On Friday 08 November 2013 11:32:23 Weedy wrote:
> Do you have any benches or log dumps that show this fixed it for you?

I am not really sure what you are expecting. I can not really show a log of 
the reset because the device just resets when the oom kicks in or the process 
is just killed (so it is really boring). And the log of a working system is 
also extreme boring. What do you want to see when nothing happens?

But I recreated a test setup with 120 clients which connect every 30 sec and 
could easily see the extreme overbooking. It didn't happen after adding this 
patch. Also a field test was done and the system had a lot less problems with 
spontaneous resets.

> Or just dropping it to 2mb or less?

The default stack size for a thread is 2 MiB in the current implementation (1 
page is used as stack overflow protector and not usable). Setting it to less 
than 2 MiB is in my eyes relative useless for the standard case. Yes, it is 
possible to change the size of the stack before creating the thread but even 
programs running on embedded systems seem to ignore that fact and just use the 
default (with fatal consequences in my case).

> I shouldn't notice this. But anyone with running LUCI probably has enough
> process churn that this will hurt.

What are you trying to tell me here? Is LUCI using pthreads? Otherwise it 
would not be relevant for it.

I am not using LUCI and this is the first time hearing an insinuation about 
threading stuff in it. So if you think that it will hurt then please test it. 
Also please check first if it uses threads and not forked processes before 
doing the tests.

Kind regards,
        Sven

Attachment: signature.asc
Description: This is a digitally signed message part.

_______________________________________________
openwrt-devel mailing list
[email protected]
https://lists.openwrt.org/cgi-bin/mailman/listinfo/openwrt-devel

Reply via email to