v4.4-rc4 regression with direct-io

v4.4-rc4 regression with direct-io

[Ross Zwisler]

Hey Jan,

In my test setup this commit:

74cedf9b6c60 ("direct-io: Fix negative return from dio read beyond eof")

causes a regression with XFS and the test generic/113.  The test used to
pass, but with the above mentioned commit it just hangs indefinitely.

This commit went in between v4.4-rc3 and v4.4-rc4, and it looks like it was
marked for stable?

I've run "echo w > /proc/sysrq-trigger", and it doesn't list any locked
processes so I don't believe that this is a deadlock.  It looks like a
livelock?

My test setup is a pair of 4GiB PMEM partitions (one as the test device and
one as the scratch) in a qemu virtual machine.  This test did not involve DAX,
and turning DAX on makes the test pass once again.

Please let me know if you need any more info or have trouble reproducing the
issue.  I'm happy to test patches.

Thanks,
- Ross

Re: v4.4-rc4 regression with direct-io

[Al Viro]

On Tue, Dec 08, 2015 at 02:39:02PM -0700, Ross Zwisler wrote:
> Hey Jan,
> 
> In my test setup this commit:
> 
> 74cedf9b6c60 ("direct-io: Fix negative return from dio read beyond eof")
> 
> causes a regression with XFS and the test generic/113.  The test used to
> pass, but with the above mentioned commit it just hangs indefinitely.
> 
> This commit went in between v4.4-rc3 and v4.4-rc4, and it looks like it was
> marked for stable?

See the tip of vfs.git#for-linus for fix...

Re: v4.4-rc4 regression with direct-io

[Ross Zwisler]

On Tue, Dec 08, 2015 at 09:51:09PM +0000, Al Viro wrote:
> On Tue, Dec 08, 2015 at 02:39:02PM -0700, Ross Zwisler wrote:
> > Hey Jan,
> > 
> > In my test setup this commit:
> > 
> > 74cedf9b6c60 ("direct-io: Fix negative return from dio read beyond eof")
> > 
> > causes a regression with XFS and the test generic/113.  The test used to
> > pass, but with the above mentioned commit it just hangs indefinitely.
> > 
> > This commit went in between v4.4-rc3 and v4.4-rc4, and it looks like it was
> > marked for stable?
> 
> See the tip of vfs.git#for-linus for fix...

I've confirmed that this fixes my test case.  Thanks!

Re: v4.4-rc4 regression with direct-io

[Jan Kara]

On Tue 08-12-15 21:51:09, Al Viro wrote:
> On Tue, Dec 08, 2015 at 02:39:02PM -0700, Ross Zwisler wrote:
> > Hey Jan,
> > 
> > In my test setup this commit:
> > 
> > 74cedf9b6c60 ("direct-io: Fix negative return from dio read beyond eof")
> > 
> > causes a regression with XFS and the test generic/113.  The test used to
> > pass, but with the above mentioned commit it just hangs indefinitely.
> > 
> > This commit went in between v4.4-rc3 and v4.4-rc4, and it looks like it was
> > marked for stable?
> 
> See the tip of vfs.git#for-linus for fix...

Ah, thanks for fixing this up! I'm just wondering how the test program I
was using could have passed because pread() I'm doing there must have
returned ENOMEM... Argh, I see. I have tested the problem with ext4 which
uses DIO_LOCKING mode and in that case 'retval' gets overwritten to 0 by
filemap_write_and_wait() call :-|.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR