[lvc-project] [PATCH net v2] net: sockmap: avoid race between sock_map_destroy() and sk_psock_put()

[Paolo Abeni]

On 9/19/24 11:26, Dmitry Antipov wrote:
> On 9/19/24 11:51 AM, Paolo Abeni wrote:
> 
>> I think there is still Cong question pending: why this could not/ should not be addressed instead in the RDS code.
> 
> Hm. Except 'rds_tcp_accept_xxx()' functions, the rest of the backtrace
> is contributed by generic TCP and IP things.

AFAICS most of RDS is build on top of TCP, most RDS-specific bugs will 
show a lot tcp/ip in their backtrace.

i.e. with mptcp we had quite a few bugs with a 'tcp mostily' or 'tcp 
only' backtrace and root caused in the mptcp code.

> I've tried to debug this
> issue starting from the closest innards and seems found an issue within
> sockmap code.
> Anyway, I'm strongly disagree with "unless otherwise proven,
> there are a lot of bugs everywhere except the subsystem I'm responsible
> to" bias, assuming that a bit more friendly and cooperative efforts
> should give us the better results.

I guess that the main point in Cong's feedback is that a sockmap update 
is not supposed to race with sock_map_destroy() (???) @Cong, @John, 
@JakubS: any comments on that?

Cheers,

Paolo