svn commit: r238990 - in head/sys: net netinet netinet6

[Kenneth D. Merry]

On Fri, Aug 03, 2012 at 01:33:44 +0400, Gleb Smirnoff wrote:
> On Thu, Aug 02, 2012 at 04:46:42PM +0000, Bjoern A. Zeeb wrote:
> B> On Thu, 2 Aug 2012, Gleb Smirnoff wrote:
> B> 
> B> > Author: glebius
> B> > Date: Thu Aug  2 13:57:49 2012
> B> > New Revision: 238990
> B> > URL: http://svn.freebsd.org/changeset/base/238990
> B> >
> B> > Log:
> B> >  Fix races between in_lltable_prefix_free(), lla_lookup(),
> B> >  llentry_free() and arptimer():
> B> >
> B> >  o Use callout_init_rw() for lle timeout, this allows us safely
> B> >    disestablish them.
> B> >    - This allows us to simplify the arptimer() and make it
> B> >      race safe.
> B> >  o Consistently use ifp->if_afdata_lock to lock access to
> B> >    linked lists in the lle hashes.
> B> >  o Introduce new lle flag LLE_LINKED, which marks an entry that
> B> >    is attached to the hash.
> B> >    - Use LLE_LINKED to avoid double unlinking via consequent
> B> >      calls to llentry_free().
> B> >    - Mark lle with LLE_DELETED via |= operation istead of =,
> B> >      so that other flags won't be lost.
> B> >  o Make LLE_ADDREF(), LLE_REMREF() and LLE_FREE_LOCKED() more
> B> >    consistent and provide more informative KASSERTs.
> B> >
> B> >  The patch is a collaborative work of all submitters and myself.
> B> 
> B> Quoting from 2 year old memory you just introduced a possible deadlock
> B> on tbale (or with that networkstack) teardown adding the extra af_data
> B> write locking to the table walk.
> 
> Can you please give more details? I didn't get it. What else needs
> afdata lock and what does it hold which is held by table walk.

Is there a deadlock in that particular change?

If so, what will it take to fix it?

Thanks,

Ken
-- 
Kenneth Merry
ken at FreeBSD.ORG