return asid + 1;
 }
 
+/*
+ * The user PCID is just the kernel one, plus the "switch bit".
+ */
+static inline u16 user_pcid(u16 asid)
+{
+       u16 ret = kern_pcid(asid);
+#ifdef CONFIG_PAGE_TABLE_ISOLATION
+       ret |= 1 << X86_CR3_PTI_SWITCH_BIT;
+#endif
+       return ret;
+}
+
 struct pgd_t;
 static inline unsigned long build_cr3(pgd_t *pgd, u16 asid)
 {
                /*
                 * Using INVPCID is considerably faster than a pair of writes
                 * to CR4 sandwiched inside an IRQ flag save/restore.
+                *
+                * Note, this works with CR4.PCIDE=0 or 1.
                 */
                invpcid_flush_all();
                return;
        if (!static_cpu_has(X86_FEATURE_PTI))
                return;
 
-       invalidate_user_asid(loaded_mm_asid);
+       /*
+        * Some platforms #GP if we call invpcid(type=1/2) before CR4.PCIDE=1.
+        * Just use invalidate_user_asid() in case we are called early.
+        */
+       if (!this_cpu_has(X86_FEATURE_INVPCID_SINGLE))
+               invalidate_user_asid(loaded_mm_asid);
+       else
+               invpcid_flush_one(user_pcid(loaded_mm_asid), addr);
 }
 
 /*
 
 
 static void setup_pcid(void)
 {
-#ifdef CONFIG_X86_64
-       if (boot_cpu_has(X86_FEATURE_PCID)) {
-               if (boot_cpu_has(X86_FEATURE_PGE)) {
-                       /*
-                        * This can't be cr4_set_bits_and_update_boot() --
-                        * the trampoline code can't handle CR4.PCIDE and
-                        * it wouldn't do any good anyway.  Despite the name,
-                        * cr4_set_bits_and_update_boot() doesn't actually
-                        * cause the bits in question to remain set all the
-                        * way through the secondary boot asm.
-                        *
-                        * Instead, we brute-force it and set CR4.PCIDE
-                        * manually in start_secondary().
-                        */
-                       cr4_set_bits(X86_CR4_PCIDE);
-               } else {
-                       /*
-                        * flush_tlb_all(), as currently implemented, won't
-                        * work if PCID is on but PGE is not.  Since that
-                        * combination doesn't exist on real hardware, there's
-                        * no reason to try to fully support it, but it's
-                        * polite to avoid corrupting data if we're on
-                        * an improperly configured VM.
-                        */
-                       setup_clear_cpu_cap(X86_FEATURE_PCID);
-               }
+       if (!IS_ENABLED(CONFIG_X86_64))
+               return;
+
+       if (!boot_cpu_has(X86_FEATURE_PCID))
+               return;
+
+       if (boot_cpu_has(X86_FEATURE_PGE)) {
+               /*
+                * This can't be cr4_set_bits_and_update_boot() -- the
+                * trampoline code can't handle CR4.PCIDE and it wouldn't
+                * do any good anyway.  Despite the name,
+                * cr4_set_bits_and_update_boot() doesn't actually cause
+                * the bits in question to remain set all the way through
+                * the secondary boot asm.
+                *
+                * Instead, we brute-force it and set CR4.PCIDE manually in
+                * start_secondary().
+                */
+               cr4_set_bits(X86_CR4_PCIDE);
+
+               /*
+                * INVPCID's single-context modes (2/3) only work if we set
+                * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
+                * on systems that have X86_CR4_PCIDE clear, or that have
+                * no INVPCID support at all.
+                */
+               if (boot_cpu_has(X86_FEATURE_INVPCID))
+                       setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
+       } else {
+               /*
+                * flush_tlb_all(), as currently implemented, won't work if
+                * PCID is on but PGE is not.  Since that combination
+                * doesn't exist on real hardware, there's no reason to try
+                * to fully support it, but it's polite to avoid corrupting
+                * data if we're on an improperly configured VM.
+                */
+               setup_clear_cpu_cap(X86_FEATURE_PCID);
        }
-#endif
 }
 
 #ifdef CONFIG_X86_32