On 17/12/18 - 08:04:08, Eric Dumazet wrote:
> On Fri, Dec 14, 2018 at 2:40 PM Christoph Paasch <cpaa...@apple.com> wrote:
> >
>
> ...
>
> > int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
> > void *key, unsigned int len)
> > {
> > @@ -96,13 +131,22 @@ error: kfree(ctx);
> > spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
> > if (sk) {
> > q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
>
> > + rcu_assign_pointer(ctx->next, q->ctx);
> At this point, ctx is not yet visible, so you do not need a barrier yet
> ctx->next = q->ctx;
Thanks, I will change that.
>
>
> > + rcu_assign_pointer(q->ctx, ctx);
>
> Note that readers could see 3 contexts in the chain, instead of maximum two.
>
> This means that proc_tcp_fastopen_key() (your 3/5 change) would
> overflow an automatic array :
>
> while (ctxt) {
> memcpy(&key[i], ctxt->key, TCP_FASTOPEN_KEY_LENGTH);
> i += 4;
> ctxt = rcu_dereference(ctxt->next);
> }
Ouch! Thanks for spotting this.
If it's ok to have a brief moment of 3 contexts for the readers, I would
protect against overflows the readers.
> > +
> > octx = rcu_dereference_protected(q->ctx,
> > lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
> > - rcu_assign_pointer(q->ctx, ctx);
> > +
> > + octx = tcp_fastopen_cut_keypool(octx,
> > &net->ipv4.tcp_fastopen_ctx_lock);
> > } else {
> > + rcu_assign_pointer(ctx->next, net->ipv4.tcp_fastopen_ctx);
>
> same remark here.
Same, will change that.
Christoph
> > + rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, ctx);
> > +
> > octx = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
> > lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
> > - rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, ctx);
> > +
> > + octx = tcp_fastopen_cut_keypool(octx,
> > +
> > &net->ipv4.tcp_fastopen_ctx_lock);
> > }
> > spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
> >
> > --
> > 2.16.2
> >
