On Wed, 28 Jun 2017, Ralf Baechle wrote: > So I think while it's a nice hack I think this patch should be reserved > for system that don't support parity or ECC or where generally a tiny bit > of performance is more important that reliability. One problem I can see here is that AFAICT we use this somewhat costly scratch setup even in cases where it is not needed (i.e. the scratch remains unused throughout the handler), defeating the intent of the TLB handler generator, which was created for the very purpose of avoiding any wasted cycles that static universal handlers necessarily incurred. I think this is what has to be addressed instead, removing the penalty from configurations that do not need it, i.e. no RIXI, no HTW, etc. Then chances are the more complex configurations will often have the required scratch resources available such as KScratch or SRS registers, which can then be used appropriately (and if some don't then it'll be them only that'll take the penalty they deserve). Maciej
