On 2021-02-18 13:52, Florian Westphal wrote:
> Richard Guy Briggs <rgb@xxxxxxxxxx> wrote:
> > On 2021-02-18 09:22, Florian Westphal wrote:
> > > > It seems I'd need to filter out the NFT_MSG_GET_* ops.
> > >
> > > No need, the GET ops do not cause changes and will not trigger a
> > > generation id change.
> >
> > Ah, so it could trigger on generation change. Would GET ops be included
> > in any other change
>
> No, GET ops are standalone, they cannot be part of a transaction.
> If you look at
>
> static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
>
> array in nf_tables_api.c, then those ops with a '.call_batch' can
> appear in transaction (i.e., can cause modification).
>
> The other ones (.call_rcu) are read-only.
>
> If they appear in a batch tehy will be ignored, if the batch consists of
> such non-modifying ops only then nf_tables_commit() returns early
> because the transaction list is empty (nothing to do/change).
Ok, one little inconvenient question: what about GETOBJ_RESET? That
looks like a hybrid that modifies kernel table counters and reports
synchronously. That could be a special case call in
nf_tables_dump_obj() and nf_tables_getobj(). Will that cause a storm
per commit?
> > such that it would be desirable to filter them out
> > to reduce noise in that single log line if it is attempted to list all
> > the change ops? It almost sounds like it would be better to do one
> > audit log line for each table for each family, and possibly for each op
> > to avoid the need to change userspace. This would already be a
> > significant improvement picking the highest ranking op.
>
> I think i understand what you'd like to do. Yes, that would reduce
> the log output a lot.
Coded, testing...
- RGB
--
Richard Guy Briggs <rgb@xxxxxxxxxx>
Sr. S/W Engineer, Kernel Security, Base Operating Systems
Remote, Ottawa, Red Hat Canada
IRC: rgb, SunRaycer
Voice: +1.647.777.2635, Internal: (81) 32635
