On Sat, Mar 09, 2019 at 11:19:11PM +0100, Michał Koc wrote:
> W dniu 09.03.2019 o 12:43, Otto Moerbeek pisze:
> > On Sat, Mar 09, 2019 at 12:10:34PM +0100, Michał Koc wrote:
> >
> > > W dniu 09.03.2019 o 08:15, Otto Moerbeek pisze:
> > > > On Fri, Mar 08, 2019 at 12:03:25PM +0100, Michał Koc wrote:
> > > >
> > > > > Hi all,
> > > > >
> > > > > We have a triple redundant vpn gateway setup with sasyncd running and
> > > > > tons
> > > > > of tunnels, about 1000 flows.
> > > > >
> > > > > Looking at the graph of memory usage, you can clearly see that
> > > > > something is
> > > > > sucking up the memory.
> > > > >
> > > > > The graph can be viewed here: https://pasteboard.co/I4sjzQ8.jpg
> > > > >
> > > > > Looking at the ps, sasyncd shows huge memory consumption:
> > > > >
> > > > > USER PID %CPU %MEM VSZ RSS TT STAT
> > > > > STARTED TIME COMMAND
> > > > > _isakmpd 33560 0.0 17.0 699264 708508 ?? S
> > > > > 26Feb19 6:58.81 /usr/sbin/sasyncd
> > > > >
> > > > > It only happens on the master node. Slaves do not show such a
> > > > > behavior.
> > > > >
> > > > > There is nothing about sasyncd in the logs.
> > > > >
> > > > > After sasyncd restart memory consumption is minimal, but tends to
> > > > > grow.
> > > > >
> > > > > Is it normal ? or am I missing something ?
> > > > >
> > > > > Best regards
> > > > > M.K.
> > > > >
> > > > This is not normal. You could try to run with -vv to see if some error
> > > > path is taken that triggers a leak.
> > > >
> > > > -Otto
> > > >
> > > Should I look for something specific ?
> > >
> > > The log grows pretty fast and it looks like it could contain some security
> > > data which I wouldn't like to post online.
> > >
> > > The statistics of the log(about 2 hours) looks like this:
> > > carp_init: 1
> > > config: 7
> > > monitor_get_pfkey_snap: 4
> > > monitor_loop: 1
> > > net: 1
> > > net_connect: 3
> > > net_ctl: 4
> > > net_disconnect_peer: 3
> > > net_handle_messages: 2
> > > net_queue: 91780
> > > net_read: 10
> > > net_send_messages: 39192
> > > pfkey_send_flush: 4
> > > pfkey_snapshot: 6832
> > > timer_add: 19
> > > timer_run: 18
> > >
> > > Best regards
> > > M.K.
> > >
> > Just the counts does not reveal anything. I did a quick audit of the
> > memory allocation logic of sasyncd and did not spot an error. If you
> > do not want to post the logs, you'll neeed to analyze them yourself.
> > This requires matching the log lines to the code and tracking where
> > stuff gets allocated and deallocated. Some digging could reveal the error.
> >
> > I used to run sasyncd, but I do no longer. Settig up a test env is
> > quite some work I do not have time for.
> >
> > -Otto
> >
> First of all, thank You for your time. I know it's one of the most valuable
> resource.
>
> We have done some analysis and we have found the problem.
>
> The problem is that the very master machine exists as a peer in it's config.
> The purpose of this is to make all of the config files to be as similar as
> possible for all of the members of the cluster.
>
> Removing it from peers fixes the problem.
>
> So there are three main roads we can follow:
> 1. Fix sasyncd to recognize self and handle it properly (a result)
> 2. It should be mentioned in manual, not to set self as a peer (an excuse)
> 3. We can change our internal config handling (no one else benefits)
>
> What would You recommend as a next step ? we will do much to follow road 1,
> but we might need help, eg. code review and some guidance to meet OpenBSD
> needs
>
> Furthermore if we follow road 1, is there any chance to get the reviewed,
> correct, accepted fix into the tree ?
>
> Best regards
> M.K.
>
Good you found the problem.
As for number 1, in general it is hard to determine if an IP is your
own (or redirects to yourself), especially since network interfaces
can be added and changed dynamically. But a starup check against
getifaddrs(3) would catch the most obvious I'm my own peer cases.
A dynamic form of loop detection would be nice, it would require
adding a hostid or equivalent. That is not normally set these days and
also would change the wire format of the messages.
But even if you start work on code, a manual page change warning
against this would be a good thing.
If you post diffs to tech@ with a proper explanation they will be
picked up and reviewed, either by me or somee other developer.
BTW, I noticed there are order dependencies in sasyncd.cnf that are
not mentioned in the man page. Also the example in
src/etc/examples/sasyncd.conf does not work if you just uncomment the
lines because of those order dependencies. So there's more room for
improvement in the man page and example config.
-Otto
