David Miller <da...@davemloft.net> wrote: > From: Florian Westphal <f...@strlen.de> > Date: Tue, 8 Aug 2017 18:02:29 +0200 > > > Unfortunately RTNL mutex is a performance issue, e.g. a cpu adding > > an ip address prevents other cpus from seemingly unrelated tasks > > such as dumping tc classifiers. > > It is related if somehow the TC entries refer to IP addresses. > > Someone could create something like that.
Actually I am not following. Why would read-only accesses need rtnl locking wrt. any other operation (provided of course rtnl lock doesn't protect the data structure)? > > Initial no-rtnl spots are ip6 fib add/del and netns new/getid. > > I could see the netns stuff being ok, but IPv6 route add/del I'm > not so sure of. [..] > There really is a hierachy of these dependencies. Device state, up > to neighbour table state, up to protocol address state, up to routes, > up to FIB tables, etc. etc. etc. > > I'd really like to make this operate more freely, but this is an > extremely delicate area which has been bottled up like this for > two decades so good luck :-) Would you accept a v2 if i don't touch ipv6 routes for the time being? I would then audit those again. At the very least inet6_rtm_getroute should be able to work without rtnl lock (i.e., use a different lock if needed to protect vs. concurrent modifications).
