On Mon, Feb 05, 2018 at 03:20:25PM -0700, Jason Gunthorpe wrote:
> On Mon, Feb 05, 2018 at 04:16:37PM -0600, Steve Wise wrote:
> > >
> > > On Mon, Feb 05, 2018 at 02:53:23PM -0600, Steve Wise wrote:
> > >
> > > > Even if we create a device-unique ID for each restrack object, to make
> > it
> > > > useful to for userspace analysis would require exposing that ID in the
> > > > various user object structures (or via query methods). EG having
> > > > pd->res_id, cq->res_id, etc...
> > >
> > > As I said, we have that, it is the fd, handle # tuple and most verbs
> > > objects expose the handle # in the user visible part of the object
> > > struct..
> >
> > Can you please show me an example of the handle # for some object like PD or
> > CQ?
>
> From rdma-core verbs.h:
>
> struct ibv_pd {
> struct ibv_context *context;
> uint32_t handle;
> };
>
> > This doesn't address kernel users' object, though, correct?
>
> Right.
>
> And id'ing a file descriptor gets kinda hard too..
And how should we handle kernel objects?
>
> Jason
Attachment:
signature.asc
Description: PGP signature
