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/pipe_fs_i.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>
38 * Passed to splice_to_pipe
40 struct splice_pipe_desc {
41 struct page **pages; /* page map */
42 struct partial_page *partial; /* pages[] may not be contig */
43 int nr_pages; /* number of pages in map */
44 unsigned int flags; /* splice flags */
45 const struct pipe_buf_operations *ops;/* ops associated with output pipe */
49 * Attempt to steal a page from a pipe buffer. This should perhaps go into
50 * a vm helper function, it's already simplified quite a bit by the
51 * addition of remove_mapping(). If success is returned, the caller may
52 * attempt to reuse this page for another destination.
54 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
55 struct pipe_buffer *buf)
57 struct page *page = buf->page;
58 struct address_space *mapping;
62 mapping = page_mapping(page);
64 WARN_ON(!PageUptodate(page));
67 * At least for ext2 with nobh option, we need to wait on
68 * writeback completing on this page, since we'll remove it
69 * from the pagecache. Otherwise truncate wont wait on the
70 * page, allowing the disk blocks to be reused by someone else
71 * before we actually wrote our data to them. fs corruption
74 wait_on_page_writeback(page);
76 if (PagePrivate(page))
77 try_to_release_page(page, GFP_KERNEL);
80 * If we succeeded in removing the mapping, set LRU flag
83 if (remove_mapping(mapping, page)) {
84 buf->flags |= PIPE_BUF_FLAG_LRU;
90 * Raced with truncate or failed to remove page from current
91 * address space, unlock and return failure.
97 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
98 struct pipe_buffer *buf)
100 page_cache_release(buf->page);
101 buf->flags &= ~PIPE_BUF_FLAG_LRU;
104 static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
105 struct pipe_buffer *buf)
107 struct page *page = buf->page;
110 if (!PageUptodate(page)) {
114 * Page got truncated/unhashed. This will cause a 0-byte
115 * splice, if this is the first page.
117 if (!page->mapping) {
123 * Uh oh, read-error from disk.
125 if (!PageUptodate(page)) {
131 * Page is ok afterall, we are done.
142 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
144 .map = generic_pipe_buf_map,
145 .unmap = generic_pipe_buf_unmap,
146 .pin = page_cache_pipe_buf_pin,
147 .release = page_cache_pipe_buf_release,
148 .steal = page_cache_pipe_buf_steal,
149 .get = generic_pipe_buf_get,
152 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
153 struct pipe_buffer *buf)
155 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
158 buf->flags |= PIPE_BUF_FLAG_LRU;
159 return generic_pipe_buf_steal(pipe, buf);
162 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
164 .map = generic_pipe_buf_map,
165 .unmap = generic_pipe_buf_unmap,
166 .pin = generic_pipe_buf_pin,
167 .release = page_cache_pipe_buf_release,
168 .steal = user_page_pipe_buf_steal,
169 .get = generic_pipe_buf_get,
173 * Pipe output worker. This sets up our pipe format with the page cache
174 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
176 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
177 struct splice_pipe_desc *spd)
179 unsigned int spd_pages = spd->nr_pages;
180 int ret, do_wakeup, page_nr;
187 mutex_lock(&pipe->inode->i_mutex);
190 if (!pipe->readers) {
191 send_sig(SIGPIPE, current, 0);
197 if (pipe->nrbufs < PIPE_BUFFERS) {
198 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
199 struct pipe_buffer *buf = pipe->bufs + newbuf;
201 buf->page = spd->pages[page_nr];
202 buf->offset = spd->partial[page_nr].offset;
203 buf->len = spd->partial[page_nr].len;
205 if (spd->flags & SPLICE_F_GIFT)
206 buf->flags |= PIPE_BUF_FLAG_GIFT;
215 if (!--spd->nr_pages)
217 if (pipe->nrbufs < PIPE_BUFFERS)
223 if (spd->flags & SPLICE_F_NONBLOCK) {
229 if (signal_pending(current)) {
237 if (waitqueue_active(&pipe->wait))
238 wake_up_interruptible_sync(&pipe->wait);
239 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
243 pipe->waiting_writers++;
245 pipe->waiting_writers--;
249 mutex_unlock(&pipe->inode->i_mutex);
253 if (waitqueue_active(&pipe->wait))
254 wake_up_interruptible(&pipe->wait);
255 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
259 while (page_nr < spd_pages)
260 page_cache_release(spd->pages[page_nr++]);
266 __generic_file_splice_read(struct file *in, loff_t *ppos,
267 struct pipe_inode_info *pipe, size_t len,
270 struct address_space *mapping = in->f_mapping;
271 unsigned int loff, nr_pages;
272 struct page *pages[PIPE_BUFFERS];
273 struct partial_page partial[PIPE_BUFFERS];
275 pgoff_t index, end_index;
278 struct splice_pipe_desc spd = {
282 .ops = &page_cache_pipe_buf_ops,
285 index = *ppos >> PAGE_CACHE_SHIFT;
286 loff = *ppos & ~PAGE_CACHE_MASK;
287 nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
289 if (nr_pages > PIPE_BUFFERS)
290 nr_pages = PIPE_BUFFERS;
293 * Don't try to 2nd guess the read-ahead logic, call into
294 * page_cache_readahead() like the page cache reads would do.
296 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
299 * Now fill in the holes:
304 * Lookup the (hopefully) full range of pages we need.
306 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
309 * If find_get_pages_contig() returned fewer pages than we needed,
312 index += spd.nr_pages;
313 while (spd.nr_pages < nr_pages) {
315 * Page could be there, find_get_pages_contig() breaks on
318 page = find_get_page(mapping, index);
321 * Make sure the read-ahead engine is notified
322 * about this failure.
324 handle_ra_miss(mapping, &in->f_ra, index);
327 * page didn't exist, allocate one.
329 page = page_cache_alloc_cold(mapping);
333 error = add_to_page_cache_lru(page, mapping, index,
335 if (unlikely(error)) {
336 page_cache_release(page);
337 if (error == -EEXIST)
342 * add_to_page_cache() locks the page, unlock it
343 * to avoid convoluting the logic below even more.
348 pages[spd.nr_pages++] = page;
353 * Now loop over the map and see if we need to start IO on any
354 * pages, fill in the partial map, etc.
356 index = *ppos >> PAGE_CACHE_SHIFT;
357 nr_pages = spd.nr_pages;
359 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
360 unsigned int this_len;
366 * this_len is the max we'll use from this page
368 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
369 page = pages[page_nr];
372 * If the page isn't uptodate, we may need to start io on it
374 if (!PageUptodate(page)) {
376 * If in nonblock mode then dont block on waiting
377 * for an in-flight io page
379 if (flags & SPLICE_F_NONBLOCK) {
380 if (TestSetPageLocked(page))
386 * page was truncated, stop here. if this isn't the
387 * first page, we'll just complete what we already
390 if (!page->mapping) {
395 * page was already under io and is now done, great
397 if (PageUptodate(page)) {
403 * need to read in the page
405 error = mapping->a_ops->readpage(in, page);
406 if (unlikely(error)) {
408 * We really should re-lookup the page here,
409 * but it complicates things a lot. Instead
410 * lets just do what we already stored, and
411 * we'll get it the next time we are called.
413 if (error == AOP_TRUNCATED_PAGE)
421 * i_size must be checked after PageUptodate.
423 isize = i_size_read(mapping->host);
424 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
425 if (unlikely(!isize || index > end_index))
429 * if this is the last page, see if we need to shrink
430 * the length and stop
432 if (end_index == index) {
436 * max good bytes in this page
438 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
443 * force quit after adding this page
445 this_len = min(this_len, plen - loff);
449 partial[page_nr].offset = loff;
450 partial[page_nr].len = this_len;
458 * Release any pages at the end, if we quit early. 'page_nr' is how far
459 * we got, 'nr_pages' is how many pages are in the map.
461 while (page_nr < nr_pages)
462 page_cache_release(pages[page_nr++]);
465 return splice_to_pipe(pipe, &spd);
471 * generic_file_splice_read - splice data from file to a pipe
472 * @in: file to splice from
473 * @pipe: pipe to splice to
474 * @len: number of bytes to splice
475 * @flags: splice modifier flags
477 * Will read pages from given file and fill them into a pipe.
479 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
480 struct pipe_inode_info *pipe, size_t len,
487 isize = i_size_read(in->f_mapping->host);
488 if (unlikely(*ppos >= isize))
491 left = isize - *ppos;
492 if (unlikely(left < len))
498 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
505 if (flags & SPLICE_F_NONBLOCK) {
522 EXPORT_SYMBOL(generic_file_splice_read);
525 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
526 * using sendpage(). Return the number of bytes sent.
528 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
529 struct pipe_buffer *buf, struct splice_desc *sd)
531 struct file *file = sd->file;
532 loff_t pos = sd->pos;
535 ret = buf->ops->pin(pipe, buf);
537 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
539 ret = file->f_op->sendpage(file, buf->page, buf->offset,
540 sd->len, &pos, more);
547 * This is a little more tricky than the file -> pipe splicing. There are
548 * basically three cases:
550 * - Destination page already exists in the address space and there
551 * are users of it. For that case we have no other option that
552 * copying the data. Tough luck.
553 * - Destination page already exists in the address space, but there
554 * are no users of it. Make sure it's uptodate, then drop it. Fall
555 * through to last case.
556 * - Destination page does not exist, we can add the pipe page to
557 * the page cache and avoid the copy.
559 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
560 * sd->flags), we attempt to migrate pages from the pipe to the output
561 * file address space page cache. This is possible if no one else has
562 * the pipe page referenced outside of the pipe and page cache. If
563 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
564 * a new page in the output file page cache and fill/dirty that.
566 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
567 struct splice_desc *sd)
569 struct file *file = sd->file;
570 struct address_space *mapping = file->f_mapping;
571 unsigned int offset, this_len;
577 * make sure the data in this buffer is uptodate
579 ret = buf->ops->pin(pipe, buf);
583 index = sd->pos >> PAGE_CACHE_SHIFT;
584 offset = sd->pos & ~PAGE_CACHE_MASK;
587 if (this_len + offset > PAGE_CACHE_SIZE)
588 this_len = PAGE_CACHE_SIZE - offset;
591 page = find_lock_page(mapping, index);
594 page = page_cache_alloc_cold(mapping);
599 * This will also lock the page
601 ret = add_to_page_cache_lru(page, mapping, index,
607 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
609 loff_t isize = i_size_read(mapping->host);
611 if (ret != AOP_TRUNCATED_PAGE)
613 page_cache_release(page);
614 if (ret == AOP_TRUNCATED_PAGE)
618 * prepare_write() may have instantiated a few blocks
619 * outside i_size. Trim these off again.
621 if (sd->pos + this_len > isize)
622 vmtruncate(mapping->host, isize);
627 if (buf->page != page) {
629 * Careful, ->map() uses KM_USER0!
631 char *src = buf->ops->map(pipe, buf, 1);
632 char *dst = kmap_atomic(page, KM_USER1);
634 memcpy(dst + offset, src + buf->offset, this_len);
635 flush_dcache_page(page);
636 kunmap_atomic(dst, KM_USER1);
637 buf->ops->unmap(pipe, buf, src);
640 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
642 if (ret == AOP_TRUNCATED_PAGE) {
643 page_cache_release(page);
649 * Partial write has happened, so 'ret' already initialized by
650 * number of bytes written, Where is nothing we have to do here.
655 * Return the number of bytes written and mark page as
656 * accessed, we are now done!
658 mark_page_accessed(page);
660 page_cache_release(page);
667 * Pipe input worker. Most of this logic works like a regular pipe, the
668 * key here is the 'actor' worker passed in that actually moves the data
669 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
671 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
674 int ret, do_wakeup, err;
681 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
682 const struct pipe_buf_operations *ops = buf->ops;
685 if (sd->len > sd->total_len)
686 sd->len = sd->total_len;
688 err = actor(pipe, buf, sd);
690 if (!ret && err != -ENODATA)
702 sd->total_len -= err;
708 ops->release(pipe, buf);
709 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
723 if (!pipe->waiting_writers) {
728 if (sd->flags & SPLICE_F_NONBLOCK) {
734 if (signal_pending(current)) {
742 if (waitqueue_active(&pipe->wait))
743 wake_up_interruptible_sync(&pipe->wait);
744 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
753 if (waitqueue_active(&pipe->wait))
754 wake_up_interruptible(&pipe->wait);
755 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
760 EXPORT_SYMBOL(__splice_from_pipe);
762 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
763 loff_t *ppos, size_t len, unsigned int flags,
767 struct inode *inode = out->f_mapping->host;
768 struct splice_desc sd = {
776 * The actor worker might be calling ->prepare_write and
777 * ->commit_write. Most of the time, these expect i_mutex to
778 * be held. Since this may result in an ABBA deadlock with
779 * pipe->inode, we have to order lock acquiry here.
781 inode_double_lock(inode, pipe->inode);
782 ret = __splice_from_pipe(pipe, &sd, actor);
783 inode_double_unlock(inode, pipe->inode);
789 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
791 * @out: file to write to
792 * @len: number of bytes to splice
793 * @flags: splice modifier flags
795 * Will either move or copy pages (determined by @flags options) from
796 * the given pipe inode to the given file. The caller is responsible
797 * for acquiring i_mutex on both inodes.
801 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
802 loff_t *ppos, size_t len, unsigned int flags)
804 struct address_space *mapping = out->f_mapping;
805 struct inode *inode = mapping->host;
806 struct splice_desc sd = {
815 err = remove_suid(out->f_path.dentry);
819 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
821 unsigned long nr_pages;
824 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
827 * If file or inode is SYNC and we actually wrote some data,
830 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
831 err = generic_osync_inode(inode, mapping,
832 OSYNC_METADATA|OSYNC_DATA);
837 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
843 EXPORT_SYMBOL(generic_file_splice_write_nolock);
846 * generic_file_splice_write - splice data from a pipe to a file
848 * @out: file to write to
849 * @len: number of bytes to splice
850 * @flags: splice modifier flags
852 * Will either move or copy pages (determined by @flags options) from
853 * the given pipe inode to the given file.
857 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
858 loff_t *ppos, size_t len, unsigned int flags)
860 struct address_space *mapping = out->f_mapping;
861 struct inode *inode = mapping->host;
865 err = should_remove_suid(out->f_path.dentry);
867 mutex_lock(&inode->i_mutex);
868 err = __remove_suid(out->f_path.dentry, err);
869 mutex_unlock(&inode->i_mutex);
874 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
876 unsigned long nr_pages;
879 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
882 * If file or inode is SYNC and we actually wrote some data,
885 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
886 mutex_lock(&inode->i_mutex);
887 err = generic_osync_inode(inode, mapping,
888 OSYNC_METADATA|OSYNC_DATA);
889 mutex_unlock(&inode->i_mutex);
894 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
900 EXPORT_SYMBOL(generic_file_splice_write);
903 * generic_splice_sendpage - splice data from a pipe to a socket
905 * @out: socket to write to
906 * @len: number of bytes to splice
907 * @flags: splice modifier flags
909 * Will send @len bytes from the pipe to a network socket. No data copying
913 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
914 loff_t *ppos, size_t len, unsigned int flags)
916 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
919 EXPORT_SYMBOL(generic_splice_sendpage);
922 * Attempt to initiate a splice from pipe to file.
924 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
925 loff_t *ppos, size_t len, unsigned int flags)
929 if (unlikely(!out->f_op || !out->f_op->splice_write))
932 if (unlikely(!(out->f_mode & FMODE_WRITE)))
935 ret = rw_verify_area(WRITE, out, ppos, len);
936 if (unlikely(ret < 0))
939 return out->f_op->splice_write(pipe, out, ppos, len, flags);
943 * Attempt to initiate a splice from a file to a pipe.
945 static long do_splice_to(struct file *in, loff_t *ppos,
946 struct pipe_inode_info *pipe, size_t len,
951 if (unlikely(!in->f_op || !in->f_op->splice_read))
954 if (unlikely(!(in->f_mode & FMODE_READ)))
957 ret = rw_verify_area(READ, in, ppos, len);
958 if (unlikely(ret < 0))
961 return in->f_op->splice_read(in, ppos, pipe, len, flags);
965 * Splices from an input file to an actor, using a 'direct' pipe.
967 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
968 splice_direct_actor *actor)
970 struct pipe_inode_info *pipe;
977 * We require the input being a regular file, as we don't want to
978 * randomly drop data for eg socket -> socket splicing. Use the
979 * piped splicing for that!
981 i_mode = in->f_path.dentry->d_inode->i_mode;
982 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
986 * neither in nor out is a pipe, setup an internal pipe attached to
987 * 'out' and transfer the wanted data from 'in' to 'out' through that
989 pipe = current->splice_pipe;
990 if (unlikely(!pipe)) {
991 pipe = alloc_pipe_info(NULL);
996 * We don't have an immediate reader, but we'll read the stuff
997 * out of the pipe right after the splice_to_pipe(). So set
998 * PIPE_READERS appropriately.
1002 current->splice_pipe = pipe;
1010 len = sd->total_len;
1014 * Don't block on output, we have to drain the direct pipe.
1016 sd->flags &= ~SPLICE_F_NONBLOCK;
1019 size_t read_len, max_read_len;
1022 * Do at most PIPE_BUFFERS pages worth of transfer:
1024 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1026 ret = do_splice_to(in, &sd->pos, pipe, max_read_len, flags);
1027 if (unlikely(ret < 0))
1031 sd->total_len = read_len;
1034 * NOTE: nonblocking mode only applies to the input. We
1035 * must not do the output in nonblocking mode as then we
1036 * could get stuck data in the internal pipe:
1038 ret = actor(pipe, sd);
1039 if (unlikely(ret < 0))
1046 * In nonblocking mode, if we got back a short read then
1047 * that was due to either an IO error or due to the
1048 * pagecache entry not being there. In the IO error case
1049 * the _next_ splice attempt will produce a clean IO error
1050 * return value (not a short read), so in both cases it's
1051 * correct to break out of the loop here:
1053 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1057 pipe->nrbufs = pipe->curbuf = 0;
1063 * If we did an incomplete transfer we must release
1064 * the pipe buffers in question:
1066 for (i = 0; i < PIPE_BUFFERS; i++) {
1067 struct pipe_buffer *buf = pipe->bufs + i;
1070 buf->ops->release(pipe, buf);
1074 pipe->nrbufs = pipe->curbuf = 0;
1077 * If we transferred some data, return the number of bytes:
1085 EXPORT_SYMBOL(splice_direct_to_actor);
1087 static int direct_splice_actor(struct pipe_inode_info *pipe,
1088 struct splice_desc *sd)
1090 struct file *file = sd->file;
1092 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1095 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1096 size_t len, unsigned int flags)
1098 struct splice_desc sd = {
1107 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1113 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1114 * location, so checking ->i_pipe is not enough to verify that this is a
1117 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1119 if (S_ISFIFO(inode->i_mode))
1120 return inode->i_pipe;
1126 * Determine where to splice to/from.
1128 static long do_splice(struct file *in, loff_t __user *off_in,
1129 struct file *out, loff_t __user *off_out,
1130 size_t len, unsigned int flags)
1132 struct pipe_inode_info *pipe;
1133 loff_t offset, *off;
1136 pipe = pipe_info(in->f_path.dentry->d_inode);
1141 if (out->f_op->llseek == no_llseek)
1143 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1149 ret = do_splice_from(pipe, out, off, len, flags);
1151 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1157 pipe = pipe_info(out->f_path.dentry->d_inode);
1162 if (in->f_op->llseek == no_llseek)
1164 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1170 ret = do_splice_to(in, off, pipe, len, flags);
1172 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1182 * Map an iov into an array of pages and offset/length tupples. With the
1183 * partial_page structure, we can map several non-contiguous ranges into
1184 * our ones pages[] map instead of splitting that operation into pieces.
1185 * Could easily be exported as a generic helper for other users, in which
1186 * case one would probably want to add a 'max_nr_pages' parameter as well.
1188 static int get_iovec_page_array(const struct iovec __user *iov,
1189 unsigned int nr_vecs, struct page **pages,
1190 struct partial_page *partial, int aligned)
1192 int buffers = 0, error = 0;
1195 * It's ok to take the mmap_sem for reading, even
1196 * across a "get_user()".
1198 down_read(¤t->mm->mmap_sem);
1201 unsigned long off, npages;
1207 * Get user address base and length for this iovec.
1209 error = get_user(base, &iov->iov_base);
1210 if (unlikely(error))
1212 error = get_user(len, &iov->iov_len);
1213 if (unlikely(error))
1217 * Sanity check this iovec. 0 read succeeds.
1222 if (unlikely(!base))
1226 * Get this base offset and number of pages, then map
1227 * in the user pages.
1229 off = (unsigned long) base & ~PAGE_MASK;
1232 * If asked for alignment, the offset must be zero and the
1233 * length a multiple of the PAGE_SIZE.
1236 if (aligned && (off || len & ~PAGE_MASK))
1239 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1240 if (npages > PIPE_BUFFERS - buffers)
1241 npages = PIPE_BUFFERS - buffers;
1243 error = get_user_pages(current, current->mm,
1244 (unsigned long) base, npages, 0, 0,
1245 &pages[buffers], NULL);
1247 if (unlikely(error <= 0))
1251 * Fill this contiguous range into the partial page map.
1253 for (i = 0; i < error; i++) {
1254 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1256 partial[buffers].offset = off;
1257 partial[buffers].len = plen;
1265 * We didn't complete this iov, stop here since it probably
1266 * means we have to move some of this into a pipe to
1267 * be able to continue.
1273 * Don't continue if we mapped fewer pages than we asked for,
1274 * or if we mapped the max number of pages that we have
1277 if (error < npages || buffers == PIPE_BUFFERS)
1284 up_read(¤t->mm->mmap_sem);
1293 * vmsplice splices a user address range into a pipe. It can be thought of
1294 * as splice-from-memory, where the regular splice is splice-from-file (or
1295 * to file). In both cases the output is a pipe, naturally.
1297 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1298 * not the other way around. Splicing from user memory is a simple operation
1299 * that can be supported without any funky alignment restrictions or nasty
1300 * vm tricks. We simply map in the user memory and fill them into a pipe.
1301 * The reverse isn't quite as easy, though. There are two possible solutions
1304 * - memcpy() the data internally, at which point we might as well just
1305 * do a regular read() on the buffer anyway.
1306 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1307 * has restriction limitations on both ends of the pipe).
1309 * Alas, it isn't here.
1312 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1313 unsigned long nr_segs, unsigned int flags)
1315 struct pipe_inode_info *pipe;
1316 struct page *pages[PIPE_BUFFERS];
1317 struct partial_page partial[PIPE_BUFFERS];
1318 struct splice_pipe_desc spd = {
1322 .ops = &user_page_pipe_buf_ops,
1325 pipe = pipe_info(file->f_path.dentry->d_inode);
1328 if (unlikely(nr_segs > UIO_MAXIOV))
1330 else if (unlikely(!nr_segs))
1333 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1334 flags & SPLICE_F_GIFT);
1335 if (spd.nr_pages <= 0)
1336 return spd.nr_pages;
1338 return splice_to_pipe(pipe, &spd);
1341 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1342 unsigned long nr_segs, unsigned int flags)
1349 file = fget_light(fd, &fput);
1351 if (file->f_mode & FMODE_WRITE)
1352 error = do_vmsplice(file, iov, nr_segs, flags);
1354 fput_light(file, fput);
1360 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1361 int fd_out, loff_t __user *off_out,
1362 size_t len, unsigned int flags)
1365 struct file *in, *out;
1366 int fput_in, fput_out;
1372 in = fget_light(fd_in, &fput_in);
1374 if (in->f_mode & FMODE_READ) {
1375 out = fget_light(fd_out, &fput_out);
1377 if (out->f_mode & FMODE_WRITE)
1378 error = do_splice(in, off_in,
1381 fput_light(out, fput_out);
1385 fput_light(in, fput_in);
1392 * Make sure there's data to read. Wait for input if we can, otherwise
1393 * return an appropriate error.
1395 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1400 * Check ->nrbufs without the inode lock first. This function
1401 * is speculative anyways, so missing one is ok.
1407 mutex_lock(&pipe->inode->i_mutex);
1409 while (!pipe->nrbufs) {
1410 if (signal_pending(current)) {
1416 if (!pipe->waiting_writers) {
1417 if (flags & SPLICE_F_NONBLOCK) {
1425 mutex_unlock(&pipe->inode->i_mutex);
1430 * Make sure there's writeable room. Wait for room if we can, otherwise
1431 * return an appropriate error.
1433 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1438 * Check ->nrbufs without the inode lock first. This function
1439 * is speculative anyways, so missing one is ok.
1441 if (pipe->nrbufs < PIPE_BUFFERS)
1445 mutex_lock(&pipe->inode->i_mutex);
1447 while (pipe->nrbufs >= PIPE_BUFFERS) {
1448 if (!pipe->readers) {
1449 send_sig(SIGPIPE, current, 0);
1453 if (flags & SPLICE_F_NONBLOCK) {
1457 if (signal_pending(current)) {
1461 pipe->waiting_writers++;
1463 pipe->waiting_writers--;
1466 mutex_unlock(&pipe->inode->i_mutex);
1471 * Link contents of ipipe to opipe.
1473 static int link_pipe(struct pipe_inode_info *ipipe,
1474 struct pipe_inode_info *opipe,
1475 size_t len, unsigned int flags)
1477 struct pipe_buffer *ibuf, *obuf;
1478 int ret = 0, i = 0, nbuf;
1481 * Potential ABBA deadlock, work around it by ordering lock
1482 * grabbing by inode address. Otherwise two different processes
1483 * could deadlock (one doing tee from A -> B, the other from B -> A).
1485 inode_double_lock(ipipe->inode, opipe->inode);
1488 if (!opipe->readers) {
1489 send_sig(SIGPIPE, current, 0);
1496 * If we have iterated all input buffers or ran out of
1497 * output room, break.
1499 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1502 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1503 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1506 * Get a reference to this pipe buffer,
1507 * so we can copy the contents over.
1509 ibuf->ops->get(ipipe, ibuf);
1511 obuf = opipe->bufs + nbuf;
1515 * Don't inherit the gift flag, we need to
1516 * prevent multiple steals of this page.
1518 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1520 if (obuf->len > len)
1529 inode_double_unlock(ipipe->inode, opipe->inode);
1532 * If we put data in the output pipe, wakeup any potential readers.
1536 if (waitqueue_active(&opipe->wait))
1537 wake_up_interruptible(&opipe->wait);
1538 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1545 * This is a tee(1) implementation that works on pipes. It doesn't copy
1546 * any data, it simply references the 'in' pages on the 'out' pipe.
1547 * The 'flags' used are the SPLICE_F_* variants, currently the only
1548 * applicable one is SPLICE_F_NONBLOCK.
1550 static long do_tee(struct file *in, struct file *out, size_t len,
1553 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1554 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1558 * Duplicate the contents of ipipe to opipe without actually
1561 if (ipipe && opipe && ipipe != opipe) {
1563 * Keep going, unless we encounter an error. The ipipe/opipe
1564 * ordering doesn't really matter.
1566 ret = link_ipipe_prep(ipipe, flags);
1568 ret = link_opipe_prep(opipe, flags);
1570 ret = link_pipe(ipipe, opipe, len, flags);
1571 if (!ret && (flags & SPLICE_F_NONBLOCK))
1580 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1589 in = fget_light(fdin, &fput_in);
1591 if (in->f_mode & FMODE_READ) {
1593 struct file *out = fget_light(fdout, &fput_out);
1596 if (out->f_mode & FMODE_WRITE)
1597 error = do_tee(in, out, len, flags);
1598 fput_light(out, fput_out);
1601 fput_light(in, fput_in);