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 tuples, along 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 spd->spd_release(spd, page_nr++);
262 static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
264 page_cache_release(spd->pages[i]);
268 __generic_file_splice_read(struct file *in, loff_t *ppos,
269 struct pipe_inode_info *pipe, size_t len,
272 struct address_space *mapping = in->f_mapping;
273 unsigned int loff, nr_pages, req_pages;
274 struct page *pages[PIPE_BUFFERS];
275 struct partial_page partial[PIPE_BUFFERS];
277 pgoff_t index, end_index;
280 struct splice_pipe_desc spd = {
284 .ops = &page_cache_pipe_buf_ops,
285 .spd_release = spd_release_page,
288 index = *ppos >> PAGE_CACHE_SHIFT;
289 loff = *ppos & ~PAGE_CACHE_MASK;
290 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
291 nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
294 * Lookup the (hopefully) full range of pages we need.
296 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
297 index += spd.nr_pages;
300 * If find_get_pages_contig() returned fewer pages than we needed,
301 * readahead/allocate the rest and fill in the holes.
303 if (spd.nr_pages < nr_pages)
304 page_cache_sync_readahead(mapping, &in->f_ra, in,
305 index, req_pages - spd.nr_pages);
308 while (spd.nr_pages < nr_pages) {
310 * Page could be there, find_get_pages_contig() breaks on
313 page = find_get_page(mapping, index);
316 * page didn't exist, allocate one.
318 page = page_cache_alloc_cold(mapping);
322 error = add_to_page_cache_lru(page, mapping, index,
324 if (unlikely(error)) {
325 page_cache_release(page);
326 if (error == -EEXIST)
331 * add_to_page_cache() locks the page, unlock it
332 * to avoid convoluting the logic below even more.
337 pages[spd.nr_pages++] = page;
342 * Now loop over the map and see if we need to start IO on any
343 * pages, fill in the partial map, etc.
345 index = *ppos >> PAGE_CACHE_SHIFT;
346 nr_pages = spd.nr_pages;
348 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
349 unsigned int this_len;
355 * this_len is the max we'll use from this page
357 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
358 page = pages[page_nr];
360 if (PageReadahead(page))
361 page_cache_async_readahead(mapping, &in->f_ra, in,
362 page, index, req_pages - page_nr);
365 * If the page isn't uptodate, we may need to start io on it
367 if (!PageUptodate(page)) {
369 * If in nonblock mode then dont block on waiting
370 * for an in-flight io page
372 if (flags & SPLICE_F_NONBLOCK) {
373 if (TestSetPageLocked(page))
379 * page was truncated, stop here. if this isn't the
380 * first page, we'll just complete what we already
383 if (!page->mapping) {
388 * page was already under io and is now done, great
390 if (PageUptodate(page)) {
396 * need to read in the page
398 error = mapping->a_ops->readpage(in, page);
399 if (unlikely(error)) {
401 * We really should re-lookup the page here,
402 * but it complicates things a lot. Instead
403 * lets just do what we already stored, and
404 * we'll get it the next time we are called.
406 if (error == AOP_TRUNCATED_PAGE)
414 * i_size must be checked after PageUptodate.
416 isize = i_size_read(mapping->host);
417 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
418 if (unlikely(!isize || index > end_index))
422 * if this is the last page, see if we need to shrink
423 * the length and stop
425 if (end_index == index) {
429 * max good bytes in this page
431 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
436 * force quit after adding this page
438 this_len = min(this_len, plen - loff);
442 partial[page_nr].offset = loff;
443 partial[page_nr].len = this_len;
451 * Release any pages at the end, if we quit early. 'page_nr' is how far
452 * we got, 'nr_pages' is how many pages are in the map.
454 while (page_nr < nr_pages)
455 page_cache_release(pages[page_nr++]);
456 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
459 return splice_to_pipe(pipe, &spd);
465 * generic_file_splice_read - splice data from file to a pipe
466 * @in: file to splice from
467 * @ppos: position in @in
468 * @pipe: pipe to splice to
469 * @len: number of bytes to splice
470 * @flags: splice modifier flags
473 * Will read pages from given file and fill them into a pipe. Can be
474 * used as long as the address_space operations for the source implements
478 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
479 struct pipe_inode_info *pipe, size_t len,
486 isize = i_size_read(in->f_mapping->host);
487 if (unlikely(*ppos >= isize))
490 left = isize - *ppos;
491 if (unlikely(left < len))
496 while (len && !spliced) {
497 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
504 if (flags & SPLICE_F_NONBLOCK) {
521 EXPORT_SYMBOL(generic_file_splice_read);
524 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
525 * using sendpage(). Return the number of bytes sent.
527 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
528 struct pipe_buffer *buf, struct splice_desc *sd)
530 struct file *file = sd->u.file;
531 loff_t pos = sd->pos;
534 ret = buf->ops->confirm(pipe, buf);
536 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
538 ret = file->f_op->sendpage(file, buf->page, buf->offset,
539 sd->len, &pos, more);
546 * This is a little more tricky than the file -> pipe splicing. There are
547 * basically three cases:
549 * - Destination page already exists in the address space and there
550 * are users of it. For that case we have no other option that
551 * copying the data. Tough luck.
552 * - Destination page already exists in the address space, but there
553 * are no users of it. Make sure it's uptodate, then drop it. Fall
554 * through to last case.
555 * - Destination page does not exist, we can add the pipe page to
556 * the page cache and avoid the copy.
558 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
559 * sd->flags), we attempt to migrate pages from the pipe to the output
560 * file address space page cache. This is possible if no one else has
561 * the pipe page referenced outside of the pipe and page cache. If
562 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
563 * a new page in the output file page cache and fill/dirty that.
565 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
566 struct splice_desc *sd)
568 struct file *file = sd->u.file;
569 struct address_space *mapping = file->f_mapping;
570 unsigned int offset, this_len;
576 * make sure the data in this buffer is uptodate
578 ret = buf->ops->confirm(pipe, buf);
582 offset = sd->pos & ~PAGE_CACHE_MASK;
585 if (this_len + offset > PAGE_CACHE_SIZE)
586 this_len = PAGE_CACHE_SIZE - offset;
588 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
589 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
593 if (buf->page != page) {
595 * Careful, ->map() uses KM_USER0!
597 char *src = buf->ops->map(pipe, buf, 1);
598 char *dst = kmap_atomic(page, KM_USER1);
600 memcpy(dst + offset, src + buf->offset, this_len);
601 flush_dcache_page(page);
602 kunmap_atomic(dst, KM_USER1);
603 buf->ops->unmap(pipe, buf, src);
605 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
612 * __splice_from_pipe - splice data from a pipe to given actor
613 * @pipe: pipe to splice from
614 * @sd: information to @actor
615 * @actor: handler that splices the data
618 * This function does little more than loop over the pipe and call
619 * @actor to do the actual moving of a single struct pipe_buffer to
620 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
624 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
627 int ret, do_wakeup, err;
634 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
635 const struct pipe_buf_operations *ops = buf->ops;
638 if (sd->len > sd->total_len)
639 sd->len = sd->total_len;
641 err = actor(pipe, buf, sd);
643 if (!ret && err != -ENODATA)
655 sd->total_len -= err;
661 ops->release(pipe, buf);
662 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
676 if (!pipe->waiting_writers) {
681 if (sd->flags & SPLICE_F_NONBLOCK) {
687 if (signal_pending(current)) {
695 if (waitqueue_active(&pipe->wait))
696 wake_up_interruptible_sync(&pipe->wait);
697 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
706 if (waitqueue_active(&pipe->wait))
707 wake_up_interruptible(&pipe->wait);
708 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
713 EXPORT_SYMBOL(__splice_from_pipe);
716 * splice_from_pipe - splice data from a pipe to a file
717 * @pipe: pipe to splice from
718 * @out: file to splice to
719 * @ppos: position in @out
720 * @len: how many bytes to splice
721 * @flags: splice modifier flags
722 * @actor: handler that splices the data
725 * See __splice_from_pipe. This function locks the input and output inodes,
726 * otherwise it's identical to __splice_from_pipe().
729 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
730 loff_t *ppos, size_t len, unsigned int flags,
734 struct inode *inode = out->f_mapping->host;
735 struct splice_desc sd = {
743 * The actor worker might be calling ->prepare_write and
744 * ->commit_write. Most of the time, these expect i_mutex to
745 * be held. Since this may result in an ABBA deadlock with
746 * pipe->inode, we have to order lock acquiry here.
748 inode_double_lock(inode, pipe->inode);
749 ret = __splice_from_pipe(pipe, &sd, actor);
750 inode_double_unlock(inode, pipe->inode);
756 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
758 * @out: file to write to
759 * @ppos: position in @out
760 * @len: number of bytes to splice
761 * @flags: splice modifier flags
764 * Will either move or copy pages (determined by @flags options) from
765 * the given pipe inode to the given file. The caller is responsible
766 * for acquiring i_mutex on both inodes.
770 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
771 loff_t *ppos, size_t len, unsigned int flags)
773 struct address_space *mapping = out->f_mapping;
774 struct inode *inode = mapping->host;
775 struct splice_desc sd = {
784 err = remove_suid(out->f_path.dentry);
788 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
790 unsigned long nr_pages;
793 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
796 * If file or inode is SYNC and we actually wrote some data,
799 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
800 err = generic_osync_inode(inode, mapping,
801 OSYNC_METADATA|OSYNC_DATA);
806 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
812 EXPORT_SYMBOL(generic_file_splice_write_nolock);
815 * generic_file_splice_write - splice data from a pipe to a file
817 * @out: file to write to
818 * @ppos: position in @out
819 * @len: number of bytes to splice
820 * @flags: splice modifier flags
823 * Will either move or copy pages (determined by @flags options) from
824 * the given pipe inode to the given file.
828 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
829 loff_t *ppos, size_t len, unsigned int flags)
831 struct address_space *mapping = out->f_mapping;
832 struct inode *inode = mapping->host;
833 int killsuid, killpriv;
837 killpriv = security_inode_need_killpriv(out->f_path.dentry);
838 killsuid = should_remove_suid(out->f_path.dentry);
839 if (unlikely(killsuid || killpriv)) {
840 mutex_lock(&inode->i_mutex);
842 err = security_inode_killpriv(out->f_path.dentry);
843 if (!err && killsuid)
844 err = __remove_suid(out->f_path.dentry, killsuid);
845 mutex_unlock(&inode->i_mutex);
850 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
852 unsigned long nr_pages;
855 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
858 * If file or inode is SYNC and we actually wrote some data,
861 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
862 mutex_lock(&inode->i_mutex);
863 err = generic_osync_inode(inode, mapping,
864 OSYNC_METADATA|OSYNC_DATA);
865 mutex_unlock(&inode->i_mutex);
870 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
876 EXPORT_SYMBOL(generic_file_splice_write);
879 * generic_splice_sendpage - splice data from a pipe to a socket
880 * @pipe: pipe to splice from
881 * @out: socket to write to
882 * @ppos: position in @out
883 * @len: number of bytes to splice
884 * @flags: splice modifier flags
887 * Will send @len bytes from the pipe to a network socket. No data copying
891 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
892 loff_t *ppos, size_t len, unsigned int flags)
894 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
897 EXPORT_SYMBOL(generic_splice_sendpage);
900 * Attempt to initiate a splice from pipe to file.
902 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
903 loff_t *ppos, size_t len, unsigned int flags)
907 if (unlikely(!out->f_op || !out->f_op->splice_write))
910 if (unlikely(!(out->f_mode & FMODE_WRITE)))
913 ret = rw_verify_area(WRITE, out, ppos, len);
914 if (unlikely(ret < 0))
917 return out->f_op->splice_write(pipe, out, ppos, len, flags);
921 * Attempt to initiate a splice from a file to a pipe.
923 static long do_splice_to(struct file *in, loff_t *ppos,
924 struct pipe_inode_info *pipe, size_t len,
929 if (unlikely(!in->f_op || !in->f_op->splice_read))
932 if (unlikely(!(in->f_mode & FMODE_READ)))
935 ret = rw_verify_area(READ, in, ppos, len);
936 if (unlikely(ret < 0))
939 return in->f_op->splice_read(in, ppos, pipe, len, flags);
943 * splice_direct_to_actor - splices data directly between two non-pipes
944 * @in: file to splice from
945 * @sd: actor information on where to splice to
946 * @actor: handles the data splicing
949 * This is a special case helper to splice directly between two
950 * points, without requiring an explicit pipe. Internally an allocated
951 * pipe is cached in the process, and reused during the lifetime of
955 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
956 splice_direct_actor *actor)
958 struct pipe_inode_info *pipe;
965 * We require the input being a regular file, as we don't want to
966 * randomly drop data for eg socket -> socket splicing. Use the
967 * piped splicing for that!
969 i_mode = in->f_path.dentry->d_inode->i_mode;
970 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
974 * neither in nor out is a pipe, setup an internal pipe attached to
975 * 'out' and transfer the wanted data from 'in' to 'out' through that
977 pipe = current->splice_pipe;
978 if (unlikely(!pipe)) {
979 pipe = alloc_pipe_info(NULL);
984 * We don't have an immediate reader, but we'll read the stuff
985 * out of the pipe right after the splice_to_pipe(). So set
986 * PIPE_READERS appropriately.
990 current->splice_pipe = pipe;
1002 * Don't block on output, we have to drain the direct pipe.
1004 sd->flags &= ~SPLICE_F_NONBLOCK;
1008 loff_t pos = sd->pos;
1010 ret = do_splice_to(in, &pos, pipe, len, flags);
1011 if (unlikely(ret <= 0))
1015 sd->total_len = read_len;
1018 * NOTE: nonblocking mode only applies to the input. We
1019 * must not do the output in nonblocking mode as then we
1020 * could get stuck data in the internal pipe:
1022 ret = actor(pipe, sd);
1023 if (unlikely(ret <= 0))
1035 pipe->nrbufs = pipe->curbuf = 0;
1041 * If we did an incomplete transfer we must release
1042 * the pipe buffers in question:
1044 for (i = 0; i < PIPE_BUFFERS; i++) {
1045 struct pipe_buffer *buf = pipe->bufs + i;
1048 buf->ops->release(pipe, buf);
1058 EXPORT_SYMBOL(splice_direct_to_actor);
1060 static int direct_splice_actor(struct pipe_inode_info *pipe,
1061 struct splice_desc *sd)
1063 struct file *file = sd->u.file;
1065 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1069 * do_splice_direct - splices data directly between two files
1070 * @in: file to splice from
1071 * @ppos: input file offset
1072 * @out: file to splice to
1073 * @len: number of bytes to splice
1074 * @flags: splice modifier flags
1077 * For use by do_sendfile(). splice can easily emulate sendfile, but
1078 * doing it in the application would incur an extra system call
1079 * (splice in + splice out, as compared to just sendfile()). So this helper
1080 * can splice directly through a process-private pipe.
1083 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1084 size_t len, unsigned int flags)
1086 struct splice_desc sd = {
1095 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1103 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1104 * location, so checking ->i_pipe is not enough to verify that this is a
1107 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1109 if (S_ISFIFO(inode->i_mode))
1110 return inode->i_pipe;
1116 * Determine where to splice to/from.
1118 static long do_splice(struct file *in, loff_t __user *off_in,
1119 struct file *out, loff_t __user *off_out,
1120 size_t len, unsigned int flags)
1122 struct pipe_inode_info *pipe;
1123 loff_t offset, *off;
1126 pipe = pipe_info(in->f_path.dentry->d_inode);
1131 if (out->f_op->llseek == no_llseek)
1133 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1139 ret = do_splice_from(pipe, out, off, len, flags);
1141 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1147 pipe = pipe_info(out->f_path.dentry->d_inode);
1152 if (in->f_op->llseek == no_llseek)
1154 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1160 ret = do_splice_to(in, off, pipe, len, flags);
1162 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1172 * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1173 * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1174 * for writing) and page faulting on the user memory pointed to by src.
1175 * This assumes that we will very rarely hit the partial != 0 path, or this
1176 * will not be a win.
1178 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1182 if (!access_ok(VERIFY_READ, src, n))
1185 pagefault_disable();
1186 partial = __copy_from_user_inatomic(dst, src, n);
1190 * Didn't copy everything, drop the mmap_sem and do a faulting copy
1192 if (unlikely(partial)) {
1193 up_read(¤t->mm->mmap_sem);
1194 partial = copy_from_user(dst, src, n);
1195 down_read(¤t->mm->mmap_sem);
1202 * Map an iov into an array of pages and offset/length tupples. With the
1203 * partial_page structure, we can map several non-contiguous ranges into
1204 * our ones pages[] map instead of splitting that operation into pieces.
1205 * Could easily be exported as a generic helper for other users, in which
1206 * case one would probably want to add a 'max_nr_pages' parameter as well.
1208 static int get_iovec_page_array(const struct iovec __user *iov,
1209 unsigned int nr_vecs, struct page **pages,
1210 struct partial_page *partial, int aligned)
1212 int buffers = 0, error = 0;
1214 down_read(¤t->mm->mmap_sem);
1217 unsigned long off, npages;
1224 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1227 base = entry.iov_base;
1228 len = entry.iov_len;
1231 * Sanity check this iovec. 0 read succeeds.
1237 if (!access_ok(VERIFY_READ, base, len))
1241 * Get this base offset and number of pages, then map
1242 * in the user pages.
1244 off = (unsigned long) base & ~PAGE_MASK;
1247 * If asked for alignment, the offset must be zero and the
1248 * length a multiple of the PAGE_SIZE.
1251 if (aligned && (off || len & ~PAGE_MASK))
1254 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1255 if (npages > PIPE_BUFFERS - buffers)
1256 npages = PIPE_BUFFERS - buffers;
1258 error = get_user_pages(current, current->mm,
1259 (unsigned long) base, npages, 0, 0,
1260 &pages[buffers], NULL);
1262 if (unlikely(error <= 0))
1266 * Fill this contiguous range into the partial page map.
1268 for (i = 0; i < error; i++) {
1269 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1271 partial[buffers].offset = off;
1272 partial[buffers].len = plen;
1280 * We didn't complete this iov, stop here since it probably
1281 * means we have to move some of this into a pipe to
1282 * be able to continue.
1288 * Don't continue if we mapped fewer pages than we asked for,
1289 * or if we mapped the max number of pages that we have
1292 if (error < npages || buffers == PIPE_BUFFERS)
1299 up_read(¤t->mm->mmap_sem);
1307 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1308 struct splice_desc *sd)
1313 ret = buf->ops->confirm(pipe, buf);
1318 * See if we can use the atomic maps, by prefaulting in the
1319 * pages and doing an atomic copy
1321 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1322 src = buf->ops->map(pipe, buf, 1);
1323 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1325 buf->ops->unmap(pipe, buf, src);
1333 * No dice, use slow non-atomic map and copy
1335 src = buf->ops->map(pipe, buf, 0);
1338 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1341 buf->ops->unmap(pipe, buf, src);
1344 sd->u.userptr += ret;
1349 * For lack of a better implementation, implement vmsplice() to userspace
1350 * as a simple copy of the pipes pages to the user iov.
1352 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1353 unsigned long nr_segs, unsigned int flags)
1355 struct pipe_inode_info *pipe;
1356 struct splice_desc sd;
1361 pipe = pipe_info(file->f_path.dentry->d_inode);
1366 mutex_lock(&pipe->inode->i_mutex);
1374 * Get user address base and length for this iovec.
1376 error = get_user(base, &iov->iov_base);
1377 if (unlikely(error))
1379 error = get_user(len, &iov->iov_len);
1380 if (unlikely(error))
1384 * Sanity check this iovec. 0 read succeeds.
1388 if (unlikely(!base)) {
1393 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1401 sd.u.userptr = base;
1404 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1422 mutex_unlock(&pipe->inode->i_mutex);
1431 * vmsplice splices a user address range into a pipe. It can be thought of
1432 * as splice-from-memory, where the regular splice is splice-from-file (or
1433 * to file). In both cases the output is a pipe, naturally.
1435 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1436 unsigned long nr_segs, unsigned int flags)
1438 struct pipe_inode_info *pipe;
1439 struct page *pages[PIPE_BUFFERS];
1440 struct partial_page partial[PIPE_BUFFERS];
1441 struct splice_pipe_desc spd = {
1445 .ops = &user_page_pipe_buf_ops,
1446 .spd_release = spd_release_page,
1449 pipe = pipe_info(file->f_path.dentry->d_inode);
1453 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1454 flags & SPLICE_F_GIFT);
1455 if (spd.nr_pages <= 0)
1456 return spd.nr_pages;
1458 return splice_to_pipe(pipe, &spd);
1462 * Note that vmsplice only really supports true splicing _from_ user memory
1463 * to a pipe, not the other way around. Splicing from user memory is a simple
1464 * operation that can be supported without any funky alignment restrictions
1465 * or nasty vm tricks. We simply map in the user memory and fill them into
1466 * a pipe. The reverse isn't quite as easy, though. There are two possible
1467 * solutions for that:
1469 * - memcpy() the data internally, at which point we might as well just
1470 * do a regular read() on the buffer anyway.
1471 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1472 * has restriction limitations on both ends of the pipe).
1474 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1477 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1478 unsigned long nr_segs, unsigned int flags)
1484 if (unlikely(nr_segs > UIO_MAXIOV))
1486 else if (unlikely(!nr_segs))
1490 file = fget_light(fd, &fput);
1492 if (file->f_mode & FMODE_WRITE)
1493 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1494 else if (file->f_mode & FMODE_READ)
1495 error = vmsplice_to_user(file, iov, nr_segs, flags);
1497 fput_light(file, fput);
1503 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1504 int fd_out, loff_t __user *off_out,
1505 size_t len, unsigned int flags)
1508 struct file *in, *out;
1509 int fput_in, fput_out;
1515 in = fget_light(fd_in, &fput_in);
1517 if (in->f_mode & FMODE_READ) {
1518 out = fget_light(fd_out, &fput_out);
1520 if (out->f_mode & FMODE_WRITE)
1521 error = do_splice(in, off_in,
1524 fput_light(out, fput_out);
1528 fput_light(in, fput_in);
1535 * Make sure there's data to read. Wait for input if we can, otherwise
1536 * return an appropriate error.
1538 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1543 * Check ->nrbufs without the inode lock first. This function
1544 * is speculative anyways, so missing one is ok.
1550 mutex_lock(&pipe->inode->i_mutex);
1552 while (!pipe->nrbufs) {
1553 if (signal_pending(current)) {
1559 if (!pipe->waiting_writers) {
1560 if (flags & SPLICE_F_NONBLOCK) {
1568 mutex_unlock(&pipe->inode->i_mutex);
1573 * Make sure there's writeable room. Wait for room if we can, otherwise
1574 * return an appropriate error.
1576 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1581 * Check ->nrbufs without the inode lock first. This function
1582 * is speculative anyways, so missing one is ok.
1584 if (pipe->nrbufs < PIPE_BUFFERS)
1588 mutex_lock(&pipe->inode->i_mutex);
1590 while (pipe->nrbufs >= PIPE_BUFFERS) {
1591 if (!pipe->readers) {
1592 send_sig(SIGPIPE, current, 0);
1596 if (flags & SPLICE_F_NONBLOCK) {
1600 if (signal_pending(current)) {
1604 pipe->waiting_writers++;
1606 pipe->waiting_writers--;
1609 mutex_unlock(&pipe->inode->i_mutex);
1614 * Link contents of ipipe to opipe.
1616 static int link_pipe(struct pipe_inode_info *ipipe,
1617 struct pipe_inode_info *opipe,
1618 size_t len, unsigned int flags)
1620 struct pipe_buffer *ibuf, *obuf;
1621 int ret = 0, i = 0, nbuf;
1624 * Potential ABBA deadlock, work around it by ordering lock
1625 * grabbing by inode address. Otherwise two different processes
1626 * could deadlock (one doing tee from A -> B, the other from B -> A).
1628 inode_double_lock(ipipe->inode, opipe->inode);
1631 if (!opipe->readers) {
1632 send_sig(SIGPIPE, current, 0);
1639 * If we have iterated all input buffers or ran out of
1640 * output room, break.
1642 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1645 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1646 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1649 * Get a reference to this pipe buffer,
1650 * so we can copy the contents over.
1652 ibuf->ops->get(ipipe, ibuf);
1654 obuf = opipe->bufs + nbuf;
1658 * Don't inherit the gift flag, we need to
1659 * prevent multiple steals of this page.
1661 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1663 if (obuf->len > len)
1672 inode_double_unlock(ipipe->inode, opipe->inode);
1675 * If we put data in the output pipe, wakeup any potential readers.
1679 if (waitqueue_active(&opipe->wait))
1680 wake_up_interruptible(&opipe->wait);
1681 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1688 * This is a tee(1) implementation that works on pipes. It doesn't copy
1689 * any data, it simply references the 'in' pages on the 'out' pipe.
1690 * The 'flags' used are the SPLICE_F_* variants, currently the only
1691 * applicable one is SPLICE_F_NONBLOCK.
1693 static long do_tee(struct file *in, struct file *out, size_t len,
1696 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1697 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1701 * Duplicate the contents of ipipe to opipe without actually
1704 if (ipipe && opipe && ipipe != opipe) {
1706 * Keep going, unless we encounter an error. The ipipe/opipe
1707 * ordering doesn't really matter.
1709 ret = link_ipipe_prep(ipipe, flags);
1711 ret = link_opipe_prep(opipe, flags);
1713 ret = link_pipe(ipipe, opipe, len, flags);
1714 if (!ret && (flags & SPLICE_F_NONBLOCK))
1723 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1732 in = fget_light(fdin, &fput_in);
1734 if (in->f_mode & FMODE_READ) {
1736 struct file *out = fget_light(fdout, &fput_out);
1739 if (out->f_mode & FMODE_WRITE)
1740 error = do_tee(in, out, len, flags);
1741 fput_light(out, fput_out);
1744 fput_light(in, fput_in);