Hey Linus,
just a wild guess, but might it be possible that the run_workqueue() function
never runs empty of work queued up because packets are added all the time?
I'm not yet sure how where the problem might be, but reading your analysis it
does not seem like send_outstanding_packet() would
Okay, this third version seems to get rid of the nasty OOMs very
well and after stopping the ping-broadcast on Node A in a setup of
A - B - C, the broadcast storm stops nearly immediately.
However, in my VM setup (no matter if qemu or kvm) I'm getting a
soft-lockup on B. After about 10 seconds
And, hmm, yeah, seems to be the dev_queue_xmit() in
send_skb_packet() that needs those 80ms on congestion.
(though I'm still not sure if those 80ms lags are the cause of
the soft-lockups anyway)
Cheers, Linus
signature.asc
Description: Digital signature
This patch limits the queue lengths of batman and broadcast packets. BATMAN
packets are held back for aggregation and jittered to avoid interferences.
Broadcast packets are stored to be sent out multiple times to increase
the probability to be received by other nodes in lossy environments.
