Re: [PATCH v1 net-next 03/15] smc: Pass kern to smc_sock_alloc().

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Wenjia Zhang <wenjia@linux.ibm.com>
Date: Mon, 9 Dec 2024 12:05:03 +0100
> On 06.12.24 08:54, Kuniyuki Iwashima wrote:
> > Since commit d7cd421da9da ("net/smc: Introduce TCP ULP support"),
> > smc_ulp_init() calls __smc_create() with kern=1.
> > 
> > However, __smc_create() does not pass it to smc_sock_alloc(),
> > which finally calls sk_alloc() with kern=0.
> > 
> > Let's pass kern down to smc_sock_alloc() and sk_alloc().
> > 
> > Note that we change kern from 1 to 0 in smc_ulp_init() not to
> > introduce any functional change.
> > 
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > ---
> >   net/smc/af_smc.c | 10 +++++-----
> >   1 file changed, 5 insertions(+), 5 deletions(-)
> > 
> Sorry, I didn't see the need to add the **kern** parameter in 
> smc_sock_alloc(). Because the **kern** parameter in sk_alloc is not used 
> others than to decide the value of sk->sk_net_refcnt, which in SMC code 
> is already set in smc_create_clcsk(). Thus, IMO removing the **kern** 
> parameter from __smc_create() makes more sense, since this parameter is 
> not used in the function.

It would introduce another consufing situation when someone calls
sock_create_kern(PF_SMC) (sock_create_net() or sock_create_net_noref()),
which alaways create a userspace socket.

As long as it's part of the kernel socket API, we shouldn't use such a
hard-coded parameter.