On Wed, 2019-02-20 at 19:09 +0000, Jason Gunthorpe wrote:
> On Wed, Feb 20, 2019 at 01:42:14PM -0500, Doug Ledford wrote:
>
> > What this really makes me think is that we don't want this alias
> > device model we have now. We want full ib_device copies
>
> No, that would break the API users have today.
That's a given. You can't implement a new feature like namespaces while
preserving non-namespace behavior, they are fundamentally diametrically
opposed behaviors. Anything you want to see the API as it is today must
stay in the init_net ns. Anything you want to put into a different ns,
needs to be able to deal with ns constraints.
> Today exactly the same
> device is visible in all the namespaces. We can't swap that out for
> some other ib_device without impacting *something* user visible. We
> have too much stuff keyed to the ib_device.
I'm dubious that this is a real concern. See below.
> > I really don't like the disconnect/reconnect model. There's no reason
> > someone with a valid namespace association at the time we make this
> > change should see anything happen. Just tear down what's invalid, and
> > leave the rest alone.
>
> We agreed this exclusive / !exclusive switch is just for
> compatibility and should be set on boot - I don't think it makes sense
> to jump through elaborate hoops to make it work a little better.
I agree. But I think we just drop it entirely.
> This really only relates to uverbs, if it is really important we could
> sweep the ufile's for ns conformance and disassociate the wrong ones.
No, as I think more about it, I think we just enable namespace support,
we default to having the real device in the init_net ns, and if we want
devices in other namespaces the admin creates them. In the long run,
that's the only model we need to support. Anything else is just a stop
gap. I say this because I don't know of any real world use involving
RDMA, containers, non-init_net namespaces, and RDMA access. Do you? Or
is this still a theoretical/research thing where only the people playing
with this prior to production viability would ever see a problem? What
I'm thinking is that all of the existing systems today that are
production systems, if they were upgraded to a namespace aware RDMA
stack, would simply continue to run all of their apps in the init_net
namespace and use the primary device instance that would still be there.
Are there any production systems, in use today, that actually put apps
in non-default namespaces and also access RDMA?
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: B826A3330E572FDD
Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
Attachment:
signature.asc
Description: This is a digitally signed message part
