Double-checked locking over a std::atomic pointer

This lazy initializer is correct as written: callers dereference the returned pointer and read p->path and p->retries with no further synchronization. Which single change keeps every one of those reads free of data races?

main.cpp
struct Config { std::string path; int retries; };

std::atomic<Config*> g_cfg{nullptr};
std::mutex g_mtx;

Config* get() {
    Config* p = g_cfg.load(std::memory_order_acquire);      // (A)
    if (!p) {
        std::lock_guard<std::mutex> lk(g_mtx);
        p = g_cfg.load(std::memory_order_acquire);          // (B)
        if (!p) {
            p = new Config{"/etc/app.conf", 3};
            g_cfg.store(p, std::memory_order_release);      // (C)
        }
    }
    return p;
}

45 of 51 in Concurrency