On Thu, Nov 06, 2014 at 05:48:16PM +0000, Wei Liu wrote:
> In b38af4721 ("x86,mm: fix pte_special versus pte_numa") pte_special()
> (SPECIAL with PRESENT or PROTNONE) was made to complement pte_numa()
> (SPECIAL with neither PRESENT nor PROTNONE). That broke Xen PV guest
> with NUMA balancing support.
>
> That's because Xen hypervisor sets _PAGE_GLOBAL (_PAGE_GLOBAL /
> _PAGE_PROTNONE in Linux) for guest user space mapping. So in a Xen PV
> guest, when NUMA balancing is enabled, a NUMA hinted PTE ends up
> "SPECIAL (in fact NUMA) with PROTNONE but not PRESENT", which makes
> pte_special() returns true when it shouldn't.
>
> Fundamentally we only need _PAGE_NUMA and _PAGE_PRESENT to tell
> difference between an unmapped entry and an entry protected for NUMA
> hinting fault. So use _PAGE_BIT_SOFTW2 as _PAGE_BIT_NUMA, adjust
> _PAGE_NUMA_MASK and SWP_OFFSET_SHIFT as needed.
>
> Suggested-by: David Vrabel <david.vrabel@xxxxxxxxxx>
> Signed-off-by: Wei Liu <wei.liu2@xxxxxxxxxx>
I suggest instead that you force automatic NUMA balancing to be disabled
on Xen PV guests until I or someone else finds time to implement Linus'
idea to remove _PAGE_NUMA entirely. It's been on my TODO list for a few
weeks but I still have not reached the point where I'm back working on
upstream material properly.
--
Mel Gorman
SUSE Labs
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>