 unsigned long nr_segs, unsigned int flags)
1384 struct pipe_inode_info *pipe;
1385 struct splice_desc sd;
1390 pipe = pipe_info(file->f_path.dentry->d_inode);
1395 mutex_lock(&pipe->inode->i_mutex);
1403 * Get user address base and length for this iovec.
1405 error = get_user(base, &iov->iov_base);
1406 if (unlikely(error))
1408 error = get_user(len, &iov->iov_len);
1409 if (unlikely(error))
1413 * Sanity check this iovec. 0 read succeeds.
1417 if (unlikely(!base)) {
1425 sd.u.userptr = base;
1428 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1446 mutex_unlock(&pipe->inode->i_mutex);
1455 * vmsplice splices a user address range into a pipe. It can be thought of
1456 * as splice-from-memory, where the regular splice is splice-from-file (or
1457 * to file). In both cases the output is a pipe, naturally.
1459 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1460 unsigned long nr_segs, unsigned int flags)
1462 struct pipe_inode_info *pipe;
1463 struct page *pages[PIPE_BUFFERS];
1464 struct partial_page partial[PIPE_BUFFERS];
1465 struct splice_pipe_desc spd = {
1469 .ops = &user_page_pipe_buf_ops,
1472 pipe = pipe_info(file->f_path.dentry->d_inode);
1476 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1477 flags & SPLICE_F_GIFT);
1478 if (spd.nr_pages <= 0)
1479 return spd.nr_pages;
1481 return splice_to_pipe(pipe, &spd);
1485 * Note that vmsplice only really supports true splicing _from_ user memory
1486 * to a pipe, not the other way around. Splicing from user memory is a simple
1487 * operation that can be supported without any funky alignment restrictions
1488 * or nasty vm tricks. We simply map in the user memory and fill them into
1489 * a pipe. The reverse isn't quite as easy, though. There are two possible
1490 * solutions for that:
1492 * - memcpy() the data internally, at which point we might as well just
1493 * do a regular read() on the buffer anyway.
1494 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1495 * has restriction limitations on both ends of the pipe).
1497 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1500 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1501 unsigned long nr_segs, unsigned int flags)
1507 if (unlikely(nr_segs > UIO_MAXIOV))
1509 else if (unlikely(!nr_segs))
1513 file = fget_light(fd, &fput);
1515 if (file->f_mode & FMODE_WRITE)
1516 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1517 else if (file->f_mode & FMODE_READ)
1518 error = vmsplice_to_user(file, iov, nr_segs, flags);
1520 fput_light(file, fput);
1526 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1527 int fd_out, loff_t __user *off_out,
1528 size_t len, unsigned int flags)
1531 struct file *in, *out;
1532 int fput_in, fput_out;
1538 in = fget_light(fd_in, &fput_in);
1540 if (in->f_mode & FMODE_READ) {
1541 out = fget_light(fd_out, &fput_out);
1543 if (out->f_mode & FMODE_WRITE)
1544 error = do_splice(in, off_in,
1547 fput_light(out, fput_out);
1551 fput_light(in, fput_in);
1558 * Make sure there's data to read. Wait for input if we can, otherwise
1559 * return an appropriate error.
1561 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1566 * Check ->nrbufs without the inode lock first. This function
1567 * is speculative anyways, so missing one is ok.
1573 mutex_lock(&pipe->inode->i_mutex);
1575 while (!pipe->nrbufs) {
1576 if (signal_pending(current)) {
1582 if (!pipe->waiting_writers) {
1583 if (flags & SPLICE_F_NONBLOCK) {
1591 mutex_unlock(&pipe->inode->i_mutex);
1596 * Make sure there's writeable room. Wait for room if we can, otherwise
1597 * return an appropriate error.
1599 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1604 * Check ->nrbufs without the inode lock first. This function
1605 * is speculative anyways, so missing one is ok.
1607 if (pipe->nrbufs < PIPE_BUFFERS)
1611 mutex_lock(&pipe->inode->i_mutex);
1613 while (pipe->nrbufs >= PIPE_BUFFERS) {
1614 if (!pipe->readers) {
1615 send_sig(SIGPIPE, current, 0);
1619 if (flags & SPLICE_F_NONBLOCK) {
1623 if (signal_pending(current)) {
1627 pipe->waiting_writers++;
1629 pipe->waiting_writers--;
1632 mutex_unlock(&pipe->inode->i_mutex);
1637 * Link contents of ipipe to opipe.
1639 static int link_pipe(struct pipe_inode_info *ipipe,
1640 struct pipe_inode_info *opipe,
1641 size_t len, unsigned int flags)
1643 struct pipe_buffer *ibuf, *obuf;
1644 int ret = 0, i = 0, nbuf;
1647 * Potential ABBA deadlock, work around it by ordering lock
1648 * grabbing by inode address. Otherwise two different processes
1649 * could deadlock (one doing tee from A -> B, the other from B -> A).
1651 inode_double_lock(ipipe->inode, opipe->inode);
1654 if (!opipe->readers) {
1655 send_sig(SIGPIPE, current, 0);
1662 * If we have iterated all input buffers or ran out of
1663 * output room, break.
1665 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1668 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1669 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1672 * Get a reference to this pipe buffer,
1673 * so we can copy the contents over.
1675 ibuf->ops->get(ipipe, ibuf);
1677 obuf = opipe->bufs + nbuf;
1681 * Don't inherit the gift flag, we need to
1682 * prevent multiple steals of this page.
1684 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1686 if (obuf->len > len)
1695 inode_double_unlock(ipipe->inode, opipe->inode);
1698 * If we put data in the output pipe, wakeup any potential readers.
1702 if (waitqueue_active(&opipe->wait))
1703 wake_up_interruptible(&opipe->wait);
1704 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1711 * This is a tee(1) implementation that works on pipes. It doesn't copy
1712 * any data, it simply references the 'in' pages on the 'out' pipe.
1713 * The 'flags' used are the SPLICE_F_* variants, currently the only
1714 * applicable one is SPLICE_F_NONBLOCK.
1716 static long do_tee(struct file *in, struct file *out, size_t len,
1719 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1720 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1724 * Duplicate the contents of ipipe to opipe without actually
1727 if (ipipe && opipe && ipipe != opipe) {
1729 * Keep going, unless we encounter an error. The ipipe/opipe
1730 * ordering doesn't really matter.
1732 ret = link_ipipe_prep(ipipe, flags);
1734 ret = link_opipe_prep(opipe, flags);
1736 ret = link_pipe(ipipe, opipe, len, flags);
1737 if (!ret && (flags & SPLICE_F_NONBLOCK))
1746 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1755 in = fget_light(fdin, &fput_in);
1757 if (in->f_mode & FMODE_READ) {
1759 struct file *out = fget_light(fdout, &fput_out);
1762 if (out->f_mode & FMODE_WRITE)
1763 error = do_tee(in, out, len, flags);
1764 fput_light(out, fput_out);
1767 fput_light(in, fput_in);