Re: kern/50186: sparc memfault panic after 7.99.21 ARP changes

[Ryota Ozaki <ozaki-r%netbsd.org@localhost>]

On Fri, Sep 4, 2015 at 5:40 PM, Christos Zoulas <christos%zoulas.com@localhost> wrote:
> On Sep 4,  1:50pm, ozaki-r%netbsd.org@localhost (Ryota Ozaki) wrote:
> -- Subject: Re: kern/50186: sparc memfault panic after 7.99.21 ARP changes
>
> | On Thu, Sep 3, 2015 at 8:38 PM, Christos Zoulas <christos%zoulas.com@localhost> wrote:
> | > I just crashed in arptimer() so there are more locking problems in the code. Can you document the locking discipline for la_rt and changing the lists?
> | >
> |
> | I'm sorry for the defect.
> |
> | An ARP cache list of an interface is protected by a rwlock
> | (IF_AFDATA_*LOCK) and each ARP cache is protected by a rwlock
> | and refernce counting (LLE_*LOCK). However, la_rt still needs
> | softnet_lock; if la_rt is accessed or modified without
> | softnet_lock, it's a bug. And I found a bug :( lltable_free
> | accesses la_rt but it's called without softnet_lock.
> |
> | Here is a patch:
> | http://www.netbsd.org/~ozaki-r/lltable_free-softnet_lock.diff
> | Could you try it?
> |
>
> Thanks, I am running with it now. Should we revert the KASSERT change
> too?

Well, yes and no. Because KASSERT was actually wrong; la_rt can be NULL
at the point according to my investigation for PR 50184. So anyway we
have to get rid of it.

I made a patch for the bug:
http://www.netbsd.org/~ozaki-r/fix-PR50184.take2.diff
which was for PR 50184. So reverting your commit and applying the patch
instead might be easy for me. Of course rebasing my patch on the HEAD
makes no difference though.

 ozaki-r