[openib-general] Re: [PATCH] [RFC] use device max_map_per_fmr in fmr pool

[Or Gerlitz]

Roland Dreier wrote:
>     Or> Remapping an FMR the maximum number of times the device allows
>     Or> should reduce the amortized cost of using fmrs to the minimum
>     Or> possible.
> 
> This change makes sense to me.  I think as a prerequisite we need to
> get all the low-level drivers that implement FMRs to report a
> max_map_per_fmr value.  Unfortunately none of them do right now :(

Indeed, looking at the code of mthca with my limited understanding of 
it, i could not conclude how hard it would be to implement it there, if 
you or mellanox can provide some info i might be able to do it.

At the fmr pool level, we can modify the code to use (say) 32 remaps
if the max_map_per_fmr attr is 0 and the device max is it is nonzero.

Or.