diff --git a/src/platform/mailbox.rs b/src/platform/mailbox.rs
index 1712bbb..795c980 100644
--- a/src/platform/mailbox.rs
+++ b/src/platform/mailbox.rs
@@ -4,6 +4,7 @@ use crate::{
     println,
 };
 use core::ops::Deref;
+use core::sync::atomic::{compiler_fence, Ordering};
 use register::mmio::*;
 
 // Public interface to the mailbox
@@ -194,6 +195,11 @@ fn write(regs: &RegisterBlock, buf_ptr: u32, channel: u32) -> Result<()> {
 
     println!("Mailbox::write {:x}/{:x}", buf_ptr, channel);
 
+    // Insert a compiler fence that ensures that all stores to the
+    // mbox buffer are finished before the GPU is signaled (which is
+    // done by a store operation as well).
+    compiler_fence(Ordering::Release);
+
     while regs.STATUS.is_set(STATUS::FULL) {
         count += 1;
         if count > (1 << 25) {