KVM: MMU: Make cmpxchg_gpte aware of nesting too
authorRoedel, Joerg <Joerg.Roedel@amd.com>
Wed, 20 Apr 2011 13:33:16 +0000 (15:33 +0200)
committerAvi Kivity <avi@redhat.com>
Sun, 22 May 2011 12:39:26 +0000 (08:39 -0400)
This patch makes the cmpxchg_gpte() function aware of the
difference between l1-gfns and l2-gfns when nested
virtualization is in use.  This fixes a potential
data-corruption problem in the l1-guest and makes the code
work correct (at least as correct as the hardware which is
emulated in this code) again.

Cc: stable@kernel.org
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/kvm/paging_tmpl.h

index 74f8567d57ace697565a1def10ff3feee197a249..1b6899088f97b5488d97cfc20a53c05c84a6f565 100644 (file)
@@ -78,15 +78,21 @@ static gfn_t gpte_to_gfn_lvl(pt_element_t gpte, int lvl)
        return (gpte & PT_LVL_ADDR_MASK(lvl)) >> PAGE_SHIFT;
 }
 
-static bool FNAME(cmpxchg_gpte)(struct kvm *kvm,
+static int FNAME(cmpxchg_gpte)(struct kvm_vcpu *vcpu, struct kvm_mmu *mmu,
                         gfn_t table_gfn, unsigned index,
                         pt_element_t orig_pte, pt_element_t new_pte)
 {
        pt_element_t ret;
        pt_element_t *table;
        struct page *page;
+       gpa_t gpa;
 
-       page = gfn_to_page(kvm, table_gfn);
+       gpa = mmu->translate_gpa(vcpu, table_gfn << PAGE_SHIFT,
+                                PFERR_USER_MASK|PFERR_WRITE_MASK);
+       if (gpa == UNMAPPED_GVA)
+               return -EFAULT;
+
+       page = gfn_to_page(vcpu->kvm, gpa_to_gfn(gpa));
 
        table = kmap_atomic(page, KM_USER0);
        ret = CMPXCHG(&table[index], orig_pte, new_pte);
@@ -192,11 +198,17 @@ walk:
 #endif
 
                if (!eperm && !rsvd_fault && !(pte & PT_ACCESSED_MASK)) {
+                       int ret;
                        trace_kvm_mmu_set_accessed_bit(table_gfn, index,
                                                       sizeof(pte));
-                       if (FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn,
-                           index, pte, pte|PT_ACCESSED_MASK))
+                       ret = FNAME(cmpxchg_gpte)(vcpu, mmu, table_gfn,
+                                       index, pte, pte|PT_ACCESSED_MASK);
+                       if (ret < 0) {
+                               present = false;
+                               break;
+                       } else if (ret)
                                goto walk;
+
                        mark_page_dirty(vcpu->kvm, table_gfn);
                        pte |= PT_ACCESSED_MASK;
                }
@@ -245,13 +257,17 @@ walk:
                goto error;
 
        if (write_fault && !is_dirty_gpte(pte)) {
-               bool ret;
+               int ret;
 
                trace_kvm_mmu_set_dirty_bit(table_gfn, index, sizeof(pte));
-               ret = FNAME(cmpxchg_gpte)(vcpu->kvm, table_gfn, index, pte,
+               ret = FNAME(cmpxchg_gpte)(vcpu, mmu, table_gfn, index, pte,
                            pte|PT_DIRTY_MASK);
-               if (ret)
+               if (ret < 0) {
+                       present = false;
+                       goto error;
+               } else if (ret)
                        goto walk;
+
                mark_page_dirty(vcpu->kvm, table_gfn);
                pte |= PT_DIRTY_MASK;
                walker->ptes[walker->level - 1] = pte;