Re: [PATCH v2] mips: mm: Call rcutree_report_cpu_starting() even earlier

[Maciej W. Rozycki]

Hi Stefan,

> >  I'm afraid I have no QEMU installation ready to use, so I can't reproduce
> > your steps without getting sidetracked.  Would you be able to explain why
> > it triggers with QEMU and possibly the I6400 CPU in particular, but not on
> > real SMP hardware I have?  From your commit description I infer this isn't
> > supposed to be a heisenbug.
> 
> The only explanation that comes to my mind is that your platform has
> TLBINV support and never calls r4k_tlb_uniquify() [1], therefore
> skipping the kmalloc() call. Would you please verify that?

 The Broadcom SiByte SB-1 core I have tried to reproduce this issue with 
(in a dual SMP configuration) is a MIPSr1 CPU and does not have the TLBINV 
instruction or the EHINV feature and neither is architectually allowed to.  
I have used it as one of the test platforms for 64-bit r4k_tlb_uniquify() 
verification and it is guaranteed to call this function.

 I'll see what I can do, but it'll take a while, so maybe someone else can 
try and reproduce it with their setup before I can get to it?

  Maciej