Re: [PATCH v1 bpf-next 03/11] tcp: Migrate TCP_ESTABLISHED/TCP_SYN_RECV sockets in accept queues.

[Martin KaFai Lau]

On Tue, Dec 08, 2020 at 03:27:14PM +0900, Kuniyuki Iwashima wrote:
> From:   Martin KaFai Lau <kafai@xxxxxx>
> Date:   Mon, 7 Dec 2020 12:14:38 -0800
> > On Sun, Dec 06, 2020 at 01:03:07AM +0900, Kuniyuki Iwashima wrote:
> > > From:   Martin KaFai Lau <kafai@xxxxxx>
> > > Date:   Fri, 4 Dec 2020 17:42:41 -0800
> > > > On Tue, Dec 01, 2020 at 11:44:10PM +0900, Kuniyuki Iwashima wrote:
> > > > [ ... ]
> > > > > diff --git a/net/core/sock_reuseport.c b/net/core/sock_reuseport.c
> > > > > index fd133516ac0e..60d7c1f28809 100644
> > > > > --- a/net/core/sock_reuseport.c
> > > > > +++ b/net/core/sock_reuseport.c
> > > > > @@ -216,9 +216,11 @@ int reuseport_add_sock(struct sock *sk, struct sock *sk2, bool bind_inany)
> > > > >  }
> > > > >  EXPORT_SYMBOL(reuseport_add_sock);
> > > > >  
> > > > > -void reuseport_detach_sock(struct sock *sk)
> > > > > +struct sock *reuseport_detach_sock(struct sock *sk)
> > > > >  {
> > > > >  	struct sock_reuseport *reuse;
> > > > > +	struct bpf_prog *prog;
> > > > > +	struct sock *nsk = NULL;
> > > > >  	int i;
> > > > >  
> > > > >  	spin_lock_bh(&reuseport_lock);
> > > > > @@ -242,8 +244,12 @@ void reuseport_detach_sock(struct sock *sk)
> > > > >  
> > > > >  		reuse->num_socks--;
> > > > >  		reuse->socks[i] = reuse->socks[reuse->num_socks];
> > > > > +		prog = rcu_dereference(reuse->prog);
> > > > Is it under rcu_read_lock() here?
> > > 
> > > reuseport_lock is locked in this function, and we do not modify the prog,
> > > but is rcu_dereference_protected() preferable?
> > > 
> > > ---8<---
> > > prog = rcu_dereference_protected(reuse->prog,
> > > 				 lockdep_is_held(&reuseport_lock));
> > > ---8<---
> > It is not only reuse->prog.  Other things also require rcu_read_lock(),
> > e.g. please take a look at __htab_map_lookup_elem().
> > 
> > The TCP_LISTEN sk (selected by bpf to be the target of the migration)
> > is also protected by rcu.
> 
> Thank you, I will use rcu_read_lock() and rcu_dereference() in v3 patchset.
> 
> 
> > I am surprised there is no WARNING in the test.
> > Do you have the needed DEBUG_LOCK* config enabled?
> 
> Yes, DEBUG_LOCK* was 'y', but rcu_dereference() without rcu_read_lock()
> does not show warnings...
I would at least expect the "WARN_ON_ONCE(!rcu_read_lock_held() ...)"
from __htab_map_lookup_elem() should fire in your test
example in the last patch.

It is better to check the config before sending v3.

[ ... ]

> > > > > diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> > > > > index 1451aa9712b0..b27241ea96bd 100644
> > > > > --- a/net/ipv4/inet_connection_sock.c
> > > > > +++ b/net/ipv4/inet_connection_sock.c
> > > > > @@ -992,6 +992,36 @@ struct sock *inet_csk_reqsk_queue_add(struct sock *sk,
> > > > >  }
> > > > >  EXPORT_SYMBOL(inet_csk_reqsk_queue_add);
> > > > >  
> > > > > +void inet_csk_reqsk_queue_migrate(struct sock *sk, struct sock *nsk)
> > > > > +{
> > > > > +	struct request_sock_queue *old_accept_queue, *new_accept_queue;
> > > > > +
> > > > > +	old_accept_queue = &inet_csk(sk)->icsk_accept_queue;
> > > > > +	new_accept_queue = &inet_csk(nsk)->icsk_accept_queue;
> > > > > +
> > > > > +	spin_lock(&old_accept_queue->rskq_lock);
> > > > > +	spin_lock(&new_accept_queue->rskq_lock);
> > > > I am also not very thrilled on this double spin_lock.
> > > > Can this be done in (or like) inet_csk_listen_stop() instead?
> > > 
> > > It will be possible to migrate sockets in inet_csk_listen_stop(), but I
> > > think it is better to do it just after reuseport_detach_sock() becuase we
> > > can select a different listener (almost) every time at a lower cost by
> > > selecting the moved socket and pass it to inet_csk_reqsk_queue_migrate()
> > > easily.
> > I don't see the "lower cost" point.  Please elaborate.
> 
> In reuseport_select_sock(), we pass sk_hash of the request socket to
> reciprocal_scale() and generate a random index for socks[] to select
> a different listener every time.
> On the other hand, we do not have request sockets in unhash path and
> sk_hash of the listener is always 0, so we have to generate a random number
> in another way. In reuseport_detach_sock(), we can use the index of the
> moved socket, but we do not have it in inet_csk_listen_stop(), so we have
> to generate a random number in inet_csk_listen_stop().
> I think it is at lower cost to use the index of the moved socket.
Generate a random number is not a big deal for the migration code path.

Also, I really still failed to see a particular way that the kernel
pick will help in the migration case.  The kernel has no clue
on how to select the right process to migrate to without
a proper policy signal from the user.  They are all as bad as
a random pick.  I am not sure this migration feature is
even useful if there is no bpf prog attached to define the policy.
That said, if it is still desired to do a random pick by kernel when
there is no bpf prog, it probably makes sense to guard it in a sysctl as
suggested in another reply.  To keep it simple, I would also keep this
kernel-pick consistent instead of request socket is doing something
different from the unhash path.

> 
> 
> > > sk_hash of the listener is 0, so we would have to generate a random number
> > > in inet_csk_listen_stop().
> > If I read it correctly, it is also passing 0 as the sk_hash to
> > bpf_run_sk_reuseport() from reuseport_detach_sock().
> > 
> > Also, how is the sk_hash expected to be used?  I don't see
> > it in the test.
> 
> I expected it should not be used in unhash path.
> We do not have the request socket in unhash path and cannot pass a proper
> sk_hash to bpf_run_sk_reuseport(). So, if u8 migration is
> BPF_SK_REUSEPORT_MIGRATE_QUEUE, we cannot use sk_hash.