KVM: MMU: introduce SPTE_MMU_WRITEABLE bit
authorXiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Wed, 20 Jun 2012 07:58:58 +0000 (15:58 +0800)
committerAvi Kivity <avi@redhat.com>
Wed, 11 Jul 2012 13:51:19 +0000 (16:51 +0300)
This bit indicates whether the spte can be writable on MMU, that means
the corresponding gpte is writable and the corresponding gfn is not
protected by shadow page protection

In the later path, SPTE_MMU_WRITEABLE will indicates whether the spte
can be locklessly updated

Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/kvm/mmu.c

index a2fc65b..b160652 100644 (file)
@@ -145,7 +145,8 @@ module_param(dbg, bool, 0644);
 #define CREATE_TRACE_POINTS
 #include "mmutrace.h"
 
-#define SPTE_HOST_WRITEABLE (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
+#define SPTE_HOST_WRITEABLE    (1ULL << PT_FIRST_AVAIL_BITS_SHIFT)
+#define SPTE_MMU_WRITEABLE     (1ULL << (PT_FIRST_AVAIL_BITS_SHIFT + 1))
 
 #define SHADOW_PT_INDEX(addr, level) PT64_INDEX(addr, level)
 
@@ -1084,34 +1085,51 @@ static void drop_large_spte(struct kvm_vcpu *vcpu, u64 *sptep)
                kvm_flush_remote_tlbs(vcpu->kvm);
 }
 
+static bool spte_is_locklessly_modifiable(u64 spte)
+{
+       return !(~spte & (SPTE_HOST_WRITEABLE | SPTE_MMU_WRITEABLE));
+}
+
 /*
- * Write-protect on the specified @sptep due to dirty page logging or
- * protecting shadow page table. @flush indicates whether tlb need be
- * flushed.
+ * Write-protect on the specified @sptep, @pt_protect indicates whether
+ * spte writ-protection is caused by protecting shadow page table.
+ * @flush indicates whether tlb need be flushed.
+ *
+ * Note: write protection is difference between drity logging and spte
+ * protection:
+ * - for dirty logging, the spte can be set to writable at anytime if
+ *   its dirty bitmap is properly set.
+ * - for spte protection, the spte can be writable only after unsync-ing
+ *   shadow page.
  *
  * Return true if the spte is dropped.
  */
-static bool spte_write_protect(struct kvm *kvm, u64 *sptep, bool *flush)
+static bool
+spte_write_protect(struct kvm *kvm, u64 *sptep, bool *flush, bool pt_protect)
 {
        u64 spte = *sptep;
 
-       if (!is_writable_pte(spte))
+       if (!is_writable_pte(spte) &&
+             !(pt_protect && spte_is_locklessly_modifiable(spte)))
                return false;
 
        rmap_printk("rmap_write_protect: spte %p %llx\n", sptep, *sptep);
 
-       *flush |= true;
-
-       if (__drop_large_spte(kvm, sptep))
+       if (__drop_large_spte(kvm, sptep)) {
+               *flush |= true;
                return true;
+       }
 
+       if (pt_protect)
+               spte &= ~SPTE_MMU_WRITEABLE;
        spte = spte & ~PT_WRITABLE_MASK;
-       mmu_spte_update(sptep, spte);
+
+       *flush |= mmu_spte_update(sptep, spte);
        return false;
 }
 
-static bool
-__rmap_write_protect(struct kvm *kvm, unsigned long *rmapp, int level)
+static bool __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp,
+                                int level, bool pt_protect)
 {
        u64 *sptep;
        struct rmap_iterator iter;
@@ -1119,7 +1137,7 @@ __rmap_write_protect(struct kvm *kvm, unsigned long *rmapp, int level)
 
        for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
                BUG_ON(!(*sptep & PT_PRESENT_MASK));
-               if (spte_write_protect(kvm, sptep, &flush)) {
+               if (spte_write_protect(kvm, sptep, &flush, pt_protect)) {
                        sptep = rmap_get_first(*rmapp, &iter);
                        continue;
                }
@@ -1148,7 +1166,7 @@ void kvm_mmu_write_protect_pt_masked(struct kvm *kvm,
 
        while (mask) {
                rmapp = &slot->rmap[gfn_offset + __ffs(mask)];
-               __rmap_write_protect(kvm, rmapp, PT_PAGE_TABLE_LEVEL);
+               __rmap_write_protect(kvm, rmapp, PT_PAGE_TABLE_LEVEL, false);
 
                /* clear the first set bit */
                mask &= mask - 1;
@@ -1167,7 +1185,7 @@ static bool rmap_write_protect(struct kvm *kvm, u64 gfn)
        for (i = PT_PAGE_TABLE_LEVEL;
             i < PT_PAGE_TABLE_LEVEL + KVM_NR_PAGE_SIZES; ++i) {
                rmapp = __gfn_to_rmap(gfn, i, slot);
-               write_protected |= __rmap_write_protect(kvm, rmapp, i);
+               write_protected |= __rmap_write_protect(kvm, rmapp, i, true);
        }
 
        return write_protected;
@@ -2296,8 +2314,10 @@ static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
                spte |= shadow_x_mask;
        else
                spte |= shadow_nx_mask;
+
        if (pte_access & ACC_USER_MASK)
                spte |= shadow_user_mask;
+
        if (level > PT_PAGE_TABLE_LEVEL)
                spte |= PT_PAGE_SIZE_MASK;
        if (tdp_enabled)
@@ -2322,7 +2342,7 @@ static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
                        goto done;
                }
 
-               spte |= PT_WRITABLE_MASK;
+               spte |= PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE;
 
                if (!vcpu->arch.mmu.direct_map
                    && !(pte_access & ACC_WRITE_MASK)) {
@@ -2351,8 +2371,7 @@ static int set_spte(struct kvm_vcpu *vcpu, u64 *sptep,
                                 __func__, gfn);
                        ret = 1;
                        pte_access &= ~ACC_WRITE_MASK;
-                       if (is_writable_pte(spte))
-                               spte &= ~PT_WRITABLE_MASK;
+                       spte &= ~(PT_WRITABLE_MASK | SPTE_MMU_WRITEABLE);
                }
        }
 
@@ -3933,7 +3952,7 @@ void kvm_mmu_slot_remove_write_access(struct kvm *kvm, int slot)
                              !is_last_spte(pt[i], sp->role.level))
                                continue;
 
-                       spte_write_protect(kvm, &pt[i], &flush);
+                       spte_write_protect(kvm, &pt[i], &flush, false);
                }
        }
        kvm_flush_remote_tlbs(kvm);