rte_exit() does not terminate the program -- is it a bug or a new feature?

[Gabor LENCSE]

Dear Stephen,

Thank you very much for your reply. Please see my answers inline.

On 9/17/2023 11:27 PM, Stephen Hemminger wrote:
> On Sun, 17 Sep 2023 21:37:30 +0200
> Gabor LENCSE <lencse at hit.bme.hu> wrote:
>
>> However, l2fwd also uses the "rte_exit()" function to terminate the
>> program. The only difference is that it calls the "rte_exit()" function
>> from the main program, and I do so in a thread started by the
>> "rte_eal_remote_launch()" function.
> Calling rte_exit in a thread other than main thread won't work because
> the cleanup code is calling rte_eal_cleanup, and inside that it ends
> up waiting for all workers.  Since the thread you are calling from
> is a worker, it ends up waiting for itself.
>
> rte_exit()
> 	rte_eal_cleanup()
> 		rte_service_finalize()
> 			rte_eal_mp_wait_lcore()
>
>
> void
> rte_eal_mp_wait_lcore(void)
> {
> 	unsigned lcore_id;
>
> 	RTE_LCORE_FOREACH_WORKER(lcore_id) {
> 		rte_eal_wait_lcore(lcore_id);
> 	}
> }

Thank you very much for the detailed explanation!

I have modified the send function of siitperf (just at one point as a 
quick hack) and also the bash shell script. Now it works well and 
produces meaningful self-test results without any problem with program 
termination. So, the issue is solved. :-)

Of course, I will review all my rte_exit calls... It'll take a while...

I am just curious, as I have no idea, why my old code worked all right 
with DPDK 16.11. Has rte_exit() been changed since then?

(But, please do not spend too much time with this question!)

> Either service handling needs to be smarter, the rte_exit() function
> check if it is called from main lcore, and/or documentation needs update.
> Not a simple fix because in order to safely do the cleanup logic
> all threads have to gone to a quiescent state.

Yes I think it definitely SHOULD be mentioned in DPDK documentation that 
the rte_exit() function may be called only from the main lcore.

Now I have learnt it for all my life, but it will be useful for a lot of 
DPDK users. :-)

Once again, thank you very much for your help!

Best regards,

Gábor

p.s.: I have another, even worse problem with siitperf running with the 
new versions of DPDK. I will report it with a different subject line.


>