x86: Remove invlpg_works_ok and invlpg only single-page regions.

The flush_area_local() interface was unclear about whether a
multi-page region (2M/4M/1G) had to be mapped by a superpage, and
indeed some callers (map_pages_to_xen()) already would specify
FLUSH_LEVEL(2) for a region actually mapped by 4kB PTEs.

The safest fix is to relax the interface and do a full TLB flush in
these cases. My suspicion is that these cases are rare enough that the
cost of INVLPG versus full flush will be unimportant.

Signed-off-by: Keir Fraser <keir@xensource.com>

1 files changed, 7 insertions, 1 deletions

diff --git a/xen/arch/x86/flushtlb.c b/xen/arch/x86/flushtlb.c
index 69aa7fa21d..808b848de5 100644
--- a/xen/arch/x86/flushtlb.c
+++ b/xen/arch/x86/flushtlb.c
@@ -108,8 +108,14 @@ void flush_area_local(const void *va, unsigned int flags)
 
     if ( flags & (FLUSH_TLB|FLUSH_TLB_GLOBAL) )
     {
-        if ( (level != 0) && test_bit(level, &c->invlpg_works_ok) )
+        if ( level == 1 )
         {
+            /*
+             * We don't INVLPG multi-page regions because the 2M/4M/1G
+             * region may not have been mapped with a superpage. Also there
+             * are various errata surrounding INVLPG usage on superpages, and
+             * a full flush is in any case not *that* expensive.
+             */
             asm volatile ( "invlpg %0"
                            : : "m" (*(const char *)(va)) : "memory" );
         }