[openib-general] Re: user_mad: large rmpp length problem

[Hal Rosenstock]

On Mon, 2005-11-21 at 15:29, Sean Hefty wrote:
> Michael S. Tsirkin wrote:
> > This approach still means that we have to allocate a potentially
> > huge amount of memory to copy all of the rmpp mad into kernel
> > in one go.
> > 
> > What I had in mind was: add a get_next_segment callback to mad,
> > which would copy the data (by copy_from_user) incrementally,
> > making forward progress without requiring us to have all of the mad
> > in kernel memory.
> 
> This may work.  I think that there may be issues with respect to providing 
> proper synchronization and handling retries, but without looking at an 
> implementation, I can't be sure.

Possibly timeouts too. I would prefer to take this a step at a time. I
think that allocating some number of page sized chunks is likely to work
and IMO we should first go down this direction. If this proves
problematic, we can then adjust the strategy.

> Is support for large RMPP transfers needed for kernel clients?

Currently no but why preclude this. Kernel clients ultimately will use
MultiPathRecords which require RMPP.

>   If not, then we 
> may be able to implement this entirely in userspace, provided an option were 
> given to disable it in the kernel.

See above.

-- Hal