Re: [PATCH] MIPS: mm: Prevent a TLB shutdown on initial uniquification

[Thomas Bogendoerfer <tsbogend@alpha.franken.de>]

On Wed, Nov 12, 2025 at 12:16:28PM +0000, Maciej W. Rozycki wrote:
> On Wed, 12 Nov 2025, Thomas Bogendoerfer wrote:
> 
> > >  Can you try the diagnostic patch below, which is what I used to verify 
> > > this change, and report the entries produced?  Otherwise I wonder whether 
> > > I haven't missed a barrier somewhere.
> > 
> > Update on the issue: Your patch is good and the segmentation faults,
> > I'm seeing, have IMHO a different reason. Instead of removing the call
> > to r4k_tlb_uniquify() I've replaced the jal in the binary with a nop.
> > And the issue is still there with this patched kernel. I've seen
> > something similair on a R12k Octanes, which comes and goes probably
> > depeding on code layout. So far I wasn't able to nail this down :-(
> 
>  Oh dear!  Something to do with the cache?  Or code alignment perhaps?

code alignment is probably the trigger. It's reproducible on an R4400SC
and R5000 Indy, but not on a R4000SC Indy. Main difference other than
clock speed is L1 cache size...

And I've missremembered the R12k Octane problem. It's not a segmentation
fault but a bus error, because of an illegal instruction. I tracked it
down to a incorrect data in I-Cache (all 0 cache line, iirc), but never
found the reason for that.

> > Do you want to send a v2 of the patch ? I'm fine with the current version
> > for applying...
> 
>  I'll send v2 with an update for the Wired register as we talked.  It may 
> take a day or two.

no problem, thank you.

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]