On Fri, May 25, 2018 at 5:20 AM Maxime Villard <max%m00nbsd.net@localhost> wrote:
Le 24/05/2018 à 21:13, Chuck Zmudzinski a écrit :
Well, the crash is repeatable on the one week old daily snapshot current
kernel. Again, here is the current kernel I am using:
NetBSD 8.99.17 (XEN3_DOMU) #0: Wed May 16 21:54:38 UTC 2018
mkrepro%mkrepro.NetBSD.org@localhost:/usr/src/sys/arch/xen/compile/XEN3_DOMU
What is happening is ... crazy.
With the current kernel, when the remote client connects, we get caught
in
an endless loop of creating ipsec security associations. The log shows
phase1
is created, then the phase2 associations, then we respond to negotiate
a new
phase1 and two new phase 2's, and I think this loop just continued
until we
ran out of memory. The windows client actually thought we were
connected and
showed it was connected in the network control panel, but the racoon log
never reported that a ppp interface was up. When you look at the
attached
snippets from the logs, I bet you will agree that many ppp interfaces
and
ipsec SAs were created and when we finally ran out of memory to create
another one, we crashed. I say this because the trace indicated the
crash
occurred at this branch. [1]. From the console at the start of the crash
report, I got this:
[ 334.5292103] panic: kernel diagnostic assertion "IFNET_LOCKED(ifp)"
failed: file "/usr/src/sys/net/if.c", line 3595
I don't understand line 3595 because if.c only has 661 lines, unless
there
was a mistake in how I copied it from the log.
You're looking at the wrong revision of if.c, yours seems to be [1].
The main issue here is that we reach this place with ifp unlocked. It's
probably not related to the system running out of memory.
That several entries get created in a loop, appears to be a separate
problem.
I know that several changes were made in netbsd-current for MPification.
It
may be that you exercise a particular condition that breaks an assumption
somewhere.
Ryota, Kengo, could you have a look?
I'm sorry I've looked the mail now.
Chuck, could you decode the backtrace of the panic? In this case the path
to the assertion (probably in if_mcast_op) is important.
Thanks,
ozaki-r