1 // SPDX-License-Identifier: GPL-2.0-only
3 * "splice": joining two ropes together by interweaving their strands.
5 * This is the "extended pipe" functionality, where a pipe is used as
6 * an arbitrary in-memory buffer. Think of a pipe as a small kernel
7 * buffer that you can use to transfer data from one end to the other.
9 * The traditional unix read/write is extended with a "splice()" operation
10 * that transfers data buffers to or from a pipe buffer.
12 * Named by Larry McVoy, original implementation from Linus, extended by
13 * Jens to support splicing to files, network, direct splicing, etc and
14 * fixing lots of bugs.
16 * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
17 * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
18 * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
21 #include <linux/bvec.h>
23 #include <linux/file.h>
24 #include <linux/pagemap.h>
25 #include <linux/splice.h>
26 #include <linux/memcontrol.h>
27 #include <linux/mm_inline.h>
28 #include <linux/swap.h>
29 #include <linux/writeback.h>
30 #include <linux/export.h>
31 #include <linux/syscalls.h>
32 #include <linux/uio.h>
33 #include <linux/fsnotify.h>
34 #include <linux/security.h>
35 #include <linux/gfp.h>
36 #include <linux/net.h>
37 #include <linux/socket.h>
38 #include <linux/sched/signal.h>
43 * Splice doesn't support FMODE_NOWAIT. Since pipes may set this flag to
44 * indicate they support non-blocking reads or writes, we must clear it
45 * here if set to avoid blocking other users of this pipe if splice is
48 static noinline void noinline pipe_clear_nowait(struct file *file)
50 fmode_t fmode = READ_ONCE(file->f_mode);
53 if (!(fmode & FMODE_NOWAIT))
55 } while (!try_cmpxchg(&file->f_mode, &fmode, fmode & ~FMODE_NOWAIT));
59 * Attempt to steal a page from a pipe buffer. This should perhaps go into
60 * a vm helper function, it's already simplified quite a bit by the
61 * addition of remove_mapping(). If success is returned, the caller may
62 * attempt to reuse this page for another destination.
64 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
65 struct pipe_buffer *buf)
67 struct folio *folio = page_folio(buf->page);
68 struct address_space *mapping;
72 mapping = folio_mapping(folio);
74 WARN_ON(!folio_test_uptodate(folio));
77 * At least for ext2 with nobh option, we need to wait on
78 * writeback completing on this folio, since we'll remove it
79 * from the pagecache. Otherwise truncate wont wait on the
80 * folio, allowing the disk blocks to be reused by someone else
81 * before we actually wrote our data to them. fs corruption
84 folio_wait_writeback(folio);
86 if (folio_has_private(folio) &&
87 !filemap_release_folio(folio, GFP_KERNEL))
91 * If we succeeded in removing the mapping, set LRU flag
94 if (remove_mapping(mapping, folio)) {
95 buf->flags |= PIPE_BUF_FLAG_LRU;
101 * Raced with truncate or failed to remove folio from current
102 * address space, unlock and return failure.
109 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
110 struct pipe_buffer *buf)
113 buf->flags &= ~PIPE_BUF_FLAG_LRU;
117 * Check whether the contents of buf is OK to access. Since the content
118 * is a page cache page, IO may be in flight.
120 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
121 struct pipe_buffer *buf)
123 struct folio *folio = page_folio(buf->page);
126 if (!folio_test_uptodate(folio)) {
130 * Folio got truncated/unhashed. This will cause a 0-byte
131 * splice, if this is the first page.
133 if (!folio->mapping) {
139 * Uh oh, read-error from disk.
141 if (!folio_test_uptodate(folio)) {
146 /* Folio is ok after all, we are done */
156 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
157 .confirm = page_cache_pipe_buf_confirm,
158 .release = page_cache_pipe_buf_release,
159 .try_steal = page_cache_pipe_buf_try_steal,
160 .get = generic_pipe_buf_get,
163 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
164 struct pipe_buffer *buf)
166 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
169 buf->flags |= PIPE_BUF_FLAG_LRU;
170 return generic_pipe_buf_try_steal(pipe, buf);
173 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
174 .release = page_cache_pipe_buf_release,
175 .try_steal = user_page_pipe_buf_try_steal,
176 .get = generic_pipe_buf_get,
179 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
182 if (waitqueue_active(&pipe->rd_wait))
183 wake_up_interruptible(&pipe->rd_wait);
184 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
188 * splice_to_pipe - fill passed data into a pipe
189 * @pipe: pipe to fill
193 * @spd contains a map of pages and len/offset tuples, along with
194 * the struct pipe_buf_operations associated with these pages. This
195 * function will link that data to the pipe.
198 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
199 struct splice_pipe_desc *spd)
201 unsigned int spd_pages = spd->nr_pages;
202 unsigned int tail = pipe->tail;
203 unsigned int head = pipe->head;
204 unsigned int mask = pipe->ring_size - 1;
205 int ret = 0, page_nr = 0;
210 if (unlikely(!pipe->readers)) {
211 send_sig(SIGPIPE, current, 0);
216 while (!pipe_full(head, tail, pipe->max_usage)) {
217 struct pipe_buffer *buf = &pipe->bufs[head & mask];
219 buf->page = spd->pages[page_nr];
220 buf->offset = spd->partial[page_nr].offset;
221 buf->len = spd->partial[page_nr].len;
222 buf->private = spd->partial[page_nr].private;
231 if (!--spd->nr_pages)
239 while (page_nr < spd_pages)
240 spd->spd_release(spd, page_nr++);
244 EXPORT_SYMBOL_GPL(splice_to_pipe);
246 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
248 unsigned int head = pipe->head;
249 unsigned int tail = pipe->tail;
250 unsigned int mask = pipe->ring_size - 1;
253 if (unlikely(!pipe->readers)) {
254 send_sig(SIGPIPE, current, 0);
256 } else if (pipe_full(head, tail, pipe->max_usage)) {
259 pipe->bufs[head & mask] = *buf;
260 pipe->head = head + 1;
263 pipe_buf_release(pipe, buf);
266 EXPORT_SYMBOL(add_to_pipe);
269 * Check if we need to grow the arrays holding pages and partial page
272 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
274 unsigned int max_usage = READ_ONCE(pipe->max_usage);
276 spd->nr_pages_max = max_usage;
277 if (max_usage <= PIPE_DEF_BUFFERS)
280 spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
281 spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
284 if (spd->pages && spd->partial)
292 void splice_shrink_spd(struct splice_pipe_desc *spd)
294 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
302 * copy_splice_read - Copy data from a file and splice the copy into a pipe
303 * @in: The file to read from
304 * @ppos: Pointer to the file position to read from
305 * @pipe: The pipe to splice into
306 * @len: The amount to splice
307 * @flags: The SPLICE_F_* flags
309 * This function allocates a bunch of pages sufficient to hold the requested
310 * amount of data (but limited by the remaining pipe capacity), passes it to
311 * the file's ->read_iter() to read into and then splices the used pages into
314 * Return: On success, the number of bytes read will be returned and *@ppos
315 * will be updated if appropriate; 0 will be returned if there is no more data
316 * to be read; -EAGAIN will be returned if the pipe had no space, and some
317 * other negative error code will be returned on error. A short read may occur
318 * if the pipe has insufficient space, we reach the end of the data or we hit a
321 ssize_t copy_splice_read(struct file *in, loff_t *ppos,
322 struct pipe_inode_info *pipe,
323 size_t len, unsigned int flags)
330 size_t used, npages, chunk, remain, keep = 0;
333 /* Work out how much data we can actually add into the pipe */
334 used = pipe_occupancy(pipe->head, pipe->tail);
335 npages = max_t(ssize_t, pipe->max_usage - used, 0);
336 len = min_t(size_t, len, npages * PAGE_SIZE);
337 npages = DIV_ROUND_UP(len, PAGE_SIZE);
339 bv = kzalloc(array_size(npages, sizeof(bv[0])) +
340 array_size(npages, sizeof(struct page *)), GFP_KERNEL);
344 pages = (struct page **)(bv + npages);
345 npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
351 remain = len = min_t(size_t, len, npages * PAGE_SIZE);
353 for (i = 0; i < npages; i++) {
354 chunk = min_t(size_t, PAGE_SIZE, remain);
355 bv[i].bv_page = pages[i];
357 bv[i].bv_len = chunk;
362 iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
363 init_sync_kiocb(&kiocb, in);
364 kiocb.ki_pos = *ppos;
365 ret = call_read_iter(in, &kiocb, &to);
368 keep = DIV_ROUND_UP(ret, PAGE_SIZE);
369 *ppos = kiocb.ki_pos;
373 * Callers of ->splice_read() expect -EAGAIN on "can't put anything in
374 * there", rather than -EFAULT.
379 /* Free any pages that didn't get touched at all. */
381 release_pages(pages + keep, npages - keep);
383 /* Push the remaining pages into the pipe. */
385 for (i = 0; i < keep; i++) {
386 struct pipe_buffer *buf = pipe_head_buf(pipe);
388 chunk = min_t(size_t, remain, PAGE_SIZE);
389 *buf = (struct pipe_buffer) {
390 .ops = &default_pipe_buf_ops,
391 .page = bv[i].bv_page,
402 EXPORT_SYMBOL(copy_splice_read);
404 const struct pipe_buf_operations default_pipe_buf_ops = {
405 .release = generic_pipe_buf_release,
406 .try_steal = generic_pipe_buf_try_steal,
407 .get = generic_pipe_buf_get,
410 /* Pipe buffer operations for a socket and similar. */
411 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
412 .release = generic_pipe_buf_release,
413 .get = generic_pipe_buf_get,
415 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
417 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
420 if (waitqueue_active(&pipe->wr_wait))
421 wake_up_interruptible(&pipe->wr_wait);
422 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
426 * splice_from_pipe_feed - feed available data from a pipe to a file
427 * @pipe: pipe to splice from
428 * @sd: information to @actor
429 * @actor: handler that splices the data
432 * This function loops over the pipe and calls @actor to do the
433 * actual moving of a single struct pipe_buffer to the desired
434 * destination. It returns when there's no more buffers left in
435 * the pipe or if the requested number of bytes (@sd->total_len)
436 * have been copied. It returns a positive number (one) if the
437 * pipe needs to be filled with more data, zero if the required
438 * number of bytes have been copied and -errno on error.
440 * This, together with splice_from_pipe_{begin,end,next}, may be
441 * used to implement the functionality of __splice_from_pipe() when
442 * locking is required around copying the pipe buffers to the
445 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
448 unsigned int head = pipe->head;
449 unsigned int tail = pipe->tail;
450 unsigned int mask = pipe->ring_size - 1;
453 while (!pipe_empty(head, tail)) {
454 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
457 if (sd->len > sd->total_len)
458 sd->len = sd->total_len;
460 ret = pipe_buf_confirm(pipe, buf);
467 ret = actor(pipe, buf, sd);
474 sd->num_spliced += ret;
477 sd->total_len -= ret;
480 pipe_buf_release(pipe, buf);
484 sd->need_wakeup = true;
494 /* We know we have a pipe buffer, but maybe it's empty? */
495 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
497 unsigned int tail = pipe->tail;
498 unsigned int mask = pipe->ring_size - 1;
499 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
501 if (unlikely(!buf->len)) {
502 pipe_buf_release(pipe, buf);
511 * splice_from_pipe_next - wait for some data to splice from
512 * @pipe: pipe to splice from
513 * @sd: information about the splice operation
516 * This function will wait for some data and return a positive
517 * value (one) if pipe buffers are available. It will return zero
518 * or -errno if no more data needs to be spliced.
520 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
523 * Check for signal early to make process killable when there are
524 * always buffers available
526 if (signal_pending(current))
530 while (pipe_empty(pipe->head, pipe->tail)) {
537 if (sd->flags & SPLICE_F_NONBLOCK)
540 if (signal_pending(current))
543 if (sd->need_wakeup) {
544 wakeup_pipe_writers(pipe);
545 sd->need_wakeup = false;
548 pipe_wait_readable(pipe);
551 if (eat_empty_buffer(pipe))
558 * splice_from_pipe_begin - start splicing from pipe
559 * @sd: information about the splice operation
562 * This function should be called before a loop containing
563 * splice_from_pipe_next() and splice_from_pipe_feed() to
564 * initialize the necessary fields of @sd.
566 static void splice_from_pipe_begin(struct splice_desc *sd)
569 sd->need_wakeup = false;
573 * splice_from_pipe_end - finish splicing from pipe
574 * @pipe: pipe to splice from
575 * @sd: information about the splice operation
578 * This function will wake up pipe writers if necessary. It should
579 * be called after a loop containing splice_from_pipe_next() and
580 * splice_from_pipe_feed().
582 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
585 wakeup_pipe_writers(pipe);
589 * __splice_from_pipe - splice data from a pipe to given actor
590 * @pipe: pipe to splice from
591 * @sd: information to @actor
592 * @actor: handler that splices the data
595 * This function does little more than loop over the pipe and call
596 * @actor to do the actual moving of a single struct pipe_buffer to
597 * the desired destination. See pipe_to_file, pipe_to_sendmsg, or
601 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
606 splice_from_pipe_begin(sd);
609 ret = splice_from_pipe_next(pipe, sd);
611 ret = splice_from_pipe_feed(pipe, sd, actor);
613 splice_from_pipe_end(pipe, sd);
615 return sd->num_spliced ? sd->num_spliced : ret;
617 EXPORT_SYMBOL(__splice_from_pipe);
620 * splice_from_pipe - splice data from a pipe to a file
621 * @pipe: pipe to splice from
622 * @out: file to splice to
623 * @ppos: position in @out
624 * @len: how many bytes to splice
625 * @flags: splice modifier flags
626 * @actor: handler that splices the data
629 * See __splice_from_pipe. This function locks the pipe inode,
630 * otherwise it's identical to __splice_from_pipe().
633 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
634 loff_t *ppos, size_t len, unsigned int flags,
638 struct splice_desc sd = {
646 ret = __splice_from_pipe(pipe, &sd, actor);
653 * iter_file_splice_write - splice data from a pipe to a file
655 * @out: file to write to
656 * @ppos: position in @out
657 * @len: number of bytes to splice
658 * @flags: splice modifier flags
661 * Will either move or copy pages (determined by @flags options) from
662 * the given pipe inode to the given file.
663 * This one is ->write_iter-based.
667 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
668 loff_t *ppos, size_t len, unsigned int flags)
670 struct splice_desc sd = {
676 int nbufs = pipe->max_usage;
677 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
681 if (unlikely(!array))
686 splice_from_pipe_begin(&sd);
687 while (sd.total_len) {
688 struct iov_iter from;
689 unsigned int head, tail, mask;
693 ret = splice_from_pipe_next(pipe, &sd);
697 if (unlikely(nbufs < pipe->max_usage)) {
699 nbufs = pipe->max_usage;
700 array = kcalloc(nbufs, sizeof(struct bio_vec),
710 mask = pipe->ring_size - 1;
712 /* build the vector */
714 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
715 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
716 size_t this_len = buf->len;
718 /* zero-length bvecs are not supported, skip them */
721 this_len = min(this_len, left);
723 ret = pipe_buf_confirm(pipe, buf);
730 bvec_set_page(&array[n], buf->page, this_len,
736 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
737 ret = vfs_iter_write(out, &from, &sd.pos, 0);
741 sd.num_spliced += ret;
745 /* dismiss the fully eaten buffers, adjust the partial one */
748 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
749 if (ret >= buf->len) {
752 pipe_buf_release(pipe, buf);
756 sd.need_wakeup = true;
766 splice_from_pipe_end(pipe, &sd);
771 ret = sd.num_spliced;
776 EXPORT_SYMBOL(iter_file_splice_write);
780 * splice_to_socket - splice data from a pipe to a socket
781 * @pipe: pipe to splice from
782 * @out: socket to write to
783 * @ppos: position in @out
784 * @len: number of bytes to splice
785 * @flags: splice modifier flags
788 * Will send @len bytes from the pipe to a network socket. No data copying
792 ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
793 loff_t *ppos, size_t len, unsigned int flags)
795 struct socket *sock = sock_from_file(out);
796 struct bio_vec bvec[16];
797 struct msghdr msg = {};
800 bool need_wakeup = false;
805 unsigned int head, tail, mask, bc = 0;
809 * Check for signal early to make process killable when there
810 * are always buffers available
813 if (signal_pending(current))
816 while (pipe_empty(pipe->head, pipe->tail)) {
825 if (flags & SPLICE_F_NONBLOCK)
829 if (signal_pending(current))
833 wakeup_pipe_writers(pipe);
837 pipe_wait_readable(pipe);
842 mask = pipe->ring_size - 1;
844 while (!pipe_empty(head, tail)) {
845 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
853 seg = min_t(size_t, remain, buf->len);
855 ret = pipe_buf_confirm(pipe, buf);
862 bvec_set_page(&bvec[bc++], buf->page, seg, buf->offset);
864 if (remain == 0 || bc >= ARRAY_SIZE(bvec))
872 msg.msg_flags = MSG_SPLICE_PAGES;
873 if (flags & SPLICE_F_MORE)
874 msg.msg_flags |= MSG_MORE;
875 if (remain && pipe_occupancy(pipe->head, tail) > 0)
876 msg.msg_flags |= MSG_MORE;
877 if (out->f_flags & O_NONBLOCK)
878 msg.msg_flags |= MSG_DONTWAIT;
880 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
882 ret = sock_sendmsg(sock, &msg);
890 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
891 size_t seg = min_t(size_t, ret, buf->len);
898 pipe_buf_release(pipe, buf);
903 if (tail != pipe->tail) {
913 wakeup_pipe_writers(pipe);
914 return spliced ?: ret;
918 static int warn_unsupported(struct file *file, const char *op)
920 pr_debug_ratelimited(
921 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
922 op, file, current->pid, current->comm);
927 * Attempt to initiate a splice from pipe to file.
929 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
930 loff_t *ppos, size_t len, unsigned int flags)
932 if (unlikely(!out->f_op->splice_write))
933 return warn_unsupported(out, "write");
934 return out->f_op->splice_write(pipe, out, ppos, len, flags);
938 * Indicate to the caller that there was a premature EOF when reading from the
939 * source and the caller didn't indicate they would be sending more data after
942 static void do_splice_eof(struct splice_desc *sd)
949 * vfs_splice_read - Read data from a file and splice it into a pipe
950 * @in: File to splice from
951 * @ppos: Input file offset
952 * @pipe: Pipe to splice to
953 * @len: Number of bytes to splice
954 * @flags: Splice modifier flags (SPLICE_F_*)
956 * Splice the requested amount of data from the input file to the pipe. This
957 * is synchronous as the caller must hold the pipe lock across the entire
960 * If successful, it returns the amount of data spliced, 0 if it hit the EOF or
961 * a hole and a negative error code otherwise.
963 long vfs_splice_read(struct file *in, loff_t *ppos,
964 struct pipe_inode_info *pipe, size_t len,
967 unsigned int p_space;
970 if (unlikely(!(in->f_mode & FMODE_READ)))
975 /* Don't try to read more the pipe has space for. */
976 p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
977 len = min_t(size_t, len, p_space << PAGE_SHIFT);
979 ret = rw_verify_area(READ, in, ppos, len);
980 if (unlikely(ret < 0))
983 if (unlikely(len > MAX_RW_COUNT))
986 if (unlikely(!in->f_op->splice_read))
987 return warn_unsupported(in, "read");
989 * O_DIRECT and DAX don't deal with the pagecache, so we allocate a
990 * buffer, copy into it and splice that into the pipe.
992 if ((in->f_flags & O_DIRECT) || IS_DAX(in->f_mapping->host))
993 return copy_splice_read(in, ppos, pipe, len, flags);
994 return in->f_op->splice_read(in, ppos, pipe, len, flags);
996 EXPORT_SYMBOL_GPL(vfs_splice_read);
999 * splice_direct_to_actor - splices data directly between two non-pipes
1000 * @in: file to splice from
1001 * @sd: actor information on where to splice to
1002 * @actor: handles the data splicing
1005 * This is a special case helper to splice directly between two
1006 * points, without requiring an explicit pipe. Internally an allocated
1007 * pipe is cached in the process, and reused during the lifetime of
1011 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1012 splice_direct_actor *actor)
1014 struct pipe_inode_info *pipe;
1020 * We require the input to be seekable, as we don't want to randomly
1021 * drop data for eg socket -> socket splicing. Use the piped splicing
1024 if (unlikely(!(in->f_mode & FMODE_LSEEK)))
1028 * neither in nor out is a pipe, setup an internal pipe attached to
1029 * 'out' and transfer the wanted data from 'in' to 'out' through that
1031 pipe = current->splice_pipe;
1032 if (unlikely(!pipe)) {
1033 pipe = alloc_pipe_info();
1038 * We don't have an immediate reader, but we'll read the stuff
1039 * out of the pipe right after the splice_to_pipe(). So set
1040 * PIPE_READERS appropriately.
1044 current->splice_pipe = pipe;
1051 len = sd->total_len;
1053 /* Don't block on output, we have to drain the direct pipe. */
1055 sd->flags &= ~SPLICE_F_NONBLOCK;
1058 * We signal MORE until we've read sufficient data to fulfill the
1059 * request and we keep signalling it if the caller set it.
1061 more = sd->flags & SPLICE_F_MORE;
1062 sd->flags |= SPLICE_F_MORE;
1064 WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
1068 loff_t pos = sd->pos, prev_pos = pos;
1070 ret = vfs_splice_read(in, &pos, pipe, len, flags);
1071 if (unlikely(ret <= 0))
1075 sd->total_len = read_len;
1078 * If we now have sufficient data to fulfill the request then
1079 * we clear SPLICE_F_MORE if it was not set initially.
1081 if (read_len >= len && !more)
1082 sd->flags &= ~SPLICE_F_MORE;
1085 * NOTE: nonblocking mode only applies to the input. We
1086 * must not do the output in nonblocking mode as then we
1087 * could get stuck data in the internal pipe:
1089 ret = actor(pipe, sd);
1090 if (unlikely(ret <= 0)) {
1099 if (ret < read_len) {
1100 sd->pos = prev_pos + ret;
1106 pipe->tail = pipe->head = 0;
1112 * If the user did *not* set SPLICE_F_MORE *and* we didn't hit that
1113 * "use all of len" case that cleared SPLICE_F_MORE, *and* we did a
1114 * "->splice_in()" that returned EOF (ie zero) *and* we have sent at
1115 * least 1 byte *then* we will also do the ->splice_eof() call.
1117 if (ret == 0 && !more && len > 0 && bytes)
1121 * If we did an incomplete transfer we must release
1122 * the pipe buffers in question:
1124 for (i = 0; i < pipe->ring_size; i++) {
1125 struct pipe_buffer *buf = &pipe->bufs[i];
1128 pipe_buf_release(pipe, buf);
1136 EXPORT_SYMBOL(splice_direct_to_actor);
1138 static int direct_splice_actor(struct pipe_inode_info *pipe,
1139 struct splice_desc *sd)
1141 struct file *file = sd->u.file;
1143 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1147 static void direct_file_splice_eof(struct splice_desc *sd)
1149 struct file *file = sd->u.file;
1151 if (file->f_op->splice_eof)
1152 file->f_op->splice_eof(file);
1156 * do_splice_direct - splices data directly between two files
1157 * @in: file to splice from
1158 * @ppos: input file offset
1159 * @out: file to splice to
1160 * @opos: output file offset
1161 * @len: number of bytes to splice
1162 * @flags: splice modifier flags
1165 * For use by do_sendfile(). splice can easily emulate sendfile, but
1166 * doing it in the application would incur an extra system call
1167 * (splice in + splice out, as compared to just sendfile()). So this helper
1168 * can splice directly through a process-private pipe.
1171 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1172 loff_t *opos, size_t len, unsigned int flags)
1174 struct splice_desc sd = {
1180 .splice_eof = direct_file_splice_eof,
1185 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1188 if (unlikely(out->f_flags & O_APPEND))
1191 ret = rw_verify_area(WRITE, out, opos, len);
1192 if (unlikely(ret < 0))
1195 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1201 EXPORT_SYMBOL(do_splice_direct);
1203 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1206 if (unlikely(!pipe->readers)) {
1207 send_sig(SIGPIPE, current, 0);
1210 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1212 if (flags & SPLICE_F_NONBLOCK)
1214 if (signal_pending(current))
1215 return -ERESTARTSYS;
1216 pipe_wait_writable(pipe);
1220 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1221 struct pipe_inode_info *opipe,
1222 size_t len, unsigned int flags);
1224 long splice_file_to_pipe(struct file *in,
1225 struct pipe_inode_info *opipe,
1227 size_t len, unsigned int flags)
1232 ret = wait_for_space(opipe, flags);
1234 ret = vfs_splice_read(in, offset, opipe, len, flags);
1237 wakeup_pipe_readers(opipe);
1242 * Determine where to splice to/from.
1244 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1245 loff_t *off_out, size_t len, unsigned int flags)
1247 struct pipe_inode_info *ipipe;
1248 struct pipe_inode_info *opipe;
1252 if (unlikely(!(in->f_mode & FMODE_READ) ||
1253 !(out->f_mode & FMODE_WRITE)))
1256 ipipe = get_pipe_info(in, true);
1257 opipe = get_pipe_info(out, true);
1259 if (ipipe && opipe) {
1260 if (off_in || off_out)
1263 /* Splicing to self would be fun, but... */
1267 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1268 flags |= SPLICE_F_NONBLOCK;
1270 ret = splice_pipe_to_pipe(ipipe, opipe, len, flags);
1275 if (!(out->f_mode & FMODE_PWRITE))
1279 offset = out->f_pos;
1282 if (unlikely(out->f_flags & O_APPEND))
1285 ret = rw_verify_area(WRITE, out, &offset, len);
1286 if (unlikely(ret < 0))
1289 if (in->f_flags & O_NONBLOCK)
1290 flags |= SPLICE_F_NONBLOCK;
1292 file_start_write(out);
1293 ret = do_splice_from(ipipe, out, &offset, len, flags);
1294 file_end_write(out);
1297 out->f_pos = offset;
1304 if (!(in->f_mode & FMODE_PREAD))
1311 if (out->f_flags & O_NONBLOCK)
1312 flags |= SPLICE_F_NONBLOCK;
1314 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1326 * Generate modify out before access in:
1327 * do_splice_from() may've already sent modify out,
1328 * and this ensures the events get merged.
1330 fsnotify_modify(out);
1331 fsnotify_access(in);
1337 static long __do_splice(struct file *in, loff_t __user *off_in,
1338 struct file *out, loff_t __user *off_out,
1339 size_t len, unsigned int flags)
1341 struct pipe_inode_info *ipipe;
1342 struct pipe_inode_info *opipe;
1343 loff_t offset, *__off_in = NULL, *__off_out = NULL;
1346 ipipe = get_pipe_info(in, true);
1347 opipe = get_pipe_info(out, true);
1352 pipe_clear_nowait(in);
1357 pipe_clear_nowait(out);
1361 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1363 __off_out = &offset;
1366 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1371 ret = do_splice(in, __off_in, out, __off_out, len, flags);
1375 if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1377 if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1383 static int iter_to_pipe(struct iov_iter *from,
1384 struct pipe_inode_info *pipe,
1387 struct pipe_buffer buf = {
1388 .ops = &user_page_pipe_buf_ops,
1394 while (iov_iter_count(from)) {
1395 struct page *pages[16];
1400 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1406 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1407 for (i = 0; i < n; i++) {
1408 int size = min_t(int, left, PAGE_SIZE - start);
1410 buf.page = pages[i];
1413 ret = add_to_pipe(pipe, &buf);
1414 if (unlikely(ret < 0)) {
1415 iov_iter_revert(from, left);
1416 // this one got dropped by add_to_pipe()
1427 return total ? total : ret;
1430 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1431 struct splice_desc *sd)
1433 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1434 return n == sd->len ? n : -EFAULT;
1438 * For lack of a better implementation, implement vmsplice() to userspace
1439 * as a simple copy of the pipes pages to the user iov.
1441 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1444 struct pipe_inode_info *pipe = get_pipe_info(file, true);
1445 struct splice_desc sd = {
1446 .total_len = iov_iter_count(iter),
1455 pipe_clear_nowait(file);
1459 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1464 fsnotify_access(file);
1470 * vmsplice splices a user address range into a pipe. It can be thought of
1471 * as splice-from-memory, where the regular splice is splice-from-file (or
1472 * to file). In both cases the output is a pipe, naturally.
1474 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1477 struct pipe_inode_info *pipe;
1479 unsigned buf_flag = 0;
1481 if (flags & SPLICE_F_GIFT)
1482 buf_flag = PIPE_BUF_FLAG_GIFT;
1484 pipe = get_pipe_info(file, true);
1488 pipe_clear_nowait(file);
1491 ret = wait_for_space(pipe, flags);
1493 ret = iter_to_pipe(iter, pipe, buf_flag);
1496 wakeup_pipe_readers(pipe);
1497 fsnotify_modify(file);
1502 static int vmsplice_type(struct fd f, int *type)
1506 if (f.file->f_mode & FMODE_WRITE) {
1507 *type = ITER_SOURCE;
1508 } else if (f.file->f_mode & FMODE_READ) {
1518 * Note that vmsplice only really supports true splicing _from_ user memory
1519 * to a pipe, not the other way around. Splicing from user memory is a simple
1520 * operation that can be supported without any funky alignment restrictions
1521 * or nasty vm tricks. We simply map in the user memory and fill them into
1522 * a pipe. The reverse isn't quite as easy, though. There are two possible
1523 * solutions for that:
1525 * - memcpy() the data internally, at which point we might as well just
1526 * do a regular read() on the buffer anyway.
1527 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1528 * has restriction limitations on both ends of the pipe).
1530 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1533 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1534 unsigned long, nr_segs, unsigned int, flags)
1536 struct iovec iovstack[UIO_FASTIOV];
1537 struct iovec *iov = iovstack;
1538 struct iov_iter iter;
1543 if (unlikely(flags & ~SPLICE_F_ALL))
1547 error = vmsplice_type(f, &type);
1551 error = import_iovec(type, uiov, nr_segs,
1552 ARRAY_SIZE(iovstack), &iov, &iter);
1556 if (!iov_iter_count(&iter))
1558 else if (type == ITER_SOURCE)
1559 error = vmsplice_to_pipe(f.file, &iter, flags);
1561 error = vmsplice_to_user(f.file, &iter, flags);
1569 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1570 int, fd_out, loff_t __user *, off_out,
1571 size_t, len, unsigned int, flags)
1579 if (unlikely(flags & ~SPLICE_F_ALL))
1585 out = fdget(fd_out);
1587 error = __do_splice(in.file, off_in, out.file, off_out,
1597 * Make sure there's data to read. Wait for input if we can, otherwise
1598 * return an appropriate error.
1600 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1605 * Check the pipe occupancy without the inode lock first. This function
1606 * is speculative anyways, so missing one is ok.
1608 if (!pipe_empty(pipe->head, pipe->tail))
1614 while (pipe_empty(pipe->head, pipe->tail)) {
1615 if (signal_pending(current)) {
1621 if (flags & SPLICE_F_NONBLOCK) {
1625 pipe_wait_readable(pipe);
1633 * Make sure there's writeable room. Wait for room if we can, otherwise
1634 * return an appropriate error.
1636 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1641 * Check pipe occupancy without the inode lock first. This function
1642 * is speculative anyways, so missing one is ok.
1644 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1650 while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1651 if (!pipe->readers) {
1652 send_sig(SIGPIPE, current, 0);
1656 if (flags & SPLICE_F_NONBLOCK) {
1660 if (signal_pending(current)) {
1664 pipe_wait_writable(pipe);
1672 * Splice contents of ipipe to opipe.
1674 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1675 struct pipe_inode_info *opipe,
1676 size_t len, unsigned int flags)
1678 struct pipe_buffer *ibuf, *obuf;
1679 unsigned int i_head, o_head;
1680 unsigned int i_tail, o_tail;
1681 unsigned int i_mask, o_mask;
1683 bool input_wakeup = false;
1687 ret = ipipe_prep(ipipe, flags);
1691 ret = opipe_prep(opipe, flags);
1696 * Potential ABBA deadlock, work around it by ordering lock
1697 * grabbing by pipe info address. Otherwise two different processes
1698 * could deadlock (one doing tee from A -> B, the other from B -> A).
1700 pipe_double_lock(ipipe, opipe);
1702 i_tail = ipipe->tail;
1703 i_mask = ipipe->ring_size - 1;
1704 o_head = opipe->head;
1705 o_mask = opipe->ring_size - 1;
1710 if (!opipe->readers) {
1711 send_sig(SIGPIPE, current, 0);
1717 i_head = ipipe->head;
1718 o_tail = opipe->tail;
1720 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1724 * Cannot make any progress, because either the input
1725 * pipe is empty or the output pipe is full.
1727 if (pipe_empty(i_head, i_tail) ||
1728 pipe_full(o_head, o_tail, opipe->max_usage)) {
1729 /* Already processed some buffers, break */
1733 if (flags & SPLICE_F_NONBLOCK) {
1739 * We raced with another reader/writer and haven't
1740 * managed to process any buffers. A zero return
1741 * value means EOF, so retry instead.
1748 ibuf = &ipipe->bufs[i_tail & i_mask];
1749 obuf = &opipe->bufs[o_head & o_mask];
1751 if (len >= ibuf->len) {
1753 * Simply move the whole buffer from ipipe to opipe
1758 ipipe->tail = i_tail;
1759 input_wakeup = true;
1762 opipe->head = o_head;
1765 * Get a reference to this pipe buffer,
1766 * so we can copy the contents over.
1768 if (!pipe_buf_get(ipipe, ibuf)) {
1776 * Don't inherit the gift and merge flags, we need to
1777 * prevent multiple steals of this page.
1779 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1780 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1783 ibuf->offset += len;
1787 opipe->head = o_head;
1797 * If we put data in the output pipe, wakeup any potential readers.
1800 wakeup_pipe_readers(opipe);
1803 wakeup_pipe_writers(ipipe);
1809 * Link contents of ipipe to opipe.
1811 static int link_pipe(struct pipe_inode_info *ipipe,
1812 struct pipe_inode_info *opipe,
1813 size_t len, unsigned int flags)
1815 struct pipe_buffer *ibuf, *obuf;
1816 unsigned int i_head, o_head;
1817 unsigned int i_tail, o_tail;
1818 unsigned int i_mask, o_mask;
1822 * Potential ABBA deadlock, work around it by ordering lock
1823 * grabbing by pipe info address. Otherwise two different processes
1824 * could deadlock (one doing tee from A -> B, the other from B -> A).
1826 pipe_double_lock(ipipe, opipe);
1828 i_tail = ipipe->tail;
1829 i_mask = ipipe->ring_size - 1;
1830 o_head = opipe->head;
1831 o_mask = opipe->ring_size - 1;
1834 if (!opipe->readers) {
1835 send_sig(SIGPIPE, current, 0);
1841 i_head = ipipe->head;
1842 o_tail = opipe->tail;
1845 * If we have iterated all input buffers or run out of
1846 * output room, break.
1848 if (pipe_empty(i_head, i_tail) ||
1849 pipe_full(o_head, o_tail, opipe->max_usage))
1852 ibuf = &ipipe->bufs[i_tail & i_mask];
1853 obuf = &opipe->bufs[o_head & o_mask];
1856 * Get a reference to this pipe buffer,
1857 * so we can copy the contents over.
1859 if (!pipe_buf_get(ipipe, ibuf)) {
1868 * Don't inherit the gift and merge flag, we need to prevent
1869 * multiple steals of this page.
1871 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1872 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1874 if (obuf->len > len)
1880 opipe->head = o_head;
1888 * If we put data in the output pipe, wakeup any potential readers.
1891 wakeup_pipe_readers(opipe);
1897 * This is a tee(1) implementation that works on pipes. It doesn't copy
1898 * any data, it simply references the 'in' pages on the 'out' pipe.
1899 * The 'flags' used are the SPLICE_F_* variants, currently the only
1900 * applicable one is SPLICE_F_NONBLOCK.
1902 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1904 struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1905 struct pipe_inode_info *opipe = get_pipe_info(out, true);
1908 if (unlikely(!(in->f_mode & FMODE_READ) ||
1909 !(out->f_mode & FMODE_WRITE)))
1913 * Duplicate the contents of ipipe to opipe without actually
1916 if (ipipe && opipe && ipipe != opipe) {
1917 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1918 flags |= SPLICE_F_NONBLOCK;
1921 * Keep going, unless we encounter an error. The ipipe/opipe
1922 * ordering doesn't really matter.
1924 ret = ipipe_prep(ipipe, flags);
1926 ret = opipe_prep(opipe, flags);
1928 ret = link_pipe(ipipe, opipe, len, flags);
1933 fsnotify_access(in);
1934 fsnotify_modify(out);
1940 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1945 if (unlikely(flags & ~SPLICE_F_ALL))
1956 error = do_tee(in.file, out.file, len, flags);