On Tue, May 20, 2014 at 08:04:23PM +0100, Mindaugas Rasiukevicius wrote: > Nicolas Joly <nj...@pasteur.fr> wrote: > > > > njoly@lanfeust [/misc/crash]> crash -M netbsd.17.core -N netbsd.17 > > Crash version 6.99.42,image version 6.99.42. > > System panicked: kernel diagnostic assertion "KERNEL_LOCKED_P()"failed: > > file "/local/src/NetBSD/src/sys/net/if_ethersubr.c", line 214 Backtrace > > from time of crash is available. > > Fixed. We may start pushing the kernel-lock to the entry points of various > subsystems (like Ethernet, in this case), but wrapping attach/detach is more > defensive way for now.
the IP stack is not supported to run under KERNEL_LOCK (and actually is not, because ip_input() is explicitely called without the KERNEL_LOCK held. The ethernet subsystem (and probably other link-layer subsystems) still needs the kernel lock, so grabing is again before sending a packet is the way to go (and is what IP is doing). This is why I added this KASSERT(). Running the whole IP stack under KERNEL_LOCK would be a regression. -- Manuel Bouyer <bou...@antioche.eu.org> NetBSD: 26 ans d'experience feront toujours la difference --