diff --git a/arch/x86_64/src/common/x86_64_exit.c b/arch/x86_64/src/common/x86_64_exit.c
index 70b210a83c7..6070cc82a60 100644
--- a/arch/x86_64/src/common/x86_64_exit.c
+++ b/arch/x86_64/src/common/x86_64_exit.c
@@ -87,23 +87,19 @@ void up_exit(int status)
 
   x86_64_restore_auxstate(tcb);
 
-  /* Restore the cpu lock */
-
-  restore_critical_section(tcb, this_cpu());
-
 #ifdef CONFIG_ARCH_KERNEL_STACK
   /* Update kernel stack top pointer */
 
   x86_64_set_ktopstk(tcb->xcp.ktopstk);
 #endif
 
-  /* Then switch contexts */
-
-  x86_64_fullcontextrestore(tcb->xcp.regs);
-
-  /* x86_64_fullcontextrestore() should not return but could if the software
-   * interrupts are disabled.
+  /* Restore the cpu lock.  This must come last and the final jump must
+   * not touch the stack (see up_switch_context()).
    */
 
-  PANIC();
+  restore_critical_section(tcb, this_cpu());
+
+  __asm__ volatile ("jmp x86_64_fullcontextrestore"
+                    :: "D" (tcb->xcp.regs) : "memory");
+  __builtin_unreachable();
 }
diff --git a/arch/x86_64/src/common/x86_64_switchcontext.c b/arch/x86_64/src/common/x86_64_switchcontext.c
index 9ddf056f37a..721d005f40d 100644
--- a/arch/x86_64/src/common/x86_64_switchcontext.c
+++ b/arch/x86_64/src/common/x86_64_switchcontext.c
@@ -86,6 +86,7 @@ void up_switch_context(struct tcb_s *tcb, struct tcb_s *rtcb)
   else if (!up_saveusercontext(rtcb->xcp.regs))
     {
       struct tcb_s **running_task;
+
       cpu = this_cpu();
 
       x86_64_restore_auxstate(tcb);
@@ -101,10 +102,6 @@ void up_switch_context(struct tcb_s *tcb, struct tcb_s *rtcb)
       tcb = this_task();
 #endif
 
-      /* Restore the cpu lock */
-
-      restore_critical_section(tcb, cpu);
-
       /* Update scheduler parameters */
 
       running_task = &g_running_tasks[cpu];
@@ -117,8 +114,16 @@ void up_switch_context(struct tcb_s *tcb, struct tcb_s *rtcb)
 
       *running_task = tcb;
 
-      /* Then switch contexts */
+      /* Restore the cpu lock.  This makes the outgoing task wakeable by
+       * other CPUs while this CPU still runs on the outgoing task's
+       * stack, so it must come last and the final jump must not touch
+       * the stack (a call would push the return address onto it).
+       */
 
-      x86_64_fullcontextrestore(tcb->xcp.regs);
+      restore_critical_section(tcb, cpu);
+
+      __asm__ volatile ("jmp x86_64_fullcontextrestore"
+                        :: "D" (tcb->xcp.regs) : "memory");
+      __builtin_unreachable();
     }
 }