Performance schema suffers from check-then act anti-pattern
For example init_setup_actor_hash:

  if ((! setup_actor_hash_inited) && (param->m_setup_actor_sizing != 0))

  {

    lf_hash_init(&setup_actor_hash, sizeof(PFS_setup_actor*), LF_HASH_UNIQUE,

                 0, 0, setup_actor_hash_get_key, &my_charset_bin);

    /* setup_actor_hash.size= param->m_setup_actor_sizing; */

    setup_actor_hash_inited= true;

  }

The compiler or CPU can reorder setup_actor_hash_inited initialization before lf_hash_init call.
At the same time, get_setup_actor_hash_pins

    if (! setup_actor_hash_inited)

      return NULL;

    thread->m_setup_actor_hash_pins= lf_hash_get_pins(&setup_actor_hash);

can be called in parallel, resulting, in fact, in having garbage in setup_actor_hash.

Same for the other modules. Basically, everything called in
initialize_performance_schema is worth checking.

Probably, PFS_buffer_scalable_container's m_initialized is not affected.