Re: [Virtio-fs] [PATCH 2/2] virtio-fs: do not removemapping if dmap will be used immediately

[Vivek Goyal <vgoyal@redhat.com>]

On Wed, Aug 14, 2019 at 01:30:12PM -0700, Liu Bo wrote:
> On Wed, Aug 14, 2019 at 03:57:18PM -0400, Vivek Goyal wrote:
> > On Tue, Aug 13, 2019 at 02:32:05AM +0800, Liu Bo wrote:
> > > According to how we implement REMOVEMAPPING in daemon, it's unnecessary to
> > > involve a REMOVEMAPPING to reuse a dmap when doing inline reclaim because
> > > dmaps got from inline reclaim get reused for another mapping without being
> > > added back to 'free' dmap pool.
> > > 
> > > This skips REMOVEMAPPING for inline reclaim only and we don't do
> > > REMOVEMAPPING unless someone has raced in to add a dmap to the range.
> > 
> > Given inline reclaims are enabled only for writes, how does this benefit
> > a random read workload.
> >
> 
> Oh, I thought your branch has done it, anyway I've made read take inline
> reclaim as well locally before these two patches, and all tests I have
> didn't complain yet.

I reverted that patch as I found an issue. I just can't remember what
was the issue. Will try to reapply the patch and see if I can see the
problem again.

> 
> I'm testing a kernel build with dmap=1, can you please elaborate more
> on why these might end up deadlock?

Can't remember right now. Generally I have faced many deadlock issues
with dmap=1. So it is a good idea to test with it. We want to make sure
that even with one dax range, we can continue to make forward progress.

Thanks
Vivek