Re: High 50us+ latencies in the process signal handling path

[Michal Schmidt <xschmi00@stud.feec.vutbr.cz>]

Ankita Garg wrote:
> Hi,
>
> There is a testcase called 'pthread_kill_latency' under the rt-test suite 
> on the the RT Wiki:
>
> http://www.kernel.org/pub/linux/kernel/people/dvhart/realtime/tests/rt-test-0.3.tar.bz2
>
> Description of the testcase: 
>
> The test measures the latency involved in sending a signal to a thread 
> using pthread_kill. Two threads are created, receiver thread (thread1) 
> and sender thread (thread2). Before sending the signal, it waits for 
> thread1 to initialize (thread1 waits by calling sigwait), notes the time and
> send pthread_kill signal to thread1. thread1 notes the time it receives the
> signal. The maximum and the minimum latency is reported. This is repeated
> about 10000 times.
>   

I'm looking at the source. The receiver thread runs with realtime 
priority (it's created with create_fifo_thread), but the sender thread 
(thread2) doesn't (create_other_thread). The sender does:

        /* Record the time just before sending the signal */
        begin = rt_gettime();
/**** The sender can get preempted by anything right here ! ****/
        if ((ret = pthread_kill(PTHREADOF(target_thread), SIGNALNUMBER))) {
            printf("pthread_kill returned %d\n", ret);
        }

Can this explain it?

Michal