[ofa-general] New proposal for memory management

[Jason Gunthorpe]

On Fri, May 01, 2009 at 09:25:33AM -0400, Tom Talpey wrote:

> Completely agree. I will add that enterprise network programmers are
> going to reject registration caching as well, because it introduces
> vulnerabilities into the data path - silent data corruption. For example,
> storage won't tolerate it, databases won't, etc.

By the same token those apps that care about data security like you
site *must* manually manage their registration to only expose the
memory that needs to be exposed at any time. That is a mandatory step
as soon as you have client initiated RDMA operations, no matter what
your protocol is.

> The problem is that userspace memory registration is slow. Let's address
> that, not address how to make a hack (registration caching) go faster.

Indeed, but how? You need to make a syscall to pin and map the pages,
which is fine, but how do you communicate the information to the HCA
in a manner that is utterly secure and doesn't let userspace 'fiddle'
it to point to arbitary random memory? You get burned pretty fast by
fact that the HCA is DMA'ing instructions out of user space directly :(

Jason