Re: preempt bug in set_pmd_pfn?

[Jeremy Fitzhardinge <jeremy@goop.org>]

Hugh Dickins wrote:
> On Wed, 5 Mar 2008, Jeremy Fitzhardinge wrote:
>   
>> If a thread goes from processor A -> B -> A, where A is first preempted
>> between a pagetable update and a tlb flush, then the second time the thread
>> runs on A may run with a stale tlb (if in the meantime A has either been idle
>> or only running kernel threads).
>>     
>
> Right, thanks, because __flush_tlb_one opts out of the full active_mm
> checking which goes on for userspace mms (which covers, for example,
> the case of preemption before dup_mmap's flush_tlb_mm).
>
> But is there actually a case where there's a problem?  So far as I can see,
> set_pmd_pfn is there solely for discontig_32's __init remap_numa_kva; and
> set_pte_pfn is there solely for set_fixmap, which operates on a per-cpu
> area of pagetable, which would already be in bigger trouble if preemption
> to another cpu were possible.

Yes, I meant set_pte_pfn; set_pmd_pfn is a typo.

Fixmap slots are global, not percpu; you may be thinking of kmap_atomic, 
which reserves percpu fixmap slots for its use.  Most uses of set_fixmap 
are early in boot, where preemption (or other CPUs) isn't a factor.  The 
exception is mapping the compat vdso global mapping.   However, that is 
special-cased anyway, since the set_fixmap is followed by an explicit 
all-cpu tlb flush.

It seems to me that the correct fix is to just make __set_fixmap disable 
preemption for its duration; it probably doesn't make much difference 
for the native case, and it makes Xen happy.

    J