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 int ret, do_wakeup, page_nr;
186 mutex_lock(&pipe->inode->i_mutex);
189 if (!pipe->readers) {
190 send_sig(SIGPIPE, current, 0);
196 if (pipe->nrbufs < PIPE_BUFFERS) {
197 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
198 struct pipe_buffer *buf = pipe->bufs + newbuf;
200 buf->page = spd->pages[page_nr];
201 buf->offset = spd->partial[page_nr].offset;
202 buf->len = spd->partial[page_nr].len;
204 if (spd->flags & SPLICE_F_GIFT)
205 buf->flags |= PIPE_BUF_FLAG_GIFT;
214 if (!--spd->nr_pages)
216 if (pipe->nrbufs < PIPE_BUFFERS)
222 if (spd->flags & SPLICE_F_NONBLOCK) {
228 if (signal_pending(current)) {
236 if (waitqueue_active(&pipe->wait))
237 wake_up_interruptible_sync(&pipe->wait);
238 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
242 pipe->waiting_writers++;
244 pipe->waiting_writers--;
248 mutex_unlock(&pipe->inode->i_mutex);
252 if (waitqueue_active(&pipe->wait))
253 wake_up_interruptible(&pipe->wait);
254 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
257 while (page_nr < spd->nr_pages)
258 page_cache_release(spd->pages[page_nr++]);
264 __generic_file_splice_read(struct file *in, loff_t *ppos,
265 struct pipe_inode_info *pipe, size_t len,
268 struct address_space *mapping = in->f_mapping;
269 unsigned int loff, nr_pages;
270 struct page *pages[PIPE_BUFFERS];
271 struct partial_page partial[PIPE_BUFFERS];
273 pgoff_t index, end_index;
277 struct splice_pipe_desc spd = {
281 .ops = &page_cache_pipe_buf_ops,
284 index = *ppos >> PAGE_CACHE_SHIFT;
285 loff = *ppos & ~PAGE_CACHE_MASK;
286 nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
288 if (nr_pages > PIPE_BUFFERS)
289 nr_pages = PIPE_BUFFERS;
292 * Initiate read-ahead on this page range. however, don't call into
293 * read-ahead if this is a non-zero offset (we are likely doing small
294 * chunk splice and the page is already there) for a single page.
296 if (!loff || nr_pages > 1)
297 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
300 * Now fill in the holes:
306 * Lookup the (hopefully) full range of pages we need.
308 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
311 * If find_get_pages_contig() returned fewer pages than we needed,
314 index += spd.nr_pages;
315 while (spd.nr_pages < nr_pages) {
317 * Page could be there, find_get_pages_contig() breaks on
320 page = find_get_page(mapping, index);
323 * Make sure the read-ahead engine is notified
324 * about this failure.
326 handle_ra_miss(mapping, &in->f_ra, index);
329 * page didn't exist, allocate one.
331 page = page_cache_alloc_cold(mapping);
335 error = add_to_page_cache_lru(page, mapping, index,
337 if (unlikely(error)) {
338 page_cache_release(page);
339 if (error == -EEXIST)
344 * add_to_page_cache() locks the page, unlock it
345 * to avoid convoluting the logic below even more.
350 pages[spd.nr_pages++] = page;
355 * Now loop over the map and see if we need to start IO on any
356 * pages, fill in the partial map, etc.
358 index = *ppos >> PAGE_CACHE_SHIFT;
359 nr_pages = spd.nr_pages;
361 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
362 unsigned int this_len;
368 * this_len is the max we'll use from this page
370 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
371 page = pages[page_nr];
374 * If the page isn't uptodate, we may need to start io on it
376 if (!PageUptodate(page)) {
378 * If in nonblock mode then dont block on waiting
379 * for an in-flight io page
381 if (flags & SPLICE_F_NONBLOCK)
387 * page was truncated, stop here. if this isn't the
388 * first page, we'll just complete what we already
391 if (!page->mapping) {
396 * page was already under io and is now done, great
398 if (PageUptodate(page)) {
404 * need to read in the page
406 error = mapping->a_ops->readpage(in, page);
407 if (unlikely(error)) {
409 * We really should re-lookup the page here,
410 * but it complicates things a lot. Instead
411 * lets just do what we already stored, and
412 * we'll get it the next time we are called.
414 if (error == AOP_TRUNCATED_PAGE)
421 * i_size must be checked after ->readpage().
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) {
433 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
434 if (total_len + loff > isize)
437 * force quit after adding this page
440 this_len = min(this_len, loff);
445 partial[page_nr].offset = loff;
446 partial[page_nr].len = this_len;
448 total_len += this_len;
455 * Release any pages at the end, if we quit early. 'i' is how far
456 * we got, 'nr_pages' is how many pages are in the map.
458 while (page_nr < nr_pages)
459 page_cache_release(pages[page_nr++]);
462 return splice_to_pipe(pipe, &spd);
468 * generic_file_splice_read - splice data from file to a pipe
469 * @in: file to splice from
470 * @pipe: pipe to splice to
471 * @len: number of bytes to splice
472 * @flags: splice modifier flags
474 * Will read pages from given file and fill them into a pipe.
476 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
477 struct pipe_inode_info *pipe, size_t len,
487 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
494 if (flags & SPLICE_F_NONBLOCK) {
511 EXPORT_SYMBOL(generic_file_splice_read);
514 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
515 * using sendpage(). Return the number of bytes sent.
517 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
518 struct pipe_buffer *buf, struct splice_desc *sd)
520 struct file *file = sd->file;
521 loff_t pos = sd->pos;
524 ret = buf->ops->pin(pipe, buf);
526 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
528 ret = file->f_op->sendpage(file, buf->page, buf->offset,
529 sd->len, &pos, more);
536 * This is a little more tricky than the file -> pipe splicing. There are
537 * basically three cases:
539 * - Destination page already exists in the address space and there
540 * are users of it. For that case we have no other option that
541 * copying the data. Tough luck.
542 * - Destination page already exists in the address space, but there
543 * are no users of it. Make sure it's uptodate, then drop it. Fall
544 * through to last case.
545 * - Destination page does not exist, we can add the pipe page to
546 * the page cache and avoid the copy.
548 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
549 * sd->flags), we attempt to migrate pages from the pipe to the output
550 * file address space page cache. This is possible if no one else has
551 * the pipe page referenced outside of the pipe and page cache. If
552 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
553 * a new page in the output file page cache and fill/dirty that.
555 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
556 struct splice_desc *sd)
558 struct file *file = sd->file;
559 struct address_space *mapping = file->f_mapping;
560 unsigned int offset, this_len;
566 * make sure the data in this buffer is uptodate
568 ret = buf->ops->pin(pipe, buf);
572 index = sd->pos >> PAGE_CACHE_SHIFT;
573 offset = sd->pos & ~PAGE_CACHE_MASK;
576 if (this_len + offset > PAGE_CACHE_SIZE)
577 this_len = PAGE_CACHE_SIZE - offset;
580 page = find_lock_page(mapping, index);
583 page = page_cache_alloc_cold(mapping);
588 * This will also lock the page
590 ret = add_to_page_cache_lru(page, mapping, index,
597 * We get here with the page locked. If the page is also
598 * uptodate, we don't need to do more. If it isn't, we
599 * may need to bring it in if we are not going to overwrite
602 if (!PageUptodate(page)) {
603 if (this_len < PAGE_CACHE_SIZE) {
604 ret = mapping->a_ops->readpage(file, page);
610 if (!PageUptodate(page)) {
612 * Page got invalidated, repeat.
614 if (!page->mapping) {
616 page_cache_release(page);
623 SetPageUptodate(page);
626 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
628 loff_t isize = i_size_read(mapping->host);
630 if (ret != AOP_TRUNCATED_PAGE)
632 page_cache_release(page);
633 if (ret == AOP_TRUNCATED_PAGE)
637 * prepare_write() may have instantiated a few blocks
638 * outside i_size. Trim these off again.
640 if (sd->pos + this_len > isize)
641 vmtruncate(mapping->host, isize);
646 if (buf->page != page) {
648 * Careful, ->map() uses KM_USER0!
650 char *src = buf->ops->map(pipe, buf, 1);
651 char *dst = kmap_atomic(page, KM_USER1);
653 memcpy(dst + offset, src + buf->offset, this_len);
654 flush_dcache_page(page);
655 kunmap_atomic(dst, KM_USER1);
656 buf->ops->unmap(pipe, buf, src);
659 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
662 * Return the number of bytes written and mark page as
663 * accessed, we are now done!
666 mark_page_accessed(page);
667 balance_dirty_pages_ratelimited(mapping);
668 } else if (ret == AOP_TRUNCATED_PAGE) {
669 page_cache_release(page);
673 page_cache_release(page);
680 * Pipe input worker. Most of this logic works like a regular pipe, the
681 * key here is the 'actor' worker passed in that actually moves the data
682 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
684 static ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
685 struct file *out, loff_t *ppos, size_t len,
686 unsigned int flags, splice_actor *actor)
688 int ret, do_wakeup, err;
689 struct splice_desc sd;
701 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
702 const struct pipe_buf_operations *ops = buf->ops;
705 if (sd.len > sd.total_len)
706 sd.len = sd.total_len;
708 err = actor(pipe, buf, &sd);
710 if (!ret && err != -ENODATA)
728 ops->release(pipe, buf);
729 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
743 if (!pipe->waiting_writers) {
748 if (flags & SPLICE_F_NONBLOCK) {
754 if (signal_pending(current)) {
762 if (waitqueue_active(&pipe->wait))
763 wake_up_interruptible_sync(&pipe->wait);
764 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
773 if (waitqueue_active(&pipe->wait))
774 wake_up_interruptible(&pipe->wait);
775 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
781 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
782 loff_t *ppos, size_t len, unsigned int flags,
786 struct inode *inode = out->f_mapping->host;
789 * The actor worker might be calling ->prepare_write and
790 * ->commit_write. Most of the time, these expect i_mutex to
791 * be held. Since this may result in an ABBA deadlock with
792 * pipe->inode, we have to order lock acquiry here.
794 inode_double_lock(inode, pipe->inode);
795 ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
796 inode_double_unlock(inode, pipe->inode);
802 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
804 * @out: file to write to
805 * @len: number of bytes to splice
806 * @flags: splice modifier flags
808 * Will either move or copy pages (determined by @flags options) from
809 * the given pipe inode to the given file. The caller is responsible
810 * for acquiring i_mutex on both inodes.
814 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
815 loff_t *ppos, size_t len, unsigned int flags)
817 struct address_space *mapping = out->f_mapping;
818 struct inode *inode = mapping->host;
822 err = remove_suid(out->f_path.dentry);
826 ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
831 * If file or inode is SYNC and we actually wrote some data,
834 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
835 err = generic_osync_inode(inode, mapping,
836 OSYNC_METADATA|OSYNC_DATA);
846 EXPORT_SYMBOL(generic_file_splice_write_nolock);
849 * generic_file_splice_write - splice data from a pipe to a file
851 * @out: file to write to
852 * @len: number of bytes to splice
853 * @flags: splice modifier flags
855 * Will either move or copy pages (determined by @flags options) from
856 * the given pipe inode to the given file.
860 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
861 loff_t *ppos, size_t len, unsigned int flags)
863 struct address_space *mapping = out->f_mapping;
864 struct inode *inode = mapping->host;
868 err = should_remove_suid(out->f_path.dentry);
870 mutex_lock(&inode->i_mutex);
871 err = __remove_suid(out->f_path.dentry, err);
872 mutex_unlock(&inode->i_mutex);
877 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
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);
899 EXPORT_SYMBOL(generic_file_splice_write);
902 * generic_splice_sendpage - splice data from a pipe to a socket
904 * @out: socket to write to
905 * @len: number of bytes to splice
906 * @flags: splice modifier flags
908 * Will send @len bytes from the pipe to a network socket. No data copying
912 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
913 loff_t *ppos, size_t len, unsigned int flags)
915 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
918 EXPORT_SYMBOL(generic_splice_sendpage);
921 * Attempt to initiate a splice from pipe to file.
923 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
924 loff_t *ppos, size_t len, unsigned int flags)
928 if (unlikely(!out->f_op || !out->f_op->splice_write))
931 if (unlikely(!(out->f_mode & FMODE_WRITE)))
934 ret = rw_verify_area(WRITE, out, ppos, len);
935 if (unlikely(ret < 0))
938 return out->f_op->splice_write(pipe, out, ppos, len, flags);
942 * Attempt to initiate a splice from a file to a pipe.
944 static long do_splice_to(struct file *in, loff_t *ppos,
945 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 isize = i_size_read(in->f_mapping->host);
962 if (unlikely(*ppos >= isize))
965 left = isize - *ppos;
966 if (unlikely(left < len))
969 return in->f_op->splice_read(in, ppos, pipe, len, flags);
972 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
973 size_t len, unsigned int flags)
975 struct pipe_inode_info *pipe;
982 * We require the input being a regular file, as we don't want to
983 * randomly drop data for eg socket -> socket splicing. Use the
984 * piped splicing for that!
986 i_mode = in->f_path.dentry->d_inode->i_mode;
987 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
991 * neither in nor out is a pipe, setup an internal pipe attached to
992 * 'out' and transfer the wanted data from 'in' to 'out' through that
994 pipe = current->splice_pipe;
995 if (unlikely(!pipe)) {
996 pipe = alloc_pipe_info(NULL);
1001 * We don't have an immediate reader, but we'll read the stuff
1002 * out of the pipe right after the splice_to_pipe(). So set
1003 * PIPE_READERS appropriately.
1007 current->splice_pipe = pipe;
1018 size_t read_len, max_read_len;
1021 * Do at most PIPE_BUFFERS pages worth of transfer:
1023 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1025 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1026 if (unlikely(ret < 0))
1032 * NOTE: nonblocking mode only applies to the input. We
1033 * must not do the output in nonblocking mode as then we
1034 * could get stuck data in the internal pipe:
1036 ret = do_splice_from(pipe, out, &out_off, read_len,
1037 flags & ~SPLICE_F_NONBLOCK);
1038 if (unlikely(ret < 0))
1045 * In nonblocking mode, if we got back a short read then
1046 * that was due to either an IO error or due to the
1047 * pagecache entry not being there. In the IO error case
1048 * the _next_ splice attempt will produce a clean IO error
1049 * return value (not a short read), so in both cases it's
1050 * correct to break out of the loop here:
1052 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1056 pipe->nrbufs = pipe->curbuf = 0;
1062 * If we did an incomplete transfer we must release
1063 * the pipe buffers in question:
1065 for (i = 0; i < PIPE_BUFFERS; i++) {
1066 struct pipe_buffer *buf = pipe->bufs + i;
1069 buf->ops->release(pipe, buf);
1073 pipe->nrbufs = pipe->curbuf = 0;
1076 * If we transferred some data, return the number of bytes:
1084 EXPORT_SYMBOL(do_splice_direct);
1087 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1088 * location, so checking ->i_pipe is not enough to verify that this is a
1091 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1093 if (S_ISFIFO(inode->i_mode))
1094 return inode->i_pipe;
1100 * Determine where to splice to/from.
1102 static long do_splice(struct file *in, loff_t __user *off_in,
1103 struct file *out, loff_t __user *off_out,
1104 size_t len, unsigned int flags)
1106 struct pipe_inode_info *pipe;
1107 loff_t offset, *off;
1110 pipe = pipe_info(in->f_path.dentry->d_inode);
1115 if (out->f_op->llseek == no_llseek)
1117 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1123 ret = do_splice_from(pipe, out, off, len, flags);
1125 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1131 pipe = pipe_info(out->f_path.dentry->d_inode);
1136 if (in->f_op->llseek == no_llseek)
1138 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1144 ret = do_splice_to(in, off, pipe, len, flags);
1146 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1156 * Map an iov into an array of pages and offset/length tupples. With the
1157 * partial_page structure, we can map several non-contiguous ranges into
1158 * our ones pages[] map instead of splitting that operation into pieces.
1159 * Could easily be exported as a generic helper for other users, in which
1160 * case one would probably want to add a 'max_nr_pages' parameter as well.
1162 static int get_iovec_page_array(const struct iovec __user *iov,
1163 unsigned int nr_vecs, struct page **pages,
1164 struct partial_page *partial, int aligned)
1166 int buffers = 0, error = 0;
1169 * It's ok to take the mmap_sem for reading, even
1170 * across a "get_user()".
1172 down_read(¤t->mm->mmap_sem);
1175 unsigned long off, npages;
1181 * Get user address base and length for this iovec.
1183 error = get_user(base, &iov->iov_base);
1184 if (unlikely(error))
1186 error = get_user(len, &iov->iov_len);
1187 if (unlikely(error))
1191 * Sanity check this iovec. 0 read succeeds.
1196 if (unlikely(!base))
1200 * Get this base offset and number of pages, then map
1201 * in the user pages.
1203 off = (unsigned long) base & ~PAGE_MASK;
1206 * If asked for alignment, the offset must be zero and the
1207 * length a multiple of the PAGE_SIZE.
1210 if (aligned && (off || len & ~PAGE_MASK))
1213 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1214 if (npages > PIPE_BUFFERS - buffers)
1215 npages = PIPE_BUFFERS - buffers;
1217 error = get_user_pages(current, current->mm,
1218 (unsigned long) base, npages, 0, 0,
1219 &pages[buffers], NULL);
1221 if (unlikely(error <= 0))
1225 * Fill this contiguous range into the partial page map.
1227 for (i = 0; i < error; i++) {
1228 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1230 partial[buffers].offset = off;
1231 partial[buffers].len = plen;
1239 * We didn't complete this iov, stop here since it probably
1240 * means we have to move some of this into a pipe to
1241 * be able to continue.
1247 * Don't continue if we mapped fewer pages than we asked for,
1248 * or if we mapped the max number of pages that we have
1251 if (error < npages || buffers == PIPE_BUFFERS)
1258 up_read(¤t->mm->mmap_sem);
1267 * vmsplice splices a user address range into a pipe. It can be thought of
1268 * as splice-from-memory, where the regular splice is splice-from-file (or
1269 * to file). In both cases the output is a pipe, naturally.
1271 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1272 * not the other way around. Splicing from user memory is a simple operation
1273 * that can be supported without any funky alignment restrictions or nasty
1274 * vm tricks. We simply map in the user memory and fill them into a pipe.
1275 * The reverse isn't quite as easy, though. There are two possible solutions
1278 * - memcpy() the data internally, at which point we might as well just
1279 * do a regular read() on the buffer anyway.
1280 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1281 * has restriction limitations on both ends of the pipe).
1283 * Alas, it isn't here.
1286 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1287 unsigned long nr_segs, unsigned int flags)
1289 struct pipe_inode_info *pipe;
1290 struct page *pages[PIPE_BUFFERS];
1291 struct partial_page partial[PIPE_BUFFERS];
1292 struct splice_pipe_desc spd = {
1296 .ops = &user_page_pipe_buf_ops,
1299 pipe = pipe_info(file->f_path.dentry->d_inode);
1302 if (unlikely(nr_segs > UIO_MAXIOV))
1304 else if (unlikely(!nr_segs))
1307 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1308 flags & SPLICE_F_GIFT);
1309 if (spd.nr_pages <= 0)
1310 return spd.nr_pages;
1312 return splice_to_pipe(pipe, &spd);
1315 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1316 unsigned long nr_segs, unsigned int flags)
1323 file = fget_light(fd, &fput);
1325 if (file->f_mode & FMODE_WRITE)
1326 error = do_vmsplice(file, iov, nr_segs, flags);
1328 fput_light(file, fput);
1334 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1335 int fd_out, loff_t __user *off_out,
1336 size_t len, unsigned int flags)
1339 struct file *in, *out;
1340 int fput_in, fput_out;
1346 in = fget_light(fd_in, &fput_in);
1348 if (in->f_mode & FMODE_READ) {
1349 out = fget_light(fd_out, &fput_out);
1351 if (out->f_mode & FMODE_WRITE)
1352 error = do_splice(in, off_in,
1355 fput_light(out, fput_out);
1359 fput_light(in, fput_in);
1366 * Make sure there's data to read. Wait for input if we can, otherwise
1367 * return an appropriate error.
1369 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1374 * Check ->nrbufs without the inode lock first. This function
1375 * is speculative anyways, so missing one is ok.
1381 mutex_lock(&pipe->inode->i_mutex);
1383 while (!pipe->nrbufs) {
1384 if (signal_pending(current)) {
1390 if (!pipe->waiting_writers) {
1391 if (flags & SPLICE_F_NONBLOCK) {
1399 mutex_unlock(&pipe->inode->i_mutex);
1404 * Make sure there's writeable room. Wait for room if we can, otherwise
1405 * return an appropriate error.
1407 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1412 * Check ->nrbufs without the inode lock first. This function
1413 * is speculative anyways, so missing one is ok.
1415 if (pipe->nrbufs < PIPE_BUFFERS)
1419 mutex_lock(&pipe->inode->i_mutex);
1421 while (pipe->nrbufs >= PIPE_BUFFERS) {
1422 if (!pipe->readers) {
1423 send_sig(SIGPIPE, current, 0);
1427 if (flags & SPLICE_F_NONBLOCK) {
1431 if (signal_pending(current)) {
1435 pipe->waiting_writers++;
1437 pipe->waiting_writers--;
1440 mutex_unlock(&pipe->inode->i_mutex);
1445 * Link contents of ipipe to opipe.
1447 static int link_pipe(struct pipe_inode_info *ipipe,
1448 struct pipe_inode_info *opipe,
1449 size_t len, unsigned int flags)
1451 struct pipe_buffer *ibuf, *obuf;
1452 int ret = 0, i = 0, nbuf;
1455 * Potential ABBA deadlock, work around it by ordering lock
1456 * grabbing by inode address. Otherwise two different processes
1457 * could deadlock (one doing tee from A -> B, the other from B -> A).
1459 inode_double_lock(ipipe->inode, opipe->inode);
1462 if (!opipe->readers) {
1463 send_sig(SIGPIPE, current, 0);
1470 * If we have iterated all input buffers or ran out of
1471 * output room, break.
1473 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1476 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1477 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1480 * Get a reference to this pipe buffer,
1481 * so we can copy the contents over.
1483 ibuf->ops->get(ipipe, ibuf);
1485 obuf = opipe->bufs + nbuf;
1489 * Don't inherit the gift flag, we need to
1490 * prevent multiple steals of this page.
1492 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1494 if (obuf->len > len)
1503 inode_double_unlock(ipipe->inode, opipe->inode);
1506 * If we put data in the output pipe, wakeup any potential readers.
1510 if (waitqueue_active(&opipe->wait))
1511 wake_up_interruptible(&opipe->wait);
1512 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1519 * This is a tee(1) implementation that works on pipes. It doesn't copy
1520 * any data, it simply references the 'in' pages on the 'out' pipe.
1521 * The 'flags' used are the SPLICE_F_* variants, currently the only
1522 * applicable one is SPLICE_F_NONBLOCK.
1524 static long do_tee(struct file *in, struct file *out, size_t len,
1527 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1528 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1532 * Duplicate the contents of ipipe to opipe without actually
1535 if (ipipe && opipe && ipipe != opipe) {
1537 * Keep going, unless we encounter an error. The ipipe/opipe
1538 * ordering doesn't really matter.
1540 ret = link_ipipe_prep(ipipe, flags);
1542 ret = link_opipe_prep(opipe, flags);
1544 ret = link_pipe(ipipe, opipe, len, flags);
1545 if (!ret && (flags & SPLICE_F_NONBLOCK))
1554 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1563 in = fget_light(fdin, &fput_in);
1565 if (in->f_mode & FMODE_READ) {
1567 struct file *out = fget_light(fdout, &fput_out);
1570 if (out->f_mode & FMODE_WRITE)
1571 error = do_tee(in, out, len, flags);
1572 fput_light(out, fput_out);
1575 fput_light(in, fput_in);