[v2] x86/ept: fix shattering of special pages

Commit Message

Roger Pau Monne June 30, 2022, 4:34 p.m. UTC

The current logic in epte_get_entry_emt() will split any page marked
as special with order greater than zero, without checking whether the
super page is all special.

Fix this by only splitting the page only if it's not all marked as
special, in order to prevent unneeded super page shuttering.

The unconditional special super page shattering has caused a
performance regression on some XenServer GPU pass through workloads.

Signed-off-by: Roger Pau Monné <[e-mail]>
Reviewed-by: Jan Beulich <[e-mail]>
Reviewed-by: Kevin Tian <[e-mail]>
---
Cc: Paul Durrant <[e-mail]>
---
 xen/arch/x86/mm/p2m-ept.c | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

Patch

diff --git a/xen/arch/x86/mm/p2m-ept.c b/xen/arch/x86/mm/p2m-ept.c
index b04ca6dbe8..b4919bad51 100644
--- a/xen/arch/x86/mm/p2m-ept.c
+++ b/xen/arch/x86/mm/p2m-ept.c
@@ -491,7 +491,7 @@  int epte_get_entry_emt(struct domain *d, gfn_t gfn, mfn_t mfn,
 {
     int gmtrr_mtype, hmtrr_mtype;
     struct vcpu *v = current;
-    unsigned long i;
+    unsigned long i, special_pgs;
 
     *ipat = false;
 
@@ -525,15 +525,17 @@  int epte_get_entry_emt(struct domain *d, gfn_t gfn, mfn_t mfn,
         return MTRR_TYPE_WRBACK;
     }
 
-    for ( i = 0; i < (1ul << order); i++ )
-    {
+    for ( special_pgs = i = 0; i < (1ul << order); i++ )
         if ( is_special_page(mfn_to_page(mfn_add(mfn, i))) )
-        {
-            if ( order )
-                return -1;
-            *ipat = true;
-            return MTRR_TYPE_WRBACK;
-        }
+            special_pgs++;
+
+    if ( special_pgs )
+    {
+        if ( special_pgs != (1ul << order) )
+            return -1;
+
+        *ipat = true;
+        return MTRR_TYPE_WRBACK;
     }
 
     switch ( type )