2 * "splice": joining two ropes together by interweaving their strands.
4 * This is the "extended pipe" functionality, where a pipe is used as
5 * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6 * buffer that you can use to transfer data from one end to the other.
8 * The traditional unix read/write is extended with a "splice()" operation
9 * that transfers data buffers to or from a pipe buffer.
11 * Named by Larry McVoy, original implementation from Linus, extended by
12 * Jens to support splicing to files, network, direct splicing, etc and
13 * fixing lots of bugs.
15 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/memcontrol.h>
25 #include <linux/mm_inline.h>
26 #include <linux/swap.h>
27 #include <linux/writeback.h>
28 #include <linux/export.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32 #include <linux/gfp.h>
33 #include <linux/socket.h>
34 #include <linux/compat.h>
35 #include <linux/aio.h>
39 * Attempt to steal a page from a pipe buffer. This should perhaps go into
40 * a vm helper function, it's already simplified quite a bit by the
41 * addition of remove_mapping(). If success is returned, the caller may
42 * attempt to reuse this page for another destination.
44 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
45 struct pipe_buffer *buf)
47 struct page *page = buf->page;
48 struct address_space *mapping;
52 mapping = page_mapping(page);
54 WARN_ON(!PageUptodate(page));
57 * At least for ext2 with nobh option, we need to wait on
58 * writeback completing on this page, since we'll remove it
59 * from the pagecache. Otherwise truncate wont wait on the
60 * page, allowing the disk blocks to be reused by someone else
61 * before we actually wrote our data to them. fs corruption
64 wait_on_page_writeback(page);
66 if (page_has_private(page) &&
67 !try_to_release_page(page, GFP_KERNEL))
71 * If we succeeded in removing the mapping, set LRU flag
74 if (remove_mapping(mapping, page)) {
75 buf->flags |= PIPE_BUF_FLAG_LRU;
81 * Raced with truncate or failed to remove page from current
82 * address space, unlock and return failure.
89 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
90 struct pipe_buffer *buf)
92 page_cache_release(buf->page);
93 buf->flags &= ~PIPE_BUF_FLAG_LRU;
97 * Check whether the contents of buf is OK to access. Since the content
98 * is a page cache page, IO may be in flight.
100 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
101 struct pipe_buffer *buf)
103 struct page *page = buf->page;
106 if (!PageUptodate(page)) {
110 * Page got truncated/unhashed. This will cause a 0-byte
111 * splice, if this is the first page.
113 if (!page->mapping) {
119 * Uh oh, read-error from disk.
121 if (!PageUptodate(page)) {
127 * Page is ok afterall, we are done.
138 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
140 .confirm = page_cache_pipe_buf_confirm,
141 .release = page_cache_pipe_buf_release,
142 .steal = page_cache_pipe_buf_steal,
143 .get = generic_pipe_buf_get,
146 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
147 struct pipe_buffer *buf)
149 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
152 buf->flags |= PIPE_BUF_FLAG_LRU;
153 return generic_pipe_buf_steal(pipe, buf);
156 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158 .confirm = generic_pipe_buf_confirm,
159 .release = page_cache_pipe_buf_release,
160 .steal = user_page_pipe_buf_steal,
161 .get = generic_pipe_buf_get,
164 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
167 if (waitqueue_active(&pipe->wait))
168 wake_up_interruptible(&pipe->wait);
169 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
173 * splice_to_pipe - fill passed data into a pipe
174 * @pipe: pipe to fill
178 * @spd contains a map of pages and len/offset tuples, along with
179 * the struct pipe_buf_operations associated with these pages. This
180 * function will link that data to the pipe.
183 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
184 struct splice_pipe_desc *spd)
186 unsigned int spd_pages = spd->nr_pages;
187 int ret, do_wakeup, page_nr;
196 if (!pipe->readers) {
197 send_sig(SIGPIPE, current, 0);
203 if (pipe->nrbufs < pipe->buffers) {
204 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
205 struct pipe_buffer *buf = pipe->bufs + newbuf;
207 buf->page = spd->pages[page_nr];
208 buf->offset = spd->partial[page_nr].offset;
209 buf->len = spd->partial[page_nr].len;
210 buf->private = spd->partial[page_nr].private;
212 if (spd->flags & SPLICE_F_GIFT)
213 buf->flags |= PIPE_BUF_FLAG_GIFT;
222 if (!--spd->nr_pages)
224 if (pipe->nrbufs < pipe->buffers)
230 if (spd->flags & SPLICE_F_NONBLOCK) {
236 if (signal_pending(current)) {
244 if (waitqueue_active(&pipe->wait))
245 wake_up_interruptible_sync(&pipe->wait);
246 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
250 pipe->waiting_writers++;
252 pipe->waiting_writers--;
258 wakeup_pipe_readers(pipe);
260 while (page_nr < spd_pages)
261 spd->spd_release(spd, page_nr++);
266 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
268 page_cache_release(spd->pages[i]);
272 * Check if we need to grow the arrays holding pages and partial page
275 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
277 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
279 spd->nr_pages_max = buffers;
280 if (buffers <= PIPE_DEF_BUFFERS)
283 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
284 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
286 if (spd->pages && spd->partial)
294 void splice_shrink_spd(struct splice_pipe_desc *spd)
296 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
304 __generic_file_splice_read(struct file *in, loff_t *ppos,
305 struct pipe_inode_info *pipe, size_t len,
308 struct address_space *mapping = in->f_mapping;
309 unsigned int loff, nr_pages, req_pages;
310 struct page *pages[PIPE_DEF_BUFFERS];
311 struct partial_page partial[PIPE_DEF_BUFFERS];
313 pgoff_t index, end_index;
316 struct splice_pipe_desc spd = {
319 .nr_pages_max = PIPE_DEF_BUFFERS,
321 .ops = &page_cache_pipe_buf_ops,
322 .spd_release = spd_release_page,
325 if (splice_grow_spd(pipe, &spd))
328 index = *ppos >> PAGE_CACHE_SHIFT;
329 loff = *ppos & ~PAGE_CACHE_MASK;
330 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
331 nr_pages = min(req_pages, spd.nr_pages_max);
334 * Lookup the (hopefully) full range of pages we need.
336 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
337 index += spd.nr_pages;
340 * If find_get_pages_contig() returned fewer pages than we needed,
341 * readahead/allocate the rest and fill in the holes.
343 if (spd.nr_pages < nr_pages)
344 page_cache_sync_readahead(mapping, &in->f_ra, in,
345 index, req_pages - spd.nr_pages);
348 while (spd.nr_pages < nr_pages) {
350 * Page could be there, find_get_pages_contig() breaks on
353 page = find_get_page(mapping, index);
356 * page didn't exist, allocate one.
358 page = page_cache_alloc_cold(mapping);
362 error = add_to_page_cache_lru(page, mapping, index,
364 if (unlikely(error)) {
365 page_cache_release(page);
366 if (error == -EEXIST)
371 * add_to_page_cache() locks the page, unlock it
372 * to avoid convoluting the logic below even more.
377 spd.pages[spd.nr_pages++] = page;
382 * Now loop over the map and see if we need to start IO on any
383 * pages, fill in the partial map, etc.
385 index = *ppos >> PAGE_CACHE_SHIFT;
386 nr_pages = spd.nr_pages;
388 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
389 unsigned int this_len;
395 * this_len is the max we'll use from this page
397 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
398 page = spd.pages[page_nr];
400 if (PageReadahead(page))
401 page_cache_async_readahead(mapping, &in->f_ra, in,
402 page, index, req_pages - page_nr);
405 * If the page isn't uptodate, we may need to start io on it
407 if (!PageUptodate(page)) {
411 * Page was truncated, or invalidated by the
412 * filesystem. Redo the find/create, but this time the
413 * page is kept locked, so there's no chance of another
414 * race with truncate/invalidate.
416 if (!page->mapping) {
418 page = find_or_create_page(mapping, index,
419 mapping_gfp_mask(mapping));
425 page_cache_release(spd.pages[page_nr]);
426 spd.pages[page_nr] = page;
429 * page was already under io and is now done, great
431 if (PageUptodate(page)) {
437 * need to read in the page
439 error = mapping->a_ops->readpage(in, page);
440 if (unlikely(error)) {
442 * We really should re-lookup the page here,
443 * but it complicates things a lot. Instead
444 * lets just do what we already stored, and
445 * we'll get it the next time we are called.
447 if (error == AOP_TRUNCATED_PAGE)
455 * i_size must be checked after PageUptodate.
457 isize = i_size_read(mapping->host);
458 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
459 if (unlikely(!isize || index > end_index))
463 * if this is the last page, see if we need to shrink
464 * the length and stop
466 if (end_index == index) {
470 * max good bytes in this page
472 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
477 * force quit after adding this page
479 this_len = min(this_len, plen - loff);
483 spd.partial[page_nr].offset = loff;
484 spd.partial[page_nr].len = this_len;
492 * Release any pages at the end, if we quit early. 'page_nr' is how far
493 * we got, 'nr_pages' is how many pages are in the map.
495 while (page_nr < nr_pages)
496 page_cache_release(spd.pages[page_nr++]);
497 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
500 error = splice_to_pipe(pipe, &spd);
502 splice_shrink_spd(&spd);
507 * generic_file_splice_read - splice data from file to a pipe
508 * @in: file to splice from
509 * @ppos: position in @in
510 * @pipe: pipe to splice to
511 * @len: number of bytes to splice
512 * @flags: splice modifier flags
515 * Will read pages from given file and fill them into a pipe. Can be
516 * used as long as the address_space operations for the source implements
520 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
521 struct pipe_inode_info *pipe, size_t len,
527 isize = i_size_read(in->f_mapping->host);
528 if (unlikely(*ppos >= isize))
531 left = isize - *ppos;
532 if (unlikely(left < len))
535 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
543 EXPORT_SYMBOL(generic_file_splice_read);
545 static const struct pipe_buf_operations default_pipe_buf_ops = {
547 .confirm = generic_pipe_buf_confirm,
548 .release = generic_pipe_buf_release,
549 .steal = generic_pipe_buf_steal,
550 .get = generic_pipe_buf_get,
553 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
554 struct pipe_buffer *buf)
559 /* Pipe buffer operations for a socket and similar. */
560 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
562 .confirm = generic_pipe_buf_confirm,
563 .release = generic_pipe_buf_release,
564 .steal = generic_pipe_buf_nosteal,
565 .get = generic_pipe_buf_get,
567 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
569 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
570 unsigned long vlen, loff_t offset)
578 /* The cast to a user pointer is valid due to the set_fs() */
579 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
585 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
593 /* The cast to a user pointer is valid due to the set_fs() */
594 res = vfs_write(file, (__force const char __user *)buf, count, &pos);
599 EXPORT_SYMBOL(kernel_write);
601 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
602 struct pipe_inode_info *pipe, size_t len,
605 unsigned int nr_pages;
606 unsigned int nr_freed;
608 struct page *pages[PIPE_DEF_BUFFERS];
609 struct partial_page partial[PIPE_DEF_BUFFERS];
610 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
615 struct splice_pipe_desc spd = {
618 .nr_pages_max = PIPE_DEF_BUFFERS,
620 .ops = &default_pipe_buf_ops,
621 .spd_release = spd_release_page,
624 if (splice_grow_spd(pipe, &spd))
629 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
630 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
635 offset = *ppos & ~PAGE_CACHE_MASK;
636 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
638 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
641 page = alloc_page(GFP_USER);
646 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
647 vec[i].iov_base = (void __user *) page_address(page);
648 vec[i].iov_len = this_len;
655 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
666 for (i = 0; i < spd.nr_pages; i++) {
667 this_len = min_t(size_t, vec[i].iov_len, res);
668 spd.partial[i].offset = 0;
669 spd.partial[i].len = this_len;
671 __free_page(spd.pages[i]);
677 spd.nr_pages -= nr_freed;
679 res = splice_to_pipe(pipe, &spd);
686 splice_shrink_spd(&spd);
690 for (i = 0; i < spd.nr_pages; i++)
691 __free_page(spd.pages[i]);
696 EXPORT_SYMBOL(default_file_splice_read);
699 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
700 * using sendpage(). Return the number of bytes sent.
702 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
703 struct pipe_buffer *buf, struct splice_desc *sd)
705 struct file *file = sd->u.file;
706 loff_t pos = sd->pos;
709 if (!likely(file->f_op->sendpage))
712 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
714 if (sd->len < sd->total_len && pipe->nrbufs > 1)
715 more |= MSG_SENDPAGE_NOTLAST;
717 return file->f_op->sendpage(file, buf->page, buf->offset,
718 sd->len, &pos, more);
722 * This is a little more tricky than the file -> pipe splicing. There are
723 * basically three cases:
725 * - Destination page already exists in the address space and there
726 * are users of it. For that case we have no other option that
727 * copying the data. Tough luck.
728 * - Destination page already exists in the address space, but there
729 * are no users of it. Make sure it's uptodate, then drop it. Fall
730 * through to last case.
731 * - Destination page does not exist, we can add the pipe page to
732 * the page cache and avoid the copy.
734 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
735 * sd->flags), we attempt to migrate pages from the pipe to the output
736 * file address space page cache. This is possible if no one else has
737 * the pipe page referenced outside of the pipe and page cache. If
738 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
739 * a new page in the output file page cache and fill/dirty that.
741 int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
742 struct splice_desc *sd)
744 struct file *file = sd->u.file;
745 struct address_space *mapping = file->f_mapping;
746 unsigned int offset, this_len;
751 offset = sd->pos & ~PAGE_CACHE_MASK;
754 if (this_len + offset > PAGE_CACHE_SIZE)
755 this_len = PAGE_CACHE_SIZE - offset;
757 ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
758 AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
762 if (buf->page != page) {
763 char *src = kmap_atomic(buf->page);
764 char *dst = kmap_atomic(page);
766 memcpy(dst + offset, src + buf->offset, this_len);
767 flush_dcache_page(page);
771 ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
776 EXPORT_SYMBOL(pipe_to_file);
778 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
781 if (waitqueue_active(&pipe->wait))
782 wake_up_interruptible(&pipe->wait);
783 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
787 * splice_from_pipe_feed - feed available data from a pipe to a file
788 * @pipe: pipe to splice from
789 * @sd: information to @actor
790 * @actor: handler that splices the data
793 * This function loops over the pipe and calls @actor to do the
794 * actual moving of a single struct pipe_buffer to the desired
795 * destination. It returns when there's no more buffers left in
796 * the pipe or if the requested number of bytes (@sd->total_len)
797 * have been copied. It returns a positive number (one) if the
798 * pipe needs to be filled with more data, zero if the required
799 * number of bytes have been copied and -errno on error.
801 * This, together with splice_from_pipe_{begin,end,next}, may be
802 * used to implement the functionality of __splice_from_pipe() when
803 * locking is required around copying the pipe buffers to the
806 int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
811 while (pipe->nrbufs) {
812 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
813 const struct pipe_buf_operations *ops = buf->ops;
816 if (sd->len > sd->total_len)
817 sd->len = sd->total_len;
819 ret = buf->ops->confirm(pipe, buf);
826 ret = actor(pipe, buf, sd);
833 sd->num_spliced += ret;
836 sd->total_len -= ret;
840 ops->release(pipe, buf);
841 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
844 sd->need_wakeup = true;
853 EXPORT_SYMBOL(splice_from_pipe_feed);
856 * splice_from_pipe_next - wait for some data to splice from
857 * @pipe: pipe to splice from
858 * @sd: information about the splice operation
861 * This function will wait for some data and return a positive
862 * value (one) if pipe buffers are available. It will return zero
863 * or -errno if no more data needs to be spliced.
865 int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
867 while (!pipe->nrbufs) {
871 if (!pipe->waiting_writers && sd->num_spliced)
874 if (sd->flags & SPLICE_F_NONBLOCK)
877 if (signal_pending(current))
880 if (sd->need_wakeup) {
881 wakeup_pipe_writers(pipe);
882 sd->need_wakeup = false;
890 EXPORT_SYMBOL(splice_from_pipe_next);
893 * splice_from_pipe_begin - start splicing from pipe
894 * @sd: information about the splice operation
897 * This function should be called before a loop containing
898 * splice_from_pipe_next() and splice_from_pipe_feed() to
899 * initialize the necessary fields of @sd.
901 void splice_from_pipe_begin(struct splice_desc *sd)
904 sd->need_wakeup = false;
906 EXPORT_SYMBOL(splice_from_pipe_begin);
909 * splice_from_pipe_end - finish splicing from pipe
910 * @pipe: pipe to splice from
911 * @sd: information about the splice operation
914 * This function will wake up pipe writers if necessary. It should
915 * be called after a loop containing splice_from_pipe_next() and
916 * splice_from_pipe_feed().
918 void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
921 wakeup_pipe_writers(pipe);
923 EXPORT_SYMBOL(splice_from_pipe_end);
926 * __splice_from_pipe - splice data from a pipe to given actor
927 * @pipe: pipe to splice from
928 * @sd: information to @actor
929 * @actor: handler that splices the data
932 * This function does little more than loop over the pipe and call
933 * @actor to do the actual moving of a single struct pipe_buffer to
934 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
938 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
943 splice_from_pipe_begin(sd);
945 ret = splice_from_pipe_next(pipe, sd);
947 ret = splice_from_pipe_feed(pipe, sd, actor);
949 splice_from_pipe_end(pipe, sd);
951 return sd->num_spliced ? sd->num_spliced : ret;
953 EXPORT_SYMBOL(__splice_from_pipe);
956 * splice_from_pipe - splice data from a pipe to a file
957 * @pipe: pipe to splice from
958 * @out: file to splice to
959 * @ppos: position in @out
960 * @len: how many bytes to splice
961 * @flags: splice modifier flags
962 * @actor: handler that splices the data
965 * See __splice_from_pipe. This function locks the pipe inode,
966 * otherwise it's identical to __splice_from_pipe().
969 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
970 loff_t *ppos, size_t len, unsigned int flags,
974 struct splice_desc sd = {
982 ret = __splice_from_pipe(pipe, &sd, actor);
989 * generic_file_splice_write - splice data from a pipe to a file
991 * @out: file to write to
992 * @ppos: position in @out
993 * @len: number of bytes to splice
994 * @flags: splice modifier flags
997 * Will either move or copy pages (determined by @flags options) from
998 * the given pipe inode to the given file.
1002 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1003 loff_t *ppos, size_t len, unsigned int flags)
1005 struct address_space *mapping = out->f_mapping;
1006 struct inode *inode = mapping->host;
1007 struct splice_desc sd = {
1017 splice_from_pipe_begin(&sd);
1019 ret = splice_from_pipe_next(pipe, &sd);
1023 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1024 ret = file_remove_suid(out);
1026 ret = file_update_time(out);
1028 ret = splice_from_pipe_feed(pipe, &sd,
1031 mutex_unlock(&inode->i_mutex);
1033 splice_from_pipe_end(pipe, &sd);
1038 ret = sd.num_spliced;
1043 err = generic_write_sync(out, *ppos, ret);
1048 balance_dirty_pages_ratelimited(mapping);
1054 EXPORT_SYMBOL(generic_file_splice_write);
1057 * iter_file_splice_write - splice data from a pipe to a file
1059 * @out: file to write to
1060 * @ppos: position in @out
1061 * @len: number of bytes to splice
1062 * @flags: splice modifier flags
1065 * Will either move or copy pages (determined by @flags options) from
1066 * the given pipe inode to the given file.
1067 * This one is ->write_iter-based.
1071 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
1072 loff_t *ppos, size_t len, unsigned int flags)
1074 struct splice_desc sd = {
1080 int nbufs = pipe->buffers;
1081 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
1085 if (unlikely(!array))
1090 splice_from_pipe_begin(&sd);
1091 while (sd.total_len) {
1092 struct iov_iter from;
1097 ret = splice_from_pipe_next(pipe, &sd);
1101 if (unlikely(nbufs < pipe->buffers)) {
1103 nbufs = pipe->buffers;
1104 array = kcalloc(nbufs, sizeof(struct bio_vec),
1112 /* build the vector */
1113 left = sd.total_len;
1114 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
1115 struct pipe_buffer *buf = pipe->bufs + idx;
1116 size_t this_len = buf->len;
1118 if (this_len > left)
1121 if (idx == pipe->buffers - 1)
1124 ret = buf->ops->confirm(pipe, buf);
1125 if (unlikely(ret)) {
1126 if (ret == -ENODATA)
1131 array[n].bv_page = buf->page;
1132 array[n].bv_len = this_len;
1133 array[n].bv_offset = buf->offset;
1138 from.type = ITER_BVEC | WRITE;
1141 from.count = sd.total_len - left;
1142 from.iov_offset = 0;
1145 init_sync_kiocb(&kiocb, out);
1146 kiocb.ki_pos = sd.pos;
1147 kiocb.ki_nbytes = sd.total_len - left;
1150 ret = out->f_op->write_iter(&kiocb, &from);
1151 if (-EIOCBQUEUED == ret)
1152 ret = wait_on_sync_kiocb(&kiocb);
1157 sd.num_spliced += ret;
1158 sd.total_len -= ret;
1159 *ppos = sd.pos = kiocb.ki_pos;
1161 /* dismiss the fully eaten buffers, adjust the partial one */
1163 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1164 if (ret >= buf->len) {
1165 const struct pipe_buf_operations *ops = buf->ops;
1169 ops->release(pipe, buf);
1170 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1173 sd.need_wakeup = true;
1183 splice_from_pipe_end(pipe, &sd);
1188 ret = sd.num_spliced;
1193 EXPORT_SYMBOL(iter_file_splice_write);
1195 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1196 struct splice_desc *sd)
1200 loff_t tmp = sd->pos;
1202 data = kmap(buf->page);
1203 ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1209 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1210 struct file *out, loff_t *ppos,
1211 size_t len, unsigned int flags)
1215 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1223 * generic_splice_sendpage - splice data from a pipe to a socket
1224 * @pipe: pipe to splice from
1225 * @out: socket to write to
1226 * @ppos: position in @out
1227 * @len: number of bytes to splice
1228 * @flags: splice modifier flags
1231 * Will send @len bytes from the pipe to a network socket. No data copying
1235 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1236 loff_t *ppos, size_t len, unsigned int flags)
1238 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1241 EXPORT_SYMBOL(generic_splice_sendpage);
1244 * Attempt to initiate a splice from pipe to file.
1246 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1247 loff_t *ppos, size_t len, unsigned int flags)
1249 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1250 loff_t *, size_t, unsigned int);
1252 if (out->f_op->splice_write)
1253 splice_write = out->f_op->splice_write;
1255 splice_write = default_file_splice_write;
1257 return splice_write(pipe, out, ppos, len, flags);
1261 * Attempt to initiate a splice from a file to a pipe.
1263 static long do_splice_to(struct file *in, loff_t *ppos,
1264 struct pipe_inode_info *pipe, size_t len,
1267 ssize_t (*splice_read)(struct file *, loff_t *,
1268 struct pipe_inode_info *, size_t, unsigned int);
1271 if (unlikely(!(in->f_mode & FMODE_READ)))
1274 ret = rw_verify_area(READ, in, ppos, len);
1275 if (unlikely(ret < 0))
1278 if (in->f_op->splice_read)
1279 splice_read = in->f_op->splice_read;
1281 splice_read = default_file_splice_read;
1283 return splice_read(in, ppos, pipe, len, flags);
1287 * splice_direct_to_actor - splices data directly between two non-pipes
1288 * @in: file to splice from
1289 * @sd: actor information on where to splice to
1290 * @actor: handles the data splicing
1293 * This is a special case helper to splice directly between two
1294 * points, without requiring an explicit pipe. Internally an allocated
1295 * pipe is cached in the process, and reused during the lifetime of
1299 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1300 splice_direct_actor *actor)
1302 struct pipe_inode_info *pipe;
1309 * We require the input being a regular file, as we don't want to
1310 * randomly drop data for eg socket -> socket splicing. Use the
1311 * piped splicing for that!
1313 i_mode = file_inode(in)->i_mode;
1314 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1318 * neither in nor out is a pipe, setup an internal pipe attached to
1319 * 'out' and transfer the wanted data from 'in' to 'out' through that
1321 pipe = current->splice_pipe;
1322 if (unlikely(!pipe)) {
1323 pipe = alloc_pipe_info();
1328 * We don't have an immediate reader, but we'll read the stuff
1329 * out of the pipe right after the splice_to_pipe(). So set
1330 * PIPE_READERS appropriately.
1334 current->splice_pipe = pipe;
1342 len = sd->total_len;
1346 * Don't block on output, we have to drain the direct pipe.
1348 sd->flags &= ~SPLICE_F_NONBLOCK;
1352 loff_t pos = sd->pos, prev_pos = pos;
1354 ret = do_splice_to(in, &pos, pipe, len, flags);
1355 if (unlikely(ret <= 0))
1359 sd->total_len = read_len;
1362 * NOTE: nonblocking mode only applies to the input. We
1363 * must not do the output in nonblocking mode as then we
1364 * could get stuck data in the internal pipe:
1366 ret = actor(pipe, sd);
1367 if (unlikely(ret <= 0)) {
1376 if (ret < read_len) {
1377 sd->pos = prev_pos + ret;
1383 pipe->nrbufs = pipe->curbuf = 0;
1389 * If we did an incomplete transfer we must release
1390 * the pipe buffers in question:
1392 for (i = 0; i < pipe->buffers; i++) {
1393 struct pipe_buffer *buf = pipe->bufs + i;
1396 buf->ops->release(pipe, buf);
1406 EXPORT_SYMBOL(splice_direct_to_actor);
1408 static int direct_splice_actor(struct pipe_inode_info *pipe,
1409 struct splice_desc *sd)
1411 struct file *file = sd->u.file;
1413 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1418 * do_splice_direct - splices data directly between two files
1419 * @in: file to splice from
1420 * @ppos: input file offset
1421 * @out: file to splice to
1422 * @opos: output file offset
1423 * @len: number of bytes to splice
1424 * @flags: splice modifier flags
1427 * For use by do_sendfile(). splice can easily emulate sendfile, but
1428 * doing it in the application would incur an extra system call
1429 * (splice in + splice out, as compared to just sendfile()). So this helper
1430 * can splice directly through a process-private pipe.
1433 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1434 loff_t *opos, size_t len, unsigned int flags)
1436 struct splice_desc sd = {
1446 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1449 if (unlikely(out->f_flags & O_APPEND))
1452 ret = rw_verify_area(WRITE, out, opos, len);
1453 if (unlikely(ret < 0))
1456 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1463 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1464 struct pipe_inode_info *opipe,
1465 size_t len, unsigned int flags);
1468 * Determine where to splice to/from.
1470 static long do_splice(struct file *in, loff_t __user *off_in,
1471 struct file *out, loff_t __user *off_out,
1472 size_t len, unsigned int flags)
1474 struct pipe_inode_info *ipipe;
1475 struct pipe_inode_info *opipe;
1479 ipipe = get_pipe_info(in);
1480 opipe = get_pipe_info(out);
1482 if (ipipe && opipe) {
1483 if (off_in || off_out)
1486 if (!(in->f_mode & FMODE_READ))
1489 if (!(out->f_mode & FMODE_WRITE))
1492 /* Splicing to self would be fun, but... */
1496 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1503 if (!(out->f_mode & FMODE_PWRITE))
1505 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1508 offset = out->f_pos;
1511 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1514 if (unlikely(out->f_flags & O_APPEND))
1517 ret = rw_verify_area(WRITE, out, &offset, len);
1518 if (unlikely(ret < 0))
1521 file_start_write(out);
1522 ret = do_splice_from(ipipe, out, &offset, len, flags);
1523 file_end_write(out);
1526 out->f_pos = offset;
1527 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1537 if (!(in->f_mode & FMODE_PREAD))
1539 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1545 ret = do_splice_to(in, &offset, opipe, len, flags);
1549 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1559 * Map an iov into an array of pages and offset/length tupples. With the
1560 * partial_page structure, we can map several non-contiguous ranges into
1561 * our ones pages[] map instead of splitting that operation into pieces.
1562 * Could easily be exported as a generic helper for other users, in which
1563 * case one would probably want to add a 'max_nr_pages' parameter as well.
1565 static int get_iovec_page_array(const struct iovec __user *iov,
1566 unsigned int nr_vecs, struct page **pages,
1567 struct partial_page *partial, bool aligned,
1568 unsigned int pipe_buffers)
1570 int buffers = 0, error = 0;
1573 unsigned long off, npages;
1580 if (copy_from_user(&entry, iov, sizeof(entry)))
1583 base = entry.iov_base;
1584 len = entry.iov_len;
1587 * Sanity check this iovec. 0 read succeeds.
1593 if (!access_ok(VERIFY_READ, base, len))
1597 * Get this base offset and number of pages, then map
1598 * in the user pages.
1600 off = (unsigned long) base & ~PAGE_MASK;
1603 * If asked for alignment, the offset must be zero and the
1604 * length a multiple of the PAGE_SIZE.
1607 if (aligned && (off || len & ~PAGE_MASK))
1610 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1611 if (npages > pipe_buffers - buffers)
1612 npages = pipe_buffers - buffers;
1614 error = get_user_pages_fast((unsigned long)base, npages,
1615 0, &pages[buffers]);
1617 if (unlikely(error <= 0))
1621 * Fill this contiguous range into the partial page map.
1623 for (i = 0; i < error; i++) {
1624 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1626 partial[buffers].offset = off;
1627 partial[buffers].len = plen;
1635 * We didn't complete this iov, stop here since it probably
1636 * means we have to move some of this into a pipe to
1637 * be able to continue.
1643 * Don't continue if we mapped fewer pages than we asked for,
1644 * or if we mapped the max number of pages that we have
1647 if (error < npages || buffers == pipe_buffers)
1660 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1661 struct splice_desc *sd)
1663 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1664 return n == sd->len ? n : -EFAULT;
1668 * For lack of a better implementation, implement vmsplice() to userspace
1669 * as a simple copy of the pipes pages to the user iov.
1671 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1672 unsigned long nr_segs, unsigned int flags)
1674 struct pipe_inode_info *pipe;
1675 struct splice_desc sd;
1677 struct iovec iovstack[UIO_FASTIOV];
1678 struct iovec *iov = iovstack;
1679 struct iov_iter iter;
1682 pipe = get_pipe_info(file);
1686 ret = rw_copy_check_uvector(READ, uiov, nr_segs,
1687 ARRAY_SIZE(iovstack), iovstack, &iov);
1691 iov_iter_init(&iter, READ, iov, nr_segs, count);
1694 sd.total_len = count;
1700 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1703 if (iov != iovstack)
1710 * vmsplice splices a user address range into a pipe. It can be thought of
1711 * as splice-from-memory, where the regular splice is splice-from-file (or
1712 * to file). In both cases the output is a pipe, naturally.
1714 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1715 unsigned long nr_segs, unsigned int flags)
1717 struct pipe_inode_info *pipe;
1718 struct page *pages[PIPE_DEF_BUFFERS];
1719 struct partial_page partial[PIPE_DEF_BUFFERS];
1720 struct splice_pipe_desc spd = {
1723 .nr_pages_max = PIPE_DEF_BUFFERS,
1725 .ops = &user_page_pipe_buf_ops,
1726 .spd_release = spd_release_page,
1730 pipe = get_pipe_info(file);
1734 if (splice_grow_spd(pipe, &spd))
1737 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1740 if (spd.nr_pages <= 0)
1743 ret = splice_to_pipe(pipe, &spd);
1745 splice_shrink_spd(&spd);
1750 * Note that vmsplice only really supports true splicing _from_ user memory
1751 * to a pipe, not the other way around. Splicing from user memory is a simple
1752 * operation that can be supported without any funky alignment restrictions
1753 * or nasty vm tricks. We simply map in the user memory and fill them into
1754 * a pipe. The reverse isn't quite as easy, though. There are two possible
1755 * solutions for that:
1757 * - memcpy() the data internally, at which point we might as well just
1758 * do a regular read() on the buffer anyway.
1759 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1760 * has restriction limitations on both ends of the pipe).
1762 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1765 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1766 unsigned long, nr_segs, unsigned int, flags)
1771 if (unlikely(nr_segs > UIO_MAXIOV))
1773 else if (unlikely(!nr_segs))
1779 if (f.file->f_mode & FMODE_WRITE)
1780 error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1781 else if (f.file->f_mode & FMODE_READ)
1782 error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1790 #ifdef CONFIG_COMPAT
1791 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1792 unsigned int, nr_segs, unsigned int, flags)
1795 struct iovec __user *iov;
1796 if (nr_segs > UIO_MAXIOV)
1798 iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1799 for (i = 0; i < nr_segs; i++) {
1800 struct compat_iovec v;
1801 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1802 get_user(v.iov_len, &iov32[i].iov_len) ||
1803 put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1804 put_user(v.iov_len, &iov[i].iov_len))
1807 return sys_vmsplice(fd, iov, nr_segs, flags);
1811 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1812 int, fd_out, loff_t __user *, off_out,
1813 size_t, len, unsigned int, flags)
1824 if (in.file->f_mode & FMODE_READ) {
1825 out = fdget(fd_out);
1827 if (out.file->f_mode & FMODE_WRITE)
1828 error = do_splice(in.file, off_in,
1840 * Make sure there's data to read. Wait for input if we can, otherwise
1841 * return an appropriate error.
1843 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1848 * Check ->nrbufs without the inode lock first. This function
1849 * is speculative anyways, so missing one is ok.
1857 while (!pipe->nrbufs) {
1858 if (signal_pending(current)) {
1864 if (!pipe->waiting_writers) {
1865 if (flags & SPLICE_F_NONBLOCK) {
1878 * Make sure there's writeable room. Wait for room if we can, otherwise
1879 * return an appropriate error.
1881 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1886 * Check ->nrbufs without the inode lock first. This function
1887 * is speculative anyways, so missing one is ok.
1889 if (pipe->nrbufs < pipe->buffers)
1895 while (pipe->nrbufs >= pipe->buffers) {
1896 if (!pipe->readers) {
1897 send_sig(SIGPIPE, current, 0);
1901 if (flags & SPLICE_F_NONBLOCK) {
1905 if (signal_pending(current)) {
1909 pipe->waiting_writers++;
1911 pipe->waiting_writers--;
1919 * Splice contents of ipipe to opipe.
1921 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1922 struct pipe_inode_info *opipe,
1923 size_t len, unsigned int flags)
1925 struct pipe_buffer *ibuf, *obuf;
1927 bool input_wakeup = false;
1931 ret = ipipe_prep(ipipe, flags);
1935 ret = opipe_prep(opipe, flags);
1940 * Potential ABBA deadlock, work around it by ordering lock
1941 * grabbing by pipe info address. Otherwise two different processes
1942 * could deadlock (one doing tee from A -> B, the other from B -> A).
1944 pipe_double_lock(ipipe, opipe);
1947 if (!opipe->readers) {
1948 send_sig(SIGPIPE, current, 0);
1954 if (!ipipe->nrbufs && !ipipe->writers)
1958 * Cannot make any progress, because either the input
1959 * pipe is empty or the output pipe is full.
1961 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1962 /* Already processed some buffers, break */
1966 if (flags & SPLICE_F_NONBLOCK) {
1972 * We raced with another reader/writer and haven't
1973 * managed to process any buffers. A zero return
1974 * value means EOF, so retry instead.
1981 ibuf = ipipe->bufs + ipipe->curbuf;
1982 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1983 obuf = opipe->bufs + nbuf;
1985 if (len >= ibuf->len) {
1987 * Simply move the whole buffer from ipipe to opipe
1992 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1994 input_wakeup = true;
1997 * Get a reference to this pipe buffer,
1998 * so we can copy the contents over.
2000 ibuf->ops->get(ipipe, ibuf);
2004 * Don't inherit the gift flag, we need to
2005 * prevent multiple steals of this page.
2007 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2011 ibuf->offset += obuf->len;
2012 ibuf->len -= obuf->len;
2022 * If we put data in the output pipe, wakeup any potential readers.
2025 wakeup_pipe_readers(opipe);
2028 wakeup_pipe_writers(ipipe);
2034 * Link contents of ipipe to opipe.
2036 static int link_pipe(struct pipe_inode_info *ipipe,
2037 struct pipe_inode_info *opipe,
2038 size_t len, unsigned int flags)
2040 struct pipe_buffer *ibuf, *obuf;
2041 int ret = 0, i = 0, nbuf;
2044 * Potential ABBA deadlock, work around it by ordering lock
2045 * grabbing by pipe info address. Otherwise two different processes
2046 * could deadlock (one doing tee from A -> B, the other from B -> A).
2048 pipe_double_lock(ipipe, opipe);
2051 if (!opipe->readers) {
2052 send_sig(SIGPIPE, current, 0);
2059 * If we have iterated all input buffers or ran out of
2060 * output room, break.
2062 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
2065 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
2066 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
2069 * Get a reference to this pipe buffer,
2070 * so we can copy the contents over.
2072 ibuf->ops->get(ipipe, ibuf);
2074 obuf = opipe->bufs + nbuf;
2078 * Don't inherit the gift flag, we need to
2079 * prevent multiple steals of this page.
2081 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2083 if (obuf->len > len)
2093 * return EAGAIN if we have the potential of some data in the
2094 * future, otherwise just return 0
2096 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
2103 * If we put data in the output pipe, wakeup any potential readers.
2106 wakeup_pipe_readers(opipe);
2112 * This is a tee(1) implementation that works on pipes. It doesn't copy
2113 * any data, it simply references the 'in' pages on the 'out' pipe.
2114 * The 'flags' used are the SPLICE_F_* variants, currently the only
2115 * applicable one is SPLICE_F_NONBLOCK.
2117 static long do_tee(struct file *in, struct file *out, size_t len,
2120 struct pipe_inode_info *ipipe = get_pipe_info(in);
2121 struct pipe_inode_info *opipe = get_pipe_info(out);
2125 * Duplicate the contents of ipipe to opipe without actually
2128 if (ipipe && opipe && ipipe != opipe) {
2130 * Keep going, unless we encounter an error. The ipipe/opipe
2131 * ordering doesn't really matter.
2133 ret = ipipe_prep(ipipe, flags);
2135 ret = opipe_prep(opipe, flags);
2137 ret = link_pipe(ipipe, opipe, len, flags);
2144 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2155 if (in.file->f_mode & FMODE_READ) {
2156 struct fd out = fdget(fdout);
2158 if (out.file->f_mode & FMODE_WRITE)
2159 error = do_tee(in.file, out.file,