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)
91 page_cache_release(buf->page);
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, do_wakeup, page_nr;
198 if (!pipe->readers) {
199 send_sig(SIGPIPE, current, 0);
205 if (pipe->nrbufs < pipe->buffers) {
206 int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
207 struct pipe_buffer *buf = pipe->bufs + newbuf;
209 buf->page = spd->pages[page_nr];
210 buf->offset = spd->partial[page_nr].offset;
211 buf->len = spd->partial[page_nr].len;
212 buf->private = spd->partial[page_nr].private;
215 if (spd->flags & SPLICE_F_GIFT)
216 buf->flags |= PIPE_BUF_FLAG_GIFT;
225 if (!--spd->nr_pages)
227 if (pipe->nrbufs < pipe->buffers)
233 if (spd->flags & SPLICE_F_NONBLOCK) {
239 if (signal_pending(current)) {
247 if (waitqueue_active(&pipe->wait))
248 wake_up_interruptible_sync(&pipe->wait);
249 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
253 pipe->waiting_writers++;
255 pipe->waiting_writers--;
261 wakeup_pipe_readers(pipe);
263 while (page_nr < spd_pages)
264 spd->spd_release(spd, page_nr++);
268 EXPORT_SYMBOL_GPL(splice_to_pipe);
270 void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
272 page_cache_release(spd->pages[i]);
276 * Check if we need to grow the arrays holding pages and partial page
279 int splice_grow_spd(const struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
281 unsigned int buffers = ACCESS_ONCE(pipe->buffers);
283 spd->nr_pages_max = buffers;
284 if (buffers <= PIPE_DEF_BUFFERS)
287 spd->pages = kmalloc(buffers * sizeof(struct page *), GFP_KERNEL);
288 spd->partial = kmalloc(buffers * sizeof(struct partial_page), GFP_KERNEL);
290 if (spd->pages && spd->partial)
298 void splice_shrink_spd(struct splice_pipe_desc *spd)
300 if (spd->nr_pages_max <= PIPE_DEF_BUFFERS)
308 __generic_file_splice_read(struct file *in, loff_t *ppos,
309 struct pipe_inode_info *pipe, size_t len,
312 struct address_space *mapping = in->f_mapping;
313 unsigned int loff, nr_pages, req_pages;
314 struct page *pages[PIPE_DEF_BUFFERS];
315 struct partial_page partial[PIPE_DEF_BUFFERS];
317 pgoff_t index, end_index;
320 struct splice_pipe_desc spd = {
323 .nr_pages_max = PIPE_DEF_BUFFERS,
325 .ops = &page_cache_pipe_buf_ops,
326 .spd_release = spd_release_page,
329 if (splice_grow_spd(pipe, &spd))
332 index = *ppos >> PAGE_CACHE_SHIFT;
333 loff = *ppos & ~PAGE_CACHE_MASK;
334 req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
335 nr_pages = min(req_pages, spd.nr_pages_max);
338 * Lookup the (hopefully) full range of pages we need.
340 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
341 index += spd.nr_pages;
344 * If find_get_pages_contig() returned fewer pages than we needed,
345 * readahead/allocate the rest and fill in the holes.
347 if (spd.nr_pages < nr_pages)
348 page_cache_sync_readahead(mapping, &in->f_ra, in,
349 index, req_pages - spd.nr_pages);
352 while (spd.nr_pages < nr_pages) {
354 * Page could be there, find_get_pages_contig() breaks on
357 page = find_get_page(mapping, index);
360 * page didn't exist, allocate one.
362 page = page_cache_alloc_cold(mapping);
366 error = add_to_page_cache_lru(page, mapping, index,
367 mapping_gfp_constraint(mapping, GFP_KERNEL));
368 if (unlikely(error)) {
369 page_cache_release(page);
370 if (error == -EEXIST)
375 * add_to_page_cache() locks the page, unlock it
376 * to avoid convoluting the logic below even more.
381 spd.pages[spd.nr_pages++] = page;
386 * Now loop over the map and see if we need to start IO on any
387 * pages, fill in the partial map, etc.
389 index = *ppos >> PAGE_CACHE_SHIFT;
390 nr_pages = spd.nr_pages;
392 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
393 unsigned int this_len;
399 * this_len is the max we'll use from this page
401 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
402 page = spd.pages[page_nr];
404 if (PageReadahead(page))
405 page_cache_async_readahead(mapping, &in->f_ra, in,
406 page, index, req_pages - page_nr);
409 * If the page isn't uptodate, we may need to start io on it
411 if (!PageUptodate(page)) {
415 * Page was truncated, or invalidated by the
416 * filesystem. Redo the find/create, but this time the
417 * page is kept locked, so there's no chance of another
418 * race with truncate/invalidate.
420 if (!page->mapping) {
422 page = find_or_create_page(mapping, index,
423 mapping_gfp_mask(mapping));
429 page_cache_release(spd.pages[page_nr]);
430 spd.pages[page_nr] = page;
433 * page was already under io and is now done, great
435 if (PageUptodate(page)) {
441 * need to read in the page
443 error = mapping->a_ops->readpage(in, page);
444 if (unlikely(error)) {
446 * We really should re-lookup the page here,
447 * but it complicates things a lot. Instead
448 * lets just do what we already stored, and
449 * we'll get it the next time we are called.
451 if (error == AOP_TRUNCATED_PAGE)
459 * i_size must be checked after PageUptodate.
461 isize = i_size_read(mapping->host);
462 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
463 if (unlikely(!isize || index > end_index))
467 * if this is the last page, see if we need to shrink
468 * the length and stop
470 if (end_index == index) {
474 * max good bytes in this page
476 plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
481 * force quit after adding this page
483 this_len = min(this_len, plen - loff);
487 spd.partial[page_nr].offset = loff;
488 spd.partial[page_nr].len = this_len;
496 * Release any pages at the end, if we quit early. 'page_nr' is how far
497 * we got, 'nr_pages' is how many pages are in the map.
499 while (page_nr < nr_pages)
500 page_cache_release(spd.pages[page_nr++]);
501 in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
504 error = splice_to_pipe(pipe, &spd);
506 splice_shrink_spd(&spd);
511 * generic_file_splice_read - splice data from file to a pipe
512 * @in: file to splice from
513 * @ppos: position in @in
514 * @pipe: pipe to splice to
515 * @len: number of bytes to splice
516 * @flags: splice modifier flags
519 * Will read pages from given file and fill them into a pipe. Can be
520 * used as long as the address_space operations for the source implements
524 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
525 struct pipe_inode_info *pipe, size_t len,
531 if (IS_DAX(in->f_mapping->host))
532 return default_file_splice_read(in, ppos, pipe, len, flags);
534 isize = i_size_read(in->f_mapping->host);
535 if (unlikely(*ppos >= isize))
538 left = isize - *ppos;
539 if (unlikely(left < len))
542 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
550 EXPORT_SYMBOL(generic_file_splice_read);
552 static const struct pipe_buf_operations default_pipe_buf_ops = {
554 .confirm = generic_pipe_buf_confirm,
555 .release = generic_pipe_buf_release,
556 .steal = generic_pipe_buf_steal,
557 .get = generic_pipe_buf_get,
560 static int generic_pipe_buf_nosteal(struct pipe_inode_info *pipe,
561 struct pipe_buffer *buf)
566 /* Pipe buffer operations for a socket and similar. */
567 const struct pipe_buf_operations nosteal_pipe_buf_ops = {
569 .confirm = generic_pipe_buf_confirm,
570 .release = generic_pipe_buf_release,
571 .steal = generic_pipe_buf_nosteal,
572 .get = generic_pipe_buf_get,
574 EXPORT_SYMBOL(nosteal_pipe_buf_ops);
576 static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
577 unsigned long vlen, loff_t offset)
585 /* The cast to a user pointer is valid due to the set_fs() */
586 res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
592 ssize_t kernel_write(struct file *file, const char *buf, size_t count,
600 /* The cast to a user pointer is valid due to the set_fs() */
601 res = vfs_write(file, (__force const char __user *)buf, count, &pos);
606 EXPORT_SYMBOL(kernel_write);
608 ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
609 struct pipe_inode_info *pipe, size_t len,
612 unsigned int nr_pages;
613 unsigned int nr_freed;
615 struct page *pages[PIPE_DEF_BUFFERS];
616 struct partial_page partial[PIPE_DEF_BUFFERS];
617 struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
622 struct splice_pipe_desc spd = {
625 .nr_pages_max = PIPE_DEF_BUFFERS,
627 .ops = &default_pipe_buf_ops,
628 .spd_release = spd_release_page,
631 if (splice_grow_spd(pipe, &spd))
636 if (spd.nr_pages_max > PIPE_DEF_BUFFERS) {
637 vec = kmalloc(spd.nr_pages_max * sizeof(struct iovec), GFP_KERNEL);
642 offset = *ppos & ~PAGE_CACHE_MASK;
643 nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
645 for (i = 0; i < nr_pages && i < spd.nr_pages_max && len; i++) {
648 page = alloc_page(GFP_USER);
653 this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
654 vec[i].iov_base = (void __user *) page_address(page);
655 vec[i].iov_len = this_len;
662 res = kernel_readv(in, vec, spd.nr_pages, *ppos);
673 for (i = 0; i < spd.nr_pages; i++) {
674 this_len = min_t(size_t, vec[i].iov_len, res);
675 spd.partial[i].offset = 0;
676 spd.partial[i].len = this_len;
678 __free_page(spd.pages[i]);
684 spd.nr_pages -= nr_freed;
686 res = splice_to_pipe(pipe, &spd);
693 splice_shrink_spd(&spd);
697 for (i = 0; i < spd.nr_pages; i++)
698 __free_page(spd.pages[i]);
703 EXPORT_SYMBOL(default_file_splice_read);
706 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
707 * using sendpage(). Return the number of bytes sent.
709 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
710 struct pipe_buffer *buf, struct splice_desc *sd)
712 struct file *file = sd->u.file;
713 loff_t pos = sd->pos;
716 if (!likely(file->f_op->sendpage))
719 more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
721 if (sd->len < sd->total_len && pipe->nrbufs > 1)
722 more |= MSG_SENDPAGE_NOTLAST;
724 return file->f_op->sendpage(file, buf->page, buf->offset,
725 sd->len, &pos, more);
728 static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
731 if (waitqueue_active(&pipe->wait))
732 wake_up_interruptible(&pipe->wait);
733 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
737 * splice_from_pipe_feed - feed available data from a pipe to a file
738 * @pipe: pipe to splice from
739 * @sd: information to @actor
740 * @actor: handler that splices the data
743 * This function loops over the pipe and calls @actor to do the
744 * actual moving of a single struct pipe_buffer to the desired
745 * destination. It returns when there's no more buffers left in
746 * the pipe or if the requested number of bytes (@sd->total_len)
747 * have been copied. It returns a positive number (one) if the
748 * pipe needs to be filled with more data, zero if the required
749 * number of bytes have been copied and -errno on error.
751 * This, together with splice_from_pipe_{begin,end,next}, may be
752 * used to implement the functionality of __splice_from_pipe() when
753 * locking is required around copying the pipe buffers to the
756 static int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
761 while (pipe->nrbufs) {
762 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
763 const struct pipe_buf_operations *ops = buf->ops;
766 if (sd->len > sd->total_len)
767 sd->len = sd->total_len;
769 ret = buf->ops->confirm(pipe, buf);
776 ret = actor(pipe, buf, sd);
783 sd->num_spliced += ret;
786 sd->total_len -= ret;
790 ops->release(pipe, buf);
791 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
794 sd->need_wakeup = true;
805 * splice_from_pipe_next - wait for some data to splice from
806 * @pipe: pipe to splice from
807 * @sd: information about the splice operation
810 * This function will wait for some data and return a positive
811 * value (one) if pipe buffers are available. It will return zero
812 * or -errno if no more data needs to be spliced.
814 static int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
817 * Check for signal early to make process killable when there are
818 * always buffers available
820 if (signal_pending(current))
823 while (!pipe->nrbufs) {
827 if (!pipe->waiting_writers && sd->num_spliced)
830 if (sd->flags & SPLICE_F_NONBLOCK)
833 if (signal_pending(current))
836 if (sd->need_wakeup) {
837 wakeup_pipe_writers(pipe);
838 sd->need_wakeup = false;
848 * splice_from_pipe_begin - start splicing from pipe
849 * @sd: information about the splice operation
852 * This function should be called before a loop containing
853 * splice_from_pipe_next() and splice_from_pipe_feed() to
854 * initialize the necessary fields of @sd.
856 static void splice_from_pipe_begin(struct splice_desc *sd)
859 sd->need_wakeup = false;
863 * splice_from_pipe_end - finish splicing from pipe
864 * @pipe: pipe to splice from
865 * @sd: information about the splice operation
868 * This function will wake up pipe writers if necessary. It should
869 * be called after a loop containing splice_from_pipe_next() and
870 * splice_from_pipe_feed().
872 static void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
875 wakeup_pipe_writers(pipe);
879 * __splice_from_pipe - splice data from a pipe to given actor
880 * @pipe: pipe to splice from
881 * @sd: information to @actor
882 * @actor: handler that splices the data
885 * This function does little more than loop over the pipe and call
886 * @actor to do the actual moving of a single struct pipe_buffer to
887 * the desired destination. See pipe_to_file, pipe_to_sendpage, or
891 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
896 splice_from_pipe_begin(sd);
899 ret = splice_from_pipe_next(pipe, sd);
901 ret = splice_from_pipe_feed(pipe, sd, actor);
903 splice_from_pipe_end(pipe, sd);
905 return sd->num_spliced ? sd->num_spliced : ret;
907 EXPORT_SYMBOL(__splice_from_pipe);
910 * splice_from_pipe - splice data from a pipe to a file
911 * @pipe: pipe to splice from
912 * @out: file to splice to
913 * @ppos: position in @out
914 * @len: how many bytes to splice
915 * @flags: splice modifier flags
916 * @actor: handler that splices the data
919 * See __splice_from_pipe. This function locks the pipe inode,
920 * otherwise it's identical to __splice_from_pipe().
923 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
924 loff_t *ppos, size_t len, unsigned int flags,
928 struct splice_desc sd = {
936 ret = __splice_from_pipe(pipe, &sd, actor);
943 * iter_file_splice_write - splice data from a pipe to a file
945 * @out: file to write to
946 * @ppos: position in @out
947 * @len: number of bytes to splice
948 * @flags: splice modifier flags
951 * Will either move or copy pages (determined by @flags options) from
952 * the given pipe inode to the given file.
953 * This one is ->write_iter-based.
957 iter_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
958 loff_t *ppos, size_t len, unsigned int flags)
960 struct splice_desc sd = {
966 int nbufs = pipe->buffers;
967 struct bio_vec *array = kcalloc(nbufs, sizeof(struct bio_vec),
971 if (unlikely(!array))
976 splice_from_pipe_begin(&sd);
977 while (sd.total_len) {
978 struct iov_iter from;
982 ret = splice_from_pipe_next(pipe, &sd);
986 if (unlikely(nbufs < pipe->buffers)) {
988 nbufs = pipe->buffers;
989 array = kcalloc(nbufs, sizeof(struct bio_vec),
997 /* build the vector */
999 for (n = 0, idx = pipe->curbuf; left && n < pipe->nrbufs; n++, idx++) {
1000 struct pipe_buffer *buf = pipe->bufs + idx;
1001 size_t this_len = buf->len;
1003 if (this_len > left)
1006 if (idx == pipe->buffers - 1)
1009 ret = buf->ops->confirm(pipe, buf);
1010 if (unlikely(ret)) {
1011 if (ret == -ENODATA)
1016 array[n].bv_page = buf->page;
1017 array[n].bv_len = this_len;
1018 array[n].bv_offset = buf->offset;
1022 iov_iter_bvec(&from, ITER_BVEC | WRITE, array, n,
1023 sd.total_len - left);
1024 ret = vfs_iter_write(out, &from, &sd.pos);
1028 sd.num_spliced += ret;
1029 sd.total_len -= ret;
1032 /* dismiss the fully eaten buffers, adjust the partial one */
1034 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
1035 if (ret >= buf->len) {
1036 const struct pipe_buf_operations *ops = buf->ops;
1040 ops->release(pipe, buf);
1041 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
1044 sd.need_wakeup = true;
1054 splice_from_pipe_end(pipe, &sd);
1059 ret = sd.num_spliced;
1064 EXPORT_SYMBOL(iter_file_splice_write);
1066 static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1067 struct splice_desc *sd)
1071 loff_t tmp = sd->pos;
1073 data = kmap(buf->page);
1074 ret = __kernel_write(sd->u.file, data + buf->offset, sd->len, &tmp);
1080 static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1081 struct file *out, loff_t *ppos,
1082 size_t len, unsigned int flags)
1086 ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1094 * generic_splice_sendpage - splice data from a pipe to a socket
1095 * @pipe: pipe to splice from
1096 * @out: socket to write to
1097 * @ppos: position in @out
1098 * @len: number of bytes to splice
1099 * @flags: splice modifier flags
1102 * Will send @len bytes from the pipe to a network socket. No data copying
1106 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1107 loff_t *ppos, size_t len, unsigned int flags)
1109 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1112 EXPORT_SYMBOL(generic_splice_sendpage);
1115 * Attempt to initiate a splice from pipe to file.
1117 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1118 loff_t *ppos, size_t len, unsigned int flags)
1120 ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1121 loff_t *, size_t, unsigned int);
1123 if (out->f_op->splice_write)
1124 splice_write = out->f_op->splice_write;
1126 splice_write = default_file_splice_write;
1128 return splice_write(pipe, out, ppos, len, flags);
1132 * Attempt to initiate a splice from a file to a pipe.
1134 static long do_splice_to(struct file *in, loff_t *ppos,
1135 struct pipe_inode_info *pipe, size_t len,
1138 ssize_t (*splice_read)(struct file *, loff_t *,
1139 struct pipe_inode_info *, size_t, unsigned int);
1142 if (unlikely(!(in->f_mode & FMODE_READ)))
1145 ret = rw_verify_area(READ, in, ppos, len);
1146 if (unlikely(ret < 0))
1149 if (in->f_op->splice_read)
1150 splice_read = in->f_op->splice_read;
1152 splice_read = default_file_splice_read;
1154 return splice_read(in, ppos, pipe, len, flags);
1158 * splice_direct_to_actor - splices data directly between two non-pipes
1159 * @in: file to splice from
1160 * @sd: actor information on where to splice to
1161 * @actor: handles the data splicing
1164 * This is a special case helper to splice directly between two
1165 * points, without requiring an explicit pipe. Internally an allocated
1166 * pipe is cached in the process, and reused during the lifetime of
1170 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1171 splice_direct_actor *actor)
1173 struct pipe_inode_info *pipe;
1180 * We require the input being a regular file, as we don't want to
1181 * randomly drop data for eg socket -> socket splicing. Use the
1182 * piped splicing for that!
1184 i_mode = file_inode(in)->i_mode;
1185 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1189 * neither in nor out is a pipe, setup an internal pipe attached to
1190 * 'out' and transfer the wanted data from 'in' to 'out' through that
1192 pipe = current->splice_pipe;
1193 if (unlikely(!pipe)) {
1194 pipe = alloc_pipe_info();
1199 * We don't have an immediate reader, but we'll read the stuff
1200 * out of the pipe right after the splice_to_pipe(). So set
1201 * PIPE_READERS appropriately.
1205 current->splice_pipe = pipe;
1213 len = sd->total_len;
1217 * Don't block on output, we have to drain the direct pipe.
1219 sd->flags &= ~SPLICE_F_NONBLOCK;
1220 more = sd->flags & SPLICE_F_MORE;
1224 loff_t pos = sd->pos, prev_pos = pos;
1226 ret = do_splice_to(in, &pos, pipe, len, flags);
1227 if (unlikely(ret <= 0))
1231 sd->total_len = read_len;
1234 * If more data is pending, set SPLICE_F_MORE
1235 * If this is the last data and SPLICE_F_MORE was not set
1236 * initially, clears it.
1239 sd->flags |= SPLICE_F_MORE;
1241 sd->flags &= ~SPLICE_F_MORE;
1243 * NOTE: nonblocking mode only applies to the input. We
1244 * must not do the output in nonblocking mode as then we
1245 * could get stuck data in the internal pipe:
1247 ret = actor(pipe, sd);
1248 if (unlikely(ret <= 0)) {
1257 if (ret < read_len) {
1258 sd->pos = prev_pos + ret;
1264 pipe->nrbufs = pipe->curbuf = 0;
1270 * If we did an incomplete transfer we must release
1271 * the pipe buffers in question:
1273 for (i = 0; i < pipe->buffers; i++) {
1274 struct pipe_buffer *buf = pipe->bufs + i;
1277 buf->ops->release(pipe, buf);
1287 EXPORT_SYMBOL(splice_direct_to_actor);
1289 static int direct_splice_actor(struct pipe_inode_info *pipe,
1290 struct splice_desc *sd)
1292 struct file *file = sd->u.file;
1294 return do_splice_from(pipe, file, sd->opos, sd->total_len,
1299 * do_splice_direct - splices data directly between two files
1300 * @in: file to splice from
1301 * @ppos: input file offset
1302 * @out: file to splice to
1303 * @opos: output file offset
1304 * @len: number of bytes to splice
1305 * @flags: splice modifier flags
1308 * For use by do_sendfile(). splice can easily emulate sendfile, but
1309 * doing it in the application would incur an extra system call
1310 * (splice in + splice out, as compared to just sendfile()). So this helper
1311 * can splice directly through a process-private pipe.
1314 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1315 loff_t *opos, size_t len, unsigned int flags)
1317 struct splice_desc sd = {
1327 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1330 if (unlikely(out->f_flags & O_APPEND))
1333 ret = rw_verify_area(WRITE, out, opos, len);
1334 if (unlikely(ret < 0))
1337 ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1343 EXPORT_SYMBOL(do_splice_direct);
1345 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1346 struct pipe_inode_info *opipe,
1347 size_t len, unsigned int flags);
1350 * Determine where to splice to/from.
1352 static long do_splice(struct file *in, loff_t __user *off_in,
1353 struct file *out, loff_t __user *off_out,
1354 size_t len, unsigned int flags)
1356 struct pipe_inode_info *ipipe;
1357 struct pipe_inode_info *opipe;
1361 ipipe = get_pipe_info(in);
1362 opipe = get_pipe_info(out);
1364 if (ipipe && opipe) {
1365 if (off_in || off_out)
1368 if (!(in->f_mode & FMODE_READ))
1371 if (!(out->f_mode & FMODE_WRITE))
1374 /* Splicing to self would be fun, but... */
1378 return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1385 if (!(out->f_mode & FMODE_PWRITE))
1387 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1390 offset = out->f_pos;
1393 if (unlikely(!(out->f_mode & FMODE_WRITE)))
1396 if (unlikely(out->f_flags & O_APPEND))
1399 ret = rw_verify_area(WRITE, out, &offset, len);
1400 if (unlikely(ret < 0))
1403 file_start_write(out);
1404 ret = do_splice_from(ipipe, out, &offset, len, flags);
1405 file_end_write(out);
1408 out->f_pos = offset;
1409 else if (copy_to_user(off_out, &offset, sizeof(loff_t)))
1419 if (!(in->f_mode & FMODE_PREAD))
1421 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1427 ret = do_splice_to(in, &offset, opipe, len, flags);
1431 else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
1441 * Map an iov into an array of pages and offset/length tupples. With the
1442 * partial_page structure, we can map several non-contiguous ranges into
1443 * our ones pages[] map instead of splitting that operation into pieces.
1444 * Could easily be exported as a generic helper for other users, in which
1445 * case one would probably want to add a 'max_nr_pages' parameter as well.
1447 static int get_iovec_page_array(const struct iovec __user *iov,
1448 unsigned int nr_vecs, struct page **pages,
1449 struct partial_page *partial, bool aligned,
1450 unsigned int pipe_buffers)
1452 int buffers = 0, error = 0;
1455 unsigned long off, npages;
1462 if (copy_from_user(&entry, iov, sizeof(entry)))
1465 base = entry.iov_base;
1466 len = entry.iov_len;
1469 * Sanity check this iovec. 0 read succeeds.
1475 if (!access_ok(VERIFY_READ, base, len))
1479 * Get this base offset and number of pages, then map
1480 * in the user pages.
1482 off = (unsigned long) base & ~PAGE_MASK;
1485 * If asked for alignment, the offset must be zero and the
1486 * length a multiple of the PAGE_SIZE.
1489 if (aligned && (off || len & ~PAGE_MASK))
1492 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1493 if (npages > pipe_buffers - buffers)
1494 npages = pipe_buffers - buffers;
1496 error = get_user_pages_fast((unsigned long)base, npages,
1497 0, &pages[buffers]);
1499 if (unlikely(error <= 0))
1503 * Fill this contiguous range into the partial page map.
1505 for (i = 0; i < error; i++) {
1506 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1508 partial[buffers].offset = off;
1509 partial[buffers].len = plen;
1517 * We didn't complete this iov, stop here since it probably
1518 * means we have to move some of this into a pipe to
1519 * be able to continue.
1525 * Don't continue if we mapped fewer pages than we asked for,
1526 * or if we mapped the max number of pages that we have
1529 if (error < npages || buffers == pipe_buffers)
1542 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1543 struct splice_desc *sd)
1545 int n = copy_page_to_iter(buf->page, buf->offset, sd->len, sd->u.data);
1546 return n == sd->len ? n : -EFAULT;
1550 * For lack of a better implementation, implement vmsplice() to userspace
1551 * as a simple copy of the pipes pages to the user iov.
1553 static long vmsplice_to_user(struct file *file, const struct iovec __user *uiov,
1554 unsigned long nr_segs, unsigned int flags)
1556 struct pipe_inode_info *pipe;
1557 struct splice_desc sd;
1559 struct iovec iovstack[UIO_FASTIOV];
1560 struct iovec *iov = iovstack;
1561 struct iov_iter iter;
1563 pipe = get_pipe_info(file);
1567 ret = import_iovec(READ, uiov, nr_segs,
1568 ARRAY_SIZE(iovstack), &iov, &iter);
1572 sd.total_len = iov_iter_count(&iter);
1580 ret = __splice_from_pipe(pipe, &sd, pipe_to_user);
1589 * vmsplice splices a user address range into a pipe. It can be thought of
1590 * as splice-from-memory, where the regular splice is splice-from-file (or
1591 * to file). In both cases the output is a pipe, naturally.
1593 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1594 unsigned long nr_segs, unsigned int flags)
1596 struct pipe_inode_info *pipe;
1597 struct page *pages[PIPE_DEF_BUFFERS];
1598 struct partial_page partial[PIPE_DEF_BUFFERS];
1599 struct splice_pipe_desc spd = {
1602 .nr_pages_max = PIPE_DEF_BUFFERS,
1604 .ops = &user_page_pipe_buf_ops,
1605 .spd_release = spd_release_page,
1609 pipe = get_pipe_info(file);
1613 if (splice_grow_spd(pipe, &spd))
1616 spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1619 if (spd.nr_pages <= 0)
1622 ret = splice_to_pipe(pipe, &spd);
1624 splice_shrink_spd(&spd);
1629 * Note that vmsplice only really supports true splicing _from_ user memory
1630 * to a pipe, not the other way around. Splicing from user memory is a simple
1631 * operation that can be supported without any funky alignment restrictions
1632 * or nasty vm tricks. We simply map in the user memory and fill them into
1633 * a pipe. The reverse isn't quite as easy, though. There are two possible
1634 * solutions for that:
1636 * - memcpy() the data internally, at which point we might as well just
1637 * do a regular read() on the buffer anyway.
1638 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1639 * has restriction limitations on both ends of the pipe).
1641 * Currently we punt and implement it as a normal copy, see pipe_to_user().
1644 SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1645 unsigned long, nr_segs, unsigned int, flags)
1650 if (unlikely(nr_segs > UIO_MAXIOV))
1652 else if (unlikely(!nr_segs))
1658 if (f.file->f_mode & FMODE_WRITE)
1659 error = vmsplice_to_pipe(f.file, iov, nr_segs, flags);
1660 else if (f.file->f_mode & FMODE_READ)
1661 error = vmsplice_to_user(f.file, iov, nr_segs, flags);
1669 #ifdef CONFIG_COMPAT
1670 COMPAT_SYSCALL_DEFINE4(vmsplice, int, fd, const struct compat_iovec __user *, iov32,
1671 unsigned int, nr_segs, unsigned int, flags)
1674 struct iovec __user *iov;
1675 if (nr_segs > UIO_MAXIOV)
1677 iov = compat_alloc_user_space(nr_segs * sizeof(struct iovec));
1678 for (i = 0; i < nr_segs; i++) {
1679 struct compat_iovec v;
1680 if (get_user(v.iov_base, &iov32[i].iov_base) ||
1681 get_user(v.iov_len, &iov32[i].iov_len) ||
1682 put_user(compat_ptr(v.iov_base), &iov[i].iov_base) ||
1683 put_user(v.iov_len, &iov[i].iov_len))
1686 return sys_vmsplice(fd, iov, nr_segs, flags);
1690 SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1691 int, fd_out, loff_t __user *, off_out,
1692 size_t, len, unsigned int, flags)
1703 if (in.file->f_mode & FMODE_READ) {
1704 out = fdget(fd_out);
1706 if (out.file->f_mode & FMODE_WRITE)
1707 error = do_splice(in.file, off_in,
1719 * Make sure there's data to read. Wait for input if we can, otherwise
1720 * return an appropriate error.
1722 static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1727 * Check ->nrbufs without the inode lock first. This function
1728 * is speculative anyways, so missing one is ok.
1736 while (!pipe->nrbufs) {
1737 if (signal_pending(current)) {
1743 if (!pipe->waiting_writers) {
1744 if (flags & SPLICE_F_NONBLOCK) {
1757 * Make sure there's writeable room. Wait for room if we can, otherwise
1758 * return an appropriate error.
1760 static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1765 * Check ->nrbufs without the inode lock first. This function
1766 * is speculative anyways, so missing one is ok.
1768 if (pipe->nrbufs < pipe->buffers)
1774 while (pipe->nrbufs >= pipe->buffers) {
1775 if (!pipe->readers) {
1776 send_sig(SIGPIPE, current, 0);
1780 if (flags & SPLICE_F_NONBLOCK) {
1784 if (signal_pending(current)) {
1788 pipe->waiting_writers++;
1790 pipe->waiting_writers--;
1798 * Splice contents of ipipe to opipe.
1800 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1801 struct pipe_inode_info *opipe,
1802 size_t len, unsigned int flags)
1804 struct pipe_buffer *ibuf, *obuf;
1806 bool input_wakeup = false;
1810 ret = ipipe_prep(ipipe, flags);
1814 ret = opipe_prep(opipe, flags);
1819 * Potential ABBA deadlock, work around it by ordering lock
1820 * grabbing by pipe info address. Otherwise two different processes
1821 * could deadlock (one doing tee from A -> B, the other from B -> A).
1823 pipe_double_lock(ipipe, opipe);
1826 if (!opipe->readers) {
1827 send_sig(SIGPIPE, current, 0);
1833 if (!ipipe->nrbufs && !ipipe->writers)
1837 * Cannot make any progress, because either the input
1838 * pipe is empty or the output pipe is full.
1840 if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1841 /* Already processed some buffers, break */
1845 if (flags & SPLICE_F_NONBLOCK) {
1851 * We raced with another reader/writer and haven't
1852 * managed to process any buffers. A zero return
1853 * value means EOF, so retry instead.
1860 ibuf = ipipe->bufs + ipipe->curbuf;
1861 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1862 obuf = opipe->bufs + nbuf;
1864 if (len >= ibuf->len) {
1866 * Simply move the whole buffer from ipipe to opipe
1871 ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1873 input_wakeup = true;
1876 * Get a reference to this pipe buffer,
1877 * so we can copy the contents over.
1879 if (!pipe_buf_get(ipipe, ibuf)) {
1887 * Don't inherit the gift flag, we need to
1888 * prevent multiple steals of this page.
1890 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1894 ibuf->offset += obuf->len;
1895 ibuf->len -= obuf->len;
1905 * If we put data in the output pipe, wakeup any potential readers.
1908 wakeup_pipe_readers(opipe);
1911 wakeup_pipe_writers(ipipe);
1917 * Link contents of ipipe to opipe.
1919 static int link_pipe(struct pipe_inode_info *ipipe,
1920 struct pipe_inode_info *opipe,
1921 size_t len, unsigned int flags)
1923 struct pipe_buffer *ibuf, *obuf;
1924 int ret = 0, i = 0, nbuf;
1927 * Potential ABBA deadlock, work around it by ordering lock
1928 * grabbing by pipe info address. Otherwise two different processes
1929 * could deadlock (one doing tee from A -> B, the other from B -> A).
1931 pipe_double_lock(ipipe, opipe);
1934 if (!opipe->readers) {
1935 send_sig(SIGPIPE, current, 0);
1942 * If we have iterated all input buffers or ran out of
1943 * output room, break.
1945 if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1948 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1949 nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1952 * Get a reference to this pipe buffer,
1953 * so we can copy the contents over.
1955 if (!pipe_buf_get(ipipe, ibuf)) {
1961 obuf = opipe->bufs + nbuf;
1965 * Don't inherit the gift flag, we need to
1966 * prevent multiple steals of this page.
1968 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1970 if (obuf->len > len)
1980 * return EAGAIN if we have the potential of some data in the
1981 * future, otherwise just return 0
1983 if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1990 * If we put data in the output pipe, wakeup any potential readers.
1993 wakeup_pipe_readers(opipe);
1999 * This is a tee(1) implementation that works on pipes. It doesn't copy
2000 * any data, it simply references the 'in' pages on the 'out' pipe.
2001 * The 'flags' used are the SPLICE_F_* variants, currently the only
2002 * applicable one is SPLICE_F_NONBLOCK.
2004 static long do_tee(struct file *in, struct file *out, size_t len,
2007 struct pipe_inode_info *ipipe = get_pipe_info(in);
2008 struct pipe_inode_info *opipe = get_pipe_info(out);
2012 * Duplicate the contents of ipipe to opipe without actually
2015 if (ipipe && opipe && ipipe != opipe) {
2017 * Keep going, unless we encounter an error. The ipipe/opipe
2018 * ordering doesn't really matter.
2020 ret = ipipe_prep(ipipe, flags);
2022 ret = opipe_prep(opipe, flags);
2024 ret = link_pipe(ipipe, opipe, len, flags);
2031 SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2042 if (in.file->f_mode & FMODE_READ) {
2043 struct fd out = fdget(fdout);
2045 if (out.file->f_mode & FMODE_WRITE)
2046 error = do_tee(in.file, out.file,