Re: ext4 dio RWF_NOWAIT change

["Theodore Y. Ts'o" <tytso@mit.edu>]

On Wed, Feb 05, 2020 at 05:13:44PM +0800, Murphy Zhou wrote:
> Hi,
> 
> Kernel commit 378f32bab3714f04c4e0c3aee4129f6703805550
> Author: Matthew Bobrowski <mbobrowski@mbobrowski.org>
> Date:   Tue Nov 5 23:02:39 2019 +1100
> 
>     ext4: introduce direct I/O write using iomap infrastructure
> 
> 
> Changed the logic of dio+RWF_NOWAIT
> 
> from:
> 
> -       if (!inode_trylock(inode)) {
> -               if (iocb->ki_flags & IOCB_NOWAIT)
> -                       return -EAGAIN;
> -               inode_lock(inode);
> -       }
> 
> 
> to:
> 
> +       if (iocb->ki_flags & IOCB_NOWAIT) {
> +               if (!inode_trylock(inode))
> +                       return -EAGAIN;
> +       } else {
> +               inode_lock(inode);
> +       }
> 
> 
> fstests generic/471 expecet EAGAIN on this situation, so it started to
> fail since than.

I don't understand why this specific change would cause the situation.
In the generic/471 test, here iocb->ki_flags will have IOCB_NOWAIT
set, and in that case I don't see how there would be a change in
behavior with respect to EAGAIN being returned.

In any case, I've been suppressing generic/471 because of concerns
that the test is bogus so I hadn't noticed.  From
kvm-xfstests/test-appliance/files/root/fs/global_exclude:

# The test generic/471 tests the RWF_NOWAIT flag; however
# how it is supposed to work with file systems is disputed,
# and not all device drivers support it.  As a result
# it doesn't work if an LVM device driver is in use (as is the
# case with gce-xfstests).  So let's suppress it for now.  For
# more details see:
# https://lore.kernel.org/linux-block/20190723220502.GX7777@dread.disaster.area/
generic/471

							- Ted