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/buffer_head.h>
29
#include <linux/module.h>
30
#include <linux/syscalls.h>
31
#include <linux/uio.h>
32
#include <linux/security.h>
33
#include <linux/gfp.h>
36
* Attempt to steal a page from a pipe buffer. This should perhaps go into
37
* a vm helper function, it's already simplified quite a bit by the
38
* addition of remove_mapping(). If success is returned, the caller may
39
* attempt to reuse this page for another destination.
41
static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
42
struct pipe_buffer *buf)
44
struct page *page = buf->page;
45
struct address_space *mapping;
49
mapping = page_mapping(page);
51
WARN_ON(!PageUptodate(page));
54
* At least for ext2 with nobh option, we need to wait on
55
* writeback completing on this page, since we'll remove it
56
* from the pagecache. Otherwise truncate wont wait on the
57
* page, allowing the disk blocks to be reused by someone else
58
* before we actually wrote our data to them. fs corruption
61
wait_on_page_writeback(page);
63
if (page_has_private(page) &&
64
!try_to_release_page(page, GFP_KERNEL))
68
* If we succeeded in removing the mapping, set LRU flag
71
if (remove_mapping(mapping, page)) {
72
buf->flags |= PIPE_BUF_FLAG_LRU;
78
* Raced with truncate or failed to remove page from current
79
* address space, unlock and return failure.
86
static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
87
struct pipe_buffer *buf)
89
page_cache_release(buf->page);
90
buf->flags &= ~PIPE_BUF_FLAG_LRU;
94
* Check whether the contents of buf is OK to access. Since the content
95
* is a page cache page, IO may be in flight.
97
static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
98
struct pipe_buffer *buf)
100
struct page *page = buf->page;
103
if (!PageUptodate(page)) {
107
* Page got truncated/unhashed. This will cause a 0-byte
108
* splice, if this is the first page.
110
if (!page->mapping) {
116
* Uh oh, read-error from disk.
118
if (!PageUptodate(page)) {
124
* Page is ok afterall, we are done.
135
const struct pipe_buf_operations page_cache_pipe_buf_ops = {
137
.map = generic_pipe_buf_map,
138
.unmap = generic_pipe_buf_unmap,
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
.map = generic_pipe_buf_map,
158
.unmap = generic_pipe_buf_unmap,
159
.confirm = generic_pipe_buf_confirm,
160
.release = page_cache_pipe_buf_release,
161
.steal = user_page_pipe_buf_steal,
162
.get = generic_pipe_buf_get,
165
static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
168
if (waitqueue_active(&pipe->wait))
169
wake_up_interruptible(&pipe->wait);
170
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
174
* splice_to_pipe - fill passed data into a pipe
175
* @pipe: pipe to fill
179
* @spd contains a map of pages and len/offset tuples, along with
180
* the struct pipe_buf_operations associated with these pages. This
181
* function will link that data to the pipe.
184
ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
185
struct splice_pipe_desc *spd)
187
unsigned int spd_pages = spd->nr_pages;
188
int ret, do_wakeup, page_nr;
197
if (!pipe->readers) {
198
send_sig(SIGPIPE, current, 0);
204
if (pipe->nrbufs < pipe->buffers) {
205
int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
206
struct pipe_buffer *buf = pipe->bufs + newbuf;
208
buf->page = spd->pages[page_nr];
209
buf->offset = spd->partial[page_nr].offset;
210
buf->len = spd->partial[page_nr].len;
211
buf->private = spd->partial[page_nr].private;
213
if (spd->flags & SPLICE_F_GIFT)
214
buf->flags |= PIPE_BUF_FLAG_GIFT;
223
if (!--spd->nr_pages)
225
if (pipe->nrbufs < pipe->buffers)
231
if (spd->flags & SPLICE_F_NONBLOCK) {
237
if (signal_pending(current)) {
245
if (waitqueue_active(&pipe->wait))
246
wake_up_interruptible_sync(&pipe->wait);
247
kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
251
pipe->waiting_writers++;
253
pipe->waiting_writers--;
259
wakeup_pipe_readers(pipe);
261
while (page_nr < spd_pages)
262
spd->spd_release(spd, page_nr++);
267
void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
269
page_cache_release(spd->pages[i]);
273
* Check if we need to grow the arrays holding pages and partial page
276
int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
278
if (pipe->buffers <= PIPE_DEF_BUFFERS)
281
spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
282
spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
284
if (spd->pages && spd->partial)
292
void splice_shrink_spd(struct pipe_inode_info *pipe,
293
struct splice_pipe_desc *spd)
295
if (pipe->buffers <= PIPE_DEF_BUFFERS)
303
__generic_file_splice_read(struct file *in, loff_t *ppos,
304
struct pipe_inode_info *pipe, size_t len,
307
struct address_space *mapping = in->f_mapping;
308
unsigned int loff, nr_pages, req_pages;
309
struct page *pages[PIPE_DEF_BUFFERS];
310
struct partial_page partial[PIPE_DEF_BUFFERS];
312
pgoff_t index, end_index;
315
struct splice_pipe_desc spd = {
319
.ops = &page_cache_pipe_buf_ops,
320
.spd_release = spd_release_page,
323
if (splice_grow_spd(pipe, &spd))
326
index = *ppos >> PAGE_CACHE_SHIFT;
327
loff = *ppos & ~PAGE_CACHE_MASK;
328
req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
329
nr_pages = min(req_pages, pipe->buffers);
332
* Lookup the (hopefully) full range of pages we need.
334
spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
335
index += spd.nr_pages;
338
* If find_get_pages_contig() returned fewer pages than we needed,
339
* readahead/allocate the rest and fill in the holes.
341
if (spd.nr_pages < nr_pages)
342
page_cache_sync_readahead(mapping, &in->f_ra, in,
343
index, req_pages - spd.nr_pages);
346
while (spd.nr_pages < nr_pages) {
348
* Page could be there, find_get_pages_contig() breaks on
351
page = find_get_page(mapping, index);
354
* page didn't exist, allocate one.
356
page = page_cache_alloc_cold(mapping);
360
error = add_to_page_cache_lru(page, mapping, index,
362
if (unlikely(error)) {
363
page_cache_release(page);
364
if (error == -EEXIST)
369
* add_to_page_cache() locks the page, unlock it
370
* to avoid convoluting the logic below even more.
375
spd.pages[spd.nr_pages++] = page;
380
* Now loop over the map and see if we need to start IO on any
381
* pages, fill in the partial map, etc.
383
index = *ppos >> PAGE_CACHE_SHIFT;
384
nr_pages = spd.nr_pages;
386
for (page_nr = 0; page_nr < nr_pages; page_nr++) {
387
unsigned int this_len;
393
* this_len is the max we'll use from this page
395
this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
396
page = spd.pages[page_nr];
398
if (PageReadahead(page))
399
page_cache_async_readahead(mapping, &in->f_ra, in,
400
page, index, req_pages - page_nr);
403
* If the page isn't uptodate, we may need to start io on it
405
if (!PageUptodate(page)) {
409
* Page was truncated, or invalidated by the
410
* filesystem. Redo the find/create, but this time the
411
* page is kept locked, so there's no chance of another
412
* race with truncate/invalidate.
414
if (!page->mapping) {
416
page = find_or_create_page(mapping, index,
417
mapping_gfp_mask(mapping));
423
page_cache_release(spd.pages[page_nr]);
424
spd.pages[page_nr] = page;
427
* page was already under io and is now done, great
429
if (PageUptodate(page)) {
435
* need to read in the page
437
error = mapping->a_ops->readpage(in, page);
438
if (unlikely(error)) {
440
* We really should re-lookup the page here,
441
* but it complicates things a lot. Instead
442
* lets just do what we already stored, and
443
* we'll get it the next time we are called.
445
if (error == AOP_TRUNCATED_PAGE)
453
* i_size must be checked after PageUptodate.
455
isize = i_size_read(mapping->host);
456
end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
457
if (unlikely(!isize || index > end_index))
461
* if this is the last page, see if we need to shrink
462
* the length and stop
464
if (end_index == index) {
468
* max good bytes in this page
470
plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
475
* force quit after adding this page
477
this_len = min(this_len, plen - loff);
481
spd.partial[page_nr].offset = loff;
482
spd.partial[page_nr].len = this_len;
490
* Release any pages at the end, if we quit early. 'page_nr' is how far
491
* we got, 'nr_pages' is how many pages are in the map.
493
while (page_nr < nr_pages)
494
page_cache_release(spd.pages[page_nr++]);
495
in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
498
error = splice_to_pipe(pipe, &spd);
500
splice_shrink_spd(pipe, &spd);
505
* generic_file_splice_read - splice data from file to a pipe
506
* @in: file to splice from
507
* @ppos: position in @in
508
* @pipe: pipe to splice to
509
* @len: number of bytes to splice
510
* @flags: splice modifier flags
513
* Will read pages from given file and fill them into a pipe. Can be
514
* used as long as the address_space operations for the source implements
518
ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
519
struct pipe_inode_info *pipe, size_t len,
525
isize = i_size_read(in->f_mapping->host);
526
if (unlikely(*ppos >= isize))
529
left = isize - *ppos;
530
if (unlikely(left < len))
533
ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
541
EXPORT_SYMBOL(generic_file_splice_read);
543
static const struct pipe_buf_operations default_pipe_buf_ops = {
545
.map = generic_pipe_buf_map,
546
.unmap = generic_pipe_buf_unmap,
547
.confirm = generic_pipe_buf_confirm,
548
.release = generic_pipe_buf_release,
549
.steal = generic_pipe_buf_steal,
550
.get = generic_pipe_buf_get,
553
static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
554
unsigned long vlen, loff_t offset)
562
/* The cast to a user pointer is valid due to the set_fs() */
563
res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
569
static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
577
/* The cast to a user pointer is valid due to the set_fs() */
578
res = vfs_write(file, (const char __user *)buf, count, &pos);
584
ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
585
struct pipe_inode_info *pipe, size_t len,
588
unsigned int nr_pages;
589
unsigned int nr_freed;
591
struct page *pages[PIPE_DEF_BUFFERS];
592
struct partial_page partial[PIPE_DEF_BUFFERS];
593
struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
598
struct splice_pipe_desc spd = {
602
.ops = &default_pipe_buf_ops,
603
.spd_release = spd_release_page,
606
if (splice_grow_spd(pipe, &spd))
611
if (pipe->buffers > PIPE_DEF_BUFFERS) {
612
vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
617
offset = *ppos & ~PAGE_CACHE_MASK;
618
nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
620
for (i = 0; i < nr_pages && i < pipe->buffers && len; i++) {
623
page = alloc_page(GFP_USER);
628
this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
629
vec[i].iov_base = (void __user *) page_address(page);
630
vec[i].iov_len = this_len;
637
res = kernel_readv(in, vec, spd.nr_pages, *ppos);
648
for (i = 0; i < spd.nr_pages; i++) {
649
this_len = min_t(size_t, vec[i].iov_len, res);
650
spd.partial[i].offset = 0;
651
spd.partial[i].len = this_len;
653
__free_page(spd.pages[i]);
659
spd.nr_pages -= nr_freed;
661
res = splice_to_pipe(pipe, &spd);
668
splice_shrink_spd(pipe, &spd);
672
for (i = 0; i < spd.nr_pages; i++)
673
__free_page(spd.pages[i]);
678
EXPORT_SYMBOL(default_file_splice_read);
681
* Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
682
* using sendpage(). Return the number of bytes sent.
684
static int pipe_to_sendpage(struct pipe_inode_info *pipe,
685
struct pipe_buffer *buf, struct splice_desc *sd)
687
struct file *file = sd->u.file;
688
loff_t pos = sd->pos;
691
if (!likely(file->f_op && file->f_op->sendpage))
694
more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
695
return file->f_op->sendpage(file, buf->page, buf->offset,
696
sd->len, &pos, more);
700
* This is a little more tricky than the file -> pipe splicing. There are
701
* basically three cases:
703
* - Destination page already exists in the address space and there
704
* are users of it. For that case we have no other option that
705
* copying the data. Tough luck.
706
* - Destination page already exists in the address space, but there
707
* are no users of it. Make sure it's uptodate, then drop it. Fall
708
* through to last case.
709
* - Destination page does not exist, we can add the pipe page to
710
* the page cache and avoid the copy.
712
* If asked to move pages to the output file (SPLICE_F_MOVE is set in
713
* sd->flags), we attempt to migrate pages from the pipe to the output
714
* file address space page cache. This is possible if no one else has
715
* the pipe page referenced outside of the pipe and page cache. If
716
* SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
717
* a new page in the output file page cache and fill/dirty that.
719
int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
720
struct splice_desc *sd)
722
struct file *file = sd->u.file;
723
struct address_space *mapping = file->f_mapping;
724
unsigned int offset, this_len;
729
offset = sd->pos & ~PAGE_CACHE_MASK;
732
if (this_len + offset > PAGE_CACHE_SIZE)
733
this_len = PAGE_CACHE_SIZE - offset;
735
ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
736
AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
740
if (buf->page != page) {
742
* Careful, ->map() uses KM_USER0!
744
char *src = buf->ops->map(pipe, buf, 1);
745
char *dst = kmap_atomic(page, KM_USER1);
747
memcpy(dst + offset, src + buf->offset, this_len);
748
flush_dcache_page(page);
749
kunmap_atomic(dst, KM_USER1);
750
buf->ops->unmap(pipe, buf, src);
752
ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
757
EXPORT_SYMBOL(pipe_to_file);
759
static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
762
if (waitqueue_active(&pipe->wait))
763
wake_up_interruptible(&pipe->wait);
764
kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
768
* splice_from_pipe_feed - feed available data from a pipe to a file
769
* @pipe: pipe to splice from
770
* @sd: information to @actor
771
* @actor: handler that splices the data
774
* This function loops over the pipe and calls @actor to do the
775
* actual moving of a single struct pipe_buffer to the desired
776
* destination. It returns when there's no more buffers left in
777
* the pipe or if the requested number of bytes (@sd->total_len)
778
* have been copied. It returns a positive number (one) if the
779
* pipe needs to be filled with more data, zero if the required
780
* number of bytes have been copied and -errno on error.
782
* This, together with splice_from_pipe_{begin,end,next}, may be
783
* used to implement the functionality of __splice_from_pipe() when
784
* locking is required around copying the pipe buffers to the
787
int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
792
while (pipe->nrbufs) {
793
struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
794
const struct pipe_buf_operations *ops = buf->ops;
797
if (sd->len > sd->total_len)
798
sd->len = sd->total_len;
800
ret = buf->ops->confirm(pipe, buf);
807
ret = actor(pipe, buf, sd);
814
sd->num_spliced += ret;
817
sd->total_len -= ret;
821
ops->release(pipe, buf);
822
pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
825
sd->need_wakeup = true;
834
EXPORT_SYMBOL(splice_from_pipe_feed);
837
* splice_from_pipe_next - wait for some data to splice from
838
* @pipe: pipe to splice from
839
* @sd: information about the splice operation
842
* This function will wait for some data and return a positive
843
* value (one) if pipe buffers are available. It will return zero
844
* or -errno if no more data needs to be spliced.
846
int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
848
while (!pipe->nrbufs) {
852
if (!pipe->waiting_writers && sd->num_spliced)
855
if (sd->flags & SPLICE_F_NONBLOCK)
858
if (signal_pending(current))
861
if (sd->need_wakeup) {
862
wakeup_pipe_writers(pipe);
863
sd->need_wakeup = false;
871
EXPORT_SYMBOL(splice_from_pipe_next);
874
* splice_from_pipe_begin - start splicing from pipe
875
* @sd: information about the splice operation
878
* This function should be called before a loop containing
879
* splice_from_pipe_next() and splice_from_pipe_feed() to
880
* initialize the necessary fields of @sd.
882
void splice_from_pipe_begin(struct splice_desc *sd)
885
sd->need_wakeup = false;
887
EXPORT_SYMBOL(splice_from_pipe_begin);
890
* splice_from_pipe_end - finish splicing from pipe
891
* @pipe: pipe to splice from
892
* @sd: information about the splice operation
895
* This function will wake up pipe writers if necessary. It should
896
* be called after a loop containing splice_from_pipe_next() and
897
* splice_from_pipe_feed().
899
void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
902
wakeup_pipe_writers(pipe);
904
EXPORT_SYMBOL(splice_from_pipe_end);
907
* __splice_from_pipe - splice data from a pipe to given actor
908
* @pipe: pipe to splice from
909
* @sd: information to @actor
910
* @actor: handler that splices the data
913
* This function does little more than loop over the pipe and call
914
* @actor to do the actual moving of a single struct pipe_buffer to
915
* the desired destination. See pipe_to_file, pipe_to_sendpage, or
919
ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
924
splice_from_pipe_begin(sd);
926
ret = splice_from_pipe_next(pipe, sd);
928
ret = splice_from_pipe_feed(pipe, sd, actor);
930
splice_from_pipe_end(pipe, sd);
932
return sd->num_spliced ? sd->num_spliced : ret;
934
EXPORT_SYMBOL(__splice_from_pipe);
937
* splice_from_pipe - splice data from a pipe to a file
938
* @pipe: pipe to splice from
939
* @out: file to splice to
940
* @ppos: position in @out
941
* @len: how many bytes to splice
942
* @flags: splice modifier flags
943
* @actor: handler that splices the data
946
* See __splice_from_pipe. This function locks the pipe inode,
947
* otherwise it's identical to __splice_from_pipe().
950
ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
951
loff_t *ppos, size_t len, unsigned int flags,
955
struct splice_desc sd = {
963
ret = __splice_from_pipe(pipe, &sd, actor);
970
* generic_file_splice_write - splice data from a pipe to a file
972
* @out: file to write to
973
* @ppos: position in @out
974
* @len: number of bytes to splice
975
* @flags: splice modifier flags
978
* Will either move or copy pages (determined by @flags options) from
979
* the given pipe inode to the given file.
983
generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
984
loff_t *ppos, size_t len, unsigned int flags)
986
struct address_space *mapping = out->f_mapping;
987
struct inode *inode = mapping->host;
988
struct splice_desc sd = {
998
splice_from_pipe_begin(&sd);
1000
ret = splice_from_pipe_next(pipe, &sd);
1004
mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
1005
ret = file_remove_suid(out);
1007
file_update_time(out);
1008
ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
1010
mutex_unlock(&inode->i_mutex);
1012
splice_from_pipe_end(pipe, &sd);
1017
ret = sd.num_spliced;
1020
unsigned long nr_pages;
1023
nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1025
err = generic_write_sync(out, *ppos, ret);
1030
balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
1036
EXPORT_SYMBOL(generic_file_splice_write);
1038
static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1039
struct splice_desc *sd)
1044
data = buf->ops->map(pipe, buf, 0);
1045
ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
1046
buf->ops->unmap(pipe, buf, data);
1051
static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
1052
struct file *out, loff_t *ppos,
1053
size_t len, unsigned int flags)
1057
ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
1065
* generic_splice_sendpage - splice data from a pipe to a socket
1066
* @pipe: pipe to splice from
1067
* @out: socket to write to
1068
* @ppos: position in @out
1069
* @len: number of bytes to splice
1070
* @flags: splice modifier flags
1073
* Will send @len bytes from the pipe to a network socket. No data copying
1077
ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
1078
loff_t *ppos, size_t len, unsigned int flags)
1080
return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
1083
EXPORT_SYMBOL(generic_splice_sendpage);
1086
* Attempt to initiate a splice from pipe to file.
1088
static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
1089
loff_t *ppos, size_t len, unsigned int flags)
1091
ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
1092
loff_t *, size_t, unsigned int);
1095
if (unlikely(!(out->f_mode & FMODE_WRITE)))
1098
if (unlikely(out->f_flags & O_APPEND))
1101
ret = rw_verify_area(WRITE, out, ppos, len);
1102
if (unlikely(ret < 0))
1105
if (out->f_op && out->f_op->splice_write)
1106
splice_write = out->f_op->splice_write;
1108
splice_write = default_file_splice_write;
1110
return splice_write(pipe, out, ppos, len, flags);
1114
* Attempt to initiate a splice from a file to a pipe.
1116
static long do_splice_to(struct file *in, loff_t *ppos,
1117
struct pipe_inode_info *pipe, size_t len,
1120
ssize_t (*splice_read)(struct file *, loff_t *,
1121
struct pipe_inode_info *, size_t, unsigned int);
1124
if (unlikely(!(in->f_mode & FMODE_READ)))
1127
ret = rw_verify_area(READ, in, ppos, len);
1128
if (unlikely(ret < 0))
1131
if (in->f_op && in->f_op->splice_read)
1132
splice_read = in->f_op->splice_read;
1134
splice_read = default_file_splice_read;
1136
return splice_read(in, ppos, pipe, len, flags);
1140
* splice_direct_to_actor - splices data directly between two non-pipes
1141
* @in: file to splice from
1142
* @sd: actor information on where to splice to
1143
* @actor: handles the data splicing
1146
* This is a special case helper to splice directly between two
1147
* points, without requiring an explicit pipe. Internally an allocated
1148
* pipe is cached in the process, and reused during the lifetime of
1152
ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1153
splice_direct_actor *actor)
1155
struct pipe_inode_info *pipe;
1162
* We require the input being a regular file, as we don't want to
1163
* randomly drop data for eg socket -> socket splicing. Use the
1164
* piped splicing for that!
1166
i_mode = in->f_path.dentry->d_inode->i_mode;
1167
if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1171
* neither in nor out is a pipe, setup an internal pipe attached to
1172
* 'out' and transfer the wanted data from 'in' to 'out' through that
1174
pipe = current->splice_pipe;
1175
if (unlikely(!pipe)) {
1176
pipe = alloc_pipe_info(NULL);
1181
* We don't have an immediate reader, but we'll read the stuff
1182
* out of the pipe right after the splice_to_pipe(). So set
1183
* PIPE_READERS appropriately.
1187
current->splice_pipe = pipe;
1195
len = sd->total_len;
1199
* Don't block on output, we have to drain the direct pipe.
1201
sd->flags &= ~SPLICE_F_NONBLOCK;
1205
loff_t pos = sd->pos, prev_pos = pos;
1207
ret = do_splice_to(in, &pos, pipe, len, flags);
1208
if (unlikely(ret <= 0))
1212
sd->total_len = read_len;
1215
* NOTE: nonblocking mode only applies to the input. We
1216
* must not do the output in nonblocking mode as then we
1217
* could get stuck data in the internal pipe:
1219
ret = actor(pipe, sd);
1220
if (unlikely(ret <= 0)) {
1229
if (ret < read_len) {
1230
sd->pos = prev_pos + ret;
1236
pipe->nrbufs = pipe->curbuf = 0;
1242
* If we did an incomplete transfer we must release
1243
* the pipe buffers in question:
1245
for (i = 0; i < pipe->buffers; i++) {
1246
struct pipe_buffer *buf = pipe->bufs + i;
1249
buf->ops->release(pipe, buf);
1259
EXPORT_SYMBOL(splice_direct_to_actor);
1261
static int direct_splice_actor(struct pipe_inode_info *pipe,
1262
struct splice_desc *sd)
1264
struct file *file = sd->u.file;
1266
return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
1271
* do_splice_direct - splices data directly between two files
1272
* @in: file to splice from
1273
* @ppos: input file offset
1274
* @out: file to splice to
1275
* @len: number of bytes to splice
1276
* @flags: splice modifier flags
1279
* For use by do_sendfile(). splice can easily emulate sendfile, but
1280
* doing it in the application would incur an extra system call
1281
* (splice in + splice out, as compared to just sendfile()). So this helper
1282
* can splice directly through a process-private pipe.
1285
long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1286
size_t len, unsigned int flags)
1288
struct splice_desc sd = {
1297
ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1304
static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1305
struct pipe_inode_info *opipe,
1306
size_t len, unsigned int flags);
1309
* Determine where to splice to/from.
1311
static long do_splice(struct file *in, loff_t __user *off_in,
1312
struct file *out, loff_t __user *off_out,
1313
size_t len, unsigned int flags)
1315
struct pipe_inode_info *ipipe;
1316
struct pipe_inode_info *opipe;
1317
loff_t offset, *off;
1320
ipipe = get_pipe_info(in);
1321
opipe = get_pipe_info(out);
1323
if (ipipe && opipe) {
1324
if (off_in || off_out)
1327
if (!(in->f_mode & FMODE_READ))
1330
if (!(out->f_mode & FMODE_WRITE))
1333
/* Splicing to self would be fun, but... */
1337
return splice_pipe_to_pipe(ipipe, opipe, len, flags);
1344
if (!(out->f_mode & FMODE_PWRITE))
1346
if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1352
ret = do_splice_from(ipipe, out, off, len, flags);
1354
if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1364
if (!(in->f_mode & FMODE_PREAD))
1366
if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1372
ret = do_splice_to(in, off, opipe, len, flags);
1374
if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1384
* Map an iov into an array of pages and offset/length tupples. With the
1385
* partial_page structure, we can map several non-contiguous ranges into
1386
* our ones pages[] map instead of splitting that operation into pieces.
1387
* Could easily be exported as a generic helper for other users, in which
1388
* case one would probably want to add a 'max_nr_pages' parameter as well.
1390
static int get_iovec_page_array(const struct iovec __user *iov,
1391
unsigned int nr_vecs, struct page **pages,
1392
struct partial_page *partial, int aligned,
1393
unsigned int pipe_buffers)
1395
int buffers = 0, error = 0;
1398
unsigned long off, npages;
1405
if (copy_from_user(&entry, iov, sizeof(entry)))
1408
base = entry.iov_base;
1409
len = entry.iov_len;
1412
* Sanity check this iovec. 0 read succeeds.
1418
if (!access_ok(VERIFY_READ, base, len))
1422
* Get this base offset and number of pages, then map
1423
* in the user pages.
1425
off = (unsigned long) base & ~PAGE_MASK;
1428
* If asked for alignment, the offset must be zero and the
1429
* length a multiple of the PAGE_SIZE.
1432
if (aligned && (off || len & ~PAGE_MASK))
1435
npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1436
if (npages > pipe_buffers - buffers)
1437
npages = pipe_buffers - buffers;
1439
error = get_user_pages_fast((unsigned long)base, npages,
1440
0, &pages[buffers]);
1442
if (unlikely(error <= 0))
1446
* Fill this contiguous range into the partial page map.
1448
for (i = 0; i < error; i++) {
1449
const int plen = min_t(size_t, len, PAGE_SIZE - off);
1451
partial[buffers].offset = off;
1452
partial[buffers].len = plen;
1460
* We didn't complete this iov, stop here since it probably
1461
* means we have to move some of this into a pipe to
1462
* be able to continue.
1468
* Don't continue if we mapped fewer pages than we asked for,
1469
* or if we mapped the max number of pages that we have
1472
if (error < npages || buffers == pipe_buffers)
1485
static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1486
struct splice_desc *sd)
1492
* See if we can use the atomic maps, by prefaulting in the
1493
* pages and doing an atomic copy
1495
if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1496
src = buf->ops->map(pipe, buf, 1);
1497
ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1499
buf->ops->unmap(pipe, buf, src);
1507
* No dice, use slow non-atomic map and copy
1509
src = buf->ops->map(pipe, buf, 0);
1512
if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1515
buf->ops->unmap(pipe, buf, src);
1518
sd->u.userptr += ret;
1523
* For lack of a better implementation, implement vmsplice() to userspace
1524
* as a simple copy of the pipes pages to the user iov.
1526
static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1527
unsigned long nr_segs, unsigned int flags)
1529
struct pipe_inode_info *pipe;
1530
struct splice_desc sd;
1535
pipe = get_pipe_info(file);
1547
* Get user address base and length for this iovec.
1549
error = get_user(base, &iov->iov_base);
1550
if (unlikely(error))
1552
error = get_user(len, &iov->iov_len);
1553
if (unlikely(error))
1557
* Sanity check this iovec. 0 read succeeds.
1561
if (unlikely(!base)) {
1566
if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
1574
sd.u.userptr = base;
1577
size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1603
* vmsplice splices a user address range into a pipe. It can be thought of
1604
* as splice-from-memory, where the regular splice is splice-from-file (or
1605
* to file). In both cases the output is a pipe, naturally.
1607
static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1608
unsigned long nr_segs, unsigned int flags)
1610
struct pipe_inode_info *pipe;
1611
struct page *pages[PIPE_DEF_BUFFERS];
1612
struct partial_page partial[PIPE_DEF_BUFFERS];
1613
struct splice_pipe_desc spd = {
1617
.ops = &user_page_pipe_buf_ops,
1618
.spd_release = spd_release_page,
1622
pipe = get_pipe_info(file);
1626
if (splice_grow_spd(pipe, &spd))
1629
spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
1630
spd.partial, flags & SPLICE_F_GIFT,
1632
if (spd.nr_pages <= 0)
1635
ret = splice_to_pipe(pipe, &spd);
1637
splice_shrink_spd(pipe, &spd);
1642
* Note that vmsplice only really supports true splicing _from_ user memory
1643
* to a pipe, not the other way around. Splicing from user memory is a simple
1644
* operation that can be supported without any funky alignment restrictions
1645
* or nasty vm tricks. We simply map in the user memory and fill them into
1646
* a pipe. The reverse isn't quite as easy, though. There are two possible
1647
* solutions for that:
1649
* - memcpy() the data internally, at which point we might as well just
1650
* do a regular read() on the buffer anyway.
1651
* - Lots of nasty vm tricks, that are neither fast nor flexible (it
1652
* has restriction limitations on both ends of the pipe).
1654
* Currently we punt and implement it as a normal copy, see pipe_to_user().
1657
SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
1658
unsigned long, nr_segs, unsigned int, flags)
1664
if (unlikely(nr_segs > UIO_MAXIOV))
1666
else if (unlikely(!nr_segs))
1670
file = fget_light(fd, &fput);
1672
if (file->f_mode & FMODE_WRITE)
1673
error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1674
else if (file->f_mode & FMODE_READ)
1675
error = vmsplice_to_user(file, iov, nr_segs, flags);
1677
fput_light(file, fput);
1683
SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
1684
int, fd_out, loff_t __user *, off_out,
1685
size_t, len, unsigned int, flags)
1688
struct file *in, *out;
1689
int fput_in, fput_out;
1695
in = fget_light(fd_in, &fput_in);
1697
if (in->f_mode & FMODE_READ) {
1698
out = fget_light(fd_out, &fput_out);
1700
if (out->f_mode & FMODE_WRITE)
1701
error = do_splice(in, off_in,
1704
fput_light(out, fput_out);
1708
fput_light(in, fput_in);
1715
* Make sure there's data to read. Wait for input if we can, otherwise
1716
* return an appropriate error.
1718
static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1723
* Check ->nrbufs without the inode lock first. This function
1724
* is speculative anyways, so missing one is ok.
1732
while (!pipe->nrbufs) {
1733
if (signal_pending(current)) {
1739
if (!pipe->waiting_writers) {
1740
if (flags & SPLICE_F_NONBLOCK) {
1753
* Make sure there's writeable room. Wait for room if we can, otherwise
1754
* return an appropriate error.
1756
static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1761
* Check ->nrbufs without the inode lock first. This function
1762
* is speculative anyways, so missing one is ok.
1764
if (pipe->nrbufs < pipe->buffers)
1770
while (pipe->nrbufs >= pipe->buffers) {
1771
if (!pipe->readers) {
1772
send_sig(SIGPIPE, current, 0);
1776
if (flags & SPLICE_F_NONBLOCK) {
1780
if (signal_pending(current)) {
1784
pipe->waiting_writers++;
1786
pipe->waiting_writers--;
1794
* Splice contents of ipipe to opipe.
1796
static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
1797
struct pipe_inode_info *opipe,
1798
size_t len, unsigned int flags)
1800
struct pipe_buffer *ibuf, *obuf;
1802
bool input_wakeup = false;
1806
ret = ipipe_prep(ipipe, flags);
1810
ret = opipe_prep(opipe, flags);
1815
* Potential ABBA deadlock, work around it by ordering lock
1816
* grabbing by pipe info address. Otherwise two different processes
1817
* could deadlock (one doing tee from A -> B, the other from B -> A).
1819
pipe_double_lock(ipipe, opipe);
1822
if (!opipe->readers) {
1823
send_sig(SIGPIPE, current, 0);
1829
if (!ipipe->nrbufs && !ipipe->writers)
1833
* Cannot make any progress, because either the input
1834
* pipe is empty or the output pipe is full.
1836
if (!ipipe->nrbufs || opipe->nrbufs >= opipe->buffers) {
1837
/* Already processed some buffers, break */
1841
if (flags & SPLICE_F_NONBLOCK) {
1847
* We raced with another reader/writer and haven't
1848
* managed to process any buffers. A zero return
1849
* value means EOF, so retry instead.
1856
ibuf = ipipe->bufs + ipipe->curbuf;
1857
nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1858
obuf = opipe->bufs + nbuf;
1860
if (len >= ibuf->len) {
1862
* Simply move the whole buffer from ipipe to opipe
1867
ipipe->curbuf = (ipipe->curbuf + 1) & (ipipe->buffers - 1);
1869
input_wakeup = true;
1872
* Get a reference to this pipe buffer,
1873
* so we can copy the contents over.
1875
ibuf->ops->get(ipipe, ibuf);
1879
* Don't inherit the gift flag, we need to
1880
* prevent multiple steals of this page.
1882
obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1886
ibuf->offset += obuf->len;
1887
ibuf->len -= obuf->len;
1897
* If we put data in the output pipe, wakeup any potential readers.
1900
wakeup_pipe_readers(opipe);
1903
wakeup_pipe_writers(ipipe);
1909
* Link contents of ipipe to opipe.
1911
static int link_pipe(struct pipe_inode_info *ipipe,
1912
struct pipe_inode_info *opipe,
1913
size_t len, unsigned int flags)
1915
struct pipe_buffer *ibuf, *obuf;
1916
int ret = 0, i = 0, nbuf;
1919
* Potential ABBA deadlock, work around it by ordering lock
1920
* grabbing by pipe info address. Otherwise two different processes
1921
* could deadlock (one doing tee from A -> B, the other from B -> A).
1923
pipe_double_lock(ipipe, opipe);
1926
if (!opipe->readers) {
1927
send_sig(SIGPIPE, current, 0);
1934
* If we have iterated all input buffers or ran out of
1935
* output room, break.
1937
if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
1940
ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
1941
nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
1944
* Get a reference to this pipe buffer,
1945
* so we can copy the contents over.
1947
ibuf->ops->get(ipipe, ibuf);
1949
obuf = opipe->bufs + nbuf;
1953
* Don't inherit the gift flag, we need to
1954
* prevent multiple steals of this page.
1956
obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1958
if (obuf->len > len)
1968
* return EAGAIN if we have the potential of some data in the
1969
* future, otherwise just return 0
1971
if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
1978
* If we put data in the output pipe, wakeup any potential readers.
1981
wakeup_pipe_readers(opipe);
1987
* This is a tee(1) implementation that works on pipes. It doesn't copy
1988
* any data, it simply references the 'in' pages on the 'out' pipe.
1989
* The 'flags' used are the SPLICE_F_* variants, currently the only
1990
* applicable one is SPLICE_F_NONBLOCK.
1992
static long do_tee(struct file *in, struct file *out, size_t len,
1995
struct pipe_inode_info *ipipe = get_pipe_info(in);
1996
struct pipe_inode_info *opipe = get_pipe_info(out);
2000
* Duplicate the contents of ipipe to opipe without actually
2003
if (ipipe && opipe && ipipe != opipe) {
2005
* Keep going, unless we encounter an error. The ipipe/opipe
2006
* ordering doesn't really matter.
2008
ret = ipipe_prep(ipipe, flags);
2010
ret = opipe_prep(opipe, flags);
2012
ret = link_pipe(ipipe, opipe, len, flags);
2019
SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
2028
in = fget_light(fdin, &fput_in);
2030
if (in->f_mode & FMODE_READ) {
2032
struct file *out = fget_light(fdout, &fput_out);
2035
if (out->f_mode & FMODE_WRITE)
2036
error = do_tee(in, out, len, flags);
2037
fput_light(out, fput_out);
2040
fput_light(in, fput_in);