On Fri, Dec 6, 2024 at 3:49 AM Dmitry Safonov <0x7f454c46@xxxxxxxxx> wrote:
>
> Hi Jakub, Eric,
>
> On Thu, 5 Dec 2024 at 09:09, Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > On Thu, Dec 5, 2024 at 2:13 AM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
> > >
> > > Hi Eric!
> > >
> > > This was posted while you were away -- any thoughts or recommendation on
> > > how to address the required nl message size changing? Or other problems
> > > pointed out by Dmitry? My suggestion in the subthread is to re-dump
> > > with a fixed, large buffer on EMSGSIZE, but that's not super clean..
> >
> > Hi Jakub
> >
> > inet_diag_dump_one_icsk() could retry, doubling the size until the
> > ~32768 byte limit is reached ?
> >
> > Also, we could make sure inet_sk_attr_size() returns at least
> > NLMSG_DEFAULT_SIZE, there is no
> > point trying to save memory for a single skb in inet_diag_dump_one_icsk().
>
> Starting from NLMSG_DEFAULT_SIZE sounds like a really sane idea! :-)
There is a consensus for this one, I will cook a patch with this part only.
>
> [..]
> > @@ -585,8 +589,14 @@ int inet_diag_dump_one_icsk(struct inet_hashinfo *hashinfo,
> >
> > err = sk_diag_fill(sk, rep, cb, req, 0, net_admin);
> > if (err < 0) {
> > - WARN_ON(err == -EMSGSIZE);
> > nlmsg_free(rep);
> > + if (err == -EMSGSIZE) {
> > + attr_size <<= 1;
> > + if (attr_size + NLMSG_HDRLEN <=
> > SKB_WITH_OVERHEAD(32768)) {
> > + cond_resched();
> > + goto retry;
> > + }
> > + }
> > goto out;
> > }
> > err = nlmsg_unicast(net->diag_nlsk, rep, NETLINK_CB(in_skb).portid);
>
> To my personal taste on larger than 327 md5 keys scale, I'd prefer to
> see "dump may be inconsistent, retry if you need consistency" than
> -EMSGSIZE fail, yet userspace potentially may use the errno as a
> "retry" signal.
>
I do not yet understand this point. I will let you send a patch for
further discussion.
Thanks.
