Re: [PATCH net-next 09/12] net/mlx5: Compare with old_dest param to modify rule destination

[Leon Romanovsky <leon@kernel.org>]

On Sun, Jul 16, 2023 at 01:39:47PM +0300, Leon Romanovsky wrote:
> On Fri, Jul 14, 2023 at 08:30:32PM -0700, Jakub Kicinski wrote:
> > On Fri, 14 Jul 2023 23:32:58 +0300 Leon Romanovsky wrote:
> > > On Fri, Jul 14, 2023 at 12:16:33PM -0700, Jakub Kicinski wrote:
> > > > On Fri, 14 Jul 2023 21:40:13 +0300 Leon Romanovsky wrote:  
> > > > > It depends on configuration order, if user configures TC first, it will
> > > > > be a), if he/she configures IPsec first, it will be b).
> > > > > 
> > > > > I just think that option b) is really matters.  
> > > > 
> > > > And only b) matches what happens in the kernel with policy based IPsec,
> > > > right?   
> > > 
> > > Can you please clarify what do you mean "policy based IPsec"?
> > 
> > I mean without a separate xfrm netdev on which you can install TC
> > rules of its own.
> 
> I call it software IPsec.
> 
> > 
> > > > IIUC what you're saying -
> > > > the result depending on order of configuration may be a major source
> > > > of surprises / hard to debug problems for the user.  
> > > 
> > > When I reviewed patches, I came exactly to an opposite conclusion :)
> > > 
> > > My rationale was that users who configure IPsec and TC are advanced
> > > users who knows their data flow and if they find a) option valuable,
> > > they can do it.
> > > 
> > > For example, a) allows to limit amount of data sent to IPsec engine.
> > > 
> > > I believe both a) and b) should be supported.
> > 
> > What does it take to switch between the modes?
> > Even if we want both modes we should have an explicit switch, I reckon.
> > Or at least a way to read back what mode we ended up in.
> 
> I had several internal discussions about how TC and IPsec should work
> together, and will need some time to think about proper implementation.
> 
> For now I'll add patch which makes TC and IPsec mutually exclusive.

Even this so called trivial patch is not so trivial in mlx5 current
implementation. Jianbo is working on it.

Thanks

> 
> Thanks
> 
> > 
>