diff --git a/vhost/src/vhost_user/connection.rs b/vhost/src/vhost_user/connection.rs
index 82cc225..92995cf 100644
--- a/vhost/src/vhost_user/connection.rs
+++ b/vhost/src/vhost_user/connection.rs
@@ -226,7 +226,7 @@ impl<H: MsgHeader> Endpoint<H> {
         body: &T,
         fds: Option<&[RawFd]>,
     ) -> Result<()> {
-        if mem::size_of::<T>() > MAX_MSG_SIZE {
+        if mem::size_of::<T>() > H::MAX_MSG_SIZE {
             return Err(Error::OversizedMsg);
         }
         let bytes = self.send_iovec_all(&[hdr.as_slice(), body.as_slice()], fds)?;
@@ -255,10 +255,10 @@ impl<H: MsgHeader> Endpoint<H> {
         fds: Option<&[RawFd]>,
     ) -> Result<()> {
         let len = payload.len();
-        if mem::size_of::<T>() > MAX_MSG_SIZE {
+        if mem::size_of::<T>() > H::MAX_MSG_SIZE {
             return Err(Error::OversizedMsg);
         }
-        if len > MAX_MSG_SIZE - mem::size_of::<T>() {
+        if len > H::MAX_MSG_SIZE - mem::size_of::<T>() {
             return Err(Error::OversizedMsg);
         }
         if let Some(fd_arr) = fds {
diff --git a/vhost/src/vhost_user/gpu_message.rs b/vhost/src/vhost_user/gpu_message.rs
index a070ce3..38ad7ae 100644
--- a/vhost/src/vhost_user/gpu_message.rs
+++ b/vhost/src/vhost_user/gpu_message.rs
@@ -168,4 +168,5 @@ impl<T: Req> VhostUserMsgValidator for VhostUserGpuMsgHeader<T> {
 
 impl<R: Req> MsgHeader for VhostUserGpuMsgHeader<R> {
     type Request = R;
+    const MAX_MSG_SIZE: usize = u32::MAX as usize;
 }
diff --git a/vhost/src/vhost_user/message.rs b/vhost/src/vhost_user/message.rs
index 8eb06ee..68333af 100644
--- a/vhost/src/vhost_user/message.rs
+++ b/vhost/src/vhost_user/message.rs
@@ -25,6 +25,15 @@ use vm_memory::{GuestAddress, MmapRange, MmapXenFlags};
 use super::{enum_value, Error, Result};
 use crate::VringConfigData;
 
+/*
+TODO: Consider deprecating this. We don't actually have any preallocated buffers except in tests,
+so we should be able to support u32::MAX normally.
+Also this doesn't need to be public api, since Endpoint is private anyway, this doesn't seem
+useful for consumers of this crate.
+
+There are GPU specific messages (GpuBackendReq::UPDATE and CURSOR_UPDATE) that are larger than 4K.
+We can use MsgHeader::MAX_MSG_SIZE, if we want to support larger messages only for GPU headers.
+*/
 /// The vhost-user specification uses a field of u32 to store message length.
 /// On the other hand, preallocated buffers are needed to receive messages from the Unix domain
 /// socket. To preallocating a 4GB buffer for each vhost-user message is really just an overhead.
@@ -58,6 +67,9 @@ pub(super) trait Req:
 
 pub(super) trait MsgHeader: ByteValued + Copy + Default + VhostUserMsgValidator {
     type Request: Req;
+
+    /// The maximum size of a msg that can be encapsulated by this MsgHeader
+    const MAX_MSG_SIZE: usize;
 }
 
 enum_value! {
@@ -228,6 +240,7 @@ pub(super) struct VhostUserMsgHeader<R: Req> {
 
 impl<R: Req> MsgHeader for VhostUserMsgHeader<R> {
     type Request = R;
+    const MAX_MSG_SIZE: usize = MAX_MSG_SIZE;
 }
 
 impl<R: Req> Debug for VhostUserMsgHeader<R> {