Actually, those lockups are not unexpected for me anymore. I can
predict them by summing up the RSS of all running processes. It usually
sums up to about 80MB, at the time it locks up, the sum of RSSs is about
10MB and lots of swap are in use. Is there any easier way to find total
kernel memory?
It seems a part of the networking code is not freeing skbufs. I
haven't find out what, yet. Andreas Degert offered a patch to track
kernel memory. As soon as I have that patch I'll inform you on where
are the problems coming from.
A hint: use shift-scroll lock and take a look at your syslog.
Check the ratio of "Networking buffers in use" and "Total network buffer
allocations". The ratio for me is always at about 1:25. The more
allocations there are the more buffers in use.
After about a day the number of skbufs is about 100 000. What is
the size of a skbuf? If they averaged 100 bytes, I would have 10 MB
taken by kernel memory.
BTW, It seems strange that so few people seemed to have
experienced that problem. I experienced it MANY times.
Regards,
Andre
-----
Andre Uratsuka Manoel Insite Internet
andre@insite.com.br +55-11-259-6979