x86, hvm: Fix domain restore bug with Intel VLAPIC acceleration.

r18383 mark video memory as ram, and make all valid pages migrated,
including vlapic page (0xFEE00), and share page(0xFFFFF).
An extra memory population for lapic page would override previous
mapping then cause HVM guest with vlapic acceleration hang.

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

diff --git a/xen/arch/x86/mm.c b/xen/arch/x86/mm.c
index f3b1fc8243c81b91c5895908a68fc2ba93c65490..2adc1ed2c7abfc9f16c61f69818573c94be2401a 100644 (file)
--- a/xen/arch/x86/mm.c
+++ b/xen/arch/x86/mm.c
@@ -1604,6 +1604,7 @@ static int mod_l1_entry(l1_pgentry_t *pl1e, l1_pgentry_t nl1e,
     struct domain *d = curr->domain;
     unsigned long mfn;
     struct page_info *l1pg = mfn_to_page(gl1mfn);
+    p2m_type_t p2mt;
     int rc = 1;
 
     page_lock(l1pg);
@@ -1621,8 +1622,8 @@ static int mod_l1_entry(l1_pgentry_t *pl1e, l1_pgentry_t nl1e,
     if ( l1e_get_flags(nl1e) & _PAGE_PRESENT )
     {
         /* Translate foreign guest addresses. */
-        mfn = gmfn_to_mfn(FOREIGNDOM, l1e_get_pfn(nl1e));
-        if ( unlikely(mfn == INVALID_MFN) )
+        mfn = mfn_x(gfn_to_mfn(FOREIGNDOM, l1e_get_pfn(nl1e), &p2mt));
+        if ( !p2m_is_ram(p2mt) || unlikely(mfn == INVALID_MFN) )
             return page_unlock(l1pg), 0;
         ASSERT((mfn & ~(PADDR_MASK >> PAGE_SHIFT)) == 0);
         nl1e = l1e_from_pfn(mfn, l1e_get_flags(nl1e));