KVM: MMU: rewrite audit_mappings_page() function
Xiao Guangrong [Sat, 28 Aug 2010 11:24:13 +0000 (19:24 +0800)]
There is a bugs in this function, we call gfn_to_pfn() and kvm_mmu_gva_to_gpa_read() in
atomic context(kvm_mmu_audit() is called under the spinlock(mmu_lock)'s protection).

This patch fix it by:
- introduce gfn_to_pfn_atomic instead of gfn_to_pfn
- get the mapping gfn from kvm_mmu_page_get_gfn()

And it adds 'notrap' ptes check in unsync/direct sps

Signed-off-by: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
Signed-off-by: Avi Kivity <avi@redhat.com>

arch/x86/kvm/mmu.c
include/linux/kvm_host.h
virt/kvm/kvm_main.c

index 68575dc..0d91f60 100644 (file)
@@ -3487,15 +3487,6 @@ EXPORT_SYMBOL_GPL(kvm_mmu_get_spte_hierarchy);
 
 static const char *audit_msg;
 
-static gva_t canonicalize(gva_t gva)
-{
-#ifdef CONFIG_X86_64
-       gva = (long long)(gva << 16) >> 16;
-#endif
-       return gva;
-}
-
-
 typedef void (*inspect_spte_fn) (struct kvm *kvm, u64 *sptep);
 
 static void __mmu_spte_walk(struct kvm *kvm, struct kvm_mmu_page *sp,
@@ -3550,39 +3541,53 @@ static void audit_mappings_page(struct kvm_vcpu *vcpu, u64 page_pte,
        gva_t va_delta = 1ul << (PAGE_SHIFT + 9 * (level - 1));
 
        for (i = 0; i < PT64_ENT_PER_PAGE; ++i, va += va_delta) {
-               u64 ent = pt[i];
+               u64 *sptep = pt + i;
+               struct kvm_mmu_page *sp;
+               gfn_t gfn;
+               pfn_t pfn;
+               hpa_t hpa;
 
-               if (ent == shadow_trap_nonpresent_pte)
-                       continue;
+               sp = page_header(__pa(sptep));
 
-               va = canonicalize(va);
-               if (is_shadow_present_pte(ent) && !is_last_spte(ent, level))
-                       audit_mappings_page(vcpu, ent, va, level - 1);
-               else {
-                       gpa_t gpa = kvm_mmu_gva_to_gpa_read(vcpu, va, NULL);
-                       gfn_t gfn = gpa >> PAGE_SHIFT;
-                       pfn_t pfn = gfn_to_pfn(vcpu->kvm, gfn);
-                       hpa_t hpa = (hpa_t)pfn << PAGE_SHIFT;
+               if (sp->unsync) {
+                       if (level != PT_PAGE_TABLE_LEVEL) {
+                               printk(KERN_ERR "audit: (%s) error: unsync sp: %p level = %d\n",
+                                               audit_msg, sp, level);
+                               return;
+                       }
 
-                       if (is_error_pfn(pfn)) {
-                               kvm_release_pfn_clean(pfn);
-                               continue;
+                       if (*sptep == shadow_notrap_nonpresent_pte) {
+                               printk(KERN_ERR "audit: (%s) error: notrap spte in unsync sp: %p\n",
+                                               audit_msg, sp);
+                               return;
                        }
+               }
 
-                       if (is_shadow_present_pte(ent)
-                           && (ent & PT64_BASE_ADDR_MASK) != hpa)
-                               printk(KERN_ERR "xx audit error: (%s) levels %d"
-                                      " gva %lx gpa %llx hpa %llx ent %llx %d\n",
-                                      audit_msg, vcpu->arch.mmu.root_level,
-                                      va, gpa, hpa, ent,
-                                      is_shadow_present_pte(ent));
-                       else if (ent == shadow_notrap_nonpresent_pte
-                                && !is_error_hpa(hpa))
-                               printk(KERN_ERR "audit: (%s) notrap shadow,"
-                                      " valid guest gva %lx\n", audit_msg, va);
-                       kvm_release_pfn_clean(pfn);
+               if (sp->role.direct && *sptep == shadow_notrap_nonpresent_pte) {
+                       printk(KERN_ERR "audit: (%s) error: notrap spte in direct sp: %p\n",
+                                       audit_msg, sp);
+                       return;
+               }
+
+               if (!is_shadow_present_pte(*sptep) ||
+                     !is_last_spte(*sptep, level))
+                       return;
+
+               gfn = kvm_mmu_page_get_gfn(sp, sptep - sp->spt);
+               pfn = gfn_to_pfn_atomic(vcpu->kvm, gfn);
 
+               if (is_error_pfn(pfn)) {
+                       kvm_release_pfn_clean(pfn);
+                       return;
                }
+
+               hpa =  pfn << PAGE_SHIFT;
+
+               if ((*sptep & PT64_BASE_ADDR_MASK) != hpa)
+                       printk(KERN_ERR "xx audit error: (%s) levels %d"
+                                          " gva %lx pfn %llx hpa %llx ent %llxn",
+                                          audit_msg, vcpu->arch.mmu.root_level,
+                                          va, pfn, hpa, *sptep);
        }
 }
 
index b837ec8..f2ecdd5 100644 (file)
@@ -300,6 +300,7 @@ void kvm_set_page_dirty(struct page *page);
 void kvm_set_page_accessed(struct page *page);
 
 pfn_t hva_to_pfn_atomic(struct kvm *kvm, unsigned long addr);
+pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn);
 pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn);
 pfn_t gfn_to_pfn_memslot(struct kvm *kvm,
                         struct kvm_memory_slot *slot, gfn_t gfn);
index 2eb0b75..c7a57b4 100644 (file)
@@ -999,7 +999,7 @@ pfn_t hva_to_pfn_atomic(struct kvm *kvm, unsigned long addr)
 }
 EXPORT_SYMBOL_GPL(hva_to_pfn_atomic);
 
-pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
+static pfn_t __gfn_to_pfn(struct kvm *kvm, gfn_t gfn, bool atomic)
 {
        unsigned long addr;
 
@@ -1009,7 +1009,18 @@ pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
                return page_to_pfn(bad_page);
        }
 
-       return hva_to_pfn(kvm, addr, false);
+       return hva_to_pfn(kvm, addr, atomic);
+}
+
+pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
+{
+       return __gfn_to_pfn(kvm, gfn, true);
+}
+EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
+
+pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
+{
+       return __gfn_to_pfn(kvm, gfn, false);
 }
 EXPORT_SYMBOL_GPL(gfn_to_pfn);