11.0 stuck on high network load

[Slawa Olhovchenkov]

On Thu, Sep 22, 2016 at 11:28:38AM +0200, Julien Charbon wrote:

> >>> What purpose to not skip locked tcptw in this loop?
> >>
> >>  If I understand your question correctly:  According to your pmcstat
> >> result, tcp_tw_2msl_scan() currently struggles with a write lock
> >> (__rw_wlock_hard) and the only write lock used tcp_tw_2msl_scan() is
> >> INP_WLOCK.  No sign of contention on TW_RLOCK(V_tw_lock) currently.
> > 
> > As I see in code, tcp_tw_2msl_scan got first node from V_twq_2msl and
> > need got RW lock on inp w/o alternates. Can tcp_tw_2msl_scan skip current node
> > and go to next node in V_twq_2msl list if current node locked by some
> > reasson?
> 
>  Interesting question indeed:  It is not optimal that all simultaneous
> calls to tcp_tw_2msl_scan() compete for the same oldest tcptw.  The next
> tcptws in the list are certainly old enough also.
> 
>  Let me see if I can make a simple change that makes kernel threads
> calling tcp_tw_2msl_scan() at same time to work on a different old
> enough tcptws.  So far, I found only solutions quite complex to implement.

Simple solution is skip in each thread ncpu elemnts and skip curent
cpu number elements at start, if I understund you correctly.