Re: [PATCH 2/5] kernel/watchdog_hld: clarify the condition in hardlockup_detector_event_create()

[Pingfan Liu]

On Thu, Sep 16, 2021 at 10:02:27AM +0200, Petr Mladek wrote:
> On Thu 2021-09-16 11:57:44, Pingfan Liu wrote:
> > On Wed, Sep 15, 2021 at 03:45:06PM +0200, Peter Zijlstra wrote:
> > > On Wed, Sep 15, 2021 at 11:51:00AM +0800, Pingfan Liu wrote:
> > > > hardlockup_detector_event_create() indirectly calls
> > > > kmem_cache_alloc_node(), which is blockable.
> > > > 
> > > > So here, the really planned context is is_percpu_thread().
> > > > 
> > > > Signed-off-by: Pingfan Liu <kernelfans@xxxxxxxxx>
> > > > Cc: Petr Mladek <pmladek@xxxxxxxx>
> > > > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > > > Cc: Wang Qing <wangqing@xxxxxxxx>
> > > > Cc: "Peter Zijlstra (Intel)" <peterz@xxxxxxxxxxxxx>
> > > > Cc: Santosh Sivaraj <santosh@xxxxxxxxxx>
> > > > Cc: Sumit Garg <sumit.garg@xxxxxxxxxx>
> > > > Cc: Will Deacon <will@xxxxxxxxxx>
> > > > Cc: Mark Rutland <mark.rutland@xxxxxxx>
> > > > To: linux-kernel@xxxxxxxxxxxxxxx
> > > > ---
> > > >  kernel/watchdog_hld.c | 5 ++++-
> > > >  1 file changed, 4 insertions(+), 1 deletion(-)
> > > > 
> > > > diff --git a/kernel/watchdog_hld.c b/kernel/watchdog_hld.c
> > > > index 247bf0b1582c..6876e796dbf5 100644
> > > > --- a/kernel/watchdog_hld.c
> > > > +++ b/kernel/watchdog_hld.c
> > > > @@ -165,10 +165,13 @@ static void watchdog_overflow_callback(struct perf_event *event,
> > > >  
> > > >  static int hardlockup_detector_event_create(void)
> > > >  {
> > > > -	unsigned int cpu = smp_processor_id();
> > > > +	unsigned int cpu;
> > > >  	struct perf_event_attr *wd_attr;
> > > >  	struct perf_event *evt;
> > > >  
> > > > +	/* This function plans to execute in cpu bound kthread */
> > > > +	BUG_ON(!is_percpu_thread());
> > > > +	cpu = raw_smp_processor_id();
> > > >  	wd_attr = &wd_hw_attr;
> > > >  	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
> > > 
> > > This patch makes no sense.
> > 
> > This patch aims to disable any attempt such as using get_cpu()/put_cpu() to
> > shut up the check_preemption_disabled().
> 
> I have to say that the description of the problem is really cryptic.
> Please, provide more context, code paths, sample code, next time.
> 
Sorry, I will be more straight forward. And I had thought commit log had
mentioned it.
> Well, I probably got it. The code might sleep. But it should run on the

And you do get it.
> same CPU even after waking up. You try to achieve this by running
> the code in a process that is bound to a single CPU.
> 
> IMHO, this is not reliable. Anyone could change the affinity
> of the process in the meantime.
> 
No, it is not. As the code says: PF_NO_SETAFFINITY.
static inline bool is_percpu_thread(void)
{
#ifdef CONFIG_SMP
	return (current->flags & PF_NO_SETAFFINITY) &&
		(current->nr_cpus_allowed  == 1);
#else
	return true;
#endif
}

This is critical for cpuhp_* (kernel/cpu.c). And
hardlockup_detector_event_create() should be planned to run on such a
kthread.

Thanks,

	Pingfan

> I see two solutions. Either avoid the sleep or making sure
> that the code access per-CPU variables on the same CPU
> all the time.
> 
> For example, you might use
> 
> 	*per_cpu_ptr(watchdog_ev, cpu) = evt;
> 
> instead of
> 
> 	this_cpu_write(watchdog_ev, evt);
> 
> Best Regards,
> Petr