RBD cache 是否开启的逻辑

void ObjectCacherObjectDispatch<I>::init() {
  auto cct = m_image_ctx->cct;
   ...
  m_writeback_handler = new LibrbdWriteback(m_image_ctx, m_cache_lock);

  uint64_t init_max_dirty = m_image_ctx->cache_max_dirty;
  if (m_image_ctx->cache_writethrough_until_flush) {
    init_max_dirty = 0;
  }
  ...
  m_object_cacher = new ObjectCacher(cct, m_image_ctx->perfcounter->get_name(),
                                     *m_writeback_handler, m_cache_lock,
                                     nullptr, nullptr, cache_size,
                                     10,  /* reset this in init */
                                     init_max_dirty, target_dirty,
                                     max_dirty_age, block_writes_upfront);

  ...
  m_object_cacher->start();
  m_image_ctx->io_object_dispatcher->register_object_dispatch(this);
}

在初始化ObjectCacherObjectDispatch的时候，先根据配置项rbd_journal_object_writethrough_until_flush是否初始化init_max_dirty。
rbd_journal_object_writethrough_until_flush默认是true，因此默认不开启cache。
因此我们有两种可能会导致rbd 的cache开启：

• 可能1：
  即直接把配置项rbd_journal_object_writethrough_until_flush设置为false
• 可能2：
  rbd_journal_object_writethrough_until_flush设置为true，等待客户端的使用者主动来调用flush接口，那么cache就会被开启，这从某种意义上来说是一种数据安全机制，毕竟主机主动调用flush，那意味着他能够忍受有部分数据在cache中，不透写到持久化层。
  逻辑如下：

template <typename I>
bool ObjectCacherObjectDispatch<I>::flush(
    io::FlushSource flush_source, const ZTracer::Trace &parent_trace,
    io::DispatchResult* dispatch_result, Context** on_finish,
    Context* on_dispatched) {
  auto cct = m_image_ctx->cct;
  ldout(cct, 20) << dendl;

  // ensure we aren't holding the cache lock post-flush
  on_dispatched = util::create_async_context_callback(*m_image_ctx,
                                                      on_dispatched);

  m_cache_lock.Lock();
  if (flush_source == io::FLUSH_SOURCE_USER && !m_user_flushed &&
      m_image_ctx->cache_writethrough_until_flush &&
      m_image_ctx->cache_max_dirty > 0) {
    m_user_flushed = true;
    m_object_cacher->set_max_dirty(m_image_ctx->cache_max_dirty);
    ldout(cct, 5) << "saw first user flush, enabling writeback" << dendl;
  }

  *dispatch_result = io::DISPATCH_RESULT_CONTINUE;
  m_object_cacher->flush_set(m_object_set, on_dispatched);
  m_cache_lock.Unlock();
  return true;
}