Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.35 as permitted sender) client-ip=23.128.96.35;
Date:   Thu, 12 Oct 2023 21:39:54 +0800
From:   Wei Gong <gongwei833x@gmail.com>
To:     Thomas Gleixner <tglx@linutronix.de>
Cc:     linux-kernel@vger.kernel.org
Subject: Re: [PATCH v3] genirq: avoid long loops in handle_edge_irq
Message-ID: <ZSf3KrAk8fpyNEZr@MacBook-Pro-3.local>
References: <20230928100638.42116-1-gongwei833x@gmail.com>
 <87r0m3q32t.ffs@tglx>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <87r0m3q32t.ffs@tglx>
Precedence: bulk

O Mon, Oct 09, 2023 at 04:32:10PM +0200, Thomas Gleixner wrote:
> On Thu, Sep 28 2023 at 18:06, Wei Gong wrote:
> > --- a/kernel/irq/chip.c
> > +++ b/kernel/irq/chip.c
> > @@ -831,7 +831,9 @@ void handle_edge_irq(struct irq_desc *desc)
> >  		handle_irq_event(desc);
> >  
> >  	} while ((desc->istate & IRQS_PENDING) &&
> > -		 !irqd_irq_disabled(&desc->irq_data));
> > +		 !irqd_irq_disabled(&desc->irq_data) &&
> > +		 cpumask_test_cpu(smp_processor_id(),
> > +				   irq_data_get_effective_affinity_mask(&desc->irq_data)));
> 
> Ok. So now that mask part is correct, but what guarantees that this does
> not lose interrupts?
> 
> Assume the following scenario:
> 
>    CPU 0                                CPU 1
> 
>    interrupt
>         set IN_PROGRESS
>         do {
>                                         change_affinity_to(CPU1);
>            handle_irq_event()
>                 ack_in_device()
>                                         interrupt
>                                            set PENDING
>         } while (COND)
> 
> Now $COND is not true due to the affinity change and the edge handler
> returns. As a consequence nothing acks the device and no further
> interrupts are sent by the device.
> 
> That might not be true for your case, but that's a generic function and the
> zoo of hardware which uses that is massive.
> 
> So no, we are not taking a risk here.
> 
> Thanks,
> 
>         tglx
> 
>     
By maintaining the original loop exit condition, if a mask mismatch is 
detected within the loop, we will not perform the unmask_irq operation. 
Instead, we will wait until the loop exits before executing unmask_irq.
Could this approach potentially solve the issue of lost interrupts?

Thanks,
Wei Gong