diff options
author | Suparna Bhattacharya <suparna@in.ibm.com> | 2005-05-21 16:33:36 -0700 |
---|---|---|
committer | Linus Torvalds <torvalds@ppc970.osdl.org> | 2005-05-21 16:45:24 -0700 |
commit | b5c44c2147a447f77e07fecdb087ae288e1f4e40 (patch) | |
tree | ec1d95eb1e03e320fc5eb5cfb40379f2f4a7267d | |
parent | f359b74c80bc76c1f6c2cb8f2837882f2335ba0c (diff) | |
download | linux-b5c44c2147a447f77e07fecdb087ae288e1f4e40.tar.gz linux-b5c44c2147a447f77e07fecdb087ae288e1f4e40.tar.bz2 linux-b5c44c2147a447f77e07fecdb087ae288e1f4e40.zip |
[PATCH] fix for __generic_file_aio_read() to return 0 on EOF
I came across the following problem while running ltp-aiodio testcases from
ltp-full-20050405 on linux-2.6.12-rc3-mm3. I tried running the tests with
EXT3 as well as JFS filesystems.
One or two fsx-linux testcases were hung after some time. These testcases
were hanging at wait_for_all_aios().
Debugging shows that there were some iocbs which were not getting completed
eventhough the last retry for those returned -EIOCBQUEUED. Also all such
pending iocbs represented READ operation.
Further debugging revealed that all such iocbs hit EOF in the DIO layer.
To be more precise, the "pos" from which they were trying to read was
greater than the "size" of the file. So the generic_file_direct_IO
returned 0.
This happens rarely as there is already a check in
__generic_file_aio_read(), for whether "pos" < "size" before calling direct
IO routine.
>size = i_size_read(inode);
>if (pos < size) {
> retval = generic_file_direct_IO(READ, iocb,
> iov, pos, nr_segs);
But for READ, we are taking the inode->i_sem only in the DIO layer. So it
is possible that some other process can change the size of the file before
we take the i_sem. In such a case ( when "pos" > "size"), the
__generic_file_aio_read() would return -EIOCBQUEUED even though there were
no I/O requests submitted by the DIO layer. This would cause the AIO layer
to expect aio_complete() for THE iocb, which doesnot happen. And thus the
test hangs forever, waiting for an I/O completion, where there are no
requests submitted at all.
The following patch makes __generic_file_aio_read() return 0 (instead of
returning -EIOCBQUEUED), on getting 0 from generic_file_direct_IO(), so
that the AIO layer does the aio_complete().
Testing:
I have tested the patch on a SMP machine(with 2 Pentium 4 (HT)) running
linux-2.6.12-rc3-mm3. I ran the ltp-aiodio testcases and none of the
fsx-linux tests hung. Also the aio-stress tests ran without any problem.
Signed-off-by: Suzuki K P <suzuki@in.ibm.com>
Signed-off-by: Suparna Bhattacharya <suparna@in.ibm.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
-rw-r--r-- | mm/filemap.c | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/mm/filemap.c b/mm/filemap.c index 47263ac3e4ea..1d33fec7bac6 100644 --- a/mm/filemap.c +++ b/mm/filemap.c @@ -1004,7 +1004,7 @@ __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov, if (pos < size) { retval = generic_file_direct_IO(READ, iocb, iov, pos, nr_segs); - if (retval >= 0 && !is_sync_kiocb(iocb)) + if (retval > 0 && !is_sync_kiocb(iocb)) retval = -EIOCBQUEUED; if (retval > 0) *ppos = pos + retval; |