[ofa-general] Fast Reg Question

[Tom Tucker]

Roland Dreier wrote:
>  > I'm a little perplexed by the fast reg WR definition. The context is
>  > that I'm using the Fast Reg verb to map the local memory that is the
>  > data source for an RDMA_WRITE.
>  > 
>  > The WR format, however, only takes an rkey. How does this all work
>  > when you're using fast reg to map local memory? Does the WR really
>  > need the mr pointer, or both the lkey and rkey? The IBTA spec seems to
>  > indicate that it needs more information about the MR than just the
>  > rkey.
>
> On Mellanox, L_Key and R_Key are always the same, 

Also true for iWARP.
> so it doesn't really
> matter.  I think in general the idea would be that the L_Key you have
> gets updatedd with any "consumer key" changes you make in the WR but
> otherwise works the same. 
Fair enough. Use the mr->lkey value in the SGE for subsequent DTO.
>  the WR processing better be able to find the
> MR by R_Key so I think it's OK.
>
>   
It just seems a little weird to be supplying the R_Key when you're 
mapping local memory.
> I'll look at the IB spec though.
>
>   
The spec refers to a bunch of verification on the L_Key. Obviously, if 
the L_Key and R_Key are the same the distinction is moot.
>  - R.
>