Newer Linux' AGP code wants to flush caches on all CPUs under certain
circumstances. Since doing this on all vCPU-s of the domain in
question doesn't yield the intended effect, this needs to be done in
the hypervisor. Add a new MMUEXT operation for this.

While looking at the pre-existing flush implementation, I also noticed
an off-by-one error in the cache flush portion of flush_area_local().

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxxxx>

--- 2010-04-22.orig/xen/arch/x86/flushtlb.c	2008-10-14 19:44:12.000000000 +0200
+++ 2010-04-22/xen/arch/x86/flushtlb.c	2010-04-29 17:47:22.000000000 +0200
@@ -149,7 +149,7 @@ void flush_area_local(const void *va, un
     {
         unsigned long i, sz = 0;
 
-        if ( order < (BITS_PER_LONG - PAGE_SHIFT - 1) )
+        if ( order < (BITS_PER_LONG - PAGE_SHIFT) )
             sz = 1UL << (order + PAGE_SHIFT);
 
         if ( c->x86_clflush_size && c->x86_cache_size && sz &&
--- 2010-04-22.orig/xen/arch/x86/mm.c	2010-04-29 17:35:20.000000000 +0200
+++ 2010-04-22/xen/arch/x86/mm.c	2010-04-29 17:51:08.000000000 +0200
@@ -2889,6 +2889,27 @@ int do_mmuext_op(
             }
             break;
 
+        case MMUEXT_FLUSH_CACHE_GLOBAL:
+            if ( unlikely(foreigndom != DOMID_SELF) )
+                okay = 0;
+            else if ( likely(cache_flush_permitted(d)) )
+            {
+                unsigned int cpu;
+                cpumask_t mask = CPU_MASK_NONE;
+
+                for_each_online_cpu(cpu)
+                    if ( !cpus_intersects(mask,
+                                          per_cpu(cpu_sibling_map, cpu)) )
+                        cpu_set(cpu, mask);
+                flush_mask(&mask, FLUSH_CACHE);
+            }
+            else
+            {
+                MEM_LOG("Non-physdev domain tried to FLUSH_CACHE_GLOBAL");
+                okay = 0;
+            }
+            break;
+
         case MMUEXT_SET_LDT:
         {
             unsigned long ptr  = op.arg1.linear_addr;
--- 2010-04-22.orig/xen/include/public/xen.h	2010-04-29 17:35:20.000000000 +0200
+++ 2010-04-22/xen/include/public/xen.h	2010-04-29 17:09:30.000000000 +0200
@@ -239,6 +239,10 @@ DEFINE_XEN_GUEST_HANDLE(xen_pfn_t);
  * 
  * cmd: MMUEXT_FLUSH_CACHE
  * No additional arguments. Writes back and flushes cache contents.
+ *
+ * cmd: MMUEXT_FLUSH_CACHE_GLOBAL
+ * No additional arguments. Writes back and flushes cache contents
+ * on all CPUs in the system.
  * 
  * cmd: MMUEXT_SET_LDT
  * linear_addr: Linear address of LDT base (NB. must be page-aligned).
@@ -268,6 +272,7 @@ DEFINE_XEN_GUEST_HANDLE(xen_pfn_t);
 #define MMUEXT_NEW_USER_BASEPTR 15
 #define MMUEXT_CLEAR_PAGE       16
 #define MMUEXT_COPY_PAGE        17
+#define MMUEXT_FLUSH_CACHE_GLOBAL 18
 
 #ifndef __ASSEMBLY__
 struct mmuext_op {