On 11/30, Andrey Ignatov wrote:
sdf@xxxxxxxxxx <sdf@xxxxxxxxxx> [Mon, 2020-11-30 08:38 -0800]:
> On 11/29, Andrey Ignatov wrote:
> > Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> [Tue, 2020-11-17
20:05
> > -0800]:
> > > On Tue, Nov 17, 2020 at 4:17 PM Stanislav Fomichev <sdf@xxxxxxxxxx>
> > 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? ;-)
