Re: [PATCH] net/smc: Add autocork support

[Stefan Raspl <raspl@linux.ibm.com>]

On 2/16/22 16:27, dust.li wrote:
> On Wed, Feb 16, 2022 at 02:58:32PM +0100, Stefan Raspl wrote:
>> On 2/16/22 04:49, Dust Li wrote:
>>> diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c
>>> index 5df3940d4543..bc737ac79805 100644
>>> --- a/net/smc/smc_tx.c
>>> +++ b/net/smc/smc_tx.c
>>> @@ -31,6 +31,7 @@
>>>    #include "smc_tracepoint.h"
>>>    #define SMC_TX_WORK_DELAY	0
>>> +#define SMC_DEFAULT_AUTOCORK_SIZE	(64 * 1024)
>>
>> Probably a matter of taste, but why not use hex here?
> 
> Yeah, I have no option on this, I will change it in the next version.
> But I think it should have no real difference since the compiler
> should do the calculation.

Agreed - this is just to make it a tiny bit easier to digest.


>> Are there any fixed plans to make SMC_DEFAULT_AUTOCORK dynamic...? 'cause
>> otherwise we could simply eliminate this parameter, and use the define within
>> smc_should_autocork() instead.
> 
> Yes! Actually I'd like it to be dynamic variable too...
> 
> I didn't do it because I also want to add a control switch for the autocork
> feature just like TCP. In that case I need to add 2 variables here.
> But I found adding dynamic variables using netlink would introduce a lot of
> redundant code and may even bring ABI compatibility issues in the future, as
> I mentioned here:
> https://lore.kernel.org/netdev/20220216114618.GA39286@linux.alibaba.com/T/#mecfcd3f8c816d07dbe35e4748d17008331c89523
> 
> I'm not sure that's the right way to do it. In this case, I prefer using
> sysctl which I think would be easier, but I would like to listen to your advice.

Extending the Netlink interface should be possible without breaking the API - 
we'd be adding further variables, not modifying or removing existing ones.
Conceptually, Netlink is the way to go for any userspace interaction with SMC, 
which includes anything config-related.
Now we understand that cloud workloads are a bit different, and the desire to be 
able to modify the environment of a container while leaving the container image 
unmodified is understandable. But then again, enabling the base image would be 
the cloud way to address this. The question to us is: How do other parts of the 
kernel address this?

Ciao,
Stefan