I think I found 2 minor bugs in the current tlb flush code:
- during crossing TLB flushes, mm->cpu_mask is not set.
- flush_tlb_current_task() contains a race between
flush_tlb_other_cpus(the IPI could set mm->cpu_mask)
& the next line.
effectively, my patch means that flush_tlb_current_task() 
is identical to flush_tlb_mm(current->mm), perhaps this should
be replaced with a #define.

I've attached a patch (untested) against 2.3.12-9,

    Manfred