[openib-general] QP from userspace used in kernel

[Caitlin Bestler]

openib-general-bounces at openib.org wrote:
> Hi,
> 
> What are the reasons that a qp allocated in user-space can't
> be passed and used by a kernel module?
> What are the steps needed to make a userspace-allocated qp usable by
> a kernel module? 
> 
> 

You would need to construct an environment such that the
device-specific verbs module, which assumes it is executing
in the user space where the QP was created, would never notice
the difference.

The device-specific verbs will typically have created shared
memory resources that are accessible by both the RDMA device
and from the creating user memory map. These resources may
include pointers that assume the original memory map. The
exact methods of remembering the locations of these resources
will vary by device, so the chance of coming up with a scheme
that works without explicit support of all device vendors is
very low. The chances of convincing all device vendors to add
a new option to support this model is similarly low unless you
can make a very compelling case as to why this is necessary.

Having the in-kernel proxy create the QP and do operations
for the end-user is a very adequate work around.

For complex cleanup purposes the kernel could simply assume
the identity of the failed process, but that would only be
required if the standard cleanup was somehow not adequate.