Re: HEADS UP: MIPS switching to common pmap (uvm/pmap)

[Matt Thomas <matt%3am-software.com@localhost>]

> On Jun 25, 2015, at 6:42 PM, Simon Burge <simonb%NetBSD.org@localhost> wrote:
> 
> Hi Matt,
> 
> Matt Thomas wrote:
> 
>> I now have MIPS using the same pmap code as PowerPC BookE in
>> sys/uvm/pmap.
>> 
>> I plan on merging it into the tree next week.
>> 
>> [ ... ]
>> 
>> By using the common pmap code, MIPS kernel now support dynamically
>> increasing kernel address space.
> 
> Have you benchmarked the new common pmap?  I'm curious if it's better,
> worse or the the same for any micro benchmarks that are pmap dependant
> (lmbench fork comes to mind, maybe something with lots of malloc/free?).

I’ve been trying to get it to work.  My kernels have DEBUG/DIAGNOSTIC
enabled so I haven’t been testing performance.  For MIPS3+ 32-bit kernels, 
both kernel and user address are both handled by the TLB miss handler.
With 8KB pages, we only need to load one PTE.  It does feel faster.  There’s
no more kern_tlb_invalidate since either both halves of the TLB entry are 
valid or neither are.  A lot of kernel-specific code has gone away now that
the kernel uses standard pmap PTE lookups.

>> If you have a MIPS system and you'd like to try a new kernel,
>> let me know and I can built you one.
> 
> If there's anything you think needs testing that you don't have access
> to, let me know.  I've still got way too much crap in the garage :).

r5k are my bane.