Jan Beulich
2005-Jul-01 14:30 UTC
[Xen-devel] fix inconsistency between change_page_attr and phys_pud_init in x86-64
Other than native x86-64 code, the XEN guest *does* use 4k mappings for the
contiguous kernel mapping of (physical) memory. Thus the code in
change_page_attr needs to work like on i386 (where large pages are used only
conditionally) rather than like native x86-64. Patch below/attached. Not doing
so triggered the BUG_ON during load of intel-agp on machines with i915 chipset.
Jan
diff -Nprux .list 2.6.11-xen/x86_64/mm/pageattr.c
nxen-guest/x86_64/mm/pageattr.c
--- 2.6.11-xen/arch/xen/x86_64/mm/pageattr.c 2005-07-01 09:14:36.606227616 -0700
+++ nxen-guest/arch/xen/x86_64/mm/pageattr.c 2005-07-01 09:15:43.387075384 -0700
@@ -166,16 +166,23 @@ __change_page_attr(unsigned long address
BUG();
/* on x86-64 the direct mapping set at boot is not using 4k pages */
- BUG_ON(PageReserved(kpte_page));
-
- switch (page_count(kpte_page)) {
- case 1:
- save_page(address, kpte_page);
- revert_page(address, ref_prot);
- break;
- case 0:
- BUG(); /* memleak and failed 2M page regeneration */
- }
+// BUG_ON(PageReserved(kpte_page));
+ /*
+ * ..., but the XEN guest kernels (currently) do:
+ * If the pte was reserved, it means it was created at boot
+ * time (not via split_large_page) and in turn we must not
+ * replace it with a large page.
+ */
+ if (!PageReserved(kpte_page)) {
+ switch (page_count(kpte_page)) {
+ case 1:
+ save_page(address, kpte_page);
+ revert_page(address, ref_prot);
+ break;
+ case 0:
+ BUG(); /* memleak and failed 2M page regeneration */
+ }
+ }
return 0;
}
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel