cvs commit: src/sys/netinet ip_divert.c
andre at freebsd.org
Tue Oct 19 16:12:20 PDT 2004
Robert Watson wrote:
> On Wed, 20 Oct 2004, Andre Oppermann wrote:
> > Robert Watson wrote:
> > >
> > > On Wed, 20 Oct 2004, Andre Oppermann wrote:
> > >
> > > > Hmm... I'll take a look at those attempts and see what I can come up
> > > > with to get some general solution for the protocol cases. The approach
> > > > of callout_drain() looks promising though.
> > >
> > > BTW, it looks like the divert pcb zone uses UMA_ZONE_NOFREE so that the
> > > memory is type-stable (presumably in particular for the sysctl), so all
> > > memory allocated by the divert module for pcbs is likely leaked on unload.
> > > I'm beginning to think we should just block unload for divert and solve
> > > the unload problem another day for another protocol...
> > It is indeed an oversight by me not to uma_zdestroy() the zone on
> > unload. The sysctl handler uses normal malloc.
> If it's like other instances of type-stable storage, it's so that weak
> consistency can be used by the monitoring sysctl in order to avoid
> allocating lots of memory and/or suspending operation while monitoring
> takes place. Other instances of type-stable storage in the socket code
> and network stack rely on the "chain'o'pcbs" list pointers not being
> cleared on allocation and free.
Sigh... What a mess... In the IPDIVERT case the sysctl can be protected
with a separate (non-adaptive) mutex. We only unload if there are no
open sockets, the protocol is unloaded so there should be no cases where
weak list pointers are used.
I am out of the office tomorrow doing consulting work. I may have time
to check into this discussion again tomorrow evening or otherwise on
More information about the cvs-src