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) && !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.
83 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
84 struct pipe_buffer *buf)
86 page_cache_release(buf->page);
87 buf->flags &= ~PIPE_BUF_FLAG_LRU;
91 * Check whether the contents of buf is OK to access. Since the content
92 * is a page cache page, IO may be in flight.
94 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
95 struct pipe_buffer *buf)
97 struct page *page = buf->page;
100 if (!PageUptodate(page)) {
104 * Page got truncated/unhashed. This will cause a 0-byte
105 * splice, if this is the first page.
107 if (!page->mapping) {
113 * Uh oh, read-error from disk.
115 if (!PageUptodate(page)) {
121 * Page is ok afterall, we are done.
132 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
134 .map = generic_pipe_buf_map,
135 .unmap = generic_pipe_buf_unmap,
136 .confirm = page_cache_pipe_buf_confirm,
137 .release = page_cache_pipe_buf_release,
138 .steal = page_cache_pipe_buf_steal,
139 .get = generic_pipe_buf_get,
142 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
143 struct pipe_buffer *buf)
145 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
148 buf->flags |= PIPE_BUF_FLAG_LRU;
149 return generic_pipe_buf_steal(pipe, buf);
152 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
154 .map = generic_pipe_buf_map,
155 .unmap = generic_pipe_buf_unmap,
156 .confirm = generic_pipe_buf_confirm,
157 .release = page_cache_pipe_buf_release,
158 .steal = user_page_pipe_buf_steal,
159 .get = generic_pipe_buf_get,
163 * splice_to_pipe - fill passed data into a pipe
164 * @pipe: pipe to fill
168 * @spd contains a map of pages and len/offset tuples, along with
169 * the struct pipe_buf_operations associated with these pages. This
170 * function will link that data to the pipe.
173 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
174 struct splice_pipe_desc *spd)
176 unsigned int spd_pages = spd->nr_pages;
177 int ret, do_wakeup, page_nr;
184 mutex_lock(&pipe->inode->i_mutex);
187 if (!pipe->readers) {
188 send_sig(SIGPIPE, current, 0);
194 if (pipe->nrbufs < PIPE_BUFFERS) {
195 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
196 struct pipe_buffer *buf = pipe->bufs + newbuf;
198 buf->page = spd->pages[page_nr];
199 buf->offset = spd->partial[page_nr].offset;
200 buf->len = spd->partial[page_nr].len;
201 buf->private = spd->partial[page_nr].private;
203 if (spd->flags & SPLICE_F_GIFT)
204 buf->flags |= PIPE_BUF_FLAG_GIFT;
213 if (!--spd->nr_pages)
215 if (pipe->nrbufs < PIPE_BUFFERS)
221 if (spd->flags & SPLICE_F_NONBLOCK) {
227 if (signal_pending(current)) {
235 if (waitqueue_active(&pipe->wait))
236 wake_up_interruptible_sync(&pipe->wait);
237 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
241 pipe->waiting_writers++;
243 pipe->waiting_writers--;
247 mutex_unlock(&pipe->inode->i_mutex);
251 if (waitqueue_active(&pipe->wait))
252 wake_up_interruptible(&pipe->wait);
253 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 __generic_file_splice_read(struct file *in, loff_t *ppos,
270 struct pipe_inode_info *pipe, size_t len,
273 struct address_space *mapping = in->f_mapping;
274 unsigned int loff, nr_pages, req_pages;
275 struct page *pages[PIPE_BUFFERS];
276 struct partial_page partial[PIPE_BUFFERS];
278 pgoff_t index, end_index;
281 struct splice_pipe_desc spd = {
285 .ops = &page_cache_pipe_buf_ops,
286 .spd_release = spd_release_page,
289 index = *ppos >> PAGE_CACHE_SHIFT;
290 loff = *ppos & ~PAGE_CACHE_MASK;
291 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
292 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
295 * Lookup the (hopefully) full range of pages we need.
297 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
298 index += spd.nr_pages;
301 * If find_get_pages_contig() returned fewer pages than we needed,
302 * readahead/allocate the rest and fill in the holes.
304 if (spd.nr_pages < nr_pages)
305 page_cache_sync_readahead(mapping, &in->f_ra, in,
306 index, req_pages - spd.nr_pages);
309 while (spd.nr_pages < nr_pages) {
311 * Page could be there, find_get_pages_contig() breaks on
314 page = find_get_page(mapping, index);
317 * page didn't exist, allocate one.
319 page = page_cache_alloc_cold(mapping);
323 error = add_to_page_cache_lru(page, mapping, index,
324 mapping_gfp_mask(mapping));
325 if (unlikely(error)) {
326 page_cache_release(page);
327 if (error == -EEXIST)
332 * add_to_page_cache() locks the page, unlock it
333 * to avoid convoluting the logic below even more.
338 pages[spd.nr_pages++] = page;
343 * Now loop over the map and see if we need to start IO on any
344 * pages, fill in the partial map, etc.
346 index = *ppos >> PAGE_CACHE_SHIFT;
347 nr_pages = spd.nr_pages;
349 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
350 unsigned int this_len;
356 * this_len is the max we'll use from this page
358 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
359 page = pages[page_nr];
361 if (PageReadahead(page))
362 page_cache_async_readahead(mapping, &in->f_ra, in,
363 page, index, req_pages - page_nr);
366 * If the page isn't uptodate, we may need to start io on it
368 if (!PageUptodate(page)) {
370 * If in nonblock mode then dont block on waiting
371 * for an in-flight io page
373 if (flags & SPLICE_F_NONBLOCK) {
374 if (!trylock_page(page)) {
382 * Page was truncated, or invalidated by the
383 * filesystem. Redo the find/create, but this time the
384 * page is kept locked, so there's no chance of another
385 * race with truncate/invalidate.
387 if (!page->mapping) {
389 page = find_or_create_page(mapping, index,
390 mapping_gfp_mask(mapping));
396 page_cache_release(pages[page_nr]);
397 pages[page_nr] = page;
400 * page was already under io and is now done, great
402 if (PageUptodate(page)) {
408 * need to read in the page
410 error = mapping->a_ops->readpage(in, page);
411 if (unlikely(error)) {
413 * We really should re-lookup the page here,
414 * but it complicates things a lot. Instead
415 * lets just do what we already stored, and
416 * we'll get it the next time we are called.
418 if (error == AOP_TRUNCATED_PAGE)
426 * i_size must be checked after PageUptodate.
428 isize = i_size_read(mapping->host);
429 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
430 if (unlikely(!isize || index > end_index))
434 * if this is the last page, see if we need to shrink
435 * the length and stop
437 if (end_index == index) {
441 * max good bytes in this page
443 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
448 * force quit after adding this page
450 this_len = min(this_len, plen - loff);
454 partial[page_nr].offset = loff;
455 partial[page_nr].len = this_len;
463 * Release any pages at the end, if we quit early. 'page_nr' is how far
464 * we got, 'nr_pages' is how many pages are in the map.
466 while (page_nr < nr_pages)
467 page_cache_release(pages[page_nr++]);
468 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
471 return splice_to_pipe(pipe, &spd);
477 * generic_file_splice_read - splice data from file to a pipe
478 * @in: file to splice from
479 * @ppos: position in @in
480 * @pipe: pipe to splice to
481 * @len: number of bytes to splice
482 * @flags: splice modifier flags
485 * Will read pages from given file and fill them into a pipe. Can be
486 * used as long as the address_space operations for the source implements
490 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
491 struct pipe_inode_info *pipe, size_t len,
497 isize = i_size_read(in->f_mapping->host);
498 if (unlikely(*ppos >= isize))
501 left = isize - *ppos;
502 if (unlikely(left < len))
505 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
512 EXPORT_SYMBOL(generic_file_splice_read);
515 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
516 * using sendpage(). Return the number of bytes sent.
518 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
519 struct pipe_buffer *buf, struct splice_desc *sd)
521 struct file *file = sd->u.file;
522 loff_t pos = sd->pos;
525 ret = buf->ops->confirm(pipe, buf);
527 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
529 ret = file->f_op->sendpage(file, buf->page, buf->offset,
530 sd->len, &pos, more);
537 * This is a little more tricky than the file -> pipe splicing. There are
538 * basically three cases:
540 * - Destination page already exists in the address space and there
541 * are users of it. For that case we have no other option that
542 * copying the data. Tough luck.
543 * - Destination page already exists in the address space, but there
544 * are no users of it. Make sure it's uptodate, then drop it. Fall
545 * through to last case.
546 * - Destination page does not exist, we can add the pipe page to
547 * the page cache and avoid the copy.
549 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
550 * sd->flags), we attempt to migrate pages from the pipe to the output
551 * file address space page cache. This is possible if no one else has
552 * the pipe page referenced outside of the pipe and page cache. If
553 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
554 * a new page in the output file page cache and fill/dirty that.
556 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
557 struct splice_desc *sd)
559 struct file *file = sd->u.file;
560 struct address_space *mapping = file->f_mapping;
561 unsigned int offset, this_len;
567 * make sure the data in this buffer is uptodate
569 ret = buf->ops->confirm(pipe, buf);
573 offset = sd->pos & ~PAGE_CACHE_MASK;
576 if (this_len + offset > PAGE_CACHE_SIZE)
577 this_len = PAGE_CACHE_SIZE - offset;
579 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
580 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
584 if (buf->page != page) {
586 * Careful, ->map() uses KM_USER0!
588 char *src = buf->ops->map(pipe, buf, 1);
589 char *dst = kmap_atomic(page, KM_USER1);
591 memcpy(dst + offset, src + buf->offset, this_len);
592 flush_dcache_page(page);
593 kunmap_atomic(dst, KM_USER1);
594 buf->ops->unmap(pipe, buf, src);
596 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
603 * __splice_from_pipe - splice data from a pipe to given actor
604 * @pipe: pipe to splice from
605 * @sd: information to @actor
606 * @actor: handler that splices the data
609 * This function does little more than loop over the pipe and call
610 * @actor to do the actual moving of a single struct pipe_buffer to
611 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
615 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
618 int ret, do_wakeup, err;
625 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
626 const struct pipe_buf_operations *ops = buf->ops;
629 if (sd->len > sd->total_len)
630 sd->len = sd->total_len;
632 err = actor(pipe, buf, sd);
634 if (!ret && err != -ENODATA)
646 sd->total_len -= err;
652 ops->release(pipe, buf);
653 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
667 if (!pipe->waiting_writers) {
672 if (sd->flags & SPLICE_F_NONBLOCK) {
678 if (signal_pending(current)) {
686 if (waitqueue_active(&pipe->wait))
687 wake_up_interruptible_sync(&pipe->wait);
688 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
697 if (waitqueue_active(&pipe->wait))
698 wake_up_interruptible(&pipe->wait);
699 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
704 EXPORT_SYMBOL(__splice_from_pipe);
707 * splice_from_pipe - splice data from a pipe to a file
708 * @pipe: pipe to splice from
709 * @out: file to splice to
710 * @ppos: position in @out
711 * @len: how many bytes to splice
712 * @flags: splice modifier flags
713 * @actor: handler that splices the data
716 * See __splice_from_pipe. This function locks the input and output inodes,
717 * otherwise it's identical to __splice_from_pipe().
720 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
721 loff_t *ppos, size_t len, unsigned int flags,
725 struct inode *inode = out->f_mapping->host;
726 struct splice_desc sd = {
734 * The actor worker might be calling ->write_begin and
735 * ->write_end. Most of the time, these expect i_mutex to
736 * be held. Since this may result in an ABBA deadlock with
737 * pipe->inode, we have to order lock acquiry here.
739 inode_double_lock(inode, pipe->inode);
740 ret = __splice_from_pipe(pipe, &sd, actor);
741 inode_double_unlock(inode, pipe->inode);
747 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
749 * @out: file to write to
750 * @ppos: position in @out
751 * @len: number of bytes to splice
752 * @flags: splice modifier flags
755 * Will either move or copy pages (determined by @flags options) from
756 * the given pipe inode to the given file. The caller is responsible
757 * for acquiring i_mutex on both inodes.
761 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
762 loff_t *ppos, size_t len, unsigned int flags)
764 struct address_space *mapping = out->f_mapping;
765 struct inode *inode = mapping->host;
766 struct splice_desc sd = {
775 err = file_remove_suid(out);
779 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
781 unsigned long nr_pages;
784 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
787 * If file or inode is SYNC and we actually wrote some data,
790 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
791 err = generic_osync_inode(inode, mapping,
792 OSYNC_METADATA|OSYNC_DATA);
797 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
803 EXPORT_SYMBOL(generic_file_splice_write_nolock);
806 * generic_file_splice_write - splice data from a pipe to a file
808 * @out: file to write to
809 * @ppos: position in @out
810 * @len: number of bytes to splice
811 * @flags: splice modifier flags
814 * Will either move or copy pages (determined by @flags options) from
815 * the given pipe inode to the given file.
819 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
820 loff_t *ppos, size_t len, unsigned int flags)
822 struct address_space *mapping = out->f_mapping;
823 struct inode *inode = mapping->host;
824 struct splice_desc sd = {
832 inode_double_lock(inode, pipe->inode);
833 ret = file_remove_suid(out);
835 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
836 inode_double_unlock(inode, pipe->inode);
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))) {
850 mutex_lock(&inode->i_mutex);
851 err = generic_osync_inode(inode, mapping,
852 OSYNC_METADATA|OSYNC_DATA);
853 mutex_unlock(&inode->i_mutex);
858 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
864 EXPORT_SYMBOL(generic_file_splice_write);
867 * generic_splice_sendpage - splice data from a pipe to a socket
868 * @pipe: pipe to splice from
869 * @out: socket to write to
870 * @ppos: position in @out
871 * @len: number of bytes to splice
872 * @flags: splice modifier flags
875 * Will send @len bytes from the pipe to a network socket. No data copying
879 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
880 loff_t *ppos, size_t len, unsigned int flags)
882 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
885 EXPORT_SYMBOL(generic_splice_sendpage);
888 * Attempt to initiate a splice from pipe to file.
890 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
891 loff_t *ppos, size_t len, unsigned int flags)
895 if (unlikely(!out->f_op || !out->f_op->splice_write))
898 if (unlikely(!(out->f_mode & FMODE_WRITE)))
901 if (unlikely(out->f_flags & O_APPEND))
904 ret = rw_verify_area(WRITE, out, ppos, len);
905 if (unlikely(ret < 0))
908 return out->f_op->splice_write(pipe, out, ppos, len, flags);
912 * Attempt to initiate a splice from a file to a pipe.
914 static long do_splice_to(struct file *in, loff_t *ppos,
915 struct pipe_inode_info *pipe, size_t len,
920 if (unlikely(!in->f_op || !in->f_op->splice_read))
923 if (unlikely(!(in->f_mode & FMODE_READ)))
926 ret = rw_verify_area(READ, in, ppos, len);
927 if (unlikely(ret < 0))
930 return in->f_op->splice_read(in, ppos, pipe, len, flags);
934 * splice_direct_to_actor - splices data directly between two non-pipes
935 * @in: file to splice from
936 * @sd: actor information on where to splice to
937 * @actor: handles the data splicing
940 * This is a special case helper to splice directly between two
941 * points, without requiring an explicit pipe. Internally an allocated
942 * pipe is cached in the process, and reused during the lifetime of
946 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
947 splice_direct_actor *actor)
949 struct pipe_inode_info *pipe;
956 * We require the input being a regular file, as we don't want to
957 * randomly drop data for eg socket -> socket splicing. Use the
958 * piped splicing for that!
960 i_mode = in->f_path.dentry->d_inode->i_mode;
961 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
965 * neither in nor out is a pipe, setup an internal pipe attached to
966 * 'out' and transfer the wanted data from 'in' to 'out' through that
968 pipe = current->splice_pipe;
969 if (unlikely(!pipe)) {
970 pipe = alloc_pipe_info(NULL);
975 * We don't have an immediate reader, but we'll read the stuff
976 * out of the pipe right after the splice_to_pipe(). So set
977 * PIPE_READERS appropriately.
981 current->splice_pipe = pipe;
993 * Don't block on output, we have to drain the direct pipe.
995 sd->flags &= ~SPLICE_F_NONBLOCK;
999 loff_t pos = sd->pos, prev_pos = pos;
1001 ret = do_splice_to(in, &pos, pipe, len, flags);
1002 if (unlikely(ret <= 0))
1006 sd->total_len = read_len;
1009 * NOTE: nonblocking mode only applies to the input. We
1010 * must not do the output in nonblocking mode as then we
1011 * could get stuck data in the internal pipe:
1013 ret = actor(pipe, sd);
1014 if (unlikely(ret <= 0)) {
1023 if (ret < read_len) {
1024 sd->pos = prev_pos + ret;
1030 pipe->nrbufs = pipe->curbuf = 0;
1036 * If we did an incomplete transfer we must release
1037 * the pipe buffers in question:
1039 for (i = 0; i < PIPE_BUFFERS; i++) {
1040 struct pipe_buffer *buf = pipe->bufs + i;
1043 buf->ops->release(pipe, buf);
1053 EXPORT_SYMBOL(splice_direct_to_actor);
1055 static int direct_splice_actor(struct pipe_inode_info *pipe,
1056 struct splice_desc *sd)
1058 struct file *file = sd->u.file;
1060 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1064 * do_splice_direct - splices data directly between two files
1065 * @in: file to splice from
1066 * @ppos: input file offset
1067 * @out: file to splice to
1068 * @len: number of bytes to splice
1069 * @flags: splice modifier flags
1072 * For use by do_sendfile(). splice can easily emulate sendfile, but
1073 * doing it in the application would incur an extra system call
1074 * (splice in + splice out, as compared to just sendfile()). So this helper
1075 * can splice directly through a process-private pipe.
1078 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1079 size_t len, unsigned int flags)
1081 struct splice_desc sd = {
1090 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1098 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1099 * location, so checking ->i_pipe is not enough to verify that this is a
1102 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1104 if (S_ISFIFO(inode->i_mode))
1105 return inode->i_pipe;
1111 * Determine where to splice to/from.
1113 static long do_splice(struct file *in, loff_t __user *off_in,
1114 struct file *out, loff_t __user *off_out,
1115 size_t len, unsigned int flags)
1117 struct pipe_inode_info *pipe;
1118 loff_t offset, *off;
1121 pipe = pipe_info(in->f_path.dentry->d_inode);
1126 if (out->f_op->llseek == no_llseek)
1128 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1134 ret = do_splice_from(pipe, out, off, len, flags);
1136 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1142 pipe = pipe_info(out->f_path.dentry->d_inode);
1147 if (in->f_op->llseek == no_llseek)
1149 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1155 ret = do_splice_to(in, off, pipe, len, flags);
1157 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1167 * Map an iov into an array of pages and offset/length tupples. With the
1168 * partial_page structure, we can map several non-contiguous ranges into
1169 * our ones pages[] map instead of splitting that operation into pieces.
1170 * Could easily be exported as a generic helper for other users, in which
1171 * case one would probably want to add a 'max_nr_pages' parameter as well.
1173 static int get_iovec_page_array(const struct iovec __user *iov,
1174 unsigned int nr_vecs, struct page **pages,
1175 struct partial_page *partial, int aligned)
1177 int buffers = 0, error = 0;
1180 unsigned long off, npages;
1187 if (copy_from_user(&entry, iov, sizeof(entry)))
1190 base = entry.iov_base;
1191 len = entry.iov_len;
1194 * Sanity check this iovec. 0 read succeeds.
1200 if (!access_ok(VERIFY_READ, base, len))
1204 * Get this base offset and number of pages, then map
1205 * in the user pages.
1207 off = (unsigned long) base & ~PAGE_MASK;
1210 * If asked for alignment, the offset must be zero and the
1211 * length a multiple of the PAGE_SIZE.
1214 if (aligned && (off || len & ~PAGE_MASK))
1217 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1218 if (npages > PIPE_BUFFERS - buffers)
1219 npages = PIPE_BUFFERS - buffers;
1221 error = get_user_pages_fast((unsigned long)base, npages,
1222 0, &pages[buffers]);
1224 if (unlikely(error <= 0))
1228 * Fill this contiguous range into the partial page map.
1230 for (i = 0; i < error; i++) {
1231 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1233 partial[buffers].offset = off;
1234 partial[buffers].len = plen;
1242 * We didn't complete this iov, stop here since it probably
1243 * means we have to move some of this into a pipe to
1244 * be able to continue.
1250 * Don't continue if we mapped fewer pages than we asked for,
1251 * or if we mapped the max number of pages that we have
1254 if (error < npages || buffers == PIPE_BUFFERS)
1267 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1268 struct splice_desc *sd)
1273 ret = buf->ops->confirm(pipe, buf);
1278 * See if we can use the atomic maps, by prefaulting in the
1279 * pages and doing an atomic copy
1281 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1282 src = buf->ops->map(pipe, buf, 1);
1283 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1285 buf->ops->unmap(pipe, buf, src);
1293 * No dice, use slow non-atomic map and copy
1295 src = buf->ops->map(pipe, buf, 0);
1298 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1301 buf->ops->unmap(pipe, buf, src);
1304 sd->u.userptr += ret;
1309 * For lack of a better implementation, implement vmsplice() to userspace
1310 * as a simple copy of the pipes pages to the user iov.
1312 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1313 unsigned long nr_segs, unsigned int flags)
1315 struct pipe_inode_info *pipe;
1316 struct splice_desc sd;
1321 pipe = pipe_info(file->f_path.dentry->d_inode);
1326 mutex_lock(&pipe->inode->i_mutex);
1334 * Get user address base and length for this iovec.
1336 error = get_user(base, &iov->iov_base);
1337 if (unlikely(error))
1339 error = get_user(len, &iov->iov_len);
1340 if (unlikely(error))
1344 * Sanity check this iovec. 0 read succeeds.
1348 if (unlikely(!base)) {
1353 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1361 sd.u.userptr = base;
1364 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1382 mutex_unlock(&pipe->inode->i_mutex);
1391 * vmsplice splices a user address range into a pipe. It can be thought of
1392 * as splice-from-memory, where the regular splice is splice-from-file (or
1393 * to file). In both cases the output is a pipe, naturally.
1395 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1396 unsigned long nr_segs, unsigned int flags)
1398 struct pipe_inode_info *pipe;
1399 struct page *pages[PIPE_BUFFERS];
1400 struct partial_page partial[PIPE_BUFFERS];
1401 struct splice_pipe_desc spd = {
1405 .ops = &user_page_pipe_buf_ops,
1406 .spd_release = spd_release_page,
1409 pipe = pipe_info(file->f_path.dentry->d_inode);
1413 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1414 flags & SPLICE_F_GIFT);
1415 if (spd.nr_pages <= 0)
1416 return spd.nr_pages;
1418 return splice_to_pipe(pipe, &spd);
1422 * Note that vmsplice only really supports true splicing _from_ user memory
1423 * to a pipe, not the other way around. Splicing from user memory is a simple
1424 * operation that can be supported without any funky alignment restrictions
1425 * or nasty vm tricks. We simply map in the user memory and fill them into
1426 * a pipe. The reverse isn't quite as easy, though. There are two possible
1427 * solutions for that:
1429 * - memcpy() the data internally, at which point we might as well just
1430 * do a regular read() on the buffer anyway.
1431 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1432 * has restriction limitations on both ends of the pipe).
1434 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1437 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1438 unsigned long nr_segs, unsigned int flags)
1444 if (unlikely(nr_segs > UIO_MAXIOV))
1446 else if (unlikely(!nr_segs))
1450 file = fget_light(fd, &fput);
1452 if (file->f_mode & FMODE_WRITE)
1453 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1454 else if (file->f_mode & FMODE_READ)
1455 error = vmsplice_to_user(file, iov, nr_segs, flags);
1457 fput_light(file, fput);
1463 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1464 int fd_out, loff_t __user *off_out,
1465 size_t len, unsigned int flags)
1468 struct file *in, *out;
1469 int fput_in, fput_out;
1475 in = fget_light(fd_in, &fput_in);
1477 if (in->f_mode & FMODE_READ) {
1478 out = fget_light(fd_out, &fput_out);
1480 if (out->f_mode & FMODE_WRITE)
1481 error = do_splice(in, off_in,
1484 fput_light(out, fput_out);
1488 fput_light(in, fput_in);
1495 * Make sure there's data to read. Wait for input if we can, otherwise
1496 * return an appropriate error.
1498 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1503 * Check ->nrbufs without the inode lock first. This function
1504 * is speculative anyways, so missing one is ok.
1510 mutex_lock(&pipe->inode->i_mutex);
1512 while (!pipe->nrbufs) {
1513 if (signal_pending(current)) {
1519 if (!pipe->waiting_writers) {
1520 if (flags & SPLICE_F_NONBLOCK) {
1528 mutex_unlock(&pipe->inode->i_mutex);
1533 * Make sure there's writeable room. Wait for room if we can, otherwise
1534 * return an appropriate error.
1536 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1541 * Check ->nrbufs without the inode lock first. This function
1542 * is speculative anyways, so missing one is ok.
1544 if (pipe->nrbufs < PIPE_BUFFERS)
1548 mutex_lock(&pipe->inode->i_mutex);
1550 while (pipe->nrbufs >= PIPE_BUFFERS) {
1551 if (!pipe->readers) {
1552 send_sig(SIGPIPE, current, 0);
1556 if (flags & SPLICE_F_NONBLOCK) {
1560 if (signal_pending(current)) {
1564 pipe->waiting_writers++;
1566 pipe->waiting_writers--;
1569 mutex_unlock(&pipe->inode->i_mutex);
1574 * Link contents of ipipe to opipe.
1576 static int link_pipe(struct pipe_inode_info *ipipe,
1577 struct pipe_inode_info *opipe,
1578 size_t len, unsigned int flags)
1580 struct pipe_buffer *ibuf, *obuf;
1581 int ret = 0, i = 0, nbuf;
1584 * Potential ABBA deadlock, work around it by ordering lock
1585 * grabbing by inode address. Otherwise two different processes
1586 * could deadlock (one doing tee from A -> B, the other from B -> A).
1588 inode_double_lock(ipipe->inode, opipe->inode);
1591 if (!opipe->readers) {
1592 send_sig(SIGPIPE, current, 0);
1599 * If we have iterated all input buffers or ran out of
1600 * output room, break.
1602 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1605 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1606 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1609 * Get a reference to this pipe buffer,
1610 * so we can copy the contents over.
1612 ibuf->ops->get(ipipe, ibuf);
1614 obuf = opipe->bufs + nbuf;
1618 * Don't inherit the gift flag, we need to
1619 * prevent multiple steals of this page.
1621 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1623 if (obuf->len > len)
1633 * return EAGAIN if we have the potential of some data in the
1634 * future, otherwise just return 0
1636 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1639 inode_double_unlock(ipipe->inode, opipe->inode);
1642 * If we put data in the output pipe, wakeup any potential readers.
1646 if (waitqueue_active(&opipe->wait))
1647 wake_up_interruptible(&opipe->wait);
1648 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1655 * This is a tee(1) implementation that works on pipes. It doesn't copy
1656 * any data, it simply references the 'in' pages on the 'out' pipe.
1657 * The 'flags' used are the SPLICE_F_* variants, currently the only
1658 * applicable one is SPLICE_F_NONBLOCK.
1660 static long do_tee(struct file *in, struct file *out, size_t len,
1663 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1664 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1668 * Duplicate the contents of ipipe to opipe without actually
1671 if (ipipe && opipe && ipipe != opipe) {
1673 * Keep going, unless we encounter an error. The ipipe/opipe
1674 * ordering doesn't really matter.
1676 ret = link_ipipe_prep(ipipe, flags);
1678 ret = link_opipe_prep(opipe, flags);
1680 ret = link_pipe(ipipe, opipe, len, flags);
1687 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1696 in = fget_light(fdin, &fput_in);
1698 if (in->f_mode & FMODE_READ) {
1700 struct file *out = fget_light(fdout, &fput_out);
1703 if (out->f_mode & FMODE_WRITE)
1704 error = do_tee(in, out, len, flags);
1705 fput_light(out, fput_out);
1708 fput_light(in, fput_in);