the "waiting" threads are not counted as active, and for that there is thd_wait_begin() / thd_wait_end() callbacks, to account for active vs waiting threads. There are at most thread_pool_size*thread_pool_oversubscribe active threads . The real maximum count worker of threads in pool is thread_pool_max_threads.

If you found a specific situation , where you think thd_wait_begin/end is missing, i.e where thread is counted as active and it is not, please provide a stacktrace of all threads, or pmp output

when thread call log_write_up_to, there is no thd_wait_begin or thd_wait_end. isn't it ?

No there is not. The idea is to put thd_wait_begin/thd_wait_end around waiting parts, not around the whole functions.
Therefore I'm asking about providing the callstacks, to find out the waiting part, which I'd guess is os_event_wait. You can add thd_wait_xx around it,and then you can evaluate how much thd_wait_begin/end helps in your benchmarks, and then we can evaluate in ours. There is unfortunately some danger that adding thd_wait_xxx can lead to creating too many threads. Adding it for long waits is no problem, but for microsecond-wait it might be counterproductive.
.