[ofa-general] Re: [PATCHv4 RFC] Scalable Reliable Connection: API and documentation

[Michael S. Tsirkin]

> Quoting Jack Morgenstein <jackm at dev.mellanox.co.il>:
> Subject: Re: [PATCHv4 RFC] Scalable Reliable Connection: API and documentation
> 
> On Sunday 12 August 2007 19:49, Michael S. Tsirkin wrote:
> > > Why not add xrc_domain to the driver-data area?  Because the xrc_domain is a ib verbs-layer construct.
> > 
> > It doesn't follow, necessarily.
> > Will we be able to void breaking kernel-user ABI if we stick domain
> > handle in driver-specific area?
> 
> Are you suggesting the following:
> 
> 1. Do NOT increment the kernel-user ABI
> 2. Put all src-related changes into the driver-specific area

I was just generally saying that if we can avoid breaking the ABI, we should.
What you describe seems OK to me.

> 3. Rely the fact that if userspace is using driver libraries which do not support SRC,
>    the src-related functions will not be present, and libibverbs will reject the src-related
>    function calls.

You mean, for devices that do not support SRC?
No, I think it's kernel's job to validate this case.

>    NOTE: This may be the case for SRC function calls.  However, there is no
>    check on qp-type in userspace during ibv_create_qp.  Its possible for the
>    user to indicate IBV_QPT_SRC, have it go all the way to kernel-space -- and
>    kernel space will take the (garbage) value for the src-domain number.
> 
> I think we will find other such holes if we don't increment the kernel-user
> ABI version.

So, kernel has to validate the SRC domain handle.
This does not look like an issue to me, at all.

-- 
MST