PERFORCE change 167260 for review

[Marko Zec]

On Thursday 13 August 2009 02:22:22 Julian Elischer wrote:
> Marko Zec wrote:
> > On Thursday 13 August 2009 00:44:49 Julian Elischer wrote:
> >> Marko Zec wrote:
> >>> On Wednesday 12 August 2009 23:58:46 Julian Elischer wrote:
> >>>> Marko Zec wrote:
> >>>
> >>> ...
> >>>
> >>>>> @@ -710,22 +715,36 @@
> >>>>>  	.pr_input =		div_input,
> >>>>>  	.pr_ctlinput =		div_ctlinput,
> >>>>>  	.pr_ctloutput =		ip_ctloutput,
> >>>>> -	.pr_init =		NULL,
> >>>>> +	.pr_init =		div_init,
> >>>>>  	.pr_usrreqs =		&div_usrreqs
> >>>>
> >>>> If you are going to make pr_init() called for every vnet then
> >>>> pr_destroy should be as well. But in fact that is not really safe.
> >>>> (either of them)
> >>>>
> >>>> The trouble is that we can not guarantee that other protocols can
> >>>> handle being called multiple times in their init and destroy methods.
> >>>> Especially 3rd party protocols.
> >>>>
> >>>> We need to ensure only protocols that have been converted to run
> >>>> with multiple vnets are ever called with multiple vnets.
> >>>>
> >>>> for this reason the only safe way to do this is via the VNET_SYSINIT
> >>>> and VNET_SYSUNINIT calls.
> >>>
> >>> That would mean you would have to convert most if not all of the
> >>> existing things that hang off of protosw-s in netinet, netinet6 etc. to
> >>> use VNET_SYSINT / VNET_SYSUNIT instead of protosw->pr_init().  So the
> >>> short answer is no.
> >>
> >> robert has done just that.
> >
> > hmm:
> >
> > tpx32% pwd
> > /u/marko/svn/head/sys
> >
> > tpx32% fgrep -R .pr_init netinet netinet6 netipsec|fgrep -v .svn
> > netinet/ip_divert.c:    .pr_init =              div_init,
> > netinet/in_proto.c:     .pr_init =              ip_init,
> > netinet/in_proto.c:     .pr_init =              udp_init,
> > netinet/in_proto.c:     .pr_init =              tcp_init,
> > netinet/in_proto.c:        .pr_init =   sctp_init,
> > netinet/in_proto.c:     .pr_init =              icmp_init,
> > netinet/in_proto.c:     .pr_init =              encap_init,
> > netinet/in_proto.c:     .pr_init =              encap_init,
> > netinet/in_proto.c:     .pr_init =              encap_init,
> > netinet/in_proto.c:     .pr_init =              encap_init,
> > netinet/in_proto.c:     .pr_init =              encap_init,
> > netinet/in_proto.c:     .pr_init =              rip_init,
> > netinet6/in6_proto.c:   .pr_init =              ip6_init,
> > netinet6/in6_proto.c:   .pr_init =              tcp_init,
> > netinet6/in6_proto.c:   .pr_init =              icmp6_init,
> > netinet6/in6_proto.c:   .pr_init =              encap_init,
> > netinet6/in6_proto.c:   .pr_init =              encap_init,
> > netinet6/ip6_mroute.c:  .pr_init =              pim6_init,
> > netipsec/keysock.c:     .pr_init =              raw_init,
>
> AND for example:
> in ./netinet/in_proto.c
> VNET_DOMAIN_SET(inet);
> includes
> VNET_SYSINIT   ##### --> called for every vnet as created ####
>     calls
> vnet_domain_init()
>     calls
> domain_init()
>     calls
> protosw_init()
>     which includes
>               if (pr->pr_init)
>                  (*pr->pr_init)();
>
> so, robert is calling the init routine from each protocol
> not the modevent.

Right.

But when we kldload ipdivert and register its protosw via pf_proto_register(), 
VNET_DOMAIN_SET() will not call ipdivert's pr_init() routine for each 
existing vnet, hence pf_proto_register() will have to do this.

For subsequently created vnets, i.e. after ipdivert has been kldloaded, 
pr_init() will indeed be called via VNET_DOMAIN_SET() mechanism on each vnet 
instantiation.  But at that point in time pf_proto_register() will not be 
called.  So, in both cases pr_init() will be called exactly once per vnet, 
but via different mechanisms.

Marko