Re: [Patch net 1/2] net_sched: defer tcf_idr_insert() in tcf_action_init_1()

[Davide Caratti <dcaratti@redhat.com>]

hello,

On Fri, 2020-09-25 at 22:45 +0300, Vlad Buslov wrote:
> On Fri 25 Sep 2020 at 22:22, Cong Wang <xiyou.wangcong@gmail.com> wrote:
> > On Fri, Sep 25, 2020 at 8:24 AM Vlad Buslov <vlad@buslov.dev> wrote:
> > > > +     if (TC_ACT_EXT_CMP(a->tcfa_action, TC_ACT_GOTO_CHAIN) &&
> > > > +         !rcu_access_pointer(a->goto_chain)) {
> > > > +             tcf_action_destroy_1(a, bind);
> > > > +             NL_SET_ERR_MSG(extack, "can't use goto chain with NULL chain");
> > > > +             return ERR_PTR(-EINVAL);
> > > > +     }
> > > 
> > > I don't think calling tcf_action_destoy_1() is enough here. Since you
> > > moved this block before assigning cookie and releasing the module, you
> > > also need to release them manually in addition to destroying the action
> > > instance.
> > > 
> > 
> > tcf_action_destoy_1() eventually calls free_tcf() which frees cookie and
> > tcf_action_destroy() which releases module refcnt.
> > 
> > What am I missing here?
> > 
> > Thanks.
> 
> The memory referenced by the function local pointer "cookie" hasn't been
> assigned yet to the a->act_cookie because in your patch you moved
> goto_chain validation code before the cookie change. That means that if
> user overwrites existing action, then action old a->act_cookie will be
> freed by tcf_action_destroy_1() but new cookie that was allocated by
> nla_memdup_cookie() will leak.

maybe we can just delete this if (TC_ACT_EXT_CMP(...)) { ... }
statement, instead of moving it? Each TC action already does the check
for NULL "goto chains" with a_o->init() -> tcf_action_check_ctrlact(),
so this if () statement looks dead code to me _ I probably forgot to
remove it after all actions were converted to validate the control
action inside their .init() function.

-- 
davide