Re: [PATCH] ice: Don't use GFP_KERNEL in atomic context

[Christophe JAILLET]

Le 16/01/2022 à 22:42, Andrew Lunn a écrit :
> On Sun, Jan 16, 2022 at 07:46:20PM +0100, Christophe JAILLET wrote:
> > ice_misc_intr() is an irq handler. It should not sleep.
> >
> > Use GFP_ATOMIC instead of GFP_KERNEL when allocating some memory.
> >
> > Fixes: 348048e724a0 ("ice: Implement iidc operations")
> > Signed-off-by: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
> > ---
> > I've never played a lot with irq handler. My understanding is that they
> > should never sleep.
>
> Hi Christophe
>
> Threaded interrupt handlers are allowed to sleep. However, this
> handler is not being used in such a way. So your are probably correct
> about GFP_KERNEL vs GFP_ATOMIC.
>
> > ---
> >   drivers/net/ethernet/intel/ice/ice_main.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/net/ethernet/intel/ice/ice_main.c b/drivers/net/ethernet/intel/ice/ice_main.c
> > index 30814435f779..65de01f3a504 100644
> > --- a/drivers/net/ethernet/intel/ice/ice_main.c
> > +++ b/drivers/net/ethernet/intel/ice/ice_main.c
> > @@ -3018,7 +3018,7 @@ static irqreturn_t ice_misc_intr(int __always_unused irq, void *data)
> >   		struct iidc_event *event;
> >   
> >   		ena_mask &= ~ICE_AUX_CRIT_ERR;
> > -		event = kzalloc(sizeof(*event), GFP_KERNEL);
> > +		event = kzalloc(sizeof(*event), GFP_ATOMIC);
> >   		if (event) {
> >   			set_bit(IIDC_EVENT_CRIT_ERR, event->type);
> >   			/* report the entire OICR value to AUX driver */
>
> What happens next is interesting...
>
>
>                          event->reg = oicr;
>                          ice_send_event_to_aux(pf, event);
>
> where:
>
> void ice_send_event_to_aux(struct ice_pf *pf, struct iidc_event *event)
> {
>          struct iidc_auxiliary_drv *iadrv;
>
>          if (!pf->adev)
>                  return;
>
>          device_lock(&pf->adev->dev);
>          iadrv = ice_get_auxiliary_drv(pf);
>          if (iadrv && iadrv->event_handler)
>                  iadrv->event_handler(pf, event);
>          device_unlock(&pf->adev->dev);
> }
>
> device_lock() takes a mutex, not something you should be doing in
> atomic context.
>
> So it looks to me, this handler really should be running in thread
> context...
>
> 	Andrew

Ok, thanks for the explanation.

ice_misc_intr() is registered with devm_request_irq(), so it is a 
handler that can't sleep.

I guess that more consideration should be taken into account than only:
  s/devm_request_irq(handler)/devm_request_threaded_irq(NULL, handler)/

So I'll leave this one to people with the expected know-how.

If my s/GFP_KERNEL/GFP_ATOMIC/ makes enough sense as-is, that's fine for 
me, but it looks that another solution is needed to fix the 2nd issue.


CJ