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/security.h>
34 #include <linux/gfp.h>
35 #include <linux/socket.h>
36 #include <linux/sched/signal.h>
41 * Attempt to steal a page from a pipe buffer. This should perhaps go into
42 * a vm helper function, it's already simplified quite a bit by the
43 * addition of remove_mapping(). If success is returned, the caller may
44 * attempt to reuse this page for another destination.
46 static bool page_cache_pipe_buf_try_steal(struct pipe_inode_info *pipe,
47 struct pipe_buffer *buf)
49 struct folio *folio = page_folio(buf->page);
50 struct address_space *mapping;
54 mapping = folio_mapping(folio);
56 WARN_ON(!folio_test_uptodate(folio));
59 * At least for ext2 with nobh option, we need to wait on
60 * writeback completing on this folio, since we'll remove it
61 * from the pagecache. Otherwise truncate wont wait on the
62 * folio, allowing the disk blocks to be reused by someone else
63 * before we actually wrote our data to them. fs corruption
66 folio_wait_writeback(folio);
68 if (folio_has_private(folio) &&
69 !filemap_release_folio(folio, GFP_KERNEL))
73 * If we succeeded in removing the mapping, set LRU flag
76 if (remove_mapping(mapping, folio)) {
77 buf->flags |= PIPE_BUF_FLAG_LRU;
83 * Raced with truncate or failed to remove folio from current
84 * address space, unlock and return failure.
91 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
92 struct pipe_buffer *buf)
95 buf->flags &= ~PIPE_BUF_FLAG_LRU;
99 * Check whether the contents of buf is OK to access. Since the content
100 * is a page cache page, IO may be in flight.
102 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
103 struct pipe_buffer *buf)
105 struct page *page = buf->page;
108 if (!PageUptodate(page)) {
112 * Page got truncated/unhashed. This will cause a 0-byte
113 * splice, if this is the first page.
115 if (!page->mapping) {
121 * Uh oh, read-error from disk.
123 if (!PageUptodate(page)) {
129 * Page is ok afterall, we are done.
140 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
141 .confirm = page_cache_pipe_buf_confirm,
142 .release = page_cache_pipe_buf_release,
143 .try_steal = page_cache_pipe_buf_try_steal,
144 .get = generic_pipe_buf_get,
147 static bool user_page_pipe_buf_try_steal(struct pipe_inode_info *pipe,
148 struct pipe_buffer *buf)
150 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
153 buf->flags |= PIPE_BUF_FLAG_LRU;
154 return generic_pipe_buf_try_steal(pipe, buf);
157 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
158 .release = page_cache_pipe_buf_release,
159 .try_steal = user_page_pipe_buf_try_steal,
160 .get = generic_pipe_buf_get,
163 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
166 if (waitqueue_active(&pipe->rd_wait))
167 wake_up_interruptible(&pipe->rd_wait);
168 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
172 * splice_to_pipe - fill passed data into a pipe
173 * @pipe: pipe to fill
177 * @spd contains a map of pages and len/offset tuples, along with
178 * the struct pipe_buf_operations associated with these pages. This
179 * function will link that data to the pipe.
182 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
183 struct splice_pipe_desc *spd)
185 unsigned int spd_pages = spd->nr_pages;
186 unsigned int tail = pipe->tail;
187 unsigned int head = pipe->head;
188 unsigned int mask = pipe->ring_size - 1;
189 int ret = 0, page_nr = 0;
194 if (unlikely(!pipe->readers)) {
195 send_sig(SIGPIPE, current, 0);
200 while (!pipe_full(head, tail, pipe->max_usage)) {
201 struct pipe_buffer *buf = &pipe->bufs[head & mask];
203 buf->page = spd->pages[page_nr];
204 buf->offset = spd->partial[page_nr].offset;
205 buf->len = spd->partial[page_nr].len;
206 buf->private = spd->partial[page_nr].private;
215 if (!--spd->nr_pages)
223 while (page_nr < spd_pages)
224 spd->spd_release(spd, page_nr++);
228 EXPORT_SYMBOL_GPL(splice_to_pipe);
230 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
232 unsigned int head = pipe->head;
233 unsigned int tail = pipe->tail;
234 unsigned int mask = pipe->ring_size - 1;
237 if (unlikely(!pipe->readers)) {
238 send_sig(SIGPIPE, current, 0);
240 } else if (pipe_full(head, tail, pipe->max_usage)) {
243 pipe->bufs[head & mask] = *buf;
244 pipe->head = head + 1;
247 pipe_buf_release(pipe, buf);
250 EXPORT_SYMBOL(add_to_pipe);
253 * Check if we need to grow the arrays holding pages and partial page
256 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
258 unsigned int max_usage = READ_ONCE(pipe->max_usage);
260 spd->nr_pages_max = max_usage;
261 if (max_usage <= PIPE_DEF_BUFFERS)
264 spd->pages = kmalloc_array(max_usage, sizeof(struct page *), GFP_KERNEL);
265 spd->partial = kmalloc_array(max_usage, sizeof(struct partial_page),
268 if (spd->pages && spd->partial)
276 void splice_shrink_spd(struct splice_pipe_desc *spd)
278 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
286 * Splice data from an O_DIRECT file into pages and then add them to the output
289 ssize_t direct_splice_read(struct file *in, loff_t *ppos,
290 struct pipe_inode_info *pipe,
291 size_t len, unsigned int flags)
298 size_t used, npages, chunk, remain, reclaim;
301 /* Work out how much data we can actually add into the pipe */
302 used = pipe_occupancy(pipe->head, pipe->tail);
303 npages = max_t(ssize_t, pipe->max_usage - used, 0);
304 len = min_t(size_t, len, npages * PAGE_SIZE);
305 npages = DIV_ROUND_UP(len, PAGE_SIZE);
307 bv = kzalloc(array_size(npages, sizeof(bv[0])) +
308 array_size(npages, sizeof(struct page *)), GFP_KERNEL);
312 pages = (void *)(bv + npages);
313 npages = alloc_pages_bulk_array(GFP_USER, npages, pages);
319 remain = len = min_t(size_t, len, npages * PAGE_SIZE);
321 for (i = 0; i < npages; i++) {
322 chunk = min_t(size_t, PAGE_SIZE, remain);
323 bv[i].bv_page = pages[i];
325 bv[i].bv_len = chunk;
330 iov_iter_bvec(&to, ITER_DEST, bv, npages, len);
331 init_sync_kiocb(&kiocb, in);
332 kiocb.ki_pos = *ppos;
333 ret = call_read_iter(in, &kiocb, &to);
335 reclaim = npages * PAGE_SIZE;
340 *ppos = kiocb.ki_pos;
342 } else if (ret < 0) {
344 * callers of ->splice_read() expect -EAGAIN on
345 * "can't put anything in there", rather than -EFAULT.
351 /* Free any pages that didn't get touched at all. */
352 reclaim /= PAGE_SIZE;
355 release_pages(pages + npages, reclaim);
358 /* Push the remaining pages into the pipe. */
359 for (i = 0; i < npages; i++) {
360 struct pipe_buffer *buf = pipe_head_buf(pipe);
362 chunk = min_t(size_t, remain, PAGE_SIZE);
363 *buf = (struct pipe_buffer) {
364 .ops = &default_pipe_buf_ops,
365 .page = bv[i].bv_page,
376 EXPORT_SYMBOL(direct_splice_read);
379 * generic_file_splice_read - splice data from file to a pipe
380 * @in: file to splice from
381 * @ppos: position in @in
382 * @pipe: pipe to splice to
383 * @len: number of bytes to splice
384 * @flags: splice modifier flags
387 * Will read pages from given file and fill them into a pipe. Can be
388 * used as long as it has more or less sane ->read_iter().
391 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
392 struct pipe_inode_info *pipe, size_t len,
399 iov_iter_pipe(&to, ITER_DEST, pipe, len);
400 init_sync_kiocb(&kiocb, in);
401 kiocb.ki_pos = *ppos;
402 ret = call_read_iter(in, &kiocb, &to);
404 *ppos = kiocb.ki_pos;
406 } else if (ret < 0) {
407 /* free what was emitted */
408 pipe_discard_from(pipe, to.start_head);
410 * callers of ->splice_read() expect -EAGAIN on
411 * "can't put anything in there", rather than -EFAULT.
419 EXPORT_SYMBOL(generic_file_splice_read);
421 const struct pipe_buf_operations default_pipe_buf_ops = {
422 .release = generic_pipe_buf_release,
423 .try_steal = generic_pipe_buf_try_steal,
424 .get = generic_pipe_buf_get,
427 /* Pipe buffer operations for a socket and similar. */
428 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
429 .release = generic_pipe_buf_release,
430 .get = generic_pipe_buf_get,
432 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
435 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
436 * using sendpage(). Return the number of bytes sent.
438 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
439 struct pipe_buffer *buf, struct splice_desc *sd)
441 struct file *file = sd->u.file;
442 loff_t pos = sd->pos;
445 if (!likely(file->f_op->sendpage))
448 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
450 if (sd->len < sd->total_len &&
451 pipe_occupancy(pipe->head, pipe->tail) > 1)
452 more |= MSG_SENDPAGE_NOTLAST;
454 return file->f_op->sendpage(file, buf->page, buf->offset,
455 sd->len, &pos, more);
458 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
461 if (waitqueue_active(&pipe->wr_wait))
462 wake_up_interruptible(&pipe->wr_wait);
463 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
467 * splice_from_pipe_feed - feed available data from a pipe to a file
468 * @pipe: pipe to splice from
469 * @sd: information to @actor
470 * @actor: handler that splices the data
473 * This function loops over the pipe and calls @actor to do the
474 * actual moving of a single struct pipe_buffer to the desired
475 * destination. It returns when there's no more buffers left in
476 * the pipe or if the requested number of bytes (@sd->total_len)
477 * have been copied. It returns a positive number (one) if the
478 * pipe needs to be filled with more data, zero if the required
479 * number of bytes have been copied and -errno on error.
481 * This, together with splice_from_pipe_{begin,end,next}, may be
482 * used to implement the functionality of __splice_from_pipe() when
483 * locking is required around copying the pipe buffers to the
486 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
489 unsigned int head = pipe->head;
490 unsigned int tail = pipe->tail;
491 unsigned int mask = pipe->ring_size - 1;
494 while (!pipe_empty(head, tail)) {
495 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
498 if (sd->len > sd->total_len)
499 sd->len = sd->total_len;
501 ret = pipe_buf_confirm(pipe, buf);
508 ret = actor(pipe, buf, sd);
515 sd->num_spliced += ret;
518 sd->total_len -= ret;
521 pipe_buf_release(pipe, buf);
525 sd->need_wakeup = true;
535 /* We know we have a pipe buffer, but maybe it's empty? */
536 static inline bool eat_empty_buffer(struct pipe_inode_info *pipe)
538 unsigned int tail = pipe->tail;
539 unsigned int mask = pipe->ring_size - 1;
540 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
542 if (unlikely(!buf->len)) {
543 pipe_buf_release(pipe, buf);
552 * splice_from_pipe_next - wait for some data to splice from
553 * @pipe: pipe to splice from
554 * @sd: information about the splice operation
557 * This function will wait for some data and return a positive
558 * value (one) if pipe buffers are available. It will return zero
559 * or -errno if no more data needs to be spliced.
561 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
564 * Check for signal early to make process killable when there are
565 * always buffers available
567 if (signal_pending(current))
571 while (pipe_empty(pipe->head, pipe->tail)) {
578 if (sd->flags & SPLICE_F_NONBLOCK)
581 if (signal_pending(current))
584 if (sd->need_wakeup) {
585 wakeup_pipe_writers(pipe);
586 sd->need_wakeup = false;
589 pipe_wait_readable(pipe);
592 if (eat_empty_buffer(pipe))
599 * splice_from_pipe_begin - start splicing from pipe
600 * @sd: information about the splice operation
603 * This function should be called before a loop containing
604 * splice_from_pipe_next() and splice_from_pipe_feed() to
605 * initialize the necessary fields of @sd.
607 static void splice_from_pipe_begin(struct splice_desc *sd)
610 sd->need_wakeup = false;
614 * splice_from_pipe_end - finish splicing from pipe
615 * @pipe: pipe to splice from
616 * @sd: information about the splice operation
619 * This function will wake up pipe writers if necessary. It should
620 * be called after a loop containing splice_from_pipe_next() and
621 * splice_from_pipe_feed().
623 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
626 wakeup_pipe_writers(pipe);
630 * __splice_from_pipe - splice data from a pipe to given actor
631 * @pipe: pipe to splice from
632 * @sd: information to @actor
633 * @actor: handler that splices the data
636 * This function does little more than loop over the pipe and call
637 * @actor to do the actual moving of a single struct pipe_buffer to
638 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
642 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
647 splice_from_pipe_begin(sd);
650 ret = splice_from_pipe_next(pipe, sd);
652 ret = splice_from_pipe_feed(pipe, sd, actor);
654 splice_from_pipe_end(pipe, sd);
656 return sd->num_spliced ? sd->num_spliced : ret;
658 EXPORT_SYMBOL(__splice_from_pipe);
661 * splice_from_pipe - splice data from a pipe to a file
662 * @pipe: pipe to splice from
663 * @out: file to splice to
664 * @ppos: position in @out
665 * @len: how many bytes to splice
666 * @flags: splice modifier flags
667 * @actor: handler that splices the data
670 * See __splice_from_pipe. This function locks the pipe inode,
671 * otherwise it's identical to __splice_from_pipe().
674 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
675 loff_t *ppos, size_t len, unsigned int flags,
679 struct splice_desc sd = {
687 ret = __splice_from_pipe(pipe, &sd, actor);
694 * iter_file_splice_write - splice data from a pipe to a file
696 * @out: file to write to
697 * @ppos: position in @out
698 * @len: number of bytes to splice
699 * @flags: splice modifier flags
702 * Will either move or copy pages (determined by @flags options) from
703 * the given pipe inode to the given file.
704 * This one is ->write_iter-based.
708 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
709 loff_t *ppos, size_t len, unsigned int flags)
711 struct splice_desc sd = {
717 int nbufs = pipe->max_usage;
718 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
722 if (unlikely(!array))
727 splice_from_pipe_begin(&sd);
728 while (sd.total_len) {
729 struct iov_iter from;
730 unsigned int head, tail, mask;
734 ret = splice_from_pipe_next(pipe, &sd);
738 if (unlikely(nbufs < pipe->max_usage)) {
740 nbufs = pipe->max_usage;
741 array = kcalloc(nbufs, sizeof(struct bio_vec),
751 mask = pipe->ring_size - 1;
753 /* build the vector */
755 for (n = 0; !pipe_empty(head, tail) && left && n < nbufs; tail++) {
756 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
757 size_t this_len = buf->len;
759 /* zero-length bvecs are not supported, skip them */
762 this_len = min(this_len, left);
764 ret = pipe_buf_confirm(pipe, buf);
771 array[n].bv_page = buf->page;
772 array[n].bv_len = this_len;
773 array[n].bv_offset = buf->offset;
778 iov_iter_bvec(&from, ITER_SOURCE, array, n, sd.total_len - left);
779 ret = vfs_iter_write(out, &from, &sd.pos, 0);
783 sd.num_spliced += ret;
787 /* dismiss the fully eaten buffers, adjust the partial one */
790 struct pipe_buffer *buf = &pipe->bufs[tail & mask];
791 if (ret >= buf->len) {
794 pipe_buf_release(pipe, buf);
798 sd.need_wakeup = true;
808 splice_from_pipe_end(pipe, &sd);
813 ret = sd.num_spliced;
818 EXPORT_SYMBOL(iter_file_splice_write);
821 * generic_splice_sendpage - splice data from a pipe to a socket
822 * @pipe: pipe to splice from
823 * @out: socket to write to
824 * @ppos: position in @out
825 * @len: number of bytes to splice
826 * @flags: splice modifier flags
829 * Will send @len bytes from the pipe to a network socket. No data copying
833 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
834 loff_t *ppos, size_t len, unsigned int flags)
836 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
839 EXPORT_SYMBOL(generic_splice_sendpage);
841 static int warn_unsupported(struct file *file, const char *op)
843 pr_debug_ratelimited(
844 "splice %s not supported for file %pD4 (pid: %d comm: %.20s)\n",
845 op, file, current->pid, current->comm);
850 * Attempt to initiate a splice from pipe to file.
852 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
853 loff_t *ppos, size_t len, unsigned int flags)
855 if (unlikely(!out->f_op->splice_write))
856 return warn_unsupported(out, "write");
857 return out->f_op->splice_write(pipe, out, ppos, len, flags);
861 * Attempt to initiate a splice from a file to a pipe.
863 static long do_splice_to(struct file *in, loff_t *ppos,
864 struct pipe_inode_info *pipe, size_t len,
867 unsigned int p_space;
870 if (unlikely(!(in->f_mode & FMODE_READ)))
873 /* Don't try to read more the pipe has space for. */
874 p_space = pipe->max_usage - pipe_occupancy(pipe->head, pipe->tail);
875 len = min_t(size_t, len, p_space << PAGE_SHIFT);
877 ret = rw_verify_area(READ, in, ppos, len);
878 if (unlikely(ret < 0))
881 if (unlikely(len > MAX_RW_COUNT))
884 if (unlikely(!in->f_op->splice_read))
885 return warn_unsupported(in, "read");
886 return in->f_op->splice_read(in, ppos, pipe, len, flags);
890 * splice_direct_to_actor - splices data directly between two non-pipes
891 * @in: file to splice from
892 * @sd: actor information on where to splice to
893 * @actor: handles the data splicing
896 * This is a special case helper to splice directly between two
897 * points, without requiring an explicit pipe. Internally an allocated
898 * pipe is cached in the process, and reused during the lifetime of
902 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
903 splice_direct_actor *actor)
905 struct pipe_inode_info *pipe;
911 * We require the input to be seekable, as we don't want to randomly
912 * drop data for eg socket -> socket splicing. Use the piped splicing
915 if (unlikely(!(in->f_mode & FMODE_LSEEK)))
919 * neither in nor out is a pipe, setup an internal pipe attached to
920 * 'out' and transfer the wanted data from 'in' to 'out' through that
922 pipe = current->splice_pipe;
923 if (unlikely(!pipe)) {
924 pipe = alloc_pipe_info();
929 * We don't have an immediate reader, but we'll read the stuff
930 * out of the pipe right after the splice_to_pipe(). So set
931 * PIPE_READERS appropriately.
935 current->splice_pipe = pipe;
947 * Don't block on output, we have to drain the direct pipe.
949 sd->flags &= ~SPLICE_F_NONBLOCK;
950 more = sd->flags & SPLICE_F_MORE;
952 WARN_ON_ONCE(!pipe_empty(pipe->head, pipe->tail));
956 loff_t pos = sd->pos, prev_pos = pos;
958 ret = do_splice_to(in, &pos, pipe, len, flags);
959 if (unlikely(ret <= 0))
963 sd->total_len = read_len;
966 * If more data is pending, set SPLICE_F_MORE
967 * If this is the last data and SPLICE_F_MORE was not set
968 * initially, clears it.
971 sd->flags |= SPLICE_F_MORE;
973 sd->flags &= ~SPLICE_F_MORE;
975 * NOTE: nonblocking mode only applies to the input. We
976 * must not do the output in nonblocking mode as then we
977 * could get stuck data in the internal pipe:
979 ret = actor(pipe, sd);
980 if (unlikely(ret <= 0)) {
989 if (ret < read_len) {
990 sd->pos = prev_pos + ret;
996 pipe->tail = pipe->head = 0;
1002 * If we did an incomplete transfer we must release
1003 * the pipe buffers in question:
1005 for (i = 0; i < pipe->ring_size; i++) {
1006 struct pipe_buffer *buf = &pipe->bufs[i];
1009 pipe_buf_release(pipe, buf);
1017 EXPORT_SYMBOL(splice_direct_to_actor);
1019 static int direct_splice_actor(struct pipe_inode_info *pipe,
1020 struct splice_desc *sd)
1022 struct file *file = sd->u.file;
1024 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1029 * do_splice_direct - splices data directly between two files
1030 * @in: file to splice from
1031 * @ppos: input file offset
1032 * @out: file to splice to
1033 * @opos: output file offset
1034 * @len: number of bytes to splice
1035 * @flags: splice modifier flags
1038 * For use by do_sendfile(). splice can easily emulate sendfile, but
1039 * doing it in the application would incur an extra system call
1040 * (splice in + splice out, as compared to just sendfile()). So this helper
1041 * can splice directly through a process-private pipe.
1044 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1045 loff_t *opos, size_t len, unsigned int flags)
1047 struct splice_desc sd = {
1057 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1060 if (unlikely(out->f_flags & O_APPEND))
1063 ret = rw_verify_area(WRITE, out, opos, len);
1064 if (unlikely(ret < 0))
1067 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1073 EXPORT_SYMBOL(do_splice_direct);
1075 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1078 if (unlikely(!pipe->readers)) {
1079 send_sig(SIGPIPE, current, 0);
1082 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1084 if (flags & SPLICE_F_NONBLOCK)
1086 if (signal_pending(current))
1087 return -ERESTARTSYS;
1088 pipe_wait_writable(pipe);
1092 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1093 struct pipe_inode_info *opipe,
1094 size_t len, unsigned int flags);
1096 long splice_file_to_pipe(struct file *in,
1097 struct pipe_inode_info *opipe,
1099 size_t len, unsigned int flags)
1104 ret = wait_for_space(opipe, flags);
1106 ret = do_splice_to(in, offset, opipe, len, flags);
1109 wakeup_pipe_readers(opipe);
1114 * Determine where to splice to/from.
1116 long do_splice(struct file *in, loff_t *off_in, struct file *out,
1117 loff_t *off_out, size_t len, unsigned int flags)
1119 struct pipe_inode_info *ipipe;
1120 struct pipe_inode_info *opipe;
1124 if (unlikely(!(in->f_mode & FMODE_READ) ||
1125 !(out->f_mode & FMODE_WRITE)))
1128 ipipe = get_pipe_info(in, true);
1129 opipe = get_pipe_info(out, true);
1131 if (ipipe && opipe) {
1132 if (off_in || off_out)
1135 /* Splicing to self would be fun, but... */
1139 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1140 flags |= SPLICE_F_NONBLOCK;
1142 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1149 if (!(out->f_mode & FMODE_PWRITE))
1153 offset = out->f_pos;
1156 if (unlikely(out->f_flags & O_APPEND))
1159 ret = rw_verify_area(WRITE, out, &offset, len);
1160 if (unlikely(ret < 0))
1163 if (in->f_flags & O_NONBLOCK)
1164 flags |= SPLICE_F_NONBLOCK;
1166 file_start_write(out);
1167 ret = do_splice_from(ipipe, out, &offset, len, flags);
1168 file_end_write(out);
1171 out->f_pos = offset;
1182 if (!(in->f_mode & FMODE_PREAD))
1189 if (out->f_flags & O_NONBLOCK)
1190 flags |= SPLICE_F_NONBLOCK;
1192 ret = splice_file_to_pipe(in, opipe, &offset, len, flags);
1204 static long __do_splice(struct file *in, loff_t __user *off_in,
1205 struct file *out, loff_t __user *off_out,
1206 size_t len, unsigned int flags)
1208 struct pipe_inode_info *ipipe;
1209 struct pipe_inode_info *opipe;
1210 loff_t offset, *__off_in = NULL, *__off_out = NULL;
1213 ipipe = get_pipe_info(in, true);
1214 opipe = get_pipe_info(out, true);
1216 if (ipipe && off_in)
1218 if (opipe && off_out)
1222 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1224 __off_out = &offset;
1227 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1232 ret = do_splice(in, __off_in, out, __off_out, len, flags);
1236 if (__off_out && copy_to_user(off_out, __off_out, sizeof(loff_t)))
1238 if (__off_in && copy_to_user(off_in, __off_in, sizeof(loff_t)))
1244 static int iter_to_pipe(struct iov_iter *from,
1245 struct pipe_inode_info *pipe,
1248 struct pipe_buffer buf = {
1249 .ops = &user_page_pipe_buf_ops,
1255 while (iov_iter_count(from)) {
1256 struct page *pages[16];
1261 left = iov_iter_get_pages2(from, pages, ~0UL, 16, &start);
1267 n = DIV_ROUND_UP(left + start, PAGE_SIZE);
1268 for (i = 0; i < n; i++) {
1269 int size = min_t(int, left, PAGE_SIZE - start);
1271 buf.page = pages[i];
1274 ret = add_to_pipe(pipe, &buf);
1275 if (unlikely(ret < 0)) {
1276 iov_iter_revert(from, left);
1277 // this one got dropped by add_to_pipe()
1288 return total ? total : ret;
1291 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1292 struct splice_desc *sd)
1294 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1295 return n == sd->len ? n : -EFAULT;
1299 * For lack of a better implementation, implement vmsplice() to userspace
1300 * as a simple copy of the pipes pages to the user iov.
1302 static long vmsplice_to_user(struct file *file, struct iov_iter *iter,
1305 struct pipe_inode_info *pipe = get_pipe_info(file, true);
1306 struct splice_desc sd = {
1307 .total_len = iov_iter_count(iter),
1318 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1326 * vmsplice splices a user address range into a pipe. It can be thought of
1327 * as splice-from-memory, where the regular splice is splice-from-file (or
1328 * to file). In both cases the output is a pipe, naturally.
1330 static long vmsplice_to_pipe(struct file *file, struct iov_iter *iter,
1333 struct pipe_inode_info *pipe;
1335 unsigned buf_flag = 0;
1337 if (flags & SPLICE_F_GIFT)
1338 buf_flag = PIPE_BUF_FLAG_GIFT;
1340 pipe = get_pipe_info(file, true);
1345 ret = wait_for_space(pipe, flags);
1347 ret = iter_to_pipe(iter, pipe, buf_flag);
1350 wakeup_pipe_readers(pipe);
1354 static int vmsplice_type(struct fd f, int *type)
1358 if (f.file->f_mode & FMODE_WRITE) {
1359 *type = ITER_SOURCE;
1360 } else if (f.file->f_mode & FMODE_READ) {
1370 * Note that vmsplice only really supports true splicing _from_ user memory
1371 * to a pipe, not the other way around. Splicing from user memory is a simple
1372 * operation that can be supported without any funky alignment restrictions
1373 * or nasty vm tricks. We simply map in the user memory and fill them into
1374 * a pipe. The reverse isn't quite as easy, though. There are two possible
1375 * solutions for that:
1377 * - memcpy() the data internally, at which point we might as well just
1378 * do a regular read() on the buffer anyway.
1379 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1380 * has restriction limitations on both ends of the pipe).
1382 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1385 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, uiov,
1386 unsigned long, nr_segs, unsigned int, flags)
1388 struct iovec iovstack[UIO_FASTIOV];
1389 struct iovec *iov = iovstack;
1390 struct iov_iter iter;
1395 if (unlikely(flags & ~SPLICE_F_ALL))
1399 error = vmsplice_type(f, &type);
1403 error = import_iovec(type, uiov, nr_segs,
1404 ARRAY_SIZE(iovstack), &iov, &iter);
1408 if (!iov_iter_count(&iter))
1410 else if (type == ITER_SOURCE)
1411 error = vmsplice_to_pipe(f.file, &iter, flags);
1413 error = vmsplice_to_user(f.file, &iter, flags);
1421 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1422 int, fd_out, loff_t __user *, off_out,
1423 size_t, len, unsigned int, flags)
1431 if (unlikely(flags & ~SPLICE_F_ALL))
1437 out = fdget(fd_out);
1439 error = __do_splice(in.file, off_in, out.file, off_out,
1449 * Make sure there's data to read. Wait for input if we can, otherwise
1450 * return an appropriate error.
1452 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1457 * Check the pipe occupancy without the inode lock first. This function
1458 * is speculative anyways, so missing one is ok.
1460 if (!pipe_empty(pipe->head, pipe->tail))
1466 while (pipe_empty(pipe->head, pipe->tail)) {
1467 if (signal_pending(current)) {
1473 if (flags & SPLICE_F_NONBLOCK) {
1477 pipe_wait_readable(pipe);
1485 * Make sure there's writeable room. Wait for room if we can, otherwise
1486 * return an appropriate error.
1488 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1493 * Check pipe occupancy without the inode lock first. This function
1494 * is speculative anyways, so missing one is ok.
1496 if (!pipe_full(pipe->head, pipe->tail, pipe->max_usage))
1502 while (pipe_full(pipe->head, pipe->tail, pipe->max_usage)) {
1503 if (!pipe->readers) {
1504 send_sig(SIGPIPE, current, 0);
1508 if (flags & SPLICE_F_NONBLOCK) {
1512 if (signal_pending(current)) {
1516 pipe_wait_writable(pipe);
1524 * Splice contents of ipipe to opipe.
1526 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1527 struct pipe_inode_info *opipe,
1528 size_t len, unsigned int flags)
1530 struct pipe_buffer *ibuf, *obuf;
1531 unsigned int i_head, o_head;
1532 unsigned int i_tail, o_tail;
1533 unsigned int i_mask, o_mask;
1535 bool input_wakeup = false;
1539 ret = ipipe_prep(ipipe, flags);
1543 ret = opipe_prep(opipe, flags);
1548 * Potential ABBA deadlock, work around it by ordering lock
1549 * grabbing by pipe info address. Otherwise two different processes
1550 * could deadlock (one doing tee from A -> B, the other from B -> A).
1552 pipe_double_lock(ipipe, opipe);
1554 i_tail = ipipe->tail;
1555 i_mask = ipipe->ring_size - 1;
1556 o_head = opipe->head;
1557 o_mask = opipe->ring_size - 1;
1562 if (!opipe->readers) {
1563 send_sig(SIGPIPE, current, 0);
1569 i_head = ipipe->head;
1570 o_tail = opipe->tail;
1572 if (pipe_empty(i_head, i_tail) && !ipipe->writers)
1576 * Cannot make any progress, because either the input
1577 * pipe is empty or the output pipe is full.
1579 if (pipe_empty(i_head, i_tail) ||
1580 pipe_full(o_head, o_tail, opipe->max_usage)) {
1581 /* Already processed some buffers, break */
1585 if (flags & SPLICE_F_NONBLOCK) {
1591 * We raced with another reader/writer and haven't
1592 * managed to process any buffers. A zero return
1593 * value means EOF, so retry instead.
1600 ibuf = &ipipe->bufs[i_tail & i_mask];
1601 obuf = &opipe->bufs[o_head & o_mask];
1603 if (len >= ibuf->len) {
1605 * Simply move the whole buffer from ipipe to opipe
1610 ipipe->tail = i_tail;
1611 input_wakeup = true;
1614 opipe->head = o_head;
1617 * Get a reference to this pipe buffer,
1618 * so we can copy the contents over.
1620 if (!pipe_buf_get(ipipe, ibuf)) {
1628 * Don't inherit the gift and merge flags, we need to
1629 * prevent multiple steals of this page.
1631 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1632 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1635 ibuf->offset += len;
1639 opipe->head = o_head;
1649 * If we put data in the output pipe, wakeup any potential readers.
1652 wakeup_pipe_readers(opipe);
1655 wakeup_pipe_writers(ipipe);
1661 * Link contents of ipipe to opipe.
1663 static int link_pipe(struct pipe_inode_info *ipipe,
1664 struct pipe_inode_info *opipe,
1665 size_t len, unsigned int flags)
1667 struct pipe_buffer *ibuf, *obuf;
1668 unsigned int i_head, o_head;
1669 unsigned int i_tail, o_tail;
1670 unsigned int i_mask, o_mask;
1674 * Potential ABBA deadlock, work around it by ordering lock
1675 * grabbing by pipe info address. Otherwise two different processes
1676 * could deadlock (one doing tee from A -> B, the other from B -> A).
1678 pipe_double_lock(ipipe, opipe);
1680 i_tail = ipipe->tail;
1681 i_mask = ipipe->ring_size - 1;
1682 o_head = opipe->head;
1683 o_mask = opipe->ring_size - 1;
1686 if (!opipe->readers) {
1687 send_sig(SIGPIPE, current, 0);
1693 i_head = ipipe->head;
1694 o_tail = opipe->tail;
1697 * If we have iterated all input buffers or run out of
1698 * output room, break.
1700 if (pipe_empty(i_head, i_tail) ||
1701 pipe_full(o_head, o_tail, opipe->max_usage))
1704 ibuf = &ipipe->bufs[i_tail & i_mask];
1705 obuf = &opipe->bufs[o_head & o_mask];
1708 * Get a reference to this pipe buffer,
1709 * so we can copy the contents over.
1711 if (!pipe_buf_get(ipipe, ibuf)) {
1720 * Don't inherit the gift and merge flag, we need to prevent
1721 * multiple steals of this page.
1723 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1724 obuf->flags &= ~PIPE_BUF_FLAG_CAN_MERGE;
1726 if (obuf->len > len)
1732 opipe->head = o_head;
1740 * If we put data in the output pipe, wakeup any potential readers.
1743 wakeup_pipe_readers(opipe);
1749 * This is a tee(1) implementation that works on pipes. It doesn't copy
1750 * any data, it simply references the 'in' pages on the 'out' pipe.
1751 * The 'flags' used are the SPLICE_F_* variants, currently the only
1752 * applicable one is SPLICE_F_NONBLOCK.
1754 long do_tee(struct file *in, struct file *out, size_t len, unsigned int flags)
1756 struct pipe_inode_info *ipipe = get_pipe_info(in, true);
1757 struct pipe_inode_info *opipe = get_pipe_info(out, true);
1760 if (unlikely(!(in->f_mode & FMODE_READ) ||
1761 !(out->f_mode & FMODE_WRITE)))
1765 * Duplicate the contents of ipipe to opipe without actually
1768 if (ipipe && opipe && ipipe != opipe) {
1769 if ((in->f_flags | out->f_flags) & O_NONBLOCK)
1770 flags |= SPLICE_F_NONBLOCK;
1773 * Keep going, unless we encounter an error. The ipipe/opipe
1774 * ordering doesn't really matter.
1776 ret = ipipe_prep(ipipe, flags);
1778 ret = opipe_prep(opipe, flags);
1780 ret = link_pipe(ipipe, opipe, len, flags);
1787 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1792 if (unlikely(flags & ~SPLICE_F_ALL))
1803 error = do_tee(in.file, out.file, len, flags);