Re: [RFC] tcp: implement new per-interface sysctl "auto_dev_bind"

[Willy Tarreau]

Hi Eric,

On Fri, Dec 13, 2019 at 08:03:52AM -0800, Eric Dumazet wrote:
> Hi Willy, thanks for working on this.

Thanks for reviewing :-)

> Could you check if your patch works with syncookies mode ?
> 
> echo 2 >/proc/sys/net/ipv4/tcp_syncookies

Good catch:

[19401.670076] Call trace:
[19401.672494]  tcp_v4_syn_recv_sock+0x168/0x250
[19401.676807]  tcp_get_cookie_sock+0x4c/0xe4
[19401.680858]  cookie_v4_check+0x470/0x4a8
[19401.684740]  tcp_v4_do_rcv+0xf4/0x1c8
[19401.688362]  tcp_v4_rcv+0x410/0x790
[19401.691815]  ip_protocol_deliver_rcu+0x8c/0x128
[19401.696298]  ip_local_deliver_finish+0x64/0x78
[19401.700696]  ip_local_deliver+0x90/0x94
[19401.704491]  ip_rcv_finish+0x3c/0x50
[19401.708027]  ip_rcv+0x5c/0x74
[19401.710961]  __netif_receive_skb_one_core+0x54/0x7c
[19401.715790]  __netif_receive_skb+0x5c/0x64
[19401.719843]  netif_receive_skb_internal+0x68/0xcc
[19401.724501]  napi_gro_receive+0x70/0xa0
[19401.728298]  gro_cell_poll+0x74/0x88
[19401.731832]  net_rx_action+0x134/0x2c8
[19401.735543]  __do_softirq+0x1bc/0x1fc
[19401.739165]  irq_exit+0x60/0xb0
[19401.742270]  __handle_domain_irq+0x6c/0x98
[19401.746322]  gic_handle_irq+0x70/0xac
[19401.749945]  el1_irq+0xb8/0x180
[19401.753052]  arch_cpu_idle+0x10/0x18
[19401.756588]  do_idle+0x134/0x22c
[19401.759778]  cpu_startup_entry+0x20/0x3c
[19401.763660]  rest_init+0xd0/0xdc
[19401.766852]  arch_call_rest_init+0xc/0x14
[19401.770818]  start_kernel+0x41c/0x448
[19401.774444] Code: 790bd261 aa1303e0 97ffbdd1 f9400ac0 (f9416000) 
[19401.780485] ---[ end trace e08862982660f052 ]---
[19401.785049] Kernel panic - not syncing: Fatal exception in interrupt

> I wonder if your patch could be simpler if you were plugging the logic for passive
> flows in inet_request_bound_dev_if() ?

I'm not sure yet. For having had a quick lookk and tried to move the
code there, I feel like I'll need to distinguish the protocols (v4/v6)
in order to look at the per-interface configuration, while it's already
done in the caller. Or maybe there are some ipv4 settings that also
apply to ipv6 and we could do the same by having a single one for the
two maybe ?

Willy