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);
258 while (page_nr < spd_pages)
259 page_cache_release(spd->pages[page_nr++]);
265 __generic_file_splice_read(struct file *in, loff_t *ppos,
266 struct pipe_inode_info *pipe, size_t len,
269 struct address_space *mapping = in->f_mapping;
270 unsigned int loff, nr_pages;
271 struct page *pages[PIPE_BUFFERS];
272 struct partial_page partial[PIPE_BUFFERS];
274 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 * Don't try to 2nd guess the read-ahead logic, call into
293 * page_cache_readahead() like the page cache reads would do.
295 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
298 * Now fill in the holes:
303 * Lookup the (hopefully) full range of pages we need.
305 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
308 * If find_get_pages_contig() returned fewer pages than we needed,
311 index += spd.nr_pages;
312 while (spd.nr_pages < nr_pages) {
314 * Page could be there, find_get_pages_contig() breaks on
317 page = find_get_page(mapping, index);
320 * Make sure the read-ahead engine is notified
321 * about this failure.
323 handle_ra_miss(mapping, &in->f_ra, index);
326 * page didn't exist, allocate one.
328 page = page_cache_alloc_cold(mapping);
332 error = add_to_page_cache_lru(page, mapping, index,
334 if (unlikely(error)) {
335 page_cache_release(page);
336 if (error == -EEXIST)
341 * add_to_page_cache() locks the page, unlock it
342 * to avoid convoluting the logic below even more.
347 pages[spd.nr_pages++] = page;
352 * Now loop over the map and see if we need to start IO on any
353 * pages, fill in the partial map, etc.
355 index = *ppos >> PAGE_CACHE_SHIFT;
356 nr_pages = spd.nr_pages;
358 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
359 unsigned int this_len;
365 * this_len is the max we'll use from this page
367 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
368 page = pages[page_nr];
371 * If the page isn't uptodate, we may need to start io on it
373 if (!PageUptodate(page)) {
375 * If in nonblock mode then dont block on waiting
376 * for an in-flight io page
378 if (flags & SPLICE_F_NONBLOCK) {
379 if (TestSetPageLocked(page))
385 * page was truncated, stop here. if this isn't the
386 * first page, we'll just complete what we already
389 if (!page->mapping) {
394 * page was already under io and is now done, great
396 if (PageUptodate(page)) {
402 * need to read in the page
404 error = mapping->a_ops->readpage(in, page);
405 if (unlikely(error)) {
407 * We really should re-lookup the page here,
408 * but it complicates things a lot. Instead
409 * lets just do what we already stored, and
410 * we'll get it the next time we are called.
412 if (error == AOP_TRUNCATED_PAGE)
420 * i_size must be checked after PageUptodate.
422 isize = i_size_read(mapping->host);
423 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
424 if (unlikely(!isize || index > end_index))
428 * if this is the last page, see if we need to shrink
429 * the length and stop
431 if (end_index == index) {
435 * max good bytes in this page
437 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
442 * force quit after adding this page
444 this_len = min(this_len, plen - loff);
448 partial[page_nr].offset = loff;
449 partial[page_nr].len = this_len;
457 * Release any pages at the end, if we quit early. 'page_nr' is how far
458 * we got, 'nr_pages' is how many pages are in the map.
460 while (page_nr < nr_pages)
461 page_cache_release(pages[page_nr++]);
464 return splice_to_pipe(pipe, &spd);
470 * generic_file_splice_read - splice data from file to a pipe
471 * @in: file to splice from
472 * @pipe: pipe to splice to
473 * @len: number of bytes to splice
474 * @flags: splice modifier flags
476 * Will read pages from given file and fill them into a pipe.
478 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
479 struct pipe_inode_info *pipe, size_t len,
486 isize = i_size_read(in->f_mapping->host);
487 if (unlikely(*ppos >= isize))
490 left = isize - *ppos;
491 if (unlikely(left < len))
497 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
504 if (flags & SPLICE_F_NONBLOCK) {
521 EXPORT_SYMBOL(generic_file_splice_read);
524 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
525 * using sendpage(). Return the number of bytes sent.
527 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
528 struct pipe_buffer *buf, struct splice_desc *sd)
530 struct file *file = sd->file;
531 loff_t pos = sd->pos;
534 ret = buf->ops->pin(pipe, buf);
536 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
538 ret = file->f_op->sendpage(file, buf->page, buf->offset,
539 sd->len, &pos, more);
546 * This is a little more tricky than the file -> pipe splicing. There are
547 * basically three cases:
549 * - Destination page already exists in the address space and there
550 * are users of it. For that case we have no other option that
551 * copying the data. Tough luck.
552 * - Destination page already exists in the address space, but there
553 * are no users of it. Make sure it's uptodate, then drop it. Fall
554 * through to last case.
555 * - Destination page does not exist, we can add the pipe page to
556 * the page cache and avoid the copy.
558 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
559 * sd->flags), we attempt to migrate pages from the pipe to the output
560 * file address space page cache. This is possible if no one else has
561 * the pipe page referenced outside of the pipe and page cache. If
562 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
563 * a new page in the output file page cache and fill/dirty that.
565 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
566 struct splice_desc *sd)
568 struct file *file = sd->file;
569 struct address_space *mapping = file->f_mapping;
570 unsigned int offset, this_len;
576 * make sure the data in this buffer is uptodate
578 ret = buf->ops->pin(pipe, buf);
582 index = sd->pos >> PAGE_CACHE_SHIFT;
583 offset = sd->pos & ~PAGE_CACHE_MASK;
586 if (this_len + offset > PAGE_CACHE_SIZE)
587 this_len = PAGE_CACHE_SIZE - offset;
590 page = find_lock_page(mapping, index);
593 page = page_cache_alloc_cold(mapping);
598 * This will also lock the page
600 ret = add_to_page_cache_lru(page, mapping, index,
606 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
608 loff_t isize = i_size_read(mapping->host);
610 if (ret != AOP_TRUNCATED_PAGE)
612 page_cache_release(page);
613 if (ret == AOP_TRUNCATED_PAGE)
617 * prepare_write() may have instantiated a few blocks
618 * outside i_size. Trim these off again.
620 if (sd->pos + this_len > isize)
621 vmtruncate(mapping->host, isize);
626 if (buf->page != page) {
628 * Careful, ->map() uses KM_USER0!
630 char *src = buf->ops->map(pipe, buf, 1);
631 char *dst = kmap_atomic(page, KM_USER1);
633 memcpy(dst + offset, src + buf->offset, this_len);
634 flush_dcache_page(page);
635 kunmap_atomic(dst, KM_USER1);
636 buf->ops->unmap(pipe, buf, src);
639 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
641 if (ret == AOP_TRUNCATED_PAGE) {
642 page_cache_release(page);
648 * Partial write has happened, so 'ret' already initialized by
649 * number of bytes written, Where is nothing we have to do here.
654 * Return the number of bytes written and mark page as
655 * accessed, we are now done!
657 mark_page_accessed(page);
659 page_cache_release(page);
666 * Pipe input worker. Most of this logic works like a regular pipe, the
667 * key here is the 'actor' worker passed in that actually moves the data
668 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
670 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
671 struct file *out, loff_t *ppos, size_t len,
672 unsigned int flags, splice_actor *actor)
674 int ret, do_wakeup, err;
675 struct splice_desc sd;
687 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
688 const struct pipe_buf_operations *ops = buf->ops;
691 if (sd.len > sd.total_len)
692 sd.len = sd.total_len;
694 err = actor(pipe, buf, &sd);
696 if (!ret && err != -ENODATA)
714 ops->release(pipe, buf);
715 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
729 if (!pipe->waiting_writers) {
734 if (flags & SPLICE_F_NONBLOCK) {
740 if (signal_pending(current)) {
748 if (waitqueue_active(&pipe->wait))
749 wake_up_interruptible_sync(&pipe->wait);
750 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
759 if (waitqueue_active(&pipe->wait))
760 wake_up_interruptible(&pipe->wait);
761 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
766 EXPORT_SYMBOL(__splice_from_pipe);
768 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
769 loff_t *ppos, size_t len, unsigned int flags,
773 struct inode *inode = out->f_mapping->host;
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, out, ppos, len, flags, 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;
809 err = remove_suid(out->f_path.dentry);
813 ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
815 unsigned long nr_pages;
818 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
821 * If file or inode is SYNC and we actually wrote some data,
824 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
825 err = generic_osync_inode(inode, mapping,
826 OSYNC_METADATA|OSYNC_DATA);
831 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
837 EXPORT_SYMBOL(generic_file_splice_write_nolock);
840 * generic_file_splice_write - splice data from a pipe to a file
842 * @out: file to write to
843 * @len: number of bytes to splice
844 * @flags: splice modifier flags
846 * Will either move or copy pages (determined by @flags options) from
847 * the given pipe inode to the given file.
851 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
852 loff_t *ppos, size_t len, unsigned int flags)
854 struct address_space *mapping = out->f_mapping;
855 struct inode *inode = mapping->host;
859 err = should_remove_suid(out->f_path.dentry);
861 mutex_lock(&inode->i_mutex);
862 err = __remove_suid(out->f_path.dentry, err);
863 mutex_unlock(&inode->i_mutex);
868 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
870 unsigned long nr_pages;
873 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
876 * If file or inode is SYNC and we actually wrote some data,
879 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
880 mutex_lock(&inode->i_mutex);
881 err = generic_osync_inode(inode, mapping,
882 OSYNC_METADATA|OSYNC_DATA);
883 mutex_unlock(&inode->i_mutex);
888 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
894 EXPORT_SYMBOL(generic_file_splice_write);
897 * generic_splice_sendpage - splice data from a pipe to a socket
899 * @out: socket to write to
900 * @len: number of bytes to splice
901 * @flags: splice modifier flags
903 * Will send @len bytes from the pipe to a network socket. No data copying
907 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
908 loff_t *ppos, size_t len, unsigned int flags)
910 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
913 EXPORT_SYMBOL(generic_splice_sendpage);
916 * Attempt to initiate a splice from pipe to file.
918 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
919 loff_t *ppos, size_t len, unsigned int flags)
923 if (unlikely(!out->f_op || !out->f_op->splice_write))
926 if (unlikely(!(out->f_mode & FMODE_WRITE)))
929 ret = rw_verify_area(WRITE, out, ppos, len);
930 if (unlikely(ret < 0))
933 return out->f_op->splice_write(pipe, out, ppos, len, flags);
937 * Attempt to initiate a splice from a file to a pipe.
939 static long do_splice_to(struct file *in, loff_t *ppos,
940 struct pipe_inode_info *pipe, size_t len,
945 if (unlikely(!in->f_op || !in->f_op->splice_read))
948 if (unlikely(!(in->f_mode & FMODE_READ)))
951 ret = rw_verify_area(READ, in, ppos, len);
952 if (unlikely(ret < 0))
955 return in->f_op->splice_read(in, ppos, pipe, len, flags);
958 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
959 size_t len, unsigned int flags)
961 struct pipe_inode_info *pipe;
968 * We require the input being a regular file, as we don't want to
969 * randomly drop data for eg socket -> socket splicing. Use the
970 * piped splicing for that!
972 i_mode = in->f_path.dentry->d_inode->i_mode;
973 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
977 * neither in nor out is a pipe, setup an internal pipe attached to
978 * 'out' and transfer the wanted data from 'in' to 'out' through that
980 pipe = current->splice_pipe;
981 if (unlikely(!pipe)) {
982 pipe = alloc_pipe_info(NULL);
987 * We don't have an immediate reader, but we'll read the stuff
988 * out of the pipe right after the splice_to_pipe(). So set
989 * PIPE_READERS appropriately.
993 current->splice_pipe = pipe;
1004 size_t read_len, max_read_len;
1007 * Do at most PIPE_BUFFERS pages worth of transfer:
1009 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1011 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1012 if (unlikely(ret < 0))
1018 * NOTE: nonblocking mode only applies to the input. We
1019 * must not do the output in nonblocking mode as then we
1020 * could get stuck data in the internal pipe:
1022 ret = do_splice_from(pipe, out, &out_off, read_len,
1023 flags & ~SPLICE_F_NONBLOCK);
1024 if (unlikely(ret < 0))
1031 * In nonblocking mode, if we got back a short read then
1032 * that was due to either an IO error or due to the
1033 * pagecache entry not being there. In the IO error case
1034 * the _next_ splice attempt will produce a clean IO error
1035 * return value (not a short read), so in both cases it's
1036 * correct to break out of the loop here:
1038 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1042 pipe->nrbufs = pipe->curbuf = 0;
1048 * If we did an incomplete transfer we must release
1049 * the pipe buffers in question:
1051 for (i = 0; i < PIPE_BUFFERS; i++) {
1052 struct pipe_buffer *buf = pipe->bufs + i;
1055 buf->ops->release(pipe, buf);
1059 pipe->nrbufs = pipe->curbuf = 0;
1062 * If we transferred some data, return the number of bytes:
1071 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1072 * location, so checking ->i_pipe is not enough to verify that this is a
1075 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1077 if (S_ISFIFO(inode->i_mode))
1078 return inode->i_pipe;
1084 * Determine where to splice to/from.
1086 static long do_splice(struct file *in, loff_t __user *off_in,
1087 struct file *out, loff_t __user *off_out,
1088 size_t len, unsigned int flags)
1090 struct pipe_inode_info *pipe;
1091 loff_t offset, *off;
1094 pipe = pipe_info(in->f_path.dentry->d_inode);
1099 if (out->f_op->llseek == no_llseek)
1101 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1107 ret = do_splice_from(pipe, out, off, len, flags);
1109 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1115 pipe = pipe_info(out->f_path.dentry->d_inode);
1120 if (in->f_op->llseek == no_llseek)
1122 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1128 ret = do_splice_to(in, off, pipe, len, flags);
1130 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1140 * Map an iov into an array of pages and offset/length tupples. With the
1141 * partial_page structure, we can map several non-contiguous ranges into
1142 * our ones pages[] map instead of splitting that operation into pieces.
1143 * Could easily be exported as a generic helper for other users, in which
1144 * case one would probably want to add a 'max_nr_pages' parameter as well.
1146 static int get_iovec_page_array(const struct iovec __user *iov,
1147 unsigned int nr_vecs, struct page **pages,
1148 struct partial_page *partial, int aligned)
1150 int buffers = 0, error = 0;
1153 * It's ok to take the mmap_sem for reading, even
1154 * across a "get_user()".
1156 down_read(¤t->mm->mmap_sem);
1159 unsigned long off, npages;
1165 * Get user address base and length for this iovec.
1167 error = get_user(base, &iov->iov_base);
1168 if (unlikely(error))
1170 error = get_user(len, &iov->iov_len);
1171 if (unlikely(error))
1175 * Sanity check this iovec. 0 read succeeds.
1180 if (unlikely(!base))
1184 * Get this base offset and number of pages, then map
1185 * in the user pages.
1187 off = (unsigned long) base & ~PAGE_MASK;
1190 * If asked for alignment, the offset must be zero and the
1191 * length a multiple of the PAGE_SIZE.
1194 if (aligned && (off || len & ~PAGE_MASK))
1197 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1198 if (npages > PIPE_BUFFERS - buffers)
1199 npages = PIPE_BUFFERS - buffers;
1201 error = get_user_pages(current, current->mm,
1202 (unsigned long) base, npages, 0, 0,
1203 &pages[buffers], NULL);
1205 if (unlikely(error <= 0))
1209 * Fill this contiguous range into the partial page map.
1211 for (i = 0; i < error; i++) {
1212 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1214 partial[buffers].offset = off;
1215 partial[buffers].len = plen;
1223 * We didn't complete this iov, stop here since it probably
1224 * means we have to move some of this into a pipe to
1225 * be able to continue.
1231 * Don't continue if we mapped fewer pages than we asked for,
1232 * or if we mapped the max number of pages that we have
1235 if (error < npages || buffers == PIPE_BUFFERS)
1242 up_read(¤t->mm->mmap_sem);
1251 * vmsplice splices a user address range into a pipe. It can be thought of
1252 * as splice-from-memory, where the regular splice is splice-from-file (or
1253 * to file). In both cases the output is a pipe, naturally.
1255 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1256 * not the other way around. Splicing from user memory is a simple operation
1257 * that can be supported without any funky alignment restrictions or nasty
1258 * vm tricks. We simply map in the user memory and fill them into a pipe.
1259 * The reverse isn't quite as easy, though. There are two possible solutions
1262 * - memcpy() the data internally, at which point we might as well just
1263 * do a regular read() on the buffer anyway.
1264 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1265 * has restriction limitations on both ends of the pipe).
1267 * Alas, it isn't here.
1270 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1271 unsigned long nr_segs, unsigned int flags)
1273 struct pipe_inode_info *pipe;
1274 struct page *pages[PIPE_BUFFERS];
1275 struct partial_page partial[PIPE_BUFFERS];
1276 struct splice_pipe_desc spd = {
1280 .ops = &user_page_pipe_buf_ops,
1283 pipe = pipe_info(file->f_path.dentry->d_inode);
1286 if (unlikely(nr_segs > UIO_MAXIOV))
1288 else if (unlikely(!nr_segs))
1291 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1292 flags & SPLICE_F_GIFT);
1293 if (spd.nr_pages <= 0)
1294 return spd.nr_pages;
1296 return splice_to_pipe(pipe, &spd);
1299 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1300 unsigned long nr_segs, unsigned int flags)
1307 file = fget_light(fd, &fput);
1309 if (file->f_mode & FMODE_WRITE)
1310 error = do_vmsplice(file, iov, nr_segs, flags);
1312 fput_light(file, fput);
1318 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1319 int fd_out, loff_t __user *off_out,
1320 size_t len, unsigned int flags)
1323 struct file *in, *out;
1324 int fput_in, fput_out;
1330 in = fget_light(fd_in, &fput_in);
1332 if (in->f_mode & FMODE_READ) {
1333 out = fget_light(fd_out, &fput_out);
1335 if (out->f_mode & FMODE_WRITE)
1336 error = do_splice(in, off_in,
1339 fput_light(out, fput_out);
1343 fput_light(in, fput_in);
1350 * Make sure there's data to read. Wait for input if we can, otherwise
1351 * return an appropriate error.
1353 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1358 * Check ->nrbufs without the inode lock first. This function
1359 * is speculative anyways, so missing one is ok.
1365 mutex_lock(&pipe->inode->i_mutex);
1367 while (!pipe->nrbufs) {
1368 if (signal_pending(current)) {
1374 if (!pipe->waiting_writers) {
1375 if (flags & SPLICE_F_NONBLOCK) {
1383 mutex_unlock(&pipe->inode->i_mutex);
1388 * Make sure there's writeable room. Wait for room if we can, otherwise
1389 * return an appropriate error.
1391 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1396 * Check ->nrbufs without the inode lock first. This function
1397 * is speculative anyways, so missing one is ok.
1399 if (pipe->nrbufs < PIPE_BUFFERS)
1403 mutex_lock(&pipe->inode->i_mutex);
1405 while (pipe->nrbufs >= PIPE_BUFFERS) {
1406 if (!pipe->readers) {
1407 send_sig(SIGPIPE, current, 0);
1411 if (flags & SPLICE_F_NONBLOCK) {
1415 if (signal_pending(current)) {
1419 pipe->waiting_writers++;
1421 pipe->waiting_writers--;
1424 mutex_unlock(&pipe->inode->i_mutex);
1429 * Link contents of ipipe to opipe.
1431 static int link_pipe(struct pipe_inode_info *ipipe,
1432 struct pipe_inode_info *opipe,
1433 size_t len, unsigned int flags)
1435 struct pipe_buffer *ibuf, *obuf;
1436 int ret = 0, i = 0, nbuf;
1439 * Potential ABBA deadlock, work around it by ordering lock
1440 * grabbing by inode address. Otherwise two different processes
1441 * could deadlock (one doing tee from A -> B, the other from B -> A).
1443 inode_double_lock(ipipe->inode, opipe->inode);
1446 if (!opipe->readers) {
1447 send_sig(SIGPIPE, current, 0);
1454 * If we have iterated all input buffers or ran out of
1455 * output room, break.
1457 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1460 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1461 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1464 * Get a reference to this pipe buffer,
1465 * so we can copy the contents over.
1467 ibuf->ops->get(ipipe, ibuf);
1469 obuf = opipe->bufs + nbuf;
1473 * Don't inherit the gift flag, we need to
1474 * prevent multiple steals of this page.
1476 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1478 if (obuf->len > len)
1487 inode_double_unlock(ipipe->inode, opipe->inode);
1490 * If we put data in the output pipe, wakeup any potential readers.
1494 if (waitqueue_active(&opipe->wait))
1495 wake_up_interruptible(&opipe->wait);
1496 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1503 * This is a tee(1) implementation that works on pipes. It doesn't copy
1504 * any data, it simply references the 'in' pages on the 'out' pipe.
1505 * The 'flags' used are the SPLICE_F_* variants, currently the only
1506 * applicable one is SPLICE_F_NONBLOCK.
1508 static long do_tee(struct file *in, struct file *out, size_t len,
1511 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1512 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1516 * Duplicate the contents of ipipe to opipe without actually
1519 if (ipipe && opipe && ipipe != opipe) {
1521 * Keep going, unless we encounter an error. The ipipe/opipe
1522 * ordering doesn't really matter.
1524 ret = link_ipipe_prep(ipipe, flags);
1526 ret = link_opipe_prep(opipe, flags);
1528 ret = link_pipe(ipipe, opipe, len, flags);
1529 if (!ret && (flags & SPLICE_F_NONBLOCK))
1538 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1547 in = fget_light(fdin, &fput_in);
1549 if (in->f_mode & FMODE_READ) {
1551 struct file *out = fget_light(fdout, &fput_out);
1554 if (out->f_mode & FMODE_WRITE)
1555 error = do_tee(in, out, len, flags);
1556 fput_light(out, fput_out);
1559 fput_light(in, fput_in);