On Thu, Dec 20, 2018 at 6:43 PM Jason Gunthorpe <jgg@xxxxxxxxxxxx> wrote:
>
> On Thu, Dec 20, 2018 at 05:00:49PM +0200, Shamir Rabinovitch wrote:
> > > The immediate step beyond this is to make rdma_get_ucontext into a
> > > driver exported function and replace all *->uobject->context with
> > > it. Maybe also adjust things so it can't fail, hard to say right now
> > > if failure will be inconvenient.
> > >
> >
> > I did something like this based on your recent work to unify the uverbs
> > & ioctl code path. Will add the patches to this email and if they are
> > fine with you we could use them to fill the ucontext in the udata from
> > uverbs & ioctl.
>
> This isn't enough, it doesn't do the ioctl path and maybe more.
> The ucontext should be stored in the attr_bundle not the udata..
>
> I'm not totally sure we need this or not - I was going to look at the
> drivers and see if accommodating failure would be easy or not..
>
> So lets go through updating the core APIs like I sent and then lets
> see what is left..
>
> Jason
I plan to make reg_user_mr() accessible from kernel. Something like this
struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
u64 virt_addr, int mr_access_flags)
{
struct ib_mr *mr;
mr = mr = pd->device->ops.reg_user_m(pd, start, length,
virt_addr, access_flags, NULL);
if (IS_ERR(mr))
return mr;
mr->device = pd->device;
mr->pd = pd;
mr->dm = NULL;
mr->uobject = NULL;
atomic_inc(&pd->usecnt);
mr->res.type = RDMA_RESTRACK_MR;
return mr;
}
For that to work I will need to eliminate the assumption that ucontext
exists inside ib_umem_get() so maybe its a good idea to take
everything that is hung on ib_ucontext which ib_umem_get() uses and
put it under pd. Than modify ib_umem_get() to take ib_pd instead of
ib_ucontext (or ib_udata as in this proposal)
