Re: preempt bug in set_pmd_pfn?

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
> actually, i think the correct approach is to remove the TLB flushing and 
> perhaps to check that the old pte is not present. Do we ever _change_ 
> mappings via __set_fixmap()? I think we only ever install them.
>   

Yes, I think that's the case.  clear_fixmap() exists for clearing out an 
existing mapping, but its only used to clear out the WP test mapping and 
in early_iounmap (if called late).  I couldn't see any instances of 
replacing a mapping.

> but if we ever change them somewhere then the correct approach is to do 
> a flush_tlb_all(). It's not just about preemption but about the fact 
> that we modified the kernel address space and we must propagate that to 
> all CPUs.
>   

Yes, I was wondering about that.  If __set_fixmap is only used at boot 
time, then a global flush isn't necessary, but if its deemed a 
general-purpose API in a normal running kernel, it needs to deal with 
cross-cpu flushes.

64-bit set_fixmap is __init only, and I'd be OK with that.  The only 
non-__init use in the 32-bit kernel is the compat vdso mapping, and that 
could easily be done by other means (though it would effectively become 
an opencoded set_fixmap, so perhaps that's not a good idea...).

> the vmalloc() backtrace you sent - how did set_pte_pfn() get into that 
> codepath - vmalloc shouldnt be using __set_fixmap().
>   

No, that's set_pte_at(), which is the real issue in both cases.  
__set_fixmap calls both set_pte_at and flush_tlb_one, which is why it 
gets two backtrackes.

    J