Re: [PATCH][RESEND] sunrpc: hold a ref on netns for tcp sockets

[Josef Bacik <josef@toxicpanda.com>]

On Wed, Mar 20, 2024 at 04:00:09PM +0100, Eric Dumazet wrote:
> On Wed, Mar 20, 2024 at 3:56 PM Josef Bacik <josef@toxicpanda.com> wrote:
> >
> > On Wed, Mar 20, 2024 at 03:28:15PM +0100, Eric Dumazet wrote:
> > > On Wed, Mar 20, 2024 at 3:10 PM Josef Bacik <josef@toxicpanda.com> wrote:
> > > >
> > > > On Tue, Mar 19, 2024 at 09:59:48PM +0000, Trond Myklebust wrote:
> > > > > On Tue, 2024-03-19 at 16:07 -0400, Josef Bacik wrote:
> > > > > > We've been seeing variations of the following panic in production
> > > > > >
> > > > > >   BUG: kernel NULL pointer dereference, address: 0000000000000000
> > > > > >   RIP: 0010:ip6_pol_route+0x59/0x7a0
> > > > > >   Call Trace:
> > > > > >    <IRQ>
> > > > > >    ? __die+0x78/0xc0
> > > > > >    ? page_fault_oops+0x286/0x380
> > > > > >    ? fib6_table_lookup+0x95/0xf40
> > > > > >    ? exc_page_fault+0x5d/0x110
> > > > > >    ? asm_exc_page_fault+0x22/0x30
> > > > > >    ? ip6_pol_route+0x59/0x7a0
> > > > > >    ? unlink_anon_vmas+0x370/0x370
> > > > > >    fib6_rule_lookup+0x56/0x1b0
> > > > > >    ? update_blocked_averages+0x2c6/0x6a0
> > > > > >    ip6_route_output_flags+0xd2/0x130
> > > > > >    ip6_dst_lookup_tail+0x3b/0x220
> > > > > >    ip6_dst_lookup_flow+0x2c/0x80
> > > > > >    inet6_sk_rebuild_header+0x14c/0x1e0
> > > > > >    ? tcp_release_cb+0x150/0x150
> > > > > >    __tcp_retransmit_skb+0x68/0x6b0
> > > > > >    ? tcp_current_mss+0xca/0x150
> > > > > >    ? tcp_release_cb+0x150/0x150
> > > > > >    tcp_send_loss_probe+0x8e/0x220
> > > > > >    tcp_write_timer+0xbe/0x2d0
> > > > > >    run_timer_softirq+0x272/0x840
> > > > > >    ? hrtimer_interrupt+0x2c9/0x5f0
> > > > > >    ? sched_clock_cpu+0xc/0x170
> > > > > >    irq_exit_rcu+0x171/0x330
> > > > > >    sysvec_apic_timer_interrupt+0x6d/0x80
> > > > > >    </IRQ>
> > > > > >    <TASK>
> > > > > >    asm_sysvec_apic_timer_interrupt+0x16/0x20
> > > > > >   RIP: 0010:cpuidle_enter_state+0xe7/0x243
> > > > > >
> > > > > > Inspecting the vmcore with drgn you can see why this is a NULL
> > > > > > pointer deref
> > > > > >
> > > > > >     >>> prog.crashed_thread().stack_trace()[0]
> > > > > >     #0 at 0xffffffff810bfa89 (ip6_pol_route+0x59/0x796) in
> > > > > > ip6_pol_route at net/ipv6/route.c:2212:40
> > > > > >
> > > > > >     2212        if (net->ipv6.devconf_all->forwarding == 0)
> > > > > >     2213              strict |= RT6_LOOKUP_F_REACHABLE;
> > > > > >
> > > > > >     >>>
> > > > > > prog.crashed_thread().stack_trace()[0]['net'].ipv6.devconf_all
> > > > > >     (struct ipv6_devconf *)0x0
> > > > > >
> > > > > > Looking at the socket you can see that it's been closed
> > > > > >
> > > > > >     >>>
> > > > > > decode_enum_type_flags(prog.crashed_thread().stack_trace()[11]['sk'].
> > > > > > __sk_common.skc_flags, prog.type('enum sock_flags'))
> > > > > >     'SOCK_DEAD|SOCK_KEEPOPEN|SOCK_ZAPPED|SOCK_USE_WRITE_QUEUE'
> > > > > >     >>> decode_enum_type_flags(1 <<
> > > > > > prog.crashed_thread().stack_trace()[11]['sk'].__sk_common.skc_state.v
> > > > > > alue_(), prog["TCPF_CLOSE"].type_, bit_numbers=False)
> > > > > >     'TCPF_FIN_WAIT1'
> > > > > >
> > > > > > This occurs in our container setup where we have an NFS mount that
> > > > > > belongs to the containers network namespace.  On container shutdown
> > > > > > our
> > > > > > netns goes away, which sets net->ipv6.defconf_all = NULL, and then we
> > > > > > panic.  In the kernel we're responsible for destroying our sockets
> > > > > > when
> > > > > > the network namespace exits, or holding a reference on the network
> > > > > > namespace for our sockets so this doesn't happen.
> > > > > >
> > > > > > Even once we shutdown the socket we can still have TCP timers that
> > > > > > fire
> > > > > > in the background, hence this panic.  SUNRPC shuts down the socket
> > > > > > and
> > > > > > throws away all knowledge of it, but it's still doing things in the
> > > > > > background.
> > > > > >
> > > > > > Fix this by grabbing a reference on the network namespace for any tcp
> > > > > > sockets we open.  With this patch I'm able to cycle my 500 node
> > > > > > stress
> > > > > > tier over and over again without panicing, whereas previously I was
> > > > > > losing 10-20 nodes every shutdown cycle.
> > > > > >
> > > > > > Signed-off-by: Josef Bacik <josef@toxicpanda.com>
> > > > > > ---
> > > > > > Apologies, I just grepped for SUNRPC in MAINTAINERS and didn't
> > > > > > realize there was
> > > > > > a division of the client and server side of SUNRPC.
> > > > > >
> > > > > >  net/sunrpc/xprtsock.c | 20 ++++++++++++++++++++
> > > > > >  1 file changed, 20 insertions(+)
> > > > > >
> > > > > > diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c
> > > > > > index bb81050c870e..f02387751a94 100644
> > > > > > --- a/net/sunrpc/xprtsock.c
> > > > > > +++ b/net/sunrpc/xprtsock.c
> > > > > > @@ -2333,6 +2333,7 @@ static int xs_tcp_finish_connecting(struct
> > > > > > rpc_xprt *xprt, struct socket *sock)
> > > > > >
> > > > > >     if (!transport->inet) {
> > > > > >             struct sock *sk = sock->sk;
> > > > > > +           struct net *net = sock_net(sk);
> > > > > >
> > > > > >             /* Avoid temporary address, they are bad for long-
> > > > > > lived
> > > > > >              * connections such as NFS mounts.
> > > > > > @@ -2350,7 +2351,26 @@ static int xs_tcp_finish_connecting(struct
> > > > > > rpc_xprt *xprt, struct socket *sock)
> > > > > >             tcp_sock_set_nodelay(sk);
> > > > > >
> > > > > >             lock_sock(sk);
> > > > > > +           /*
> > > > > > +            * Because timers can fire after the fact we need to
> > > > > > hold a
> > > > > > +            * reference on the netns for this socket.
> > > > > > +            */
> > > > > > +           if (!sk->sk_net_refcnt) {
> > > > > > +                   if (!maybe_get_net(net)) {
> > > > > > +                          release_sock(sk);
> > > > > > +                          return -ENOTCONN;
> > > > > > +                  }
> > > > > > +                  /*
> > > > > > +                   * For kernel sockets we have a tracker put
> > > > > > in place for
> > > > > > +                   * the tracing, we need to free this to
> > > > > > maintaine
> > > > > > +                   * consistent tracking info.
> > > > > > +                   */
> > > > > > +                  __netns_tracker_free(net, &sk->ns_tracker,
> > > > > > false);
> > > > > >
> > > > > > +                  sk->sk_net_refcnt = 1;
> > > > > > +                  netns_tracker_alloc(net, &sk->ns_tracker,
> > > > > > GFP_KERNEL);
> > > > > > +                  sock_inuse_add(net, 1);
> > > > > > +           }
> > > > > >             xs_save_old_callbacks(transport, sk);
> > > > > >
> > > > > >             sk->sk_user_data = xprt;
> > > > >
> > > > > Hmm... Doesn't this end up being more or less equivalent to calling
> > > > > __sock_create() with the kernel flag being set to 0?
> > > >
> > > > AFAICT yes, but there are a lot of other things that happen with kern being set
> > > > to 1, so I think this is a safer bet, and is analagous to this other fix
> > > > 3a58f13a881e ("net: rds: acquire refcount on TCP sockets").  Thanks,
> > > >
> > >
> > > Hmm... this would prevent a netns with one or more TCP flows owned by
> > > this layer to be dismantled,
> > > even if all other processes/sockets disappeared ?
> >
> > Yeah but if sockets are still in use then we want the netns to still be up
> > right?  I personally am very confused about how the lifetime stuff works for
> > sockets, I don't understand how shutting down the socket means it gets to stick
> > around after the fact forever, but feels like if it's tied to a netns then it's
> > completely valid to hold the netns open until we're done with the socket.
> >
> > >
> > > Have you looked at my suggestion instead ?
> > >
> > > https://lore.kernel.org/bpf/CANn89i+484ffqb93aQm1N-tjxxvb3WDKX0EbD7318RwRgsatjw@mail.gmail.com/
> > >
> > > I never formally submitted this patch because I got no feedback.
> >
> > I did something similar, tho not with _sync so maybe that was the problem, but
> > this is what I did originally in production before I emailed you the first time
> 
> 
> The _sync part is mandatory really for this context.
> 
> Not that it needs to be done while the socket is not locked, or risk a deadlock.
> 
> Note that modern trees have timer_shutdown_sync() which might even be better.
> 

Sounds good, I've reverted my patches and I've applied this patch, I should have
results by the end of the day on wether or not it fixed the problem.  Thanks,

Josef