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 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, mapping_gfp_mask(mapping));
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 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 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,
336 mapping_gfp_mask(mapping));
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 gfp_t gfp_mask = mapping_gfp_mask(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 * Reuse buf page, if SPLICE_F_MOVE is set and we are doing a full
584 if ((sd->flags & SPLICE_F_MOVE) && this_len == PAGE_CACHE_SIZE) {
586 * If steal succeeds, buf->page is now pruned from the
587 * pagecache and we can reuse it. The page will also be
588 * locked on successful return.
590 if (buf->ops->steal(pipe, buf))
594 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
599 page_cache_get(page);
601 if (!(buf->flags & PIPE_BUF_FLAG_LRU))
605 page = find_lock_page(mapping, index);
608 page = page_cache_alloc_cold(mapping);
613 * This will also lock the page
615 ret = add_to_page_cache_lru(page, mapping, index,
622 * We get here with the page locked. If the page is also
623 * uptodate, we don't need to do more. If it isn't, we
624 * may need to bring it in if we are not going to overwrite
627 if (!PageUptodate(page)) {
628 if (this_len < PAGE_CACHE_SIZE) {
629 ret = mapping->a_ops->readpage(file, page);
635 if (!PageUptodate(page)) {
637 * Page got invalidated, repeat.
639 if (!page->mapping) {
641 page_cache_release(page);
648 SetPageUptodate(page);
652 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
654 loff_t isize = i_size_read(mapping->host);
656 if (ret != AOP_TRUNCATED_PAGE)
658 page_cache_release(page);
659 if (ret == AOP_TRUNCATED_PAGE)
663 * prepare_write() may have instantiated a few blocks
664 * outside i_size. Trim these off again.
666 if (sd->pos + this_len > isize)
667 vmtruncate(mapping->host, isize);
672 if (buf->page != page) {
674 * Careful, ->map() uses KM_USER0!
676 char *src = buf->ops->map(pipe, buf, 1);
677 char *dst = kmap_atomic(page, KM_USER1);
679 memcpy(dst + offset, src + buf->offset, this_len);
680 flush_dcache_page(page);
681 kunmap_atomic(dst, KM_USER1);
682 buf->ops->unmap(pipe, buf, src);
685 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
688 * Return the number of bytes written and mark page as
689 * accessed, we are now done!
692 mark_page_accessed(page);
693 balance_dirty_pages_ratelimited(mapping);
694 } else if (ret == AOP_TRUNCATED_PAGE) {
695 page_cache_release(page);
699 page_cache_release(page);
706 * Pipe input worker. Most of this logic works like a regular pipe, the
707 * key here is the 'actor' worker passed in that actually moves the data
708 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
710 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
711 loff_t *ppos, size_t len, unsigned int flags,
714 int ret, do_wakeup, err;
715 struct splice_desc sd;
716 struct inode *inode = out->f_mapping->host;
727 * The actor worker might be calling ->prepare_write and
728 * ->commit_write. Most of the time, these expect i_mutex to
729 * be held. Since this may result in an ABBA deadlock with
730 * pipe->inode, we have to order lock acquiry here.
732 inode_double_lock(inode, pipe->inode);
736 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
737 struct pipe_buf_operations *ops = buf->ops;
740 if (sd.len > sd.total_len)
741 sd.len = sd.total_len;
743 err = actor(pipe, buf, &sd);
745 if (!ret && err != -ENODATA)
763 ops->release(pipe, buf);
764 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
778 if (!pipe->waiting_writers) {
783 if (flags & SPLICE_F_NONBLOCK) {
789 if (signal_pending(current)) {
797 if (waitqueue_active(&pipe->wait))
798 wake_up_interruptible_sync(&pipe->wait);
799 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
806 inode_double_unlock(inode, pipe->inode);
810 if (waitqueue_active(&pipe->wait))
811 wake_up_interruptible(&pipe->wait);
812 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
819 * generic_file_splice_write - splice data from a pipe to a file
821 * @out: file to write to
822 * @len: number of bytes to splice
823 * @flags: splice modifier flags
825 * Will either move or copy pages (determined by @flags options) from
826 * the given pipe inode to the given file.
830 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
831 loff_t *ppos, size_t len, unsigned int flags)
833 struct address_space *mapping = out->f_mapping;
836 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
838 struct inode *inode = mapping->host;
843 * If file or inode is SYNC and we actually wrote some data,
846 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
849 mutex_lock(&inode->i_mutex);
850 err = generic_osync_inode(inode, mapping,
851 OSYNC_METADATA|OSYNC_DATA);
852 mutex_unlock(&inode->i_mutex);
862 EXPORT_SYMBOL(generic_file_splice_write);
865 * generic_splice_sendpage - splice data from a pipe to a socket
867 * @out: socket to write to
868 * @len: number of bytes to splice
869 * @flags: splice modifier flags
871 * Will send @len bytes from the pipe to a network socket. No data copying
875 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
876 loff_t *ppos, size_t len, unsigned int flags)
878 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
881 EXPORT_SYMBOL(generic_splice_sendpage);
884 * Attempt to initiate a splice from pipe to file.
886 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
887 loff_t *ppos, size_t len, unsigned int flags)
891 if (unlikely(!out->f_op || !out->f_op->splice_write))
894 if (unlikely(!(out->f_mode & FMODE_WRITE)))
897 ret = rw_verify_area(WRITE, out, ppos, len);
898 if (unlikely(ret < 0))
901 return out->f_op->splice_write(pipe, out, ppos, len, flags);
905 * Attempt to initiate a splice from a file to a pipe.
907 static long do_splice_to(struct file *in, loff_t *ppos,
908 struct pipe_inode_info *pipe, size_t len,
914 if (unlikely(!in->f_op || !in->f_op->splice_read))
917 if (unlikely(!(in->f_mode & FMODE_READ)))
920 ret = rw_verify_area(READ, in, ppos, len);
921 if (unlikely(ret < 0))
924 isize = i_size_read(in->f_mapping->host);
925 if (unlikely(*ppos >= isize))
928 left = isize - *ppos;
929 if (unlikely(left < len))
932 return in->f_op->splice_read(in, ppos, pipe, len, flags);
935 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
936 size_t len, unsigned int flags)
938 struct pipe_inode_info *pipe;
945 * We require the input being a regular file, as we don't want to
946 * randomly drop data for eg socket -> socket splicing. Use the
947 * piped splicing for that!
949 i_mode = in->f_dentry->d_inode->i_mode;
950 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
954 * neither in nor out is a pipe, setup an internal pipe attached to
955 * 'out' and transfer the wanted data from 'in' to 'out' through that
957 pipe = current->splice_pipe;
958 if (unlikely(!pipe)) {
959 pipe = alloc_pipe_info(NULL);
964 * We don't have an immediate reader, but we'll read the stuff
965 * out of the pipe right after the splice_to_pipe(). So set
966 * PIPE_READERS appropriately.
970 current->splice_pipe = pipe;
981 size_t read_len, max_read_len;
984 * Do at most PIPE_BUFFERS pages worth of transfer:
986 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
988 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
989 if (unlikely(ret < 0))
995 * NOTE: nonblocking mode only applies to the input. We
996 * must not do the output in nonblocking mode as then we
997 * could get stuck data in the internal pipe:
999 ret = do_splice_from(pipe, out, &out_off, read_len,
1000 flags & ~SPLICE_F_NONBLOCK);
1001 if (unlikely(ret < 0))
1008 * In nonblocking mode, if we got back a short read then
1009 * that was due to either an IO error or due to the
1010 * pagecache entry not being there. In the IO error case
1011 * the _next_ splice attempt will produce a clean IO error
1012 * return value (not a short read), so in both cases it's
1013 * correct to break out of the loop here:
1015 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1019 pipe->nrbufs = pipe->curbuf = 0;
1025 * If we did an incomplete transfer we must release
1026 * the pipe buffers in question:
1028 for (i = 0; i < PIPE_BUFFERS; i++) {
1029 struct pipe_buffer *buf = pipe->bufs + i;
1032 buf->ops->release(pipe, buf);
1036 pipe->nrbufs = pipe->curbuf = 0;
1039 * If we transferred some data, return the number of bytes:
1047 EXPORT_SYMBOL(do_splice_direct);
1050 * Determine where to splice to/from.
1052 static long do_splice(struct file *in, loff_t __user *off_in,
1053 struct file *out, loff_t __user *off_out,
1054 size_t len, unsigned int flags)
1056 struct pipe_inode_info *pipe;
1057 loff_t offset, *off;
1060 pipe = in->f_dentry->d_inode->i_pipe;
1065 if (out->f_op->llseek == no_llseek)
1067 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1073 ret = do_splice_from(pipe, out, off, len, flags);
1075 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1081 pipe = out->f_dentry->d_inode->i_pipe;
1086 if (in->f_op->llseek == no_llseek)
1088 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1094 ret = do_splice_to(in, off, pipe, len, flags);
1096 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1106 * Map an iov into an array of pages and offset/length tupples. With the
1107 * partial_page structure, we can map several non-contiguous ranges into
1108 * our ones pages[] map instead of splitting that operation into pieces.
1109 * Could easily be exported as a generic helper for other users, in which
1110 * case one would probably want to add a 'max_nr_pages' parameter as well.
1112 static int get_iovec_page_array(const struct iovec __user *iov,
1113 unsigned int nr_vecs, struct page **pages,
1114 struct partial_page *partial, int aligned)
1116 int buffers = 0, error = 0;
1119 * It's ok to take the mmap_sem for reading, even
1120 * across a "get_user()".
1122 down_read(¤t->mm->mmap_sem);
1125 unsigned long off, npages;
1131 * Get user address base and length for this iovec.
1133 error = get_user(base, &iov->iov_base);
1134 if (unlikely(error))
1136 error = get_user(len, &iov->iov_len);
1137 if (unlikely(error))
1141 * Sanity check this iovec. 0 read succeeds.
1146 if (unlikely(!base))
1150 * Get this base offset and number of pages, then map
1151 * in the user pages.
1153 off = (unsigned long) base & ~PAGE_MASK;
1156 * If asked for alignment, the offset must be zero and the
1157 * length a multiple of the PAGE_SIZE.
1160 if (aligned && (off || len & ~PAGE_MASK))
1163 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1164 if (npages > PIPE_BUFFERS - buffers)
1165 npages = PIPE_BUFFERS - buffers;
1167 error = get_user_pages(current, current->mm,
1168 (unsigned long) base, npages, 0, 0,
1169 &pages[buffers], NULL);
1171 if (unlikely(error <= 0))
1175 * Fill this contiguous range into the partial page map.
1177 for (i = 0; i < error; i++) {
1178 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1180 partial[buffers].offset = off;
1181 partial[buffers].len = plen;
1189 * We didn't complete this iov, stop here since it probably
1190 * means we have to move some of this into a pipe to
1191 * be able to continue.
1197 * Don't continue if we mapped fewer pages than we asked for,
1198 * or if we mapped the max number of pages that we have
1201 if (error < npages || buffers == PIPE_BUFFERS)
1208 up_read(¤t->mm->mmap_sem);
1217 * vmsplice splices a user address range into a pipe. It can be thought of
1218 * as splice-from-memory, where the regular splice is splice-from-file (or
1219 * to file). In both cases the output is a pipe, naturally.
1221 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1222 * not the other way around. Splicing from user memory is a simple operation
1223 * that can be supported without any funky alignment restrictions or nasty
1224 * vm tricks. We simply map in the user memory and fill them into a pipe.
1225 * The reverse isn't quite as easy, though. There are two possible solutions
1228 * - memcpy() the data internally, at which point we might as well just
1229 * do a regular read() on the buffer anyway.
1230 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1231 * has restriction limitations on both ends of the pipe).
1233 * Alas, it isn't here.
1236 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1237 unsigned long nr_segs, unsigned int flags)
1239 struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1240 struct page *pages[PIPE_BUFFERS];
1241 struct partial_page partial[PIPE_BUFFERS];
1242 struct splice_pipe_desc spd = {
1246 .ops = &user_page_pipe_buf_ops,
1249 if (unlikely(!pipe))
1251 if (unlikely(nr_segs > UIO_MAXIOV))
1253 else if (unlikely(!nr_segs))
1256 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1257 flags & SPLICE_F_GIFT);
1258 if (spd.nr_pages <= 0)
1259 return spd.nr_pages;
1261 return splice_to_pipe(pipe, &spd);
1264 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1265 unsigned long nr_segs, unsigned int flags)
1272 file = fget_light(fd, &fput);
1274 if (file->f_mode & FMODE_WRITE)
1275 error = do_vmsplice(file, iov, nr_segs, flags);
1277 fput_light(file, fput);
1283 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1284 int fd_out, loff_t __user *off_out,
1285 size_t len, unsigned int flags)
1288 struct file *in, *out;
1289 int fput_in, fput_out;
1295 in = fget_light(fd_in, &fput_in);
1297 if (in->f_mode & FMODE_READ) {
1298 out = fget_light(fd_out, &fput_out);
1300 if (out->f_mode & FMODE_WRITE)
1301 error = do_splice(in, off_in,
1304 fput_light(out, fput_out);
1308 fput_light(in, fput_in);
1315 * Make sure there's data to read. Wait for input if we can, otherwise
1316 * return an appropriate error.
1318 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1323 * Check ->nrbufs without the inode lock first. This function
1324 * is speculative anyways, so missing one is ok.
1330 mutex_lock(&pipe->inode->i_mutex);
1332 while (!pipe->nrbufs) {
1333 if (signal_pending(current)) {
1339 if (!pipe->waiting_writers) {
1340 if (flags & SPLICE_F_NONBLOCK) {
1348 mutex_unlock(&pipe->inode->i_mutex);
1353 * Make sure there's writeable room. Wait for room if we can, otherwise
1354 * return an appropriate error.
1356 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1361 * Check ->nrbufs without the inode lock first. This function
1362 * is speculative anyways, so missing one is ok.
1364 if (pipe->nrbufs < PIPE_BUFFERS)
1368 mutex_lock(&pipe->inode->i_mutex);
1370 while (pipe->nrbufs >= PIPE_BUFFERS) {
1371 if (!pipe->readers) {
1372 send_sig(SIGPIPE, current, 0);
1376 if (flags & SPLICE_F_NONBLOCK) {
1380 if (signal_pending(current)) {
1384 pipe->waiting_writers++;
1386 pipe->waiting_writers--;
1389 mutex_unlock(&pipe->inode->i_mutex);
1394 * Link contents of ipipe to opipe.
1396 static int link_pipe(struct pipe_inode_info *ipipe,
1397 struct pipe_inode_info *opipe,
1398 size_t len, unsigned int flags)
1400 struct pipe_buffer *ibuf, *obuf;
1401 int ret = 0, i = 0, nbuf;
1404 * Potential ABBA deadlock, work around it by ordering lock
1405 * grabbing by inode address. Otherwise two different processes
1406 * could deadlock (one doing tee from A -> B, the other from B -> A).
1408 inode_double_lock(ipipe->inode, opipe->inode);
1411 if (!opipe->readers) {
1412 send_sig(SIGPIPE, current, 0);
1419 * If we have iterated all input buffers or ran out of
1420 * output room, break.
1422 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1425 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1426 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1429 * Get a reference to this pipe buffer,
1430 * so we can copy the contents over.
1432 ibuf->ops->get(ipipe, ibuf);
1434 obuf = opipe->bufs + nbuf;
1438 * Don't inherit the gift flag, we need to
1439 * prevent multiple steals of this page.
1441 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1443 if (obuf->len > len)
1452 inode_double_unlock(ipipe->inode, opipe->inode);
1455 * If we put data in the output pipe, wakeup any potential readers.
1459 if (waitqueue_active(&opipe->wait))
1460 wake_up_interruptible(&opipe->wait);
1461 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1468 * This is a tee(1) implementation that works on pipes. It doesn't copy
1469 * any data, it simply references the 'in' pages on the 'out' pipe.
1470 * The 'flags' used are the SPLICE_F_* variants, currently the only
1471 * applicable one is SPLICE_F_NONBLOCK.
1473 static long do_tee(struct file *in, struct file *out, size_t len,
1476 struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1477 struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1481 * Duplicate the contents of ipipe to opipe without actually
1484 if (ipipe && opipe && ipipe != opipe) {
1486 * Keep going, unless we encounter an error. The ipipe/opipe
1487 * ordering doesn't really matter.
1489 ret = link_ipipe_prep(ipipe, flags);
1491 ret = link_opipe_prep(opipe, flags);
1493 ret = link_pipe(ipipe, opipe, len, flags);
1494 if (!ret && (flags & SPLICE_F_NONBLOCK))
1503 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1512 in = fget_light(fdin, &fput_in);
1514 if (in->f_mode & FMODE_READ) {
1516 struct file *out = fget_light(fdout, &fput_out);
1519 if (out->f_mode & FMODE_WRITE)
1520 error = do_tee(in, out, len, flags);
1521 fput_light(out, fput_out);
1524 fput_light(in, fput_in);