KVM: MMU: Fix SMP shadow instantiation race
authorMarcelo Tosatti <marcelo@kvack.org>
Wed, 12 Dec 2007 00:12:27 +0000 (19:12 -0500)
committerAvi Kivity <avi@qumranet.com>
Wed, 30 Jan 2008 15:53:21 +0000 (17:53 +0200)
There is a race where VCPU0 is shadowing a pagetable entry while VCPU1
is updating it, which results in a stale shadow copy.

Fix that by comparing the contents of the cached guest pte with the
current guest pte after write-protecting the guest pagetable.

Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
drivers/kvm/mmu.c
drivers/kvm/paging_tmpl.h

index ba71e8d66761e4451c12e373907b334d67d54430..92ac0d1106b417aeb2884833f530bf75cde649b2 100644 (file)
@@ -681,7 +681,8 @@ static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
                                             unsigned level,
                                             int metaphysical,
                                             unsigned access,
-                                            u64 *parent_pte)
+                                            u64 *parent_pte,
+                                            bool *new_page)
 {
        union kvm_mmu_page_role role;
        unsigned index;
@@ -720,6 +721,8 @@ static struct kvm_mmu_page *kvm_mmu_get_page(struct kvm_vcpu *vcpu,
        vcpu->mmu.prefetch_page(vcpu, sp);
        if (!metaphysical)
                rmap_write_protect(vcpu->kvm, gfn);
+       if (new_page)
+               *new_page = 1;
        return sp;
 }
 
@@ -993,7 +996,8 @@ static int nonpaging_map(struct kvm_vcpu *vcpu, gva_t v, int write, gfn_t gfn)
                                >> PAGE_SHIFT;
                        new_table = kvm_mmu_get_page(vcpu, pseudo_gfn,
                                                     v, level - 1,
-                                                    1, ACC_ALL, &table[index]);
+                                                    1, ACC_ALL, &table[index],
+                                                    NULL);
                        if (!new_table) {
                                pgprintk("nonpaging_map: ENOMEM\n");
                                return -ENOMEM;
@@ -1059,7 +1063,7 @@ static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
 
                ASSERT(!VALID_PAGE(root));
                sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
-                                     PT64_ROOT_LEVEL, 0, ACC_ALL, NULL);
+                                     PT64_ROOT_LEVEL, 0, ACC_ALL, NULL, NULL);
                root = __pa(sp->spt);
                ++sp->root_count;
                vcpu->mmu.root_hpa = root;
@@ -1080,7 +1084,7 @@ static void mmu_alloc_roots(struct kvm_vcpu *vcpu)
                        root_gfn = 0;
                sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
                                      PT32_ROOT_LEVEL, !is_paging(vcpu),
-                                     ACC_ALL, NULL);
+                                     ACC_ALL, NULL, NULL);
                root = __pa(sp->spt);
                ++sp->root_count;
                vcpu->mmu.pae_root[i] = root | PT_PRESENT_MASK;
index 3ab3fb635e1627267aff7720c44fac11828e5b1b..fb19596c9589829669d701aa36849cc558b932a8 100644 (file)
@@ -65,7 +65,8 @@
 struct guest_walker {
        int level;
        gfn_t table_gfn[PT_MAX_FULL_LEVELS];
-       pt_element_t pte;
+       pt_element_t ptes[PT_MAX_FULL_LEVELS];
+       gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
        unsigned pt_access;
        unsigned pte_access;
        gfn_t gfn;
@@ -150,6 +151,7 @@ walk:
                pte_gpa = gfn_to_gpa(table_gfn);
                pte_gpa += index * sizeof(pt_element_t);
                walker->table_gfn[walker->level - 1] = table_gfn;
+               walker->pte_gpa[walker->level - 1] = pte_gpa;
                pgprintk("%s: table_gfn[%d] %lx\n", __FUNCTION__,
                         walker->level - 1, table_gfn);
 
@@ -180,6 +182,8 @@ walk:
 
                pte_access = pt_access & FNAME(gpte_access)(vcpu, pte);
 
+               walker->ptes[walker->level - 1] = pte;
+
                if (walker->level == PT_PAGE_TABLE_LEVEL) {
                        walker->gfn = gpte_to_gfn(pte);
                        break;
@@ -209,9 +213,9 @@ walk:
                        goto walk;
                pte |= PT_DIRTY_MASK;
                kvm_mmu_pte_write(vcpu, pte_gpa, (u8 *)&pte, sizeof(pte));
+               walker->ptes[walker->level - 1] = pte;
        }
 
-       walker->pte = pte;
        walker->pt_access = pt_access;
        walker->pte_access = pte_access;
        pgprintk("%s: pte %llx pte_access %x pt_access %x\n",
@@ -268,7 +272,7 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
        u64 *shadow_ent;
        unsigned access = walker->pt_access;
 
-       if (!is_present_pte(walker->pte))
+       if (!is_present_pte(walker->ptes[walker->level - 1]))
                return NULL;
 
        shadow_addr = vcpu->mmu.root_hpa;
@@ -285,6 +289,7 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
                u64 shadow_pte;
                int metaphysical;
                gfn_t table_gfn;
+               bool new_page = 0;
 
                shadow_ent = ((u64 *)__va(shadow_addr)) + index;
                if (is_shadow_present_pte(*shadow_ent)) {
@@ -300,16 +305,23 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
                if (level - 1 == PT_PAGE_TABLE_LEVEL
                    && walker->level == PT_DIRECTORY_LEVEL) {
                        metaphysical = 1;
-                       if (!is_dirty_pte(walker->pte))
+                       if (!is_dirty_pte(walker->ptes[level - 1]))
                                access &= ~ACC_WRITE_MASK;
-                       table_gfn = gpte_to_gfn(walker->pte);
+                       table_gfn = gpte_to_gfn(walker->ptes[level - 1]);
                } else {
                        metaphysical = 0;
                        table_gfn = walker->table_gfn[level - 2];
                }
                shadow_page = kvm_mmu_get_page(vcpu, table_gfn, addr, level-1,
                                               metaphysical, access,
-                                              shadow_ent);
+                                              shadow_ent, &new_page);
+               if (new_page && !metaphysical) {
+                       pt_element_t curr_pte;
+                       kvm_read_guest(vcpu->kvm, walker->pte_gpa[level - 2],
+                                      &curr_pte, sizeof(curr_pte));
+                       if (curr_pte != walker->ptes[level - 2])
+                               return NULL;
+               }
                shadow_addr = __pa(shadow_page->spt);
                shadow_pte = shadow_addr | PT_PRESENT_MASK | PT_ACCESSED_MASK
                        | PT_WRITABLE_MASK | PT_USER_MASK;
@@ -317,7 +329,8 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr,
        }
 
        mmu_set_spte(vcpu, shadow_ent, access, walker->pte_access & access,
-                    user_fault, write_fault, walker->pte & PT_DIRTY_MASK,
+                    user_fault, write_fault,
+                    walker->ptes[walker->level-1] & PT_DIRTY_MASK,
                     ptwrite, walker->gfn);
 
        return shadow_ent;
@@ -382,7 +395,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,
        /*
         * mmio: emulate if accessible, otherwise its a guest fault.
         */
-       if (is_io_pte(*shadow_pte))
+       if (shadow_pte && is_io_pte(*shadow_pte))
                return 1;
 
        ++vcpu->stat.pf_fixed;