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,
323 mapping_gfp_mask(mapping));
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)) {
381 * page was truncated, stop here. if this isn't the
382 * first page, we'll just complete what we already
385 if (!page->mapping) {
390 * page was already under io and is now done, great
392 if (PageUptodate(page)) {
398 * need to read in the page
400 error = mapping->a_ops->readpage(in, page);
401 if (unlikely(error)) {
403 * We really should re-lookup the page here,
404 * but it complicates things a lot. Instead
405 * lets just do what we already stored, and
406 * we'll get it the next time we are called.
408 if (error == AOP_TRUNCATED_PAGE)
416 * i_size must be checked after PageUptodate.
418 isize = i_size_read(mapping->host);
419 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
420 if (unlikely(!isize || index > end_index))
424 * if this is the last page, see if we need to shrink
425 * the length and stop
427 if (end_index == index) {
431 * max good bytes in this page
433 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
438 * force quit after adding this page
440 this_len = min(this_len, plen - loff);
444 partial[page_nr].offset = loff;
445 partial[page_nr].len = this_len;
453 * Release any pages at the end, if we quit early. 'page_nr' is how far
454 * we got, 'nr_pages' is how many pages are in the map.
456 while (page_nr < nr_pages)
457 page_cache_release(pages[page_nr++]);
458 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
461 return splice_to_pipe(pipe, &spd);
467 * generic_file_splice_read - splice data from file to a pipe
468 * @in: file to splice from
469 * @ppos: position in @in
470 * @pipe: pipe to splice to
471 * @len: number of bytes to splice
472 * @flags: splice modifier flags
475 * Will read pages from given file and fill them into a pipe. Can be
476 * used as long as the address_space operations for the source implements
480 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
481 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))
495 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
502 EXPORT_SYMBOL(generic_file_splice_read);
505 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
506 * using sendpage(). Return the number of bytes sent.
508 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
509 struct pipe_buffer *buf, struct splice_desc *sd)
511 struct file *file = sd->u.file;
512 loff_t pos = sd->pos;
515 ret = buf->ops->confirm(pipe, buf);
517 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
519 ret = file->f_op->sendpage(file, buf->page, buf->offset,
520 sd->len, &pos, more);
527 * This is a little more tricky than the file -> pipe splicing. There are
528 * basically three cases:
530 * - Destination page already exists in the address space and there
531 * are users of it. For that case we have no other option that
532 * copying the data. Tough luck.
533 * - Destination page already exists in the address space, but there
534 * are no users of it. Make sure it's uptodate, then drop it. Fall
535 * through to last case.
536 * - Destination page does not exist, we can add the pipe page to
537 * the page cache and avoid the copy.
539 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
540 * sd->flags), we attempt to migrate pages from the pipe to the output
541 * file address space page cache. This is possible if no one else has
542 * the pipe page referenced outside of the pipe and page cache. If
543 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
544 * a new page in the output file page cache and fill/dirty that.
546 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
547 struct splice_desc *sd)
549 struct file *file = sd->u.file;
550 struct address_space *mapping = file->f_mapping;
551 unsigned int offset, this_len;
557 * make sure the data in this buffer is uptodate
559 ret = buf->ops->confirm(pipe, buf);
563 offset = sd->pos & ~PAGE_CACHE_MASK;
566 if (this_len + offset > PAGE_CACHE_SIZE)
567 this_len = PAGE_CACHE_SIZE - offset;
569 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
570 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
574 if (buf->page != page) {
576 * Careful, ->map() uses KM_USER0!
578 char *src = buf->ops->map(pipe, buf, 1);
579 char *dst = kmap_atomic(page, KM_USER1);
581 memcpy(dst + offset, src + buf->offset, this_len);
582 flush_dcache_page(page);
583 kunmap_atomic(dst, KM_USER1);
584 buf->ops->unmap(pipe, buf, src);
586 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
593 * __splice_from_pipe - splice data from a pipe to given actor
594 * @pipe: pipe to splice from
595 * @sd: information to @actor
596 * @actor: handler that splices the data
599 * This function does little more than loop over the pipe and call
600 * @actor to do the actual moving of a single struct pipe_buffer to
601 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
605 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
608 int ret, do_wakeup, err;
615 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
616 const struct pipe_buf_operations *ops = buf->ops;
619 if (sd->len > sd->total_len)
620 sd->len = sd->total_len;
622 err = actor(pipe, buf, sd);
624 if (!ret && err != -ENODATA)
636 sd->total_len -= err;
642 ops->release(pipe, buf);
643 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
657 if (!pipe->waiting_writers) {
662 if (sd->flags & SPLICE_F_NONBLOCK) {
668 if (signal_pending(current)) {
676 if (waitqueue_active(&pipe->wait))
677 wake_up_interruptible_sync(&pipe->wait);
678 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
687 if (waitqueue_active(&pipe->wait))
688 wake_up_interruptible(&pipe->wait);
689 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
694 EXPORT_SYMBOL(__splice_from_pipe);
697 * splice_from_pipe - splice data from a pipe to a file
698 * @pipe: pipe to splice from
699 * @out: file to splice to
700 * @ppos: position in @out
701 * @len: how many bytes to splice
702 * @flags: splice modifier flags
703 * @actor: handler that splices the data
706 * See __splice_from_pipe. This function locks the input and output inodes,
707 * otherwise it's identical to __splice_from_pipe().
710 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
711 loff_t *ppos, size_t len, unsigned int flags,
715 struct inode *inode = out->f_mapping->host;
716 struct splice_desc sd = {
724 * The actor worker might be calling ->prepare_write and
725 * ->commit_write. Most of the time, these expect i_mutex to
726 * be held. Since this may result in an ABBA deadlock with
727 * pipe->inode, we have to order lock acquiry here.
729 inode_double_lock(inode, pipe->inode);
730 ret = __splice_from_pipe(pipe, &sd, actor);
731 inode_double_unlock(inode, pipe->inode);
737 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
739 * @out: file to write to
740 * @ppos: position in @out
741 * @len: number of bytes to splice
742 * @flags: splice modifier flags
745 * Will either move or copy pages (determined by @flags options) from
746 * the given pipe inode to the given file. The caller is responsible
747 * for acquiring i_mutex on both inodes.
751 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
752 loff_t *ppos, size_t len, unsigned int flags)
754 struct address_space *mapping = out->f_mapping;
755 struct inode *inode = mapping->host;
756 struct splice_desc sd = {
765 err = remove_suid(out->f_path.dentry);
769 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
771 unsigned long nr_pages;
774 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
777 * If file or inode is SYNC and we actually wrote some data,
780 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
781 err = generic_osync_inode(inode, mapping,
782 OSYNC_METADATA|OSYNC_DATA);
787 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
793 EXPORT_SYMBOL(generic_file_splice_write_nolock);
796 * generic_file_splice_write - splice data from a pipe to a file
798 * @out: file to write to
799 * @ppos: position in @out
800 * @len: number of bytes to splice
801 * @flags: splice modifier flags
804 * Will either move or copy pages (determined by @flags options) from
805 * the given pipe inode to the given file.
809 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
810 loff_t *ppos, size_t len, unsigned int flags)
812 struct address_space *mapping = out->f_mapping;
813 struct inode *inode = mapping->host;
814 struct splice_desc sd = {
822 inode_double_lock(inode, pipe->inode);
823 ret = remove_suid(out->f_path.dentry);
825 ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
826 inode_double_unlock(inode, pipe->inode);
828 unsigned long nr_pages;
831 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
834 * If file or inode is SYNC and we actually wrote some data,
837 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
840 mutex_lock(&inode->i_mutex);
841 err = generic_osync_inode(inode, mapping,
842 OSYNC_METADATA|OSYNC_DATA);
843 mutex_unlock(&inode->i_mutex);
848 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
854 EXPORT_SYMBOL(generic_file_splice_write);
857 * generic_splice_sendpage - splice data from a pipe to a socket
858 * @pipe: pipe to splice from
859 * @out: socket to write to
860 * @ppos: position in @out
861 * @len: number of bytes to splice
862 * @flags: splice modifier flags
865 * Will send @len bytes from the pipe to a network socket. No data copying
869 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
870 loff_t *ppos, size_t len, unsigned int flags)
872 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
875 EXPORT_SYMBOL(generic_splice_sendpage);
878 * Attempt to initiate a splice from pipe to file.
880 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
881 loff_t *ppos, size_t len, unsigned int flags)
885 if (unlikely(!out->f_op || !out->f_op->splice_write))
888 if (unlikely(!(out->f_mode & FMODE_WRITE)))
891 ret = rw_verify_area(WRITE, out, ppos, len);
892 if (unlikely(ret < 0))
895 return out->f_op->splice_write(pipe, out, ppos, len, flags);
899 * Attempt to initiate a splice from a file to a pipe.
901 static long do_splice_to(struct file *in, loff_t *ppos,
902 struct pipe_inode_info *pipe, size_t len,
907 if (unlikely(!in->f_op || !in->f_op->splice_read))
910 if (unlikely(!(in->f_mode & FMODE_READ)))
913 ret = rw_verify_area(READ, in, ppos, len);
914 if (unlikely(ret < 0))
917 return in->f_op->splice_read(in, ppos, pipe, len, flags);
921 * splice_direct_to_actor - splices data directly between two non-pipes
922 * @in: file to splice from
923 * @sd: actor information on where to splice to
924 * @actor: handles the data splicing
927 * This is a special case helper to splice directly between two
928 * points, without requiring an explicit pipe. Internally an allocated
929 * pipe is cached in the process, and reused during the lifetime of
933 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
934 splice_direct_actor *actor)
936 struct pipe_inode_info *pipe;
943 * We require the input being a regular file, as we don't want to
944 * randomly drop data for eg socket -> socket splicing. Use the
945 * piped splicing for that!
947 i_mode = in->f_path.dentry->d_inode->i_mode;
948 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
952 * neither in nor out is a pipe, setup an internal pipe attached to
953 * 'out' and transfer the wanted data from 'in' to 'out' through that
955 pipe = current->splice_pipe;
956 if (unlikely(!pipe)) {
957 pipe = alloc_pipe_info(NULL);
962 * We don't have an immediate reader, but we'll read the stuff
963 * out of the pipe right after the splice_to_pipe(). So set
964 * PIPE_READERS appropriately.
968 current->splice_pipe = pipe;
980 * Don't block on output, we have to drain the direct pipe.
982 sd->flags &= ~SPLICE_F_NONBLOCK;
986 loff_t pos = sd->pos, prev_pos = pos;
988 ret = do_splice_to(in, &pos, pipe, len, flags);
989 if (unlikely(ret <= 0))
993 sd->total_len = read_len;
996 * NOTE: nonblocking mode only applies to the input. We
997 * must not do the output in nonblocking mode as then we
998 * could get stuck data in the internal pipe:
1000 ret = actor(pipe, sd);
1001 if (unlikely(ret <= 0)) {
1010 if (ret < read_len) {
1011 sd->pos = prev_pos + ret;
1017 pipe->nrbufs = pipe->curbuf = 0;
1023 * If we did an incomplete transfer we must release
1024 * the pipe buffers in question:
1026 for (i = 0; i < PIPE_BUFFERS; i++) {
1027 struct pipe_buffer *buf = pipe->bufs + i;
1030 buf->ops->release(pipe, buf);
1040 EXPORT_SYMBOL(splice_direct_to_actor);
1042 static int direct_splice_actor(struct pipe_inode_info *pipe,
1043 struct splice_desc *sd)
1045 struct file *file = sd->u.file;
1047 return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1051 * do_splice_direct - splices data directly between two files
1052 * @in: file to splice from
1053 * @ppos: input file offset
1054 * @out: file to splice to
1055 * @len: number of bytes to splice
1056 * @flags: splice modifier flags
1059 * For use by do_sendfile(). splice can easily emulate sendfile, but
1060 * doing it in the application would incur an extra system call
1061 * (splice in + splice out, as compared to just sendfile()). So this helper
1062 * can splice directly through a process-private pipe.
1065 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1066 size_t len, unsigned int flags)
1068 struct splice_desc sd = {
1077 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1085 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1086 * location, so checking ->i_pipe is not enough to verify that this is a
1089 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1091 if (S_ISFIFO(inode->i_mode))
1092 return inode->i_pipe;
1098 * Determine where to splice to/from.
1100 static long do_splice(struct file *in, loff_t __user *off_in,
1101 struct file *out, loff_t __user *off_out,
1102 size_t len, unsigned int flags)
1104 struct pipe_inode_info *pipe;
1105 loff_t offset, *off;
1108 pipe = pipe_info(in->f_path.dentry->d_inode);
1113 if (out->f_op->llseek == no_llseek)
1115 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1121 ret = do_splice_from(pipe, out, off, len, flags);
1123 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1129 pipe = pipe_info(out->f_path.dentry->d_inode);
1134 if (in->f_op->llseek == no_llseek)
1136 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1142 ret = do_splice_to(in, off, pipe, len, flags);
1144 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1154 * Do a copy-from-user while holding the mmap_semaphore for reading, in a
1155 * manner safe from deadlocking with simultaneous mmap() (grabbing mmap_sem
1156 * for writing) and page faulting on the user memory pointed to by src.
1157 * This assumes that we will very rarely hit the partial != 0 path, or this
1158 * will not be a win.
1160 static int copy_from_user_mmap_sem(void *dst, const void __user *src, size_t n)
1164 if (!access_ok(VERIFY_READ, src, n))
1167 pagefault_disable();
1168 partial = __copy_from_user_inatomic(dst, src, n);
1172 * Didn't copy everything, drop the mmap_sem and do a faulting copy
1174 if (unlikely(partial)) {
1175 up_read(¤t->mm->mmap_sem);
1176 partial = copy_from_user(dst, src, n);
1177 down_read(¤t->mm->mmap_sem);
1184 * Map an iov into an array of pages and offset/length tupples. With the
1185 * partial_page structure, we can map several non-contiguous ranges into
1186 * our ones pages[] map instead of splitting that operation into pieces.
1187 * Could easily be exported as a generic helper for other users, in which
1188 * case one would probably want to add a 'max_nr_pages' parameter as well.
1190 static int get_iovec_page_array(const struct iovec __user *iov,
1191 unsigned int nr_vecs, struct page **pages,
1192 struct partial_page *partial, int aligned)
1194 int buffers = 0, error = 0;
1196 down_read(¤t->mm->mmap_sem);
1199 unsigned long off, npages;
1206 if (copy_from_user_mmap_sem(&entry, iov, sizeof(entry)))
1209 base = entry.iov_base;
1210 len = entry.iov_len;
1213 * Sanity check this iovec. 0 read succeeds.
1219 if (!access_ok(VERIFY_READ, base, len))
1223 * Get this base offset and number of pages, then map
1224 * in the user pages.
1226 off = (unsigned long) base & ~PAGE_MASK;
1229 * If asked for alignment, the offset must be zero and the
1230 * length a multiple of the PAGE_SIZE.
1233 if (aligned && (off || len & ~PAGE_MASK))
1236 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1237 if (npages > PIPE_BUFFERS - buffers)
1238 npages = PIPE_BUFFERS - buffers;
1240 error = get_user_pages(current, current->mm,
1241 (unsigned long) base, npages, 0, 0,
1242 &pages[buffers], NULL);
1244 if (unlikely(error <= 0))
1248 * Fill this contiguous range into the partial page map.
1250 for (i = 0; i < error; i++) {
1251 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1253 partial[buffers].offset = off;
1254 partial[buffers].len = plen;
1262 * We didn't complete this iov, stop here since it probably
1263 * means we have to move some of this into a pipe to
1264 * be able to continue.
1270 * Don't continue if we mapped fewer pages than we asked for,
1271 * or if we mapped the max number of pages that we have
1274 if (error < npages || buffers == PIPE_BUFFERS)
1281 up_read(¤t->mm->mmap_sem);
1289 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1290 struct splice_desc *sd)
1295 ret = buf->ops->confirm(pipe, buf);
1300 * See if we can use the atomic maps, by prefaulting in the
1301 * pages and doing an atomic copy
1303 if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1304 src = buf->ops->map(pipe, buf, 1);
1305 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1307 buf->ops->unmap(pipe, buf, src);
1315 * No dice, use slow non-atomic map and copy
1317 src = buf->ops->map(pipe, buf, 0);
1320 if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1323 buf->ops->unmap(pipe, buf, src);
1326 sd->u.userptr += ret;
1331 * For lack of a better implementation, implement vmsplice() to userspace
1332 * as a simple copy of the pipes pages to the user iov.
1334 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1335 unsigned long nr_segs, unsigned int flags)
1337 struct pipe_inode_info *pipe;
1338 struct splice_desc sd;
1343 pipe = pipe_info(file->f_path.dentry->d_inode);
1348 mutex_lock(&pipe->inode->i_mutex);
1356 * Get user address base and length for this iovec.
1358 error = get_user(base, &iov->iov_base);
1359 if (unlikely(error))
1361 error = get_user(len, &iov->iov_len);
1362 if (unlikely(error))
1366 * Sanity check this iovec. 0 read succeeds.
1370 if (unlikely(!base)) {
1375 if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1383 sd.u.userptr = base;
1386 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1404 mutex_unlock(&pipe->inode->i_mutex);
1413 * vmsplice splices a user address range into a pipe. It can be thought of
1414 * as splice-from-memory, where the regular splice is splice-from-file (or
1415 * to file). In both cases the output is a pipe, naturally.
1417 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1418 unsigned long nr_segs, unsigned int flags)
1420 struct pipe_inode_info *pipe;
1421 struct page *pages[PIPE_BUFFERS];
1422 struct partial_page partial[PIPE_BUFFERS];
1423 struct splice_pipe_desc spd = {
1427 .ops = &user_page_pipe_buf_ops,
1428 .spd_release = spd_release_page,
1431 pipe = pipe_info(file->f_path.dentry->d_inode);
1435 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1436 flags & SPLICE_F_GIFT);
1437 if (spd.nr_pages <= 0)
1438 return spd.nr_pages;
1440 return splice_to_pipe(pipe, &spd);
1444 * Note that vmsplice only really supports true splicing _from_ user memory
1445 * to a pipe, not the other way around. Splicing from user memory is a simple
1446 * operation that can be supported without any funky alignment restrictions
1447 * or nasty vm tricks. We simply map in the user memory and fill them into
1448 * a pipe. The reverse isn't quite as easy, though. There are two possible
1449 * solutions for that:
1451 * - memcpy() the data internally, at which point we might as well just
1452 * do a regular read() on the buffer anyway.
1453 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1454 * has restriction limitations on both ends of the pipe).
1456 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1459 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1460 unsigned long nr_segs, unsigned int flags)
1466 if (unlikely(nr_segs > UIO_MAXIOV))
1468 else if (unlikely(!nr_segs))
1472 file = fget_light(fd, &fput);
1474 if (file->f_mode & FMODE_WRITE)
1475 error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1476 else if (file->f_mode & FMODE_READ)
1477 error = vmsplice_to_user(file, iov, nr_segs, flags);
1479 fput_light(file, fput);
1485 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1486 int fd_out, loff_t __user *off_out,
1487 size_t len, unsigned int flags)
1490 struct file *in, *out;
1491 int fput_in, fput_out;
1497 in = fget_light(fd_in, &fput_in);
1499 if (in->f_mode & FMODE_READ) {
1500 out = fget_light(fd_out, &fput_out);
1502 if (out->f_mode & FMODE_WRITE)
1503 error = do_splice(in, off_in,
1506 fput_light(out, fput_out);
1510 fput_light(in, fput_in);
1517 * Make sure there's data to read. Wait for input if we can, otherwise
1518 * return an appropriate error.
1520 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1525 * Check ->nrbufs without the inode lock first. This function
1526 * is speculative anyways, so missing one is ok.
1532 mutex_lock(&pipe->inode->i_mutex);
1534 while (!pipe->nrbufs) {
1535 if (signal_pending(current)) {
1541 if (!pipe->waiting_writers) {
1542 if (flags & SPLICE_F_NONBLOCK) {
1550 mutex_unlock(&pipe->inode->i_mutex);
1555 * Make sure there's writeable room. Wait for room if we can, otherwise
1556 * return an appropriate error.
1558 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1563 * Check ->nrbufs without the inode lock first. This function
1564 * is speculative anyways, so missing one is ok.
1566 if (pipe->nrbufs < PIPE_BUFFERS)
1570 mutex_lock(&pipe->inode->i_mutex);
1572 while (pipe->nrbufs >= PIPE_BUFFERS) {
1573 if (!pipe->readers) {
1574 send_sig(SIGPIPE, current, 0);
1578 if (flags & SPLICE_F_NONBLOCK) {
1582 if (signal_pending(current)) {
1586 pipe->waiting_writers++;
1588 pipe->waiting_writers--;
1591 mutex_unlock(&pipe->inode->i_mutex);
1596 * Link contents of ipipe to opipe.
1598 static int link_pipe(struct pipe_inode_info *ipipe,
1599 struct pipe_inode_info *opipe,
1600 size_t len, unsigned int flags)
1602 struct pipe_buffer *ibuf, *obuf;
1603 int ret = 0, i = 0, nbuf;
1606 * Potential ABBA deadlock, work around it by ordering lock
1607 * grabbing by inode address. Otherwise two different processes
1608 * could deadlock (one doing tee from A -> B, the other from B -> A).
1610 inode_double_lock(ipipe->inode, opipe->inode);
1613 if (!opipe->readers) {
1614 send_sig(SIGPIPE, current, 0);
1621 * If we have iterated all input buffers or ran out of
1622 * output room, break.
1624 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1627 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1628 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1631 * Get a reference to this pipe buffer,
1632 * so we can copy the contents over.
1634 ibuf->ops->get(ipipe, ibuf);
1636 obuf = opipe->bufs + nbuf;
1640 * Don't inherit the gift flag, we need to
1641 * prevent multiple steals of this page.
1643 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1645 if (obuf->len > len)
1655 * return EAGAIN if we have the potential of some data in the
1656 * future, otherwise just return 0
1658 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1661 inode_double_unlock(ipipe->inode, opipe->inode);
1664 * If we put data in the output pipe, wakeup any potential readers.
1668 if (waitqueue_active(&opipe->wait))
1669 wake_up_interruptible(&opipe->wait);
1670 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1677 * This is a tee(1) implementation that works on pipes. It doesn't copy
1678 * any data, it simply references the 'in' pages on the 'out' pipe.
1679 * The 'flags' used are the SPLICE_F_* variants, currently the only
1680 * applicable one is SPLICE_F_NONBLOCK.
1682 static long do_tee(struct file *in, struct file *out, size_t len,
1685 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1686 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1690 * Duplicate the contents of ipipe to opipe without actually
1693 if (ipipe && opipe && ipipe != opipe) {
1695 * Keep going, unless we encounter an error. The ipipe/opipe
1696 * ordering doesn't really matter.
1698 ret = link_ipipe_prep(ipipe, flags);
1700 ret = link_opipe_prep(opipe, flags);
1702 ret = link_pipe(ipipe, opipe, len, flags);
1709 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1718 in = fget_light(fdin, &fput_in);
1720 if (in->f_mode & FMODE_READ) {
1722 struct file *out = fget_light(fdout, &fput_out);
1725 if (out->f_mode & FMODE_WRITE)
1726 error = do_tee(in, out, len, flags);
1727 fput_light(out, fput_out);
1730 fput_light(in, fput_in);