Re: [PATCH] net: sched: fix memory leak in tcindex_set_parms

[Hawkins Jiawei]

On Mon, 7 Nov 2022 at 01:49, Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
>
> On Sun, Nov 06, 2022 at 10:55:31PM +0800, Hawkins Jiawei wrote:
> > Hi Cong,
> >
> > >
> > >
> > > diff --git a/net/sched/cls_tcindex.c b/net/sched/cls_tcindex.c
> > > index 1c9eeb98d826..00a6c04a4b42 100644
> > > --- a/net/sched/cls_tcindex.c
> > > +++ b/net/sched/cls_tcindex.c
> > > @@ -479,6 +479,7 @@ tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
> > >         }
> > >
> > >         if (old_r && old_r != r) {
> > > +               tcf_exts_destroy(&old_r->exts);
> > >                 err = tcindex_filter_result_init(old_r, cp, net);
> > >                 if (err < 0) {
> > >                         kfree(f);
> >
> > As for the position of the tcf_exts_destroy(), should we
> > call it after the RCU updating, after
> > `rcu_assign_pointer(tp->root, cp)` ?
> >
> > Or the concurrent RCU readers may derefer this freed memory
> > (Please correct me If I am wrong).
>
> I don't think so, because we already have tcf_exts_change() in multiple
> places within tcindex_set_parms(). Even if this is really a problem,

Do you mean that, if this is a problem, then these tcf_exts_change()
should have already triggered the Use-after-Free?(Please correct me
if I get wrong)

But it seems that these tcf_exts_change() don't destory the old_r,
so it doesn't face the above concurrent problems.

I find there are two tcf_exts_chang() in tcindex_set_parms().
One is

	oldp = p;
	r->res = cr;
	tcf_exts_change(&r->exts, &e);

	rcu_assign_pointer(tp->root, cp);

the other is

	f->result.res = r->res;
	tcf_exts_change(&f->result.exts, &r->exts);

	fp = cp->h + (handle % cp->hash);
	for (nfp = rtnl_dereference(*fp);
	     nfp;
	     fp = &nfp->next, nfp = rtnl_dereference(*fp))
			; /* nothing */

	rcu_assign_pointer(*fp, f);

*r->exts* or *f->result.exts*, both are newly allocated in
`tcindex_set_params()`, so the concurrent RCU readers won't read them
before RCU updating.

> moving it after rcu_assign_pointer() does not help, you need to wait for
> a grace period.

Yes, you are right. So if this is really a problem, I wonder if we can
add the synchronize_rcu() before freeing the old->exts, like:

diff --git a/net/sched/cls_tcindex.c b/net/sched/cls_tcindex.c
index 1c9eeb98d826..57d900c664cf 100644
--- a/net/sched/cls_tcindex.c
+++ b/net/sched/cls_tcindex.c
@@ -338,6 +338,7 @@ tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
        struct tcf_result cr = {};
        int err, balloc = 0;
        struct tcf_exts e;
+       struct tcf_exts old_e = {};
 
        err = tcf_exts_init(&e, net, TCA_TCINDEX_ACT, TCA_TCINDEX_POLICE);
        if (err < 0)
@@ -479,6 +480,7 @@ tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
        }
 
        if (old_r && old_r != r) {
+               old_e = old_r->exts;
                err = tcindex_filter_result_init(old_r, cp, net);
                if (err < 0) {
                        kfree(f);
@@ -510,6 +512,9 @@ tcindex_set_parms(struct net *net, struct tcf_proto *tp, unsigned long base,
                tcf_exts_destroy(&new_filter_result.exts);
        }
 
+       synchronize_rcu();
+       tcf_exts_destroy(&old_e);
+
        if (oldp)
                tcf_queue_work(&oldp->rwork, tcindex_partial_destroy_work);
        return 0;

>
> Thanks.