RE: 2.6.16 hugetlbfs problem - DEBUG_PAGEALLOC

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

Andrew Morton wrote on Friday, March 24, 2006 6:54 PM
> "Chen, Kenneth W" <kenneth.w.chen@intel.com> wrote:
> >
> > Mark Rustad wrote on Friday, March 24, 2006 9:52 AM
> >  > I have narrowed this down to DEBUG_PAGEALLOC. If that option is  
> >  > enabled, attempts to reference areas mmap-ed from hugetlbfs files  
> >  > fault forever. You can see that I had that set in the failing config  
> >  > I reported below.
> > 
> >  Yeah, it turns out that the debug option is not compatible with hugetlb
> >  page support. That debug option turns off PSE. Once it is turned off in
> >  CR4, cpu will ignore pse bit in the pmd and causing infinite page-not-
> >  present fault :-(
> 
> I wonder if any of the other architectures which implement both these
> features might have problems too.


Only 32-bit x86 arch implements both. We get away by not having DEBUG_PAGEALLOC
feature on any other arch.

I read the ia32 processor manual, it says the pse flag in CR4 controls
4K/4M mapping in addition to pse bit in page directory entry.  But in
PAE mode, cr4.pse is ignored and pse bit in pmd controls 4K/2M mapping.
I was going to verify that on my ia32 box, but apparently, turning on
64G highmem gives me machine reset at boot. (is it a known regression?)

X86-64 surly ignores cr4.pse for 4K/2M mapping. It is solely controlled
by pse bit in pmd.

- Ken