Re: [git head] Should X86_PAT really default to yes?

[Jesse Barnes <jesse.barnes@intel.com>]

On Friday, May 02, 2008 2:55 pm Pallipadi, Venkatesh wrote:
> >-----Original Message-----
> >From: Barnes, Jesse
> >Sent: Friday, May 02, 2008 1:40 PM
> >To: Pallipadi, Venkatesh
> >Cc: Frans Pop; linux-kernel@vger.kernel.org; Ingo Molnar;
> >Packard, Keith
> >Subject: Re: [git head] Should X86_PAT really default to yes?
> >
> >On Friday, May 02, 2008 12:37 pm Pallipadi, Venkatesh wrote:
> >> >-----Original Message-----
> >> >From: Frans Pop [mailto:elendil@planet.nl]
> >> >Sent: Friday, May 02, 2008 12:22 PM
> >> >To: linux-kernel@vger.kernel.org
> >> >Cc: Pallipadi, Venkatesh; Ingo Molnar
> >> >Subject: [git head] Should X86_PAT really default to yes?
> >> >
> >> >With X86_PAT enabled, when X is started I get about 40 lines
> >> >(with varying
> >> >addresses) like:
> >> >kernel: Xorg:3358 /dev/mem expected mapping type write-back for
> >> >807bf000-81000000, got uncached-minus
> >
> >These messages?  They're coming from the kernel it looks like,
> >from the
> >map_devmem routine in pat.c.  I'm not sure they're accurate
> >though; for PCI
> >regions /dev/mem is *supposed* to map with UC- and not WB, so
> >maybe this
> >function needs to be updated?
>
> Indeed.
> I think these messages are due to X using the mprotect workaround to
> change UC_MINUS to WB.
> I don't see these error messages on my 965 here. May be I have different
> x version.

Hm, yeah that could be.  strace would tell us.

>
> What may be happening:
> 1) process A mmaps /dev/mem and gets UC_MINUS
> 2) Changes the page table to make pg_prot WB
> 3) Does a fork to create process B
> 4) While copying the vma, we go through map_devmem request WB, but get
> UC_MINUS back
> 5) We are not changing vma pg_prot to new value at this point (we should
> change this), so one more round of errors will be there when forked
> process exits.
>
> Again, this should not have any side-effect like the band etc. It just a
> "friendly warning". It should go away when X moves to using WC or does
> not use the mprotect workaround to make pg_prot WB.

More recent versions of X will use sysfs rather than /dev/mem (though we're 
still using the mprotect hack there to give us UC-), so yeah this warning 
should already be gone in more recent builds.

Jesse