Re: [PATCH net] net: keep the NAPI_DISABLE flag while napi is disabled

[Maxime Chevallier]

On 9/30/26 22:41, Jakub Kicinski wrote:
> On Mon, 28 Sep 2026 09:08:29 +0200 Maxime Chevallier wrote:
>> On 9/28/26 00:41, Jakub Kicinski wrote:
>>> On Sat, 26 Sep 2026 21:47:13 +0200 Maxime Chevallier wrote:  
>>>> This seems to be more general than stmmac though, there are lots of
>>>> drivers that create the napi instances in .probe(), so they still live
>>>> outside of .open()/.close(). It was verified by running the following on
>>>> an mvneta board :  
>>>
>>> Not sure this works, you assume ownership if DISABLE is set but we set
>>> it to start the shutdown, not once we stopped the poller.
>>>
>>> We've been tempted to "fix" this multiple times. I'd prefer to fix
>>> drivers.  
>>
>> There's 2 classes of problems then
>>
>> - drivers that create a napi instance but don't use it even when
>> admin up (stmmac, I think it used to be the case for i40e looking
>> at the history). Fixing the driver makes sense then
>>
>> - drivers that add napi instances at .probe() time, in which
>> case setting threaded to 1 -> 0 when the interface is admin down
>> is always going to hang. There's quite a few of them :
>>
>> git grep -p netif_napi_add | grep probe | wc -l
>> 75
> 
> TIL git grep -p, nice!
> 
>> So at least 75 drivers call netif_napi_add from their probe
>> functions.
> 
> Mostly old drivers, half of which we should probably delete :S
> 
> Can we not refuse disabling the threaded state if the device
> is down and has NAPIs?

I don't see why not :) TBH this is a bit outside of my comfort zone,
I'll spin something and give it a test on as much HW as possible.

Thanks,

Maxime