diff --git a/src/mesa/main/glthread.c b/src/mesa/main/glthread.c
index fb153af5ee9..a6e62516d02 100644
--- a/src/mesa/main/glthread.c
+++ b/src/mesa/main/glthread.c
@@ -340,6 +340,11 @@ _mesa_glthread_flush_batch(struct gl_context *ctx)
 
    struct glthread_batch *next = glthread->next_batch;
 
+   /* Mark the end of the batch, but don't increment "used". */
+   struct marshal_cmd_base *last =
+      (struct marshal_cmd_base *)&next->buffer[glthread->used];
+   last->cmd_id = NUM_DISPATCH_CMD;
+
    p_atomic_add(&glthread->stats.num_offloaded_items, glthread->used);
    next->used = glthread->used;
 
diff --git a/src/mesa/main/glthread.h b/src/mesa/main/glthread.h
index 773c2f6e7c0..1a5399e33df 100644
--- a/src/mesa/main/glthread.h
+++ b/src/mesa/main/glthread.h
@@ -33,7 +33,12 @@
  *   chance of experiencing CPU cache thrashing
  * but it should be high enough so that u_queue overhead remains negligible.
  */
-#define MARSHAL_MAX_CMD_SIZE (8 * 1024)
+#define MARSHAL_MAX_CMD_BUFFER_SIZE (8 * 1024)
+
+/* We need to leave 1 slot at the end to insert the END marker for unmarshal
+ * calls that look ahead to know where the batch ends.
+ */
+#define MARSHAL_MAX_CMD_SIZE (MARSHAL_MAX_CMD_BUFFER_SIZE - 8)
 
 /* The number of batch slots in memory.
  *
@@ -136,7 +141,7 @@ struct glthread_batch
    unsigned used;
 
    /** Data contained in the command buffer. */
-   uint64_t buffer[MARSHAL_MAX_CMD_SIZE / 8];
+   uint64_t buffer[MARSHAL_MAX_CMD_BUFFER_SIZE / 8];
 };
 
 struct glthread_client_attrib {