Re: [Xen-devel] PAE PV guest kernel regression

[Andrew Jones <drjones@xxxxxxxxxx>]

----- Original Message ----- 
> Was the fix for the git commit 26c191788f18 (which I understand is in
> RHEL6)
> posted?

Nope, it got rolled into 26c191788f18 before 26c191788f18 was
committed. The fix was to not use __pmd, which is paravirtualized,
when doing the final creation of the pmd_t retval. Here it is
(the 2f88dfd2 referenced there is a rhel6 commit).

    Commit 2f88dfd2 introduces read_pmd_atomic which calls __pmd. __pmd is
    paravirtualized and for xen it calls xen_make_pmd(). xen_make_pmd may
    return a zero pmd in the case that the pfn's associated mfn is still
    invalid. However at the point when read_pmd_atomic is called the mfn can
    appear invalid when it isn't, if its creation was lazily deferred. If
    read_pmd_atomic returns zero in these cases then it can eventually lead to
    a crash. To resolve the issue we can avoid calling __pmd and just return
    the open coded compound literal (as native_make_pmd would do) instead.
    
diff --git a/arch/x86/include/asm/pgtable-3level.h b/arch/x86/include/asm/pgtabl
index f2d473b..4f325a1 100644
--- a/arch/x86/include/asm/pgtable-3level.h
+++ b/arch/x86/include/asm/pgtable-3level.h
@@ -73,12 +73,13 @@ static inline pmd_t read_pmd_atomic(pmd_t *pmdp)
                ret |= ((pmdval_t)*(tmp + 1)) << 32;
        }
 
-       return __pmd(ret);
+       return (pmd_t) { ret };
 }
 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 static inline pmd_t read_pmd_atomic(pmd_t *pmdp)
 {
-       return __pmd(atomic64_read((atomic64_t *)pmdp));
+       pmdval_t val = (pmdval_t)atomic64_read((atomic64_t *)pmdp);
+       return (pmd_t) { val };
 }
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel