Re: [PATCH v2 net] af_unix: Do not call kmemdup() for init_net's sysctl table.

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From:   Eric Dumazet <edumazet@google.com>
Date:   Mon, 27 Jun 2022 21:36:18 +0200
> On Mon, Jun 27, 2022 at 9:16 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From:   Eric Dumazet <edumazet@google.com>
> > Date:   Mon, 27 Jun 2022 21:06:14 +0200
> > > On Mon, Jun 27, 2022 at 8:59 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > >
> > > > From:   Eric Dumazet <edumazet@google.com>
> > > > Date:   Mon, 27 Jun 2022 20:40:24 +0200
> > > > > On Mon, Jun 27, 2022 at 8:30 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > > > >
> > > > > > From:   Jakub Kicinski <kuba@kernel.org>
> > > > > > Date:   Mon, 27 Jun 2022 10:58:59 -0700
> > > > > > > On Sun, 26 Jun 2022 11:43:27 -0500 Eric W. Biederman wrote:
> > > > > > > > Kuniyuki Iwashima <kuniyu@amazon.com> writes:
> > > > > > > >
> > > > > > > > > While setting up init_net's sysctl table, we need not duplicate the global
> > > > > > > > > table and can use it directly.
> > > > > > > >
> > > > > > > > Acked-by: "Eric W. Biederman" <ebiederm@xmission.com>
> > > > > > > >
> > > > > > > > I am not quite certain the savings of a single entry table justivies
> > > > > > > > the complexity.  But the looks correct.
> > > > > > >
> > > > > > > Yeah, the commit message is a little sparse. The "why" is not addressed.
> > > > > > > Could you add more details to explain the motivation?
> > > > > >
> > > > > > I was working on a series which converts UDP/TCP hash tables into per-netns
> > > > > > ones like AF_UNIX to speed up looking up sockets.  It will consume much
> > > > > > memory on a host with thousands of netns, but it can be waste if we do not
> > > > > > have its protocol family's sockets.
> > > > >
> > > > > For the record, I doubt we will accept such a patch (per net-ns
> > > > > TCP/UDP hash tables)
> > > >
> > > > Is it because it's risky?
> > >
> > > Because it will be very expensive. TCP hash tables are quite big.
> >
> > Yes, so I'm wondering if changing the size by sysctl makes sense.  If we
> > have per-netns hash tables, each table should have smaller amount of
> > sockets and smaller size should be enough, I think.
> 
> How can a sysctl be safely used if two different threads call "unshare
> -n" at the same time ?

I think two unshare are safe. Each of them reads its parent netns's sysctl
knob.  Even when the parent is the same, they can read the same value.

But I think we need READ_ONCE()/WRITE_ONCE() in such a sysctl.  While
creating a child netns, another one can change the value and there can be
a data-race.  So we have to use custome handler and pass write/read handler
as conv of do_proc_douintvec(), like do_proc_douintvec_conv_lockless().

If there are some sysctls missing READ_ONCE/WRITE_ONCE(), I will add
more general one, proc_douintvec_lockless().


> > > [    4.917080] tcp_listen_portaddr_hash hash table entries: 65536
> > > (order: 8, 1048576 bytes, vmalloc)
> > > [    4.917260] TCP established hash table entries: 524288 (order: 10,
> > > 4194304 bytes, vmalloc hugepage)
> > > [    4.917760] TCP bind hash table entries: 65536 (order: 8, 1048576
> > > bytes, vmalloc)
> > > [    4.917881] TCP: Hash tables configured (established 524288 bind 65536)
> > >
> > >
> > >
> > > > IIRC, you said we need per netns table for TCP in the future.
> > >
> > > Which ones exactly ? I guess you misunderstood.
> >
> > I think this.
> > https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net.git/commit/?id=04c494e68a13
> 
> "might" is very different than "will"
> 
> I would rather use the list of time_wait, instead of adding huge
> memory costs for hosts with hundreds of netns.

Sorry, my bad.
I would give it a try only for TIME_WAIT.