Re: [PATCH RFC net] net/smc: Ensure the active closing peer first closes clcsock

[Tony Lu <tonylu@linux.alibaba.com>]

On Wed, Nov 24, 2021 at 11:08:23AM +0100, Karsten Graul wrote:
> On 24/11/2021 09:57, Tony Lu wrote:
> > IMHO, given that, it is better to not ignore smc_close_final(), and move 
> > kernel_sock_shutdown() to __smc_release(), because smc_shutdown() also
> > calls kernel_sock_shutdown() after smc_close_active() and
> > smc_close_shutdown_write(), then enters SMC_PEERCLOSEWAIT1. It's no need
> > to call it twice with SHUT_WR and SHUT_RDWR. 
> 
> Since the idea is to shutdown the socket before the remote peer shutdowns it
> first, are you sure that this shutdown in smc_release() is not too late?

Hi Graul,

Yes, I have tested this idea, it will be too late sometime. I won't fix
this issue.

> Is it sure that smc_release() is called in time for this processing?
> 
> Maybe its better to keep the shutdown in smc_close_active() and to use an rc1
> just like shown in your proposal, and return either the rc of smc_close_final() 
> or the rc of kernel_sock_shutdown().

Yep, I am testing this approach in my environment. I am going to keep
these return codes and return the available one.

> I see the possibility of calling shutdown twice for the clcsocket, but does it
> harm enough to give a reason to check it before in smc_shutdown()? I expect TCP
> to handle this already.

TCP could handle this already, but it doesn't make much sense to call it twice. When
call smc_shutdown(), we can check sk_shutdown before call kernel_sock_shutdown(),
so that it can slightly speed up the release process.

I will send this soon, thanks for your advice.

Tony Lu