[ofiwg] fork support and MR cache

[Jason Gunthorpe]

On Mon, Oct 12, 2020 at 07:51:36PM +0300, Gal Pressman wrote:
> > On Thu, Oct 08, 2020 at 07:28:50PM +0000, Hefty, Sean wrote:
> >> > Please don't recode the kernel interface outside rdma-core, that stuff
> >> > is hard to do right.
> >>
> >> I'm was referring to calling:
> >>
> >> mr->context->ops.reg_mr(...)
> >>
> >> instead of indirectly through ibv_reg_mr().
> > 
> > I deleted those function pointers from the public header a long time
> > ago. It was leaking internal ABI.
> > 
> >> I'm actually fine disabling the cache if fork is enabled.  But how can that be detected?
> > 
> > If you wan to propose an verbs API to show if the fork support MADV
> > stuff has been enabled that seems reasonable to me.
> 
> When you say DONTFORK will be obsolete I assume you refer to:
> https://lore.kernel.org/lkml/20200925222600.6832-1-peterx@redhat.com/
> 
> Right?
> 
> How will userspace rdma-core know whether DONTFORK is needed or not? Are there
> any RDMA kernel changes I'm missing?

We will need to add something, I am waiting to know that it works as
expected.

Jason