x86: don't use VA for cache flush when also flushing TLB

Doing both flushes at once is a strong indication for the address
mapping to either having got dropped (in which case the cache flush,
when done via INVLPG, would fault) or its physical address having
changed (in which case the cache flush would end up being done on the
wrong address range). There is no adverse effect (other than the
obvious performance one) using WBINVD in this case regardless of the
range's size; only map_pages_to_xen() uses combined flushes at present.

This problem was observed with the 2nd try backport of d6cb14b3 ("VT-d:
suppress UR signaling for desktop chipsets") to 4.2 (where ioremap()
needs to be replaced with set_fixmap_nocache(); the now commented out
__set_fixmap(, 0, 0) there to undo the mapping resulted in the first of
the above two scenarios).

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

--- a/xen/arch/x86/flushtlb.c
+++ b/xen/arch/x86/flushtlb.c
@@ -152,7 +152,8 @@ void flush_area_local(const void *va, un
         if ( order < (BITS_PER_LONG - PAGE_SHIFT) )
             sz = 1UL << (order + PAGE_SHIFT);
 
-        if ( c->x86_clflush_size && c->x86_cache_size && sz &&
+        if ( !(flags & (FLUSH_TLB|FLUSH_TLB_GLOBAL)) &&
+             c->x86_clflush_size && c->x86_cache_size && sz &&
              ((sz >> 10) < c->x86_cache_size) )
         {
             va = (const void *)((unsigned long)va & ~(sz - 1));