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>
38 * Attempt to steal a page from a pipe buffer. This should perhaps go into
39 * a vm helper function, it's already simplified quite a bit by the
40 * addition of remove_mapping(). If success is returned, the caller may
41 * attempt to reuse this page for another destination.
43 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
44 struct pipe_buffer *buf)
46 struct page *page = buf->page;
47 struct address_space *mapping;
51 mapping = page_mapping(page);
53 WARN_ON(!PageUptodate(page));
56 * At least for ext2 with nobh option, we need to wait on
57 * writeback completing on this page, since we'll remove it
58 * from the pagecache. Otherwise truncate wont wait on the
59 * page, allowing the disk blocks to be reused by someone else
60 * before we actually wrote our data to them. fs corruption
63 wait_on_page_writeback(page);
65 if (page_has_private(page) &&
66 !try_to_release_page(page, GFP_KERNEL))
70 * If we succeeded in removing the mapping, set LRU flag
73 if (remove_mapping(mapping, page)) {
74 buf->flags |= PIPE_BUF_FLAG_LRU;
80 * Raced with truncate or failed to remove page from current
81 * address space, unlock and return failure.
88 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
89 struct pipe_buffer *buf)
92 buf->flags &= ~PIPE_BUF_FLAG_LRU;
96 * Check whether the contents of buf is OK to access. Since the content
97 * is a page cache page, IO may be in flight.
99 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
100 struct pipe_buffer *buf)
102 struct page *page = buf->page;
105 if (!PageUptodate(page)) {
109 * Page got truncated/unhashed. This will cause a 0-byte
110 * splice, if this is the first page.
112 if (!page->mapping) {
118 * Uh oh, read-error from disk.
120 if (!PageUptodate(page)) {
126 * Page is ok afterall, we are done.
137 const struct pipe_buf_operations page_cache_pipe_buf_ops = {
139 .confirm = page_cache_pipe_buf_confirm,
140 .release = page_cache_pipe_buf_release,
141 .steal = page_cache_pipe_buf_steal,
142 .get = generic_pipe_buf_get,
145 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
146 struct pipe_buffer *buf)
148 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
151 buf->flags |= PIPE_BUF_FLAG_LRU;
152 return generic_pipe_buf_steal(pipe, buf);
155 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
157 .confirm = generic_pipe_buf_confirm,
158 .release = page_cache_pipe_buf_release,
159 .steal = user_page_pipe_buf_steal,
160 .get = generic_pipe_buf_get,
163 static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
166 if (waitqueue_active(&pipe->wait))
167 wake_up_interruptible(&pipe->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 int ret = 0, page_nr = 0;
191 if (unlikely(!pipe->readers)) {
192 send_sig(SIGPIPE, current, 0);
197 while (pipe->nrbufs < pipe->buffers) {
198 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
199 struct pipe_buffer *buf = pipe->bufs + newbuf;
201 buf->page = spd->pages[page_nr];
202 buf->offset = spd->partial[page_nr].offset;
203 buf->len = spd->partial[page_nr].len;
204 buf->private = spd->partial[page_nr].private;
211 if (!--spd->nr_pages)
219 while (page_nr < spd_pages)
220 spd->spd_release(spd, page_nr++);
224 EXPORT_SYMBOL_GPL(splice_to_pipe);
226 ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct pipe_buffer *buf)
230 if (unlikely(!pipe->readers)) {
231 send_sig(SIGPIPE, current, 0);
233 } else if (pipe->nrbufs == pipe->buffers) {
236 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
237 pipe->bufs[newbuf] = *buf;
241 pipe_buf_release(pipe, buf);
244 EXPORT_SYMBOL(add_to_pipe);
246 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
248 put_page(spd->pages[i]);
252 * Check if we need to grow the arrays holding pages and partial page
255 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
257 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
259 spd->nr_pages_max = buffers;
260 if (buffers <= PIPE_DEF_BUFFERS)
263 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
264 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
266 if (spd->pages && spd->partial)
274 void splice_shrink_spd(struct splice_pipe_desc *spd)
276 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
284 * generic_file_splice_read - splice data from file to a pipe
285 * @in: file to splice from
286 * @ppos: position in @in
287 * @pipe: pipe to splice to
288 * @len: number of bytes to splice
289 * @flags: splice modifier flags
292 * Will read pages from given file and fill them into a pipe. Can be
293 * used as long as it has more or less sane ->read_iter().
296 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
297 struct pipe_inode_info *pipe, size_t len,
305 isize = i_size_read(in->f_mapping->host);
306 if (unlikely(*ppos >= isize))
309 iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len);
311 init_sync_kiocb(&kiocb, in);
312 kiocb.ki_pos = *ppos;
313 ret = in->f_op->read_iter(&kiocb, &to);
315 *ppos = kiocb.ki_pos;
317 } else if (ret < 0) {
318 if (WARN_ON(to.idx != idx || to.iov_offset)) {
320 * a bogus ->read_iter() has copied something and still
321 * returned an error instead of a short read.
325 iov_iter_advance(&to, 0); /* to free what was emitted */
328 * callers of ->splice_read() expect -EAGAIN on
329 * "can't put anything in there", rather than -EFAULT.
337 EXPORT_SYMBOL(generic_file_splice_read);
339 const struct pipe_buf_operations default_pipe_buf_ops = {
341 .confirm = generic_pipe_buf_confirm,
342 .release = generic_pipe_buf_release,
343 .steal = generic_pipe_buf_steal,
344 .get = generic_pipe_buf_get,
347 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
348 struct pipe_buffer *buf)
353 /* Pipe buffer operations for a socket and similar. */
354 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
356 .confirm = generic_pipe_buf_confirm,
357 .release = generic_pipe_buf_release,
358 .steal = generic_pipe_buf_nosteal,
359 .get = generic_pipe_buf_get,
361 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
363 static ssize_t kernel_readv(struct file *file, const struct kvec *vec,
364 unsigned long vlen, loff_t offset)
372 /* The cast to a user pointer is valid due to the set_fs() */
373 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos, 0);
379 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
387 /* The cast to a user pointer is valid due to the set_fs() */
388 res = vfs_write(file, (__force const char __user *)buf, count, &pos);
393 EXPORT_SYMBOL(kernel_write);
395 static ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
396 struct pipe_inode_info *pipe, size_t len,
399 struct kvec *vec, __vec[PIPE_DEF_BUFFERS];
402 unsigned int nr_pages;
403 size_t offset, dummy, copied = 0;
407 if (pipe->nrbufs == pipe->buffers)
411 * Try to keep page boundaries matching to source pagecache ones -
412 * it probably won't be much help, but...
414 offset = *ppos & ~PAGE_MASK;
416 iov_iter_pipe(&to, ITER_PIPE | READ, pipe, len + offset);
418 res = iov_iter_get_pages_alloc(&to, &pages, len + offset, &dummy);
422 nr_pages = res / PAGE_SIZE;
425 if (nr_pages > PIPE_DEF_BUFFERS) {
426 vec = kmalloc(nr_pages * sizeof(struct kvec), GFP_KERNEL);
427 if (unlikely(!vec)) {
433 pipe->bufs[to.idx].offset = offset;
434 pipe->bufs[to.idx].len -= offset;
436 for (i = 0; i < nr_pages; i++) {
437 size_t this_len = min_t(size_t, len, PAGE_SIZE - offset);
438 vec[i].iov_base = page_address(pages[i]) + offset;
439 vec[i].iov_len = this_len;
444 res = kernel_readv(in, vec, nr_pages, *ppos);
453 for (i = 0; i < nr_pages; i++)
456 iov_iter_advance(&to, copied); /* truncates and discards */
461 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
462 * using sendpage(). Return the number of bytes sent.
464 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
465 struct pipe_buffer *buf, struct splice_desc *sd)
467 struct file *file = sd->u.file;
468 loff_t pos = sd->pos;
471 if (!likely(file->f_op->sendpage))
474 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
476 if (sd->len < sd->total_len && pipe->nrbufs > 1)
477 more |= MSG_SENDPAGE_NOTLAST;
479 return file->f_op->sendpage(file, buf->page, buf->offset,
480 sd->len, &pos, more);
483 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
486 if (waitqueue_active(&pipe->wait))
487 wake_up_interruptible(&pipe->wait);
488 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
492 * splice_from_pipe_feed - feed available data from a pipe to a file
493 * @pipe: pipe to splice from
494 * @sd: information to @actor
495 * @actor: handler that splices the data
498 * This function loops over the pipe and calls @actor to do the
499 * actual moving of a single struct pipe_buffer to the desired
500 * destination. It returns when there's no more buffers left in
501 * the pipe or if the requested number of bytes (@sd->total_len)
502 * have been copied. It returns a positive number (one) if the
503 * pipe needs to be filled with more data, zero if the required
504 * number of bytes have been copied and -errno on error.
506 * This, together with splice_from_pipe_{begin,end,next}, may be
507 * used to implement the functionality of __splice_from_pipe() when
508 * locking is required around copying the pipe buffers to the
511 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
516 while (pipe->nrbufs) {
517 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
520 if (sd->len > sd->total_len)
521 sd->len = sd->total_len;
523 ret = buf->ops->confirm(pipe, buf);
530 ret = actor(pipe, buf, sd);
537 sd->num_spliced += ret;
540 sd->total_len -= ret;
543 pipe_buf_release(pipe, buf);
544 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
547 sd->need_wakeup = true;
558 * splice_from_pipe_next - wait for some data to splice from
559 * @pipe: pipe to splice from
560 * @sd: information about the splice operation
563 * This function will wait for some data and return a positive
564 * value (one) if pipe buffers are available. It will return zero
565 * or -errno if no more data needs to be spliced.
567 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
570 * Check for signal early to make process killable when there are
571 * always buffers available
573 if (signal_pending(current))
576 while (!pipe->nrbufs) {
580 if (!pipe->waiting_writers && sd->num_spliced)
583 if (sd->flags & SPLICE_F_NONBLOCK)
586 if (signal_pending(current))
589 if (sd->need_wakeup) {
590 wakeup_pipe_writers(pipe);
591 sd->need_wakeup = false;
601 * splice_from_pipe_begin - start splicing from pipe
602 * @sd: information about the splice operation
605 * This function should be called before a loop containing
606 * splice_from_pipe_next() and splice_from_pipe_feed() to
607 * initialize the necessary fields of @sd.
609 static void splice_from_pipe_begin(struct splice_desc *sd)
612 sd->need_wakeup = false;
616 * splice_from_pipe_end - finish splicing from pipe
617 * @pipe: pipe to splice from
618 * @sd: information about the splice operation
621 * This function will wake up pipe writers if necessary. It should
622 * be called after a loop containing splice_from_pipe_next() and
623 * splice_from_pipe_feed().
625 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
628 wakeup_pipe_writers(pipe);
632 * __splice_from_pipe - splice data from a pipe to given actor
633 * @pipe: pipe to splice from
634 * @sd: information to @actor
635 * @actor: handler that splices the data
638 * This function does little more than loop over the pipe and call
639 * @actor to do the actual moving of a single struct pipe_buffer to
640 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
644 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
649 splice_from_pipe_begin(sd);
652 ret = splice_from_pipe_next(pipe, sd);
654 ret = splice_from_pipe_feed(pipe, sd, actor);
656 splice_from_pipe_end(pipe, sd);
658 return sd->num_spliced ? sd->num_spliced : ret;
660 EXPORT_SYMBOL(__splice_from_pipe);
663 * splice_from_pipe - splice data from a pipe to a file
664 * @pipe: pipe to splice from
665 * @out: file to splice to
666 * @ppos: position in @out
667 * @len: how many bytes to splice
668 * @flags: splice modifier flags
669 * @actor: handler that splices the data
672 * See __splice_from_pipe. This function locks the pipe inode,
673 * otherwise it's identical to __splice_from_pipe().
676 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
677 loff_t *ppos, size_t len, unsigned int flags,
681 struct splice_desc sd = {
689 ret = __splice_from_pipe(pipe, &sd, actor);
696 * iter_file_splice_write - splice data from a pipe to a file
698 * @out: file to write to
699 * @ppos: position in @out
700 * @len: number of bytes to splice
701 * @flags: splice modifier flags
704 * Will either move or copy pages (determined by @flags options) from
705 * the given pipe inode to the given file.
706 * This one is ->write_iter-based.
710 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
711 loff_t *ppos, size_t len, unsigned int flags)
713 struct splice_desc sd = {
719 int nbufs = pipe->buffers;
720 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
724 if (unlikely(!array))
729 splice_from_pipe_begin(&sd);
730 while (sd.total_len) {
731 struct iov_iter from;
735 ret = splice_from_pipe_next(pipe, &sd);
739 if (unlikely(nbufs < pipe->buffers)) {
741 nbufs = pipe->buffers;
742 array = kcalloc(nbufs, sizeof(struct bio_vec),
750 /* build the vector */
752 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
753 struct pipe_buffer *buf = pipe->bufs + idx;
754 size_t this_len = buf->len;
759 if (idx == pipe->buffers - 1)
762 ret = buf->ops->confirm(pipe, buf);
769 array[n].bv_page = buf->page;
770 array[n].bv_len = this_len;
771 array[n].bv_offset = buf->offset;
775 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
776 sd.total_len - left);
777 ret = vfs_iter_write(out, &from, &sd.pos);
781 sd.num_spliced += ret;
785 /* dismiss the fully eaten buffers, adjust the partial one */
787 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
788 if (ret >= buf->len) {
791 pipe_buf_release(pipe, buf);
792 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
795 sd.need_wakeup = true;
805 splice_from_pipe_end(pipe, &sd);
810 ret = sd.num_spliced;
815 EXPORT_SYMBOL(iter_file_splice_write);
817 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
818 struct splice_desc *sd)
822 loff_t tmp = sd->pos;
824 data = kmap(buf->page);
825 ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
831 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
832 struct file *out, loff_t *ppos,
833 size_t len, unsigned int flags)
837 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
845 * generic_splice_sendpage - splice data from a pipe to a socket
846 * @pipe: pipe to splice from
847 * @out: socket to write to
848 * @ppos: position in @out
849 * @len: number of bytes to splice
850 * @flags: splice modifier flags
853 * Will send @len bytes from the pipe to a network socket. No data copying
857 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
858 loff_t *ppos, size_t len, unsigned int flags)
860 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
863 EXPORT_SYMBOL(generic_splice_sendpage);
866 * Attempt to initiate a splice from pipe to file.
868 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
869 loff_t *ppos, size_t len, unsigned int flags)
871 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
872 loff_t *, size_t, unsigned int);
874 if (out->f_op->splice_write)
875 splice_write = out->f_op->splice_write;
877 splice_write = default_file_splice_write;
879 return splice_write(pipe, out, ppos, len, flags);
883 * Attempt to initiate a splice from a file to a pipe.
885 static long do_splice_to(struct file *in, loff_t *ppos,
886 struct pipe_inode_info *pipe, size_t len,
889 ssize_t (*splice_read)(struct file *, loff_t *,
890 struct pipe_inode_info *, size_t, unsigned int);
893 if (unlikely(!(in->f_mode & FMODE_READ)))
896 ret = rw_verify_area(READ, in, ppos, len);
897 if (unlikely(ret < 0))
900 if (unlikely(len > MAX_RW_COUNT))
903 if (in->f_op->splice_read)
904 splice_read = in->f_op->splice_read;
906 splice_read = default_file_splice_read;
908 return splice_read(in, ppos, pipe, len, flags);
912 * splice_direct_to_actor - splices data directly between two non-pipes
913 * @in: file to splice from
914 * @sd: actor information on where to splice to
915 * @actor: handles the data splicing
918 * This is a special case helper to splice directly between two
919 * points, without requiring an explicit pipe. Internally an allocated
920 * pipe is cached in the process, and reused during the lifetime of
924 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
925 splice_direct_actor *actor)
927 struct pipe_inode_info *pipe;
934 * We require the input being a regular file, as we don't want to
935 * randomly drop data for eg socket -> socket splicing. Use the
936 * piped splicing for that!
938 i_mode = file_inode(in)->i_mode;
939 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
943 * neither in nor out is a pipe, setup an internal pipe attached to
944 * 'out' and transfer the wanted data from 'in' to 'out' through that
946 pipe = current->splice_pipe;
947 if (unlikely(!pipe)) {
948 pipe = alloc_pipe_info();
953 * We don't have an immediate reader, but we'll read the stuff
954 * out of the pipe right after the splice_to_pipe(). So set
955 * PIPE_READERS appropriately.
959 current->splice_pipe = pipe;
971 * Don't block on output, we have to drain the direct pipe.
973 sd->flags &= ~SPLICE_F_NONBLOCK;
974 more = sd->flags & SPLICE_F_MORE;
978 loff_t pos = sd->pos, prev_pos = pos;
980 ret = do_splice_to(in, &pos, pipe, len, flags);
981 if (unlikely(ret <= 0))
985 sd->total_len = read_len;
988 * If more data is pending, set SPLICE_F_MORE
989 * If this is the last data and SPLICE_F_MORE was not set
990 * initially, clears it.
993 sd->flags |= SPLICE_F_MORE;
995 sd->flags &= ~SPLICE_F_MORE;
997 * NOTE: nonblocking mode only applies to the input. We
998 * must not do the output in nonblocking mode as then we
999 * could get stuck data in the internal pipe:
1001 ret = actor(pipe, sd);
1002 if (unlikely(ret <= 0)) {
1011 if (ret < read_len) {
1012 sd->pos = prev_pos + ret;
1018 pipe->nrbufs = pipe->curbuf = 0;
1024 * If we did an incomplete transfer we must release
1025 * the pipe buffers in question:
1027 for (i = 0; i < pipe->buffers; i++) {
1028 struct pipe_buffer *buf = pipe->bufs + i;
1031 pipe_buf_release(pipe, buf);
1039 EXPORT_SYMBOL(splice_direct_to_actor);
1041 static int direct_splice_actor(struct pipe_inode_info *pipe,
1042 struct splice_desc *sd)
1044 struct file *file = sd->u.file;
1046 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1051 * do_splice_direct - splices data directly between two files
1052 * @in: file to splice from
1053 * @ppos: input file offset
1054 * @out: file to splice to
1055 * @opos: output file offset
1056 * @len: number of bytes to splice
1057 * @flags: splice modifier flags
1060 * For use by do_sendfile(). splice can easily emulate sendfile, but
1061 * doing it in the application would incur an extra system call
1062 * (splice in + splice out, as compared to just sendfile()). So this helper
1063 * can splice directly through a process-private pipe.
1066 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1067 loff_t *opos, size_t len, unsigned int flags)
1069 struct splice_desc sd = {
1079 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1082 if (unlikely(out->f_flags & O_APPEND))
1085 ret = rw_verify_area(WRITE, out, opos, len);
1086 if (unlikely(ret < 0))
1089 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1095 EXPORT_SYMBOL(do_splice_direct);
1097 static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
1099 while (pipe->nrbufs == pipe->buffers) {
1100 if (flags & SPLICE_F_NONBLOCK)
1102 if (signal_pending(current))
1103 return -ERESTARTSYS;
1104 pipe->waiting_writers++;
1106 pipe->waiting_writers--;
1111 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1112 struct pipe_inode_info *opipe,
1113 size_t len, unsigned int flags);
1116 * Determine where to splice to/from.
1118 static long do_splice(struct file *in, loff_t __user *off_in,
1119 struct file *out, loff_t __user *off_out,
1120 size_t len, unsigned int flags)
1122 struct pipe_inode_info *ipipe;
1123 struct pipe_inode_info *opipe;
1127 ipipe = get_pipe_info(in);
1128 opipe = get_pipe_info(out);
1130 if (ipipe && opipe) {
1131 if (off_in || off_out)
1134 if (!(in->f_mode & FMODE_READ))
1137 if (!(out->f_mode & FMODE_WRITE))
1140 /* Splicing to self would be fun, but... */
1144 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1151 if (!(out->f_mode & FMODE_PWRITE))
1153 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1156 offset = out->f_pos;
1159 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1162 if (unlikely(out->f_flags & O_APPEND))
1165 ret = rw_verify_area(WRITE, out, &offset, len);
1166 if (unlikely(ret < 0))
1169 file_start_write(out);
1170 ret = do_splice_from(ipipe, out, &offset, len, flags);
1171 file_end_write(out);
1174 out->f_pos = offset;
1175 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1185 if (!(in->f_mode & FMODE_PREAD))
1187 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1194 ret = wait_for_space(opipe, flags);
1196 ret = do_splice_to(in, &offset, opipe, len, flags);
1199 wakeup_pipe_readers(opipe);
1202 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1211 static int iter_to_pipe(struct iov_iter *from,
1212 struct pipe_inode_info *pipe,
1215 struct pipe_buffer buf = {
1216 .ops = &user_page_pipe_buf_ops,
1221 bool failed = false;
1223 while (iov_iter_count(from) && !failed) {
1224 struct page *pages[16];
1229 copied = iov_iter_get_pages(from, pages, ~0UL, 16, &start);
1235 for (n = 0; copied; n++, start = 0) {
1236 int size = min_t(int, copied, PAGE_SIZE - start);
1238 buf.page = pages[n];
1241 ret = add_to_pipe(pipe, &buf);
1242 if (unlikely(ret < 0)) {
1245 iov_iter_advance(from, ret);
1254 return total ? total : ret;
1257 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1258 struct splice_desc *sd)
1260 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1261 return n == sd->len ? n : -EFAULT;
1265 * For lack of a better implementation, implement vmsplice() to userspace
1266 * as a simple copy of the pipes pages to the user iov.
1268 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1269 unsigned long nr_segs, unsigned int flags)
1271 struct pipe_inode_info *pipe;
1272 struct splice_desc sd;
1274 struct iovec iovstack[UIO_FASTIOV];
1275 struct iovec *iov = iovstack;
1276 struct iov_iter iter;
1278 pipe = get_pipe_info(file);
1282 ret = import_iovec(READ, uiov, nr_segs,
1283 ARRAY_SIZE(iovstack), &iov, &iter);
1287 sd.total_len = iov_iter_count(&iter);
1295 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1304 * vmsplice splices a user address range into a pipe. It can be thought of
1305 * as splice-from-memory, where the regular splice is splice-from-file (or
1306 * to file). In both cases the output is a pipe, naturally.
1308 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
1309 unsigned long nr_segs, unsigned int flags)
1311 struct pipe_inode_info *pipe;
1312 struct iovec iovstack[UIO_FASTIOV];
1313 struct iovec *iov = iovstack;
1314 struct iov_iter from;
1316 unsigned buf_flag = 0;
1318 if (flags & SPLICE_F_GIFT)
1319 buf_flag = PIPE_BUF_FLAG_GIFT;
1321 pipe = get_pipe_info(file);
1325 ret = import_iovec(WRITE, uiov, nr_segs,
1326 ARRAY_SIZE(iovstack), &iov, &from);
1331 ret = wait_for_space(pipe, flags);
1333 ret = iter_to_pipe(&from, pipe, buf_flag);
1336 wakeup_pipe_readers(pipe);
1342 * Note that vmsplice only really supports true splicing _from_ user memory
1343 * to a pipe, not the other way around. Splicing from user memory is a simple
1344 * operation that can be supported without any funky alignment restrictions
1345 * or nasty vm tricks. We simply map in the user memory and fill them into
1346 * a pipe. The reverse isn't quite as easy, though. There are two possible
1347 * solutions for that:
1349 * - memcpy() the data internally, at which point we might as well just
1350 * do a regular read() on the buffer anyway.
1351 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1352 * has restriction limitations on both ends of the pipe).
1354 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1357 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1358 unsigned long, nr_segs, unsigned int, flags)
1363 if (unlikely(nr_segs > UIO_MAXIOV))
1365 else if (unlikely(!nr_segs))
1371 if (f.file->f_mode & FMODE_WRITE)
1372 error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1373 else if (f.file->f_mode & FMODE_READ)
1374 error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1382 #ifdef CONFIG_COMPAT
1383 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1384 unsigned int, nr_segs, unsigned int, flags)
1387 struct iovec __user *iov;
1388 if (nr_segs > UIO_MAXIOV)
1390 iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1391 for (i = 0; i < nr_segs; i++) {
1392 struct compat_iovec v;
1393 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1394 get_user(v.iov_len, &iov32[i].iov_len) ||
1395 put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1396 put_user(v.iov_len, &iov[i].iov_len))
1399 return sys_vmsplice(fd, iov, nr_segs, flags);
1403 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1404 int, fd_out, loff_t __user *, off_out,
1405 size_t, len, unsigned int, flags)
1416 if (in.file->f_mode & FMODE_READ) {
1417 out = fdget(fd_out);
1419 if (out.file->f_mode & FMODE_WRITE)
1420 error = do_splice(in.file, off_in,
1432 * Make sure there's data to read. Wait for input if we can, otherwise
1433 * return an appropriate error.
1435 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1440 * Check ->nrbufs without the inode lock first. This function
1441 * is speculative anyways, so missing one is ok.
1449 while (!pipe->nrbufs) {
1450 if (signal_pending(current)) {
1456 if (!pipe->waiting_writers) {
1457 if (flags & SPLICE_F_NONBLOCK) {
1470 * Make sure there's writeable room. Wait for room if we can, otherwise
1471 * return an appropriate error.
1473 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1478 * Check ->nrbufs without the inode lock first. This function
1479 * is speculative anyways, so missing one is ok.
1481 if (pipe->nrbufs < pipe->buffers)
1487 while (pipe->nrbufs >= pipe->buffers) {
1488 if (!pipe->readers) {
1489 send_sig(SIGPIPE, current, 0);
1493 if (flags & SPLICE_F_NONBLOCK) {
1497 if (signal_pending(current)) {
1501 pipe->waiting_writers++;
1503 pipe->waiting_writers--;
1511 * Splice contents of ipipe to opipe.
1513 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1514 struct pipe_inode_info *opipe,
1515 size_t len, unsigned int flags)
1517 struct pipe_buffer *ibuf, *obuf;
1519 bool input_wakeup = false;
1523 ret = ipipe_prep(ipipe, flags);
1527 ret = opipe_prep(opipe, flags);
1532 * Potential ABBA deadlock, work around it by ordering lock
1533 * grabbing by pipe info address. Otherwise two different processes
1534 * could deadlock (one doing tee from A -> B, the other from B -> A).
1536 pipe_double_lock(ipipe, opipe);
1539 if (!opipe->readers) {
1540 send_sig(SIGPIPE, current, 0);
1546 if (!ipipe->nrbufs && !ipipe->writers)
1550 * Cannot make any progress, because either the input
1551 * pipe is empty or the output pipe is full.
1553 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1554 /* Already processed some buffers, break */
1558 if (flags & SPLICE_F_NONBLOCK) {
1564 * We raced with another reader/writer and haven't
1565 * managed to process any buffers. A zero return
1566 * value means EOF, so retry instead.
1573 ibuf = ipipe->bufs + ipipe->curbuf;
1574 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1575 obuf = opipe->bufs + nbuf;
1577 if (len >= ibuf->len) {
1579 * Simply move the whole buffer from ipipe to opipe
1584 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1586 input_wakeup = true;
1589 * Get a reference to this pipe buffer,
1590 * so we can copy the contents over.
1592 pipe_buf_get(ipipe, ibuf);
1596 * Don't inherit the gift flag, we need to
1597 * prevent multiple steals of this page.
1599 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1603 ibuf->offset += obuf->len;
1604 ibuf->len -= obuf->len;
1614 * If we put data in the output pipe, wakeup any potential readers.
1617 wakeup_pipe_readers(opipe);
1620 wakeup_pipe_writers(ipipe);
1626 * Link contents of ipipe to opipe.
1628 static int link_pipe(struct pipe_inode_info *ipipe,
1629 struct pipe_inode_info *opipe,
1630 size_t len, unsigned int flags)
1632 struct pipe_buffer *ibuf, *obuf;
1633 int ret = 0, i = 0, nbuf;
1636 * Potential ABBA deadlock, work around it by ordering lock
1637 * grabbing by pipe info address. Otherwise two different processes
1638 * could deadlock (one doing tee from A -> B, the other from B -> A).
1640 pipe_double_lock(ipipe, opipe);
1643 if (!opipe->readers) {
1644 send_sig(SIGPIPE, current, 0);
1651 * If we have iterated all input buffers or ran out of
1652 * output room, break.
1654 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1657 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1658 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1661 * Get a reference to this pipe buffer,
1662 * so we can copy the contents over.
1664 pipe_buf_get(ipipe, ibuf);
1666 obuf = opipe->bufs + nbuf;
1670 * Don't inherit the gift flag, we need to
1671 * prevent multiple steals of this page.
1673 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1675 if (obuf->len > len)
1685 * return EAGAIN if we have the potential of some data in the
1686 * future, otherwise just return 0
1688 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1695 * If we put data in the output pipe, wakeup any potential readers.
1698 wakeup_pipe_readers(opipe);
1704 * This is a tee(1) implementation that works on pipes. It doesn't copy
1705 * any data, it simply references the 'in' pages on the 'out' pipe.
1706 * The 'flags' used are the SPLICE_F_* variants, currently the only
1707 * applicable one is SPLICE_F_NONBLOCK.
1709 static long do_tee(struct file *in, struct file *out, size_t len,
1712 struct pipe_inode_info *ipipe = get_pipe_info(in);
1713 struct pipe_inode_info *opipe = get_pipe_info(out);
1717 * Duplicate the contents of ipipe to opipe without actually
1720 if (ipipe && opipe && ipipe != opipe) {
1722 * Keep going, unless we encounter an error. The ipipe/opipe
1723 * ordering doesn't really matter.
1725 ret = ipipe_prep(ipipe, flags);
1727 ret = opipe_prep(opipe, flags);
1729 ret = link_pipe(ipipe, opipe, len, flags);
1736 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
1747 if (in.file->f_mode & FMODE_READ) {
1748 struct fd out = fdget(fdout);
1750 if (out.file->f_mode & FMODE_WRITE)
1751 error = do_tee(in.file, out.file,