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) {
382 if (TestSetPageLocked(page))
388 * page was truncated, stop here. if this isn't the
389 * first page, we'll just complete what we already
392 if (!page->mapping) {
397 * page was already under io and is now done, great
399 if (PageUptodate(page)) {
405 * need to read in the page
407 error = mapping->a_ops->readpage(in, page);
408 if (unlikely(error)) {
410 * We really should re-lookup the page here,
411 * but it complicates things a lot. Instead
412 * lets just do what we already stored, and
413 * we'll get it the next time we are called.
415 if (error == AOP_TRUNCATED_PAGE)
422 * i_size must be checked after ->readpage().
424 isize = i_size_read(mapping->host);
425 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
426 if (unlikely(!isize || index > end_index))
430 * if this is the last page, see if we need to shrink
431 * the length and stop
433 if (end_index == index) {
434 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
435 if (total_len + loff > isize)
438 * force quit after adding this page
441 this_len = min(this_len, loff);
446 partial[page_nr].offset = loff;
447 partial[page_nr].len = this_len;
449 total_len += this_len;
456 * Release any pages at the end, if we quit early. 'i' is how far
457 * we got, 'nr_pages' is how many pages are in the map.
459 while (page_nr < nr_pages)
460 page_cache_release(pages[page_nr++]);
463 return splice_to_pipe(pipe, &spd);
469 * generic_file_splice_read - splice data from file to a pipe
470 * @in: file to splice from
471 * @pipe: pipe to splice to
472 * @len: number of bytes to splice
473 * @flags: splice modifier flags
475 * Will read pages from given file and fill them into a pipe.
477 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
478 struct pipe_inode_info *pipe, size_t len,
488 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
495 if (flags & SPLICE_F_NONBLOCK) {
512 EXPORT_SYMBOL(generic_file_splice_read);
515 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
516 * using sendpage(). Return the number of bytes sent.
518 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
519 struct pipe_buffer *buf, struct splice_desc *sd)
521 struct file *file = sd->file;
522 loff_t pos = sd->pos;
525 ret = buf->ops->pin(pipe, buf);
527 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
529 ret = file->f_op->sendpage(file, buf->page, buf->offset,
530 sd->len, &pos, more);
537 * This is a little more tricky than the file -> pipe splicing. There are
538 * basically three cases:
540 * - Destination page already exists in the address space and there
541 * are users of it. For that case we have no other option that
542 * copying the data. Tough luck.
543 * - Destination page already exists in the address space, but there
544 * are no users of it. Make sure it's uptodate, then drop it. Fall
545 * through to last case.
546 * - Destination page does not exist, we can add the pipe page to
547 * the page cache and avoid the copy.
549 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
550 * sd->flags), we attempt to migrate pages from the pipe to the output
551 * file address space page cache. This is possible if no one else has
552 * the pipe page referenced outside of the pipe and page cache. If
553 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
554 * a new page in the output file page cache and fill/dirty that.
556 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
557 struct splice_desc *sd)
559 struct file *file = sd->file;
560 struct address_space *mapping = file->f_mapping;
561 unsigned int offset, this_len;
567 * make sure the data in this buffer is uptodate
569 ret = buf->ops->pin(pipe, buf);
573 index = sd->pos >> PAGE_CACHE_SHIFT;
574 offset = sd->pos & ~PAGE_CACHE_MASK;
577 if (this_len + offset > PAGE_CACHE_SIZE)
578 this_len = PAGE_CACHE_SIZE - offset;
581 page = find_lock_page(mapping, index);
584 page = page_cache_alloc_cold(mapping);
589 * This will also lock the page
591 ret = add_to_page_cache_lru(page, mapping, index,
597 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
599 loff_t isize = i_size_read(mapping->host);
601 if (ret != AOP_TRUNCATED_PAGE)
603 page_cache_release(page);
604 if (ret == AOP_TRUNCATED_PAGE)
608 * prepare_write() may have instantiated a few blocks
609 * outside i_size. Trim these off again.
611 if (sd->pos + this_len > isize)
612 vmtruncate(mapping->host, isize);
617 if (buf->page != page) {
619 * Careful, ->map() uses KM_USER0!
621 char *src = buf->ops->map(pipe, buf, 1);
622 char *dst = kmap_atomic(page, KM_USER1);
624 memcpy(dst + offset, src + buf->offset, this_len);
625 flush_dcache_page(page);
626 kunmap_atomic(dst, KM_USER1);
627 buf->ops->unmap(pipe, buf, src);
630 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
632 if (ret == AOP_TRUNCATED_PAGE) {
633 page_cache_release(page);
639 * Partial write has happened, so 'ret' already initialized by
640 * number of bytes written, Where is nothing we have to do here.
645 * Return the number of bytes written and mark page as
646 * accessed, we are now done!
648 mark_page_accessed(page);
649 balance_dirty_pages_ratelimited(mapping);
651 page_cache_release(page);
658 * Pipe input worker. Most of this logic works like a regular pipe, the
659 * key here is the 'actor' worker passed in that actually moves the data
660 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
662 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
663 struct file *out, loff_t *ppos, size_t len,
664 unsigned int flags, splice_actor *actor)
666 int ret, do_wakeup, err;
667 struct splice_desc sd;
679 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
680 const struct pipe_buf_operations *ops = buf->ops;
683 if (sd.len > sd.total_len)
684 sd.len = sd.total_len;
686 err = actor(pipe, buf, &sd);
688 if (!ret && err != -ENODATA)
706 ops->release(pipe, buf);
707 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
721 if (!pipe->waiting_writers) {
726 if (flags & SPLICE_F_NONBLOCK) {
732 if (signal_pending(current)) {
740 if (waitqueue_active(&pipe->wait))
741 wake_up_interruptible_sync(&pipe->wait);
742 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
751 if (waitqueue_active(&pipe->wait))
752 wake_up_interruptible(&pipe->wait);
753 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
758 EXPORT_SYMBOL(__splice_from_pipe);
760 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
761 loff_t *ppos, size_t len, unsigned int flags,
765 struct inode *inode = out->f_mapping->host;
768 * The actor worker might be calling ->prepare_write and
769 * ->commit_write. Most of the time, these expect i_mutex to
770 * be held. Since this may result in an ABBA deadlock with
771 * pipe->inode, we have to order lock acquiry here.
773 inode_double_lock(inode, pipe->inode);
774 ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
775 inode_double_unlock(inode, pipe->inode);
781 * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
783 * @out: file to write to
784 * @len: number of bytes to splice
785 * @flags: splice modifier flags
787 * Will either move or copy pages (determined by @flags options) from
788 * the given pipe inode to the given file. The caller is responsible
789 * for acquiring i_mutex on both inodes.
793 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
794 loff_t *ppos, size_t len, unsigned int flags)
796 struct address_space *mapping = out->f_mapping;
797 struct inode *inode = mapping->host;
801 err = remove_suid(out->f_path.dentry);
805 ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
810 * If file or inode is SYNC and we actually wrote some data,
813 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
814 err = generic_osync_inode(inode, mapping,
815 OSYNC_METADATA|OSYNC_DATA);
825 EXPORT_SYMBOL(generic_file_splice_write_nolock);
828 * generic_file_splice_write - splice data from a pipe to a file
830 * @out: file to write to
831 * @len: number of bytes to splice
832 * @flags: splice modifier flags
834 * Will either move or copy pages (determined by @flags options) from
835 * the given pipe inode to the given file.
839 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
840 loff_t *ppos, size_t len, unsigned int flags)
842 struct address_space *mapping = out->f_mapping;
843 struct inode *inode = mapping->host;
847 err = should_remove_suid(out->f_path.dentry);
849 mutex_lock(&inode->i_mutex);
850 err = __remove_suid(out->f_path.dentry, err);
851 mutex_unlock(&inode->i_mutex);
856 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
861 * If file or inode is SYNC and we actually wrote some data,
864 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
865 mutex_lock(&inode->i_mutex);
866 err = generic_osync_inode(inode, mapping,
867 OSYNC_METADATA|OSYNC_DATA);
868 mutex_unlock(&inode->i_mutex);
878 EXPORT_SYMBOL(generic_file_splice_write);
881 * generic_splice_sendpage - splice data from a pipe to a socket
883 * @out: socket to write to
884 * @len: number of bytes to splice
885 * @flags: splice modifier flags
887 * Will send @len bytes from the pipe to a network socket. No data copying
891 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
892 loff_t *ppos, size_t len, unsigned int flags)
894 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
897 EXPORT_SYMBOL(generic_splice_sendpage);
900 * Attempt to initiate a splice from pipe to file.
902 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
903 loff_t *ppos, size_t len, unsigned int flags)
907 if (unlikely(!out->f_op || !out->f_op->splice_write))
910 if (unlikely(!(out->f_mode & FMODE_WRITE)))
913 ret = rw_verify_area(WRITE, out, ppos, len);
914 if (unlikely(ret < 0))
917 return out->f_op->splice_write(pipe, out, ppos, len, flags);
921 * Attempt to initiate a splice from a file to a pipe.
923 static long do_splice_to(struct file *in, loff_t *ppos,
924 struct pipe_inode_info *pipe, size_t len,
930 if (unlikely(!in->f_op || !in->f_op->splice_read))
933 if (unlikely(!(in->f_mode & FMODE_READ)))
936 ret = rw_verify_area(READ, in, ppos, len);
937 if (unlikely(ret < 0))
940 isize = i_size_read(in->f_mapping->host);
941 if (unlikely(*ppos >= isize))
944 left = isize - *ppos;
945 if (unlikely(left < len))
948 return in->f_op->splice_read(in, ppos, pipe, len, flags);
951 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
952 size_t len, unsigned int flags)
954 struct pipe_inode_info *pipe;
961 * We require the input being a regular file, as we don't want to
962 * randomly drop data for eg socket -> socket splicing. Use the
963 * piped splicing for that!
965 i_mode = in->f_path.dentry->d_inode->i_mode;
966 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
970 * neither in nor out is a pipe, setup an internal pipe attached to
971 * 'out' and transfer the wanted data from 'in' to 'out' through that
973 pipe = current->splice_pipe;
974 if (unlikely(!pipe)) {
975 pipe = alloc_pipe_info(NULL);
980 * We don't have an immediate reader, but we'll read the stuff
981 * out of the pipe right after the splice_to_pipe(). So set
982 * PIPE_READERS appropriately.
986 current->splice_pipe = pipe;
997 size_t read_len, max_read_len;
1000 * Do at most PIPE_BUFFERS pages worth of transfer:
1002 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
1004 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
1005 if (unlikely(ret < 0))
1011 * NOTE: nonblocking mode only applies to the input. We
1012 * must not do the output in nonblocking mode as then we
1013 * could get stuck data in the internal pipe:
1015 ret = do_splice_from(pipe, out, &out_off, read_len,
1016 flags & ~SPLICE_F_NONBLOCK);
1017 if (unlikely(ret < 0))
1024 * In nonblocking mode, if we got back a short read then
1025 * that was due to either an IO error or due to the
1026 * pagecache entry not being there. In the IO error case
1027 * the _next_ splice attempt will produce a clean IO error
1028 * return value (not a short read), so in both cases it's
1029 * correct to break out of the loop here:
1031 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1035 pipe->nrbufs = pipe->curbuf = 0;
1041 * If we did an incomplete transfer we must release
1042 * the pipe buffers in question:
1044 for (i = 0; i < PIPE_BUFFERS; i++) {
1045 struct pipe_buffer *buf = pipe->bufs + i;
1048 buf->ops->release(pipe, buf);
1052 pipe->nrbufs = pipe->curbuf = 0;
1055 * If we transferred some data, return the number of bytes:
1063 EXPORT_SYMBOL(do_splice_direct);
1066 * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1067 * location, so checking ->i_pipe is not enough to verify that this is a
1070 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1072 if (S_ISFIFO(inode->i_mode))
1073 return inode->i_pipe;
1079 * Determine where to splice to/from.
1081 static long do_splice(struct file *in, loff_t __user *off_in,
1082 struct file *out, loff_t __user *off_out,
1083 size_t len, unsigned int flags)
1085 struct pipe_inode_info *pipe;
1086 loff_t offset, *off;
1089 pipe = pipe_info(in->f_path.dentry->d_inode);
1094 if (out->f_op->llseek == no_llseek)
1096 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1102 ret = do_splice_from(pipe, out, off, len, flags);
1104 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1110 pipe = pipe_info(out->f_path.dentry->d_inode);
1115 if (in->f_op->llseek == no_llseek)
1117 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1123 ret = do_splice_to(in, off, pipe, len, flags);
1125 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1135 * Map an iov into an array of pages and offset/length tupples. With the
1136 * partial_page structure, we can map several non-contiguous ranges into
1137 * our ones pages[] map instead of splitting that operation into pieces.
1138 * Could easily be exported as a generic helper for other users, in which
1139 * case one would probably want to add a 'max_nr_pages' parameter as well.
1141 static int get_iovec_page_array(const struct iovec __user *iov,
1142 unsigned int nr_vecs, struct page **pages,
1143 struct partial_page *partial, int aligned)
1145 int buffers = 0, error = 0;
1148 * It's ok to take the mmap_sem for reading, even
1149 * across a "get_user()".
1151 down_read(¤t->mm->mmap_sem);
1154 unsigned long off, npages;
1160 * Get user address base and length for this iovec.
1162 error = get_user(base, &iov->iov_base);
1163 if (unlikely(error))
1165 error = get_user(len, &iov->iov_len);
1166 if (unlikely(error))
1170 * Sanity check this iovec. 0 read succeeds.
1175 if (unlikely(!base))
1179 * Get this base offset and number of pages, then map
1180 * in the user pages.
1182 off = (unsigned long) base & ~PAGE_MASK;
1185 * If asked for alignment, the offset must be zero and the
1186 * length a multiple of the PAGE_SIZE.
1189 if (aligned && (off || len & ~PAGE_MASK))
1192 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1193 if (npages > PIPE_BUFFERS - buffers)
1194 npages = PIPE_BUFFERS - buffers;
1196 error = get_user_pages(current, current->mm,
1197 (unsigned long) base, npages, 0, 0,
1198 &pages[buffers], NULL);
1200 if (unlikely(error <= 0))
1204 * Fill this contiguous range into the partial page map.
1206 for (i = 0; i < error; i++) {
1207 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1209 partial[buffers].offset = off;
1210 partial[buffers].len = plen;
1218 * We didn't complete this iov, stop here since it probably
1219 * means we have to move some of this into a pipe to
1220 * be able to continue.
1226 * Don't continue if we mapped fewer pages than we asked for,
1227 * or if we mapped the max number of pages that we have
1230 if (error < npages || buffers == PIPE_BUFFERS)
1237 up_read(¤t->mm->mmap_sem);
1246 * vmsplice splices a user address range into a pipe. It can be thought of
1247 * as splice-from-memory, where the regular splice is splice-from-file (or
1248 * to file). In both cases the output is a pipe, naturally.
1250 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1251 * not the other way around. Splicing from user memory is a simple operation
1252 * that can be supported without any funky alignment restrictions or nasty
1253 * vm tricks. We simply map in the user memory and fill them into a pipe.
1254 * The reverse isn't quite as easy, though. There are two possible solutions
1257 * - memcpy() the data internally, at which point we might as well just
1258 * do a regular read() on the buffer anyway.
1259 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1260 * has restriction limitations on both ends of the pipe).
1262 * Alas, it isn't here.
1265 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1266 unsigned long nr_segs, unsigned int flags)
1268 struct pipe_inode_info *pipe;
1269 struct page *pages[PIPE_BUFFERS];
1270 struct partial_page partial[PIPE_BUFFERS];
1271 struct splice_pipe_desc spd = {
1275 .ops = &user_page_pipe_buf_ops,
1278 pipe = pipe_info(file->f_path.dentry->d_inode);
1281 if (unlikely(nr_segs > UIO_MAXIOV))
1283 else if (unlikely(!nr_segs))
1286 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1287 flags & SPLICE_F_GIFT);
1288 if (spd.nr_pages <= 0)
1289 return spd.nr_pages;
1291 return splice_to_pipe(pipe, &spd);
1294 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1295 unsigned long nr_segs, unsigned int flags)
1302 file = fget_light(fd, &fput);
1304 if (file->f_mode & FMODE_WRITE)
1305 error = do_vmsplice(file, iov, nr_segs, flags);
1307 fput_light(file, fput);
1313 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1314 int fd_out, loff_t __user *off_out,
1315 size_t len, unsigned int flags)
1318 struct file *in, *out;
1319 int fput_in, fput_out;
1325 in = fget_light(fd_in, &fput_in);
1327 if (in->f_mode & FMODE_READ) {
1328 out = fget_light(fd_out, &fput_out);
1330 if (out->f_mode & FMODE_WRITE)
1331 error = do_splice(in, off_in,
1334 fput_light(out, fput_out);
1338 fput_light(in, fput_in);
1345 * Make sure there's data to read. Wait for input if we can, otherwise
1346 * return an appropriate error.
1348 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1353 * Check ->nrbufs without the inode lock first. This function
1354 * is speculative anyways, so missing one is ok.
1360 mutex_lock(&pipe->inode->i_mutex);
1362 while (!pipe->nrbufs) {
1363 if (signal_pending(current)) {
1369 if (!pipe->waiting_writers) {
1370 if (flags & SPLICE_F_NONBLOCK) {
1378 mutex_unlock(&pipe->inode->i_mutex);
1383 * Make sure there's writeable room. Wait for room if we can, otherwise
1384 * return an appropriate error.
1386 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1391 * Check ->nrbufs without the inode lock first. This function
1392 * is speculative anyways, so missing one is ok.
1394 if (pipe->nrbufs < PIPE_BUFFERS)
1398 mutex_lock(&pipe->inode->i_mutex);
1400 while (pipe->nrbufs >= PIPE_BUFFERS) {
1401 if (!pipe->readers) {
1402 send_sig(SIGPIPE, current, 0);
1406 if (flags & SPLICE_F_NONBLOCK) {
1410 if (signal_pending(current)) {
1414 pipe->waiting_writers++;
1416 pipe->waiting_writers--;
1419 mutex_unlock(&pipe->inode->i_mutex);
1424 * Link contents of ipipe to opipe.
1426 static int link_pipe(struct pipe_inode_info *ipipe,
1427 struct pipe_inode_info *opipe,
1428 size_t len, unsigned int flags)
1430 struct pipe_buffer *ibuf, *obuf;
1431 int ret = 0, i = 0, nbuf;
1434 * Potential ABBA deadlock, work around it by ordering lock
1435 * grabbing by inode address. Otherwise two different processes
1436 * could deadlock (one doing tee from A -> B, the other from B -> A).
1438 inode_double_lock(ipipe->inode, opipe->inode);
1441 if (!opipe->readers) {
1442 send_sig(SIGPIPE, current, 0);
1449 * If we have iterated all input buffers or ran out of
1450 * output room, break.
1452 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1455 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1456 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1459 * Get a reference to this pipe buffer,
1460 * so we can copy the contents over.
1462 ibuf->ops->get(ipipe, ibuf);
1464 obuf = opipe->bufs + nbuf;
1468 * Don't inherit the gift flag, we need to
1469 * prevent multiple steals of this page.
1471 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1473 if (obuf->len > len)
1482 inode_double_unlock(ipipe->inode, opipe->inode);
1485 * If we put data in the output pipe, wakeup any potential readers.
1489 if (waitqueue_active(&opipe->wait))
1490 wake_up_interruptible(&opipe->wait);
1491 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1498 * This is a tee(1) implementation that works on pipes. It doesn't copy
1499 * any data, it simply references the 'in' pages on the 'out' pipe.
1500 * The 'flags' used are the SPLICE_F_* variants, currently the only
1501 * applicable one is SPLICE_F_NONBLOCK.
1503 static long do_tee(struct file *in, struct file *out, size_t len,
1506 struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1507 struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1511 * Duplicate the contents of ipipe to opipe without actually
1514 if (ipipe && opipe && ipipe != opipe) {
1516 * Keep going, unless we encounter an error. The ipipe/opipe
1517 * ordering doesn't really matter.
1519 ret = link_ipipe_prep(ipipe, flags);
1521 ret = link_opipe_prep(opipe, flags);
1523 ret = link_pipe(ipipe, opipe, len, flags);
1524 if (!ret && (flags & SPLICE_F_NONBLOCK))
1533 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1542 in = fget_light(fdin, &fput_in);
1544 if (in->f_mode & FMODE_READ) {
1546 struct file *out = fget_light(fdout, &fput_out);
1549 if (out->f_mode & FMODE_WRITE)
1550 error = do_tee(in, out, len, flags);
1551 fput_light(out, fput_out);
1554 fput_light(in, fput_in);