netdev - Re: [PATCH bpf-next 2/3] bpf: allow bpf_{s,g}etsockopt from cgroup bind{4,6} hooks

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20201201192201.GB27988@rdna-mbp.dhcp.thefacebook.com>
Date:   Tue, 1 Dec 2020 11:22:01 -0800
From:   Andrey Ignatov <rdna@...com>
To:     <sdf@...gle.com>
CC:     Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Network Development <netdev@...r.kernel.org>,
        bpf <bpf@...r.kernel.org>,
        "David S. Miller" <davem@...emloft.net>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>
Subject: Re: [PATCH bpf-next 2/3] bpf: allow bpf_{s,g}etsockopt from cgroup
 bind{4,6} hooks

sdf@...gle.com <sdf@...gle.com> [Tue, 2020-12-01 10:43 -0800]:
> On 11/30, Andrey Ignatov wrote:
> > sdf@...gle.com <sdf@...gle.com> [Mon, 2020-11-30 08:38 -0800]:
> > > On 11/29, Andrey Ignatov wrote:
> > > > Alexei Starovoitov <alexei.starovoitov@...il.com> [Tue, 2020-11-17
> > 20:05
> > > > -0800]:
> > > > > On Tue, Nov 17, 2020 at 4:17 PM Stanislav Fomichev <sdf@...gle.com>
> > > > wrote:
> > > [..]
> > > > >
> > > > > I think it is ok, but I need to go through the locking paths more.
> > > > > Andrey,
> > > > > please take a look as well.
> > >
> > > > Sorry for delay, I was offline for the last two weeks.
> > > No worries, I was OOO myself last week, thanks for the feedback!
> > >
> > > >  From the correctness perspective it looks fine to me.
> > >
> > > >  From the performance perspective I can think of one relevant
> > scenario.
> > > > Quite common use-case in applications is to use bind(2) not before
> > > > listen(2) but before connect(2) for client sockets so that connection
> > > > can be set up from specific source IP and, optionally, port.
> > >
> > > > Binding to both IP and port case is not interesting since it's already
> > > > slow due to get_port().
> > >
> > > > But some applications do care about connection setup performance and
> > at
> > > > the same time need to set source IP only (no port). In this case they
> > > > use IP_BIND_ADDRESS_NO_PORT socket option, what makes bind(2) fast
> > > > (we've discussed it with Stanislav earlier in [0]).
> > >
> > > > I can imagine some pathological case when an application sets up
> > tons of
> > > > connections with bind(2) before connect(2) for sockets with
> > > > IP_BIND_ADDRESS_NO_PORT enabled (that by itself requires setsockopt(2)
> > > > though, i.e. socket lock/unlock) and that another lock/unlock to run
> > > > bind hook may add some overhead. Though I do not know how critical
> > that
> > > > overhead may be and whether it's worth to benchmark or not (maybe too
> > > > much paranoia).
> > >
> > > > [0] https://lore.kernel.org/bpf/20200505182010.GB55644@rdna-mbp/
> > > Even in case of IP_BIND_ADDRESS_NO_PORT, inet[6]_bind() does
> > > lock_sock down the line, so it's not like we are switching
> > > a lockless path to the one with the lock, right?
> 
> > Right, I understand that it's going from one lock/unlock to two (not
> > from zero to one), that's what I meant by "another". My point was about
> > this one more lock.
> 
> > > And in this case, similar to listen, the socket is still uncontended and
> > > owned by the userspace. So that extra lock/unlock should be cheap
> > > enough to be ignored (spin_lock_bh on the warm cache line).
> > >
> > > Am I missing something?
> 
> > As I mentioned it may come up only in "pathological case" what is
> > probably fine to ignore, i.e. I'd rather agree with "cheap enough to be
> > ignored" and benchmark would likely confirm it, I just couldn't say that
> > for sure w/o numbers so brought this point.
> 
> > Given that we both agree that it should be fine to ignore this +1 lock,
> > IMO it should be good to go unless someone else has objections.
> Thanks, agreed. Do you mind giving it an acked-by so it gets some
> attention in the patchwork? ;-)

Sure. Acked this one.

-- 
Andrey Ignatov