Starting from 4.9-rc1 kernel, I started noticing some test failures
of sendfile(2) and splice(2) (sendfile0N and splice01 from LTP) when
testing on sub-page block size filesystems (tested both XFS and
ext4), these syscalls start to return EIO in the tests. e.g.
sendfile02 1 TFAIL : sendfile02.c:133: sendfile(2) failed to return
expected value, expected: 26, got: -1
sendfile02 2 TFAIL : sendfile02.c:133: sendfile(2) failed to return
expected value, expected: 24, got: -1
sendfile02 3 TFAIL : sendfile02.c:133: sendfile(2) failed to return
expected value, expected: 22, got: -1
sendfile02 4 TFAIL : sendfile02.c:133: sendfile(2) failed to return
expected value, expected: 20, got: -1
This is because that in sub-page block size cases, we don't need the
whole page to be uptodate, only the part we care about is uptodate
is OK (if fs has ->is_partially_uptodate defined). But
page_cache_pipe_buf_confirm() doesn't have the ability to check the
partially-uptodate case, it needs the whole page to be uptodate. So
it returns EIO in this case.
This is a regression introduced by commit 82c156f85384 ("switch
generic_file_splice_read() to use of ->read_iter()"). Prior to the
change, generic_file_splice_read() doesn't allow partially-uptodate
page either, so it worked fine.
Fix it by skipping the partially-uptodate check if we're working on
a pipe in do_generic_file_read(), so we read the whole page from
disk as long as the page is not uptodate.
Signed-off-by: Eryu Guan <guane...@gmail.com>
---
I think the other way to fix it is to add the ability to check & allow
partially-uptodate page to page_cache_pipe_buf_confirm(), but that is much
harder to do and seems gain little.
v2:
- Update summary a little bit
- Update commit log
- Add comment to the code
- Add more people/list to cc
v1: http://marc.info/?l=linux-mm&m=147756897431777&w=2
mm/filemap.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/mm/filemap.c b/mm/filemap.c
index 849f459..670264d 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -1734,6 +1734,9 @@ static ssize_t do_generic_file_read(struct file *filp,
loff_t *ppos,
if (inode->i_blkbits == PAGE_SHIFT ||
!mapping->a_ops->is_partially_uptodate)
goto page_not_up_to_date;
+ /* pipes can't handle partially uptodate pages */
+ if (unlikely(iter->type & ITER_PIPE))
+ goto page_not_up_to_date;
if (!trylock_page(page))
goto page_not_up_to_date;
/* Did it get truncated before we got the lock? */
--
2.7.4
