On Fri, Jul 26, 2019 at 7:18 AM Nikolay Aleksandrov <[email protected]> wrote:
> > You've said it yourself - it has linear traversal time, but doesn't this > > patch allow any netns on the > > system to increase its limit to any value, thus possibly affecting others ? > > Though the socket limit will kick in at some point. I think that's where > > David > > was going with his suggestion back in 2018: > > https://www.spinics.net/lists/netdev/msg514543.html > > > > If we add this sysctl now, we'll be stuck with it. I'd prefer David's > > suggestion > > so we can rely only on the receive queue queue limit which is already > > configurable. > > We still need to be careful with the defaults though, the NOCACHE entry is > > 128 bytes > > and with the skb overhead currently on my setup we end up at about 277 > > entries default limit. > > I mean that people might be surprised if they increased that limit by > default, that's the > only problem I'm not sure how to handle. Maybe we need some hard limit anyway. > Have you done any tests what value works for your setup ? FYI: for ours, it is 2048.
