On Thu, 28 Sep 2023, Pablo Neira Ayuso wrote:
> On Thu, Sep 28, 2023 at 08:57:45PM +0200, Florian Westphal wrote:
> > Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote:
> > > On Thu, Sep 28, 2023 at 07:46:30PM +0200, Florian Westphal wrote:
> > > > Phil Sutter <phil@xxxxxx> wrote:
> > > > > +static int nf_tables_dumpreset_set(struct sk_buff *skb,
> > > > > + struct netlink_callback *cb)
> > > > > +{
> > > > > + struct nftables_pernet *nft_net = nft_pernet(sock_net(skb->sk));
> > > > > + struct nft_set_dump_ctx *dump_ctx = cb->data;
> > > > > + int ret, skip = cb->args[0];
> > > > > +
> > > > > + mutex_lock(&nft_net->commit_mutex);
> > > > > + ret = nf_tables_dump_set(skb, cb);
> > > > > + mutex_unlock(&nft_net->commit_mutex);
> > > > > +
> > > > > + if (cb->args[0] > skip)
> > > > > + audit_log_nft_set_reset(dump_ctx->ctx.table, cb->seq,
> > > > > + cb->args[0] - skip);
> > > > > +
> > > >
> > > > Once commit_mutex is dropped, parallel user can
> > > > delete table, and ctx.table references garbage.
> > >
> > > This path should hold rcu read lock.
> >
> > Then it would splat on first mutex_lock().
> >
> > > I think spinlock would be better, we would just spin for very little
> > > time here for another thread to complete the reset, and the race is
> > > fixed.
> > >
> > > The use of commit_mutex here is confusing is really misleading to the
> > > reader, this is also not the commit path.
> >
> > I'd say its semantically the same thing, we alter state.
> >
> > We even emit audit records to tell userspace that there is state
> > change.
>
> This is a netlink event, how does the mutex help in that regard?
>
> > Also, are you sure spinlock is appropriate here?
> >
> > For full-ruleset resets we might be doing quite some
> > traverals.
>
> I said several times, we grab and release this for each
> netlink_recvmsg(), commit_mutex helps us in no way to fix the "two
> concurrent dump-and-reset problem".
>
> One concern might be deadlock due to reordering, but I don't see how
> that can happen.
The same problem exists ipset: when a set is listed/saved (dumped),
concurrent destroy/rename/swap for the same set must be excluded. As
neither spinlock nor mutex helps, a reference counter is used: the start
of the dump increases it and by checking it all concurrent events can
safely be rejected by returning EBUSY.
Best regards,
Jozsef
-
E-mail : kadlec@xxxxxxxxxxxxxxxxx, kadlecsik.jozsef@xxxxxxxxx
PGP key : https://wigner.hu/~kadlec/pgp_public_key.txt
Address : Wigner Research Centre for Physics
H-1525 Budapest 114, POB. 49, Hungary
