Re: [PATCH net-next v4 4/4] net: wangxun: support to use adaptive RX/TX coalescing

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 20 Aug 2025 09:57:43 +0800 Jiawen Wu wrote:
> On Sat, Aug 16, 2025 2:19 AM, Jakub Kicinski wrote:
> > On Tue, 12 Aug 2025 09:50:23 +0800 Jiawen Wu wrote:  
> > > @@ -878,6 +909,8 @@ static int wx_poll(struct napi_struct *napi, int budget)
> > >
> > >  	/* all work done, exit the polling mode */
> > >  	if (likely(napi_complete_done(napi, work_done))) {
> > > +		if (wx->adaptive_itr)
> > > +			wx_update_dim_sample(q_vector);  
> > 
> > this is racy, napi is considered released after napi_complete_done()
> > returns. So napi_disable() can succeed right after that point...
> >   
> > > @@ -1611,6 +1708,8 @@ void wx_napi_disable_all(struct wx *wx)
> > >  	for (q_idx = 0; q_idx < wx->num_q_vectors; q_idx++) {
> > >  		q_vector = wx->q_vector[q_idx];
> > >  		napi_disable(&q_vector->napi);
> > > +		cancel_work_sync(&q_vector->rx.dim.work);
> > > +		cancel_work_sync(&q_vector->tx.dim.work);  
> > 
> > so you may end up with the DIM work scheduled after the device is
> > stopped.  
> 
> But the DIM work doesn't seem to be concerned about the status of napi.
> And even if the device is stopped, setting itr would not cause any errors.
> 
> I can't fully grasp this point...
> Should I move cancel_work_sync() in front of napi_disable()?

My point is that this is possible today:

     CPU 0                     CPU 1

  napi_complete_done()
                            napi_disable()
                            cancel_work()
  wx_update_dim_sample()
    schedule_work()

You can probably use disable_work_sync() and enable_work..
to fix it.