Messages in this thread | | | Subject | Re: getaddrinfo slowdown in 3.17.1, due to getifaddrs | From | Eric Dumazet <> | Date | Fri, 17 Oct 2014 09:28:03 -0700 |
| |
On Fri, 2014-10-17 at 07:25 +0100, Thomas Graf wrote: > On 10/17/14 at 02:34am, Steinar H. Gunderson wrote: > > On Fri, Oct 17, 2014 at 02:21:32AM +0200, Steinar H. Gunderson wrote: > > > Hi, > > > > > > We recently upgraded a machine from 3.14.5 to 3.17.1, and a Perl script we're > > > running to poll SNMP suddenly needed ten times as much time to complete. > > > > e341694e3eb57fcda9f1adc7bfea42fe080d8d7a looks like it might cause something > > like this (it certainly added the synchronize_net() call). Cc-ing people on > > that commit; quoting the entire rest of the message for reference. > > I think the only option at this point is to re-add the nltable lock to > netlink_lookup() so we can drop the synchronize_net() until we find a > way to RCUify socket destruction. I will cook up a patch today unless > somebody can come up with a smarter way to work around needing the > synchronize_net().
I had a patch to add conditional RCUify socket destruction for some kind of sockets (opt-in at protocol level).
I needed this to remove too expensive UDP socket refcount inc/dec (so get rid of SLAB_DESTROY_BY_RCU, and instead use call_rcu() or kfree_rcu thing)
But I couldn't finish and submit this for 3.18, and I believe I got it wrong anyway. Cant remember exact details right now, maybe later in the day, once my headache is better.
| |