Re: [PATCH net] net: lwtunnel: disable preemption when required

[Jakub Kicinski <kuba@kernel.org>]

On Thu, 3 Apr 2025 21:08:12 +0200 Justin Iurman wrote:
> On 4/3/25 18:24, Stanislav Fomichev wrote:
> > On 04/03, Justin Iurman wrote:  
> >> In lwtunnel_{input|output|xmit}(), dev_xmit_recursion() may be called in
> >> preemptible scope for PREEMPT kernels. This patch disables preemption
> >> before calling dev_xmit_recursion(). Preemption is re-enabled only at
> >> the end, since we must ensure the same CPU is used for both
> >> dev_xmit_recursion_inc() and dev_xmit_recursion_dec() (and any other
> >> recursion levels in some cases) in order to maintain valid per-cpu
> >> counters.  
> > 
> > Dummy question: CONFIG_PREEMPT_RT uses current->net_xmit.recursion to
> > track the recursion. Any reason not to do it in the generic PREEMPT case?  
> 
> I'd say PREEMPT_RT is a different beast. IMO, softirqs can be 
> preempted/migrated in RT kernels, which is not true for non-RT kernels. 
> Maybe RT kernels could use __this_cpu_* instead of "current" though, but 
> it would be less trivial. For example, see commit ecefbc09e8ee ("net: 
> softnet_data: Make xmit per task.") on why it makes sense to use 
> "current" in RT kernels. I guess the opposite as you suggest (i.e., 
> non-RT kernels using "current") would be technically possible, but there 
> must be a reason it is defined the way it is... so probably incorrect or 
> inefficient?

I suspect it's to avoid the performance overhead.
IIUC you would be better off using local_bh_disable() here.
It doesn't disable preemption on RT.
I don't believe "disable preemption if !RT" primitive exists.