Re: [PATCH 28/28] xfs: rework unreferenced inode lookups

[Christoph Hellwig]

On Fri, Nov 15, 2019 at 09:16:02AM +1100, Dave Chinner wrote:
> > Can we tie these into the proper locking interface using flags? For
> > example, something like xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_ILOCK_NONOWNER)
> > or xfs_ilock(ip, XFS_ILOCK_EXCL_NONOWNER) perhaps?
> 
> I'd prefer not to make this part of the common locking interface -
> it's a one off special use case, not something we want to progate
> elsewhere into the code.
> 
> Now that I think over it, I probably should have tagged this with
> patch with [RFC]. I think we should just get rid of the mrlock
> wrappers rather than add more, and that would simplify this a lot.

Yes, killing off the mrlock wrappers would be very helpful.  The only
thing we use them for is asserts on the locking state.  We could either
switch to lockdep_assert_held*, or just open code the write locked bit.
While it is a little more ugly I'd tend towards the latter given that
the locking asserts are too useful to require lockdep builds with their
performance impact.