diff --git a/src/mesa/main/glthread.c b/src/mesa/main/glthread.c
index 09de7f436d0..68bf9427260 100644
--- a/src/mesa/main/glthread.c
+++ b/src/mesa/main/glthread.c
@@ -121,6 +121,7 @@ _mesa_glthread_init(struct gl_context *ctx)
       util_queue_fence_init(&glthread->batches[i].fence);
    }
    glthread->next_batch = &glthread->batches[glthread->next];
+   glthread->used = 0;
 
    glthread->enabled = true;
    glthread->stats.queue = &glthread->queue;
@@ -207,8 +208,7 @@ _mesa_glthread_flush_batch(struct gl_context *ctx)
    if (!glthread->enabled)
       return;
 
-   struct glthread_batch *next = glthread->next_batch;
-   if (!next->used)
+   if (!glthread->used)
       return;
 
    /* Pin threads regularly to the same Zen CCX that the main thread is
@@ -230,6 +230,8 @@ _mesa_glthread_flush_batch(struct gl_context *ctx)
       }
    }
 
+   struct glthread_batch *next = glthread->next_batch;
+
    /* Debug: execute the batch immediately from this thread.
     *
     * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
@@ -241,13 +243,15 @@ _mesa_glthread_flush_batch(struct gl_context *ctx)
       return;
    }
 
-   p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
+   p_atomic_add(&glthread->stats.num_offloaded_items, glthread->used);
+   next->used = glthread->used;
 
    util_queue_add_job(&glthread->queue, next, &next->fence,
                       glthread_unmarshal_batch, NULL, 0);
    glthread->last = glthread->next;
    glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
    glthread->next_batch = &glthread->batches[glthread->next];
+   glthread->used = 0;
 }
 
 /**
@@ -280,8 +284,10 @@ _mesa_glthread_finish(struct gl_context *ctx)
       synced = true;
    }
 
-   if (next->used) {
-      p_atomic_add(&glthread->stats.num_direct_items, next->used);
+   if (glthread->used) {
+      p_atomic_add(&glthread->stats.num_direct_items, glthread->used);
+      next->used = glthread->used;
+      glthread->used = 0;
 
       /* Since glthread_unmarshal_batch changes the dispatch to direct,
        * restore it after it's done.
diff --git a/src/mesa/main/glthread.h b/src/mesa/main/glthread.h
index 677cbe672ea..73d1b22c02a 100644
--- a/src/mesa/main/glthread.h
+++ b/src/mesa/main/glthread.h
@@ -100,7 +100,12 @@ struct glthread_batch
    /** The worker thread will access the context with this. */
    struct gl_context *ctx;
 
-   /** Amount of data used by batch commands, in bytes. */
+   /**
+    * Amount of data used by batch commands, in bytes.
+    * This is 0 when it's being filled because glthread::used holds the real
+    * value temporarily, and glthread::used is copied to this variable when
+    * the batch is submitted.
+    */
    int used;
 
    /** Data contained in the command buffer. */
@@ -153,6 +158,9 @@ struct glthread_state
    /** Index of the batch being filled and about to be submitted. */
    unsigned next;
 
+   /** Amount of data filled in next_batch, in bytes. */
+   int used;
+
    /** Upload buffer. */
    struct gl_buffer_object *upload_buffer;
    uint8_t *upload_ptr;
diff --git a/src/mesa/main/glthread_marshal.h b/src/mesa/main/glthread_marshal.h
index 5acc36716d9..1ffb4df2e53 100644
--- a/src/mesa/main/glthread_marshal.h
+++ b/src/mesa/main/glthread_marshal.h
@@ -57,17 +57,15 @@ _mesa_glthread_allocate_command(struct gl_context *ctx,
                                 int size)
 {
    struct glthread_state *glthread = &ctx->GLThread;
-   struct glthread_batch *next = glthread->next_batch;
-   struct marshal_cmd_base *cmd_base;
 
-   if (unlikely(next->used + size > MARSHAL_MAX_CMD_SIZE)) {
+   if (unlikely(glthread->used + size > MARSHAL_MAX_CMD_SIZE))
       _mesa_glthread_flush_batch(ctx);
-      next = glthread->next_batch;
-   }
 
+   struct glthread_batch *next = glthread->next_batch;
    const int aligned_size = align(size, 8);
-   cmd_base = (struct marshal_cmd_base *)&next->buffer[next->used];
-   next->used += aligned_size;
+   struct marshal_cmd_base *cmd_base =
+      (struct marshal_cmd_base *)&next->buffer[glthread->used];
+   glthread->used += aligned_size;
    cmd_base->cmd_id = cmd_id;
    cmd_base->cmd_size = aligned_size;
    return cmd_base;