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@suse.de>
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/pipe_fs_i.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
38 * Passed to splice_to_pipe
40 struct splice_pipe_desc {
41 struct page **pages; /* page map */
42 struct partial_page *partial; /* pages[] may not be contig */
43 int nr_pages; /* number of pages in map */
44 unsigned int flags; /* splice flags */
45 struct pipe_buf_operations *ops;/* ops associated with output pipe */
49 * Attempt to steal a page from a pipe buffer. This should perhaps go into
50 * a vm helper function, it's already simplified quite a bit by the
51 * addition of remove_mapping(). If success is returned, the caller may
52 * attempt to reuse this page for another destination.
54 static int page_cache_pipe_buf_steal(struct pipe_inode_info *info,
55 struct pipe_buffer *buf)
57 struct page *page = buf->page;
58 struct address_space *mapping = page_mapping(page);
62 WARN_ON(!PageUptodate(page));
65 * At least for ext2 with nobh option, we need to wait on writeback
66 * completing on this page, since we'll remove it from the pagecache.
67 * Otherwise truncate wont wait on the page, allowing the disk
68 * blocks to be reused by someone else before we actually wrote our
69 * data to them. fs corruption ensues.
71 wait_on_page_writeback(page);
73 if (PagePrivate(page))
74 try_to_release_page(page, mapping_gfp_mask(mapping));
76 if (!remove_mapping(mapping, page)) {
84 static void page_cache_pipe_buf_release(struct pipe_inode_info *info,
85 struct pipe_buffer *buf)
87 page_cache_release(buf->page);
90 static int page_cache_pipe_buf_pin(struct pipe_inode_info *info,
91 struct pipe_buffer *buf)
93 struct page *page = buf->page;
96 if (!PageUptodate(page)) {
100 * Page got truncated/unhashed. This will cause a 0-byte
101 * splice, if this is the first page.
103 if (!page->mapping) {
109 * Uh oh, read-error from disk.
111 if (!PageUptodate(page)) {
117 * Page is ok afterall, we are done.
128 static struct pipe_buf_operations page_cache_pipe_buf_ops = {
130 .map = generic_pipe_buf_map,
131 .unmap = generic_pipe_buf_unmap,
132 .pin = page_cache_pipe_buf_pin,
133 .release = page_cache_pipe_buf_release,
134 .steal = page_cache_pipe_buf_steal,
135 .get = generic_pipe_buf_get,
138 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
139 struct pipe_buffer *buf)
141 if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
144 return generic_pipe_buf_steal(pipe, buf);
147 static struct pipe_buf_operations user_page_pipe_buf_ops = {
149 .map = generic_pipe_buf_map,
150 .unmap = generic_pipe_buf_unmap,
151 .pin = generic_pipe_buf_pin,
152 .release = page_cache_pipe_buf_release,
153 .steal = user_page_pipe_buf_steal,
154 .get = generic_pipe_buf_get,
158 * Pipe output worker. This sets up our pipe format with the page cache
159 * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
161 static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
162 struct splice_pipe_desc *spd)
164 int ret, do_wakeup, page_nr;
171 mutex_lock(&pipe->inode->i_mutex);
174 if (!pipe->readers) {
175 send_sig(SIGPIPE, current, 0);
181 if (pipe->nrbufs < PIPE_BUFFERS) {
182 int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
183 struct pipe_buffer *buf = pipe->bufs + newbuf;
185 buf->page = spd->pages[page_nr];
186 buf->offset = spd->partial[page_nr].offset;
187 buf->len = spd->partial[page_nr].len;
189 if (spd->flags & SPLICE_F_GIFT)
190 buf->flags |= PIPE_BUF_FLAG_GIFT;
199 if (!--spd->nr_pages)
201 if (pipe->nrbufs < PIPE_BUFFERS)
207 if (spd->flags & SPLICE_F_NONBLOCK) {
213 if (signal_pending(current)) {
221 if (waitqueue_active(&pipe->wait))
222 wake_up_interruptible_sync(&pipe->wait);
223 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
227 pipe->waiting_writers++;
229 pipe->waiting_writers--;
233 mutex_unlock(&pipe->inode->i_mutex);
237 if (waitqueue_active(&pipe->wait))
238 wake_up_interruptible(&pipe->wait);
239 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
242 while (page_nr < spd->nr_pages)
243 page_cache_release(spd->pages[page_nr++]);
249 __generic_file_splice_read(struct file *in, loff_t *ppos,
250 struct pipe_inode_info *pipe, size_t len,
253 struct address_space *mapping = in->f_mapping;
254 unsigned int loff, nr_pages;
255 struct page *pages[PIPE_BUFFERS];
256 struct partial_page partial[PIPE_BUFFERS];
258 pgoff_t index, end_index;
262 struct splice_pipe_desc spd = {
266 .ops = &page_cache_pipe_buf_ops,
269 index = *ppos >> PAGE_CACHE_SHIFT;
270 loff = *ppos & ~PAGE_CACHE_MASK;
271 nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
273 if (nr_pages > PIPE_BUFFERS)
274 nr_pages = PIPE_BUFFERS;
277 * Initiate read-ahead on this page range. however, don't call into
278 * read-ahead if this is a non-zero offset (we are likely doing small
279 * chunk splice and the page is already there) for a single page.
281 if (!loff || nr_pages > 1)
282 page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
285 * Now fill in the holes:
291 * Lookup the (hopefully) full range of pages we need.
293 spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
296 * If find_get_pages_contig() returned fewer pages than we needed,
299 index += spd.nr_pages;
300 while (spd.nr_pages < nr_pages) {
302 * Page could be there, find_get_pages_contig() breaks on
305 page = find_get_page(mapping, index);
308 * Make sure the read-ahead engine is notified
309 * about this failure.
311 handle_ra_miss(mapping, &in->f_ra, index);
314 * page didn't exist, allocate one.
316 page = page_cache_alloc_cold(mapping);
320 error = add_to_page_cache_lru(page, mapping, index,
321 mapping_gfp_mask(mapping));
322 if (unlikely(error)) {
323 page_cache_release(page);
327 * add_to_page_cache() locks the page, unlock it
328 * to avoid convoluting the logic below even more.
333 pages[spd.nr_pages++] = page;
338 * Now loop over the map and see if we need to start IO on any
339 * pages, fill in the partial map, etc.
341 index = *ppos >> PAGE_CACHE_SHIFT;
342 nr_pages = spd.nr_pages;
344 for (page_nr = 0; page_nr < nr_pages; page_nr++) {
345 unsigned int this_len;
351 * this_len is the max we'll use from this page
353 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
354 page = pages[page_nr];
357 * If the page isn't uptodate, we may need to start io on it
359 if (!PageUptodate(page)) {
361 * If in nonblock mode then dont block on waiting
362 * for an in-flight io page
364 if (flags & SPLICE_F_NONBLOCK)
370 * page was truncated, stop here. if this isn't the
371 * first page, we'll just complete what we already
374 if (!page->mapping) {
379 * page was already under io and is now done, great
381 if (PageUptodate(page)) {
387 * need to read in the page
389 error = mapping->a_ops->readpage(in, page);
390 if (unlikely(error)) {
392 * We really should re-lookup the page here,
393 * but it complicates things a lot. Instead
394 * lets just do what we already stored, and
395 * we'll get it the next time we are called.
397 if (error == AOP_TRUNCATED_PAGE)
404 * i_size must be checked after ->readpage().
406 isize = i_size_read(mapping->host);
407 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
408 if (unlikely(!isize || index > end_index))
412 * if this is the last page, see if we need to shrink
413 * the length and stop
415 if (end_index == index) {
416 loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
417 if (total_len + loff > isize)
420 * force quit after adding this page
423 this_len = min(this_len, loff);
428 partial[page_nr].offset = loff;
429 partial[page_nr].len = this_len;
431 total_len += this_len;
438 * Release any pages at the end, if we quit early. 'i' is how far
439 * we got, 'nr_pages' is how many pages are in the map.
441 while (page_nr < nr_pages)
442 page_cache_release(pages[page_nr++]);
445 return splice_to_pipe(pipe, &spd);
451 * generic_file_splice_read - splice data from file to a pipe
452 * @in: file to splice from
453 * @pipe: pipe to splice to
454 * @len: number of bytes to splice
455 * @flags: splice modifier flags
457 * Will read pages from given file and fill them into a pipe.
459 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
460 struct pipe_inode_info *pipe, size_t len,
470 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
477 if (flags & SPLICE_F_NONBLOCK) {
494 EXPORT_SYMBOL(generic_file_splice_read);
497 * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
498 * using sendpage(). Return the number of bytes sent.
500 static int pipe_to_sendpage(struct pipe_inode_info *info,
501 struct pipe_buffer *buf, struct splice_desc *sd)
503 struct file *file = sd->file;
504 loff_t pos = sd->pos;
507 ret = buf->ops->pin(info, buf);
509 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
511 ret = file->f_op->sendpage(file, buf->page, buf->offset,
512 sd->len, &pos, more);
519 * This is a little more tricky than the file -> pipe splicing. There are
520 * basically three cases:
522 * - Destination page already exists in the address space and there
523 * are users of it. For that case we have no other option that
524 * copying the data. Tough luck.
525 * - Destination page already exists in the address space, but there
526 * are no users of it. Make sure it's uptodate, then drop it. Fall
527 * through to last case.
528 * - Destination page does not exist, we can add the pipe page to
529 * the page cache and avoid the copy.
531 * If asked to move pages to the output file (SPLICE_F_MOVE is set in
532 * sd->flags), we attempt to migrate pages from the pipe to the output
533 * file address space page cache. This is possible if no one else has
534 * the pipe page referenced outside of the pipe and page cache. If
535 * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
536 * a new page in the output file page cache and fill/dirty that.
538 static int pipe_to_file(struct pipe_inode_info *info, struct pipe_buffer *buf,
539 struct splice_desc *sd)
541 struct file *file = sd->file;
542 struct address_space *mapping = file->f_mapping;
543 gfp_t gfp_mask = mapping_gfp_mask(mapping);
544 unsigned int offset, this_len;
550 * make sure the data in this buffer is uptodate
552 ret = buf->ops->pin(info, buf);
556 index = sd->pos >> PAGE_CACHE_SHIFT;
557 offset = sd->pos & ~PAGE_CACHE_MASK;
560 if (this_len + offset > PAGE_CACHE_SIZE)
561 this_len = PAGE_CACHE_SIZE - offset;
564 * Reuse buf page, if SPLICE_F_MOVE is set and we are doing a full
567 if ((sd->flags & SPLICE_F_MOVE) && this_len == PAGE_CACHE_SIZE) {
569 * If steal succeeds, buf->page is now pruned from the vm
570 * side (page cache) and we can reuse it. The page will also
571 * be locked on successful return.
573 if (buf->ops->steal(info, buf))
577 page_cache_get(page);
580 * page must be on the LRU for adding to the pagecache.
581 * Check this without grabbing the zone lock, if it isn't
582 * the do grab the zone lock, recheck, and add if necessary.
584 if (!PageLRU(page)) {
585 struct zone *zone = page_zone(page);
587 spin_lock_irq(&zone->lru_lock);
588 if (!PageLRU(page)) {
590 add_page_to_inactive_list(zone, page);
592 spin_unlock_irq(&zone->lru_lock);
595 if (add_to_page_cache(page, mapping, index, gfp_mask)) {
596 page_cache_release(page);
602 page = find_lock_page(mapping, index);
605 page = page_cache_alloc_cold(mapping);
610 * This will also lock the page
612 ret = add_to_page_cache_lru(page, mapping, index,
619 * We get here with the page locked. If the page is also
620 * uptodate, we don't need to do more. If it isn't, we
621 * may need to bring it in if we are not going to overwrite
624 if (!PageUptodate(page)) {
625 if (this_len < PAGE_CACHE_SIZE) {
626 ret = mapping->a_ops->readpage(file, page);
632 if (!PageUptodate(page)) {
634 * Page got invalidated, repeat.
636 if (!page->mapping) {
638 page_cache_release(page);
645 SetPageUptodate(page);
649 ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
651 loff_t isize = i_size_read(mapping->host);
653 if (ret != AOP_TRUNCATED_PAGE)
655 page_cache_release(page);
656 if (ret == AOP_TRUNCATED_PAGE)
660 * prepare_write() may have instantiated a few blocks
661 * outside i_size. Trim these off again.
663 if (sd->pos + this_len > isize)
664 vmtruncate(mapping->host, isize);
669 if (buf->page != page) {
671 * Careful, ->map() uses KM_USER0!
673 char *src = buf->ops->map(info, buf, 1);
674 char *dst = kmap_atomic(page, KM_USER1);
676 memcpy(dst + offset, src + buf->offset, this_len);
677 flush_dcache_page(page);
678 kunmap_atomic(dst, KM_USER1);
679 buf->ops->unmap(info, buf, src);
682 ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
685 * Return the number of bytes written and mark page as
686 * accessed, we are now done!
689 mark_page_accessed(page);
690 balance_dirty_pages_ratelimited(mapping);
691 } else if (ret == AOP_TRUNCATED_PAGE) {
692 page_cache_release(page);
696 page_cache_release(page);
703 * Pipe input worker. Most of this logic works like a regular pipe, the
704 * key here is the 'actor' worker passed in that actually moves the data
705 * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
707 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
708 loff_t *ppos, size_t len, unsigned int flags,
711 int ret, do_wakeup, err;
712 struct splice_desc sd;
723 mutex_lock(&pipe->inode->i_mutex);
727 struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
728 struct pipe_buf_operations *ops = buf->ops;
731 if (sd.len > sd.total_len)
732 sd.len = sd.total_len;
734 err = actor(pipe, buf, &sd);
736 if (!ret && err != -ENODATA)
754 ops->release(pipe, buf);
755 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
769 if (!pipe->waiting_writers) {
774 if (flags & SPLICE_F_NONBLOCK) {
780 if (signal_pending(current)) {
788 if (waitqueue_active(&pipe->wait))
789 wake_up_interruptible_sync(&pipe->wait);
790 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
798 mutex_unlock(&pipe->inode->i_mutex);
802 if (waitqueue_active(&pipe->wait))
803 wake_up_interruptible(&pipe->wait);
804 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
811 * generic_file_splice_write - splice data from a pipe to a file
813 * @out: file to write to
814 * @len: number of bytes to splice
815 * @flags: splice modifier flags
817 * Will either move or copy pages (determined by @flags options) from
818 * the given pipe inode to the given file.
822 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
823 loff_t *ppos, size_t len, unsigned int flags)
825 struct address_space *mapping = out->f_mapping;
828 ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
830 struct inode *inode = mapping->host;
835 * If file or inode is SYNC and we actually wrote some data,
838 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
841 mutex_lock(&inode->i_mutex);
842 err = generic_osync_inode(inode, mapping,
843 OSYNC_METADATA|OSYNC_DATA);
844 mutex_unlock(&inode->i_mutex);
854 EXPORT_SYMBOL(generic_file_splice_write);
857 * generic_splice_sendpage - splice data from a pipe to a socket
859 * @out: socket to write to
860 * @len: number of bytes to splice
861 * @flags: splice modifier flags
863 * Will send @len bytes from the pipe to a network socket. No data copying
867 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
868 loff_t *ppos, size_t len, unsigned int flags)
870 return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
873 EXPORT_SYMBOL(generic_splice_sendpage);
876 * Attempt to initiate a splice from pipe to file.
878 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
879 loff_t *ppos, size_t len, unsigned int flags)
883 if (unlikely(!out->f_op || !out->f_op->splice_write))
886 if (unlikely(!(out->f_mode & FMODE_WRITE)))
889 ret = rw_verify_area(WRITE, out, ppos, len);
890 if (unlikely(ret < 0))
893 return out->f_op->splice_write(pipe, out, ppos, len, flags);
897 * Attempt to initiate a splice from a file to a pipe.
899 static long do_splice_to(struct file *in, loff_t *ppos,
900 struct pipe_inode_info *pipe, size_t len,
906 if (unlikely(!in->f_op || !in->f_op->splice_read))
909 if (unlikely(!(in->f_mode & FMODE_READ)))
912 ret = rw_verify_area(READ, in, ppos, len);
913 if (unlikely(ret < 0))
916 isize = i_size_read(in->f_mapping->host);
917 if (unlikely(*ppos >= isize))
920 left = isize - *ppos;
921 if (unlikely(left < len))
924 return in->f_op->splice_read(in, ppos, pipe, len, flags);
927 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
928 size_t len, unsigned int flags)
930 struct pipe_inode_info *pipe;
937 * We require the input being a regular file, as we don't want to
938 * randomly drop data for eg socket -> socket splicing. Use the
939 * piped splicing for that!
941 i_mode = in->f_dentry->d_inode->i_mode;
942 if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
946 * neither in nor out is a pipe, setup an internal pipe attached to
947 * 'out' and transfer the wanted data from 'in' to 'out' through that
949 pipe = current->splice_pipe;
950 if (unlikely(!pipe)) {
951 pipe = alloc_pipe_info(NULL);
956 * We don't have an immediate reader, but we'll read the stuff
957 * out of the pipe right after the splice_to_pipe(). So set
958 * PIPE_READERS appropriately.
962 current->splice_pipe = pipe;
973 size_t read_len, max_read_len;
976 * Do at most PIPE_BUFFERS pages worth of transfer:
978 max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
980 ret = do_splice_to(in, ppos, pipe, max_read_len, flags);
981 if (unlikely(ret < 0))
987 * NOTE: nonblocking mode only applies to the input. We
988 * must not do the output in nonblocking mode as then we
989 * could get stuck data in the internal pipe:
991 ret = do_splice_from(pipe, out, &out_off, read_len,
992 flags & ~SPLICE_F_NONBLOCK);
993 if (unlikely(ret < 0))
1000 * In nonblocking mode, if we got back a short read then
1001 * that was due to either an IO error or due to the
1002 * pagecache entry not being there. In the IO error case
1003 * the _next_ splice attempt will produce a clean IO error
1004 * return value (not a short read), so in both cases it's
1005 * correct to break out of the loop here:
1007 if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
1011 pipe->nrbufs = pipe->curbuf = 0;
1017 * If we did an incomplete transfer we must release
1018 * the pipe buffers in question:
1020 for (i = 0; i < PIPE_BUFFERS; i++) {
1021 struct pipe_buffer *buf = pipe->bufs + i;
1024 buf->ops->release(pipe, buf);
1028 pipe->nrbufs = pipe->curbuf = 0;
1031 * If we transferred some data, return the number of bytes:
1039 EXPORT_SYMBOL(do_splice_direct);
1042 * Determine where to splice to/from.
1044 static long do_splice(struct file *in, loff_t __user *off_in,
1045 struct file *out, loff_t __user *off_out,
1046 size_t len, unsigned int flags)
1048 struct pipe_inode_info *pipe;
1049 loff_t offset, *off;
1052 pipe = in->f_dentry->d_inode->i_pipe;
1057 if (out->f_op->llseek == no_llseek)
1059 if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1065 ret = do_splice_from(pipe, out, off, len, flags);
1067 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1073 pipe = out->f_dentry->d_inode->i_pipe;
1078 if (in->f_op->llseek == no_llseek)
1080 if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1086 ret = do_splice_to(in, off, pipe, len, flags);
1088 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1098 * Map an iov into an array of pages and offset/length tupples. With the
1099 * partial_page structure, we can map several non-contiguous ranges into
1100 * our ones pages[] map instead of splitting that operation into pieces.
1101 * Could easily be exported as a generic helper for other users, in which
1102 * case one would probably want to add a 'max_nr_pages' parameter as well.
1104 static int get_iovec_page_array(const struct iovec __user *iov,
1105 unsigned int nr_vecs, struct page **pages,
1106 struct partial_page *partial, int aligned)
1108 int buffers = 0, error = 0;
1111 * It's ok to take the mmap_sem for reading, even
1112 * across a "get_user()".
1114 down_read(¤t->mm->mmap_sem);
1117 unsigned long off, npages;
1123 * Get user address base and length for this iovec.
1125 error = get_user(base, &iov->iov_base);
1126 if (unlikely(error))
1128 error = get_user(len, &iov->iov_len);
1129 if (unlikely(error))
1133 * Sanity check this iovec. 0 read succeeds.
1138 if (unlikely(!base))
1142 * Get this base offset and number of pages, then map
1143 * in the user pages.
1145 off = (unsigned long) base & ~PAGE_MASK;
1148 * If asked for alignment, the offset must be zero and the
1149 * length a multiple of the PAGE_SIZE.
1152 if (aligned && (off || len & ~PAGE_MASK))
1155 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1156 if (npages > PIPE_BUFFERS - buffers)
1157 npages = PIPE_BUFFERS - buffers;
1159 error = get_user_pages(current, current->mm,
1160 (unsigned long) base, npages, 0, 0,
1161 &pages[buffers], NULL);
1163 if (unlikely(error <= 0))
1167 * Fill this contiguous range into the partial page map.
1169 for (i = 0; i < error; i++) {
1170 const int plen = min_t(size_t, len, PAGE_SIZE - off);
1172 partial[buffers].offset = off;
1173 partial[buffers].len = plen;
1181 * We didn't complete this iov, stop here since it probably
1182 * means we have to move some of this into a pipe to
1183 * be able to continue.
1189 * Don't continue if we mapped fewer pages than we asked for,
1190 * or if we mapped the max number of pages that we have
1193 if (error < npages || buffers == PIPE_BUFFERS)
1200 up_read(¤t->mm->mmap_sem);
1209 * vmsplice splices a user address range into a pipe. It can be thought of
1210 * as splice-from-memory, where the regular splice is splice-from-file (or
1211 * to file). In both cases the output is a pipe, naturally.
1213 * Note that vmsplice only supports splicing _from_ user memory to a pipe,
1214 * not the other way around. Splicing from user memory is a simple operation
1215 * that can be supported without any funky alignment restrictions or nasty
1216 * vm tricks. We simply map in the user memory and fill them into a pipe.
1217 * The reverse isn't quite as easy, though. There are two possible solutions
1220 * - memcpy() the data internally, at which point we might as well just
1221 * do a regular read() on the buffer anyway.
1222 * - Lots of nasty vm tricks, that are neither fast nor flexible (it
1223 * has restriction limitations on both ends of the pipe).
1225 * Alas, it isn't here.
1228 static long do_vmsplice(struct file *file, const struct iovec __user *iov,
1229 unsigned long nr_segs, unsigned int flags)
1231 struct pipe_inode_info *pipe = file->f_dentry->d_inode->i_pipe;
1232 struct page *pages[PIPE_BUFFERS];
1233 struct partial_page partial[PIPE_BUFFERS];
1234 struct splice_pipe_desc spd = {
1238 .ops = &user_page_pipe_buf_ops,
1241 if (unlikely(!pipe))
1243 if (unlikely(nr_segs > UIO_MAXIOV))
1245 else if (unlikely(!nr_segs))
1248 spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1249 flags & SPLICE_F_GIFT);
1250 if (spd.nr_pages <= 0)
1251 return spd.nr_pages;
1253 return splice_to_pipe(pipe, &spd);
1256 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1257 unsigned long nr_segs, unsigned int flags)
1264 file = fget_light(fd, &fput);
1266 if (file->f_mode & FMODE_WRITE)
1267 error = do_vmsplice(file, iov, nr_segs, flags);
1269 fput_light(file, fput);
1275 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1276 int fd_out, loff_t __user *off_out,
1277 size_t len, unsigned int flags)
1280 struct file *in, *out;
1281 int fput_in, fput_out;
1287 in = fget_light(fd_in, &fput_in);
1289 if (in->f_mode & FMODE_READ) {
1290 out = fget_light(fd_out, &fput_out);
1292 if (out->f_mode & FMODE_WRITE)
1293 error = do_splice(in, off_in,
1296 fput_light(out, fput_out);
1300 fput_light(in, fput_in);
1307 * Link contents of ipipe to opipe.
1309 static int link_pipe(struct pipe_inode_info *ipipe,
1310 struct pipe_inode_info *opipe,
1311 size_t len, unsigned int flags)
1313 struct pipe_buffer *ibuf, *obuf;
1314 int ret, do_wakeup, i, ipipe_first;
1316 ret = do_wakeup = ipipe_first = 0;
1319 * Potential ABBA deadlock, work around it by ordering lock
1320 * grabbing by inode address. Otherwise two different processes
1321 * could deadlock (one doing tee from A -> B, the other from B -> A).
1323 if (ipipe->inode < opipe->inode) {
1325 mutex_lock(&ipipe->inode->i_mutex);
1326 mutex_lock(&opipe->inode->i_mutex);
1328 mutex_lock(&opipe->inode->i_mutex);
1329 mutex_lock(&ipipe->inode->i_mutex);
1333 if (!opipe->readers) {
1334 send_sig(SIGPIPE, current, 0);
1339 if (ipipe->nrbufs - i) {
1340 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1343 * If we have room, fill this buffer
1345 if (opipe->nrbufs < PIPE_BUFFERS) {
1346 int nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1349 * Get a reference to this pipe buffer,
1350 * so we can copy the contents over.
1352 ibuf->ops->get(ipipe, ibuf);
1354 obuf = opipe->bufs + nbuf;
1358 * Don't inherit the gift flag, we need to
1359 * prevent multiple steals of this page.
1361 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1363 if (obuf->len > len)
1373 if (opipe->nrbufs < PIPE_BUFFERS)
1378 * We have input available, but no output room.
1379 * If we already copied data, return that. If we
1380 * need to drop the opipe lock, it must be ordered
1381 * last to avoid deadlocks.
1383 if ((flags & SPLICE_F_NONBLOCK) || !ipipe_first) {
1388 if (signal_pending(current)) {
1395 if (waitqueue_active(&opipe->wait))
1396 wake_up_interruptible(&opipe->wait);
1397 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1401 opipe->waiting_writers++;
1403 opipe->waiting_writers--;
1408 * No input buffers, do the usual checks for available
1409 * writers and blocking and wait if necessary
1411 if (!ipipe->writers)
1413 if (!ipipe->waiting_writers) {
1418 * pipe_wait() drops the ipipe mutex. To avoid deadlocks
1419 * with another process, we can only safely do that if
1420 * the ipipe lock is ordered last.
1422 if ((flags & SPLICE_F_NONBLOCK) || ipipe_first) {
1427 if (signal_pending(current)) {
1433 if (waitqueue_active(&ipipe->wait))
1434 wake_up_interruptible_sync(&ipipe->wait);
1435 kill_fasync(&ipipe->fasync_writers, SIGIO, POLL_OUT);
1440 mutex_unlock(&ipipe->inode->i_mutex);
1441 mutex_unlock(&opipe->inode->i_mutex);
1445 if (waitqueue_active(&opipe->wait))
1446 wake_up_interruptible(&opipe->wait);
1447 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1454 * This is a tee(1) implementation that works on pipes. It doesn't copy
1455 * any data, it simply references the 'in' pages on the 'out' pipe.
1456 * The 'flags' used are the SPLICE_F_* variants, currently the only
1457 * applicable one is SPLICE_F_NONBLOCK.
1459 static long do_tee(struct file *in, struct file *out, size_t len,
1462 struct pipe_inode_info *ipipe = in->f_dentry->d_inode->i_pipe;
1463 struct pipe_inode_info *opipe = out->f_dentry->d_inode->i_pipe;
1466 * Link ipipe to the two output pipes, consuming as we go along.
1469 return link_pipe(ipipe, opipe, len, flags);
1474 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1483 in = fget_light(fdin, &fput_in);
1485 if (in->f_mode & FMODE_READ) {
1487 struct file *out = fget_light(fdout, &fput_out);
1490 if (out->f_mode & FMODE_WRITE)
1491 error = do_tee(in, out, len, flags);
1492 fput_light(out, fput_out);
1495 fput_light(in, fput_in);