[SSSD] [PATCHES] HBAC performance enhancements

Jakub Hrozek jhrozek at redhat.com
Fri Oct 14 12:18:49 UTC 2011


On Thu, Oct 13, 2011 at 04:03:48PM -0400, Stephen Gallagher wrote:
> On Mon, 2011-10-10 at 13:25 +0200, Jakub Hrozek wrote:
> > On Fri, Oct 07, 2011 at 02:12:08PM -0400, Stephen Gallagher wrote:
> > > https://fedorahosted.org/sssd/ticket/1029
> > > 
> > > The problem here is that we were trying to perform an "optimization" by
> > > bulk-deleting the contents of the service and host lists in the sysdb
> > > before dumping into it the new data we received from LDAP.
> > > 
> > > This was causing a major performance hit on large deployments, because
> > > this recursive delete was repeatedly hitting a weak point of the
> > > memberOf plugin. However, upon closer analysis, Sumit pointed out that
> > > we don't actually need to rely on the local memberOf plugin in this
> > > situation.
> > > 
> > > These patches remove the member/memberOf relationship from
> > > host/hostgroup and service/servicegroup entries in the SSSD. As a
> > > result, we don't invoke the memberOf plugin during the mass-delete and
> > > we see a significant performance increase.
> > > 
> > > The patches [ab]use the fact that we know the DN structure of the hosts,
> > > service and groups so that we don't need to go and look them up when
> > > constructing the requests. Instead we take the originalMemberOf object
> > > and interpret the value directly from it. This is much faster than
> > > searching the sysdb for the original object to get its fqdn or cn value.
> > 
> > Patch 1/3 - Ack
> > 
> > Patch 2/3 - Nack,
> > In hbac_eval_service_element(), I assume you meant to loop until
> > i < el->num_values, count is always 1 there.
> > 
> > > +    for (i = j = 0; i < count; i++) {
> > > +        ret = get_ipa_servicegroupname(tmp_ctx, sysdb,
> > > +                                       (const char *)el->values[i].data,
> > > +                                       &name);
> > > +        if (ret != EOK && ret != ENOENT) goto done;
> > >  
> > > -        DEBUG(6, ("Added service group [%s] to the eval request\n",
> > > -                  svc->groups[i]));
> > > +        /* ENOENT means we had a memberOf entry that wasn't a
> > > +         * service group. We'll just ignore those (could be
> > > +         * HBAC rules)
> > > +         */
> > > +
> > > +        if (ret == EOK) {
> > > +            svc->groups[j] = talloc_steal(svc->groups, name);
> > > +            j++;
> > > +        }
> > 
> > Patch 3/3 - Nack, same comment as above. Also please change the comment in
> > get_ipa_hostgroupname() from "It's not a service." to "It's not a hostgroup."
> 
> Thanks, good catch.
> 
> New patches attached.

Ack to all three patches



More information about the sssd-devel mailing list