The following reply was made to PR kern/53043; it has been noted by GNATS.
From: Ryota Ozaki <ozaki-r%netbsd.org@localhost>
To: "gnats-bugs%NetBSD.org@localhost" <gnats-bugs%netbsd.org@localhost>
Cc: kern-bug-people%netbsd.org@localhost, gnats-admin%netbsd.org@localhost, netbsd-bugs%netbsd.org@localhost
Subject: Re: kern/53043: deadlock on evbarm/TEGRA with netbsd-8
Date: Thu, 22 Feb 2018 17:24:13 +0900
This is a deadlock that occurs between softnet_lock and IFNET_LOCK;
the locks can be held in different orders.
softnet_lock in in6_control (and in_control) had been introduced to
address PR 51356 that was a race condition between ioctls and packet
inputs and forwarding (ipintr and ip6intr). However holding it
in6_control and in_control was not a good idea because softnet_lock
should be basically held at the very beginning of call paths to comply
the locking order and the functions were in the middle of call paths.
I think we have two options to solve the issue:
(1) Give up relying on softnet_lock to protect the network stack and
remove softnet_lock from in6_control/in_control and add some
KERNEL_LOCK to the network stack, e.g., ipintr and ip6intr.
(2) Just get rid of softnet_lock from in6_control/in_control.
(1) is safer than (2) but add some performance penalty. (2) sounds
awful but the situation is the same as netbsd-7 and netbsd-6, i.e.,
it's enough safe in practice...
A patch for (1) is here: http://www.netbsd.org/~ozaki-r/fix-pr53043.diff
Note that I gave up moving softnet_lock to doifioctl because it
just introduced other deadlocks and was more problematic.
Comments?