Re: [PATCH] mm/vmscan: fix data races at kswapd_classzone_idx

[Qian Cai]

> On Feb 25, 2020, at 8:29 PM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> 
> On Tue, Feb 25, 2020 at 11:55:26AM -0500, Qian Cai wrote:
>> -	if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES)
>> -		pgdat->kswapd_classzone_idx = classzone_idx;
>> -	else
>> -		pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx,
>> -						  classzone_idx);
>> +	if (READ_ONCE(pgdat->kswapd_classzone_idx) == MAX_NR_ZONES ||
>> +	    READ_ONCE(pgdat->kswapd_classzone_idx) < classzone_idx)
>> +		WRITE_ONCE(pgdat->kswapd_classzone_idx, classzone_idx);
>> +
>> 	pgdat->kswapd_order = max(pgdat->kswapd_order, order);
> 
> Doesn't this line have the exact same problem you're "solving" above?

Good catch. I missed that.

> 
> Also, why would you do this crazy "f(READ_ONCE(x)) || g(READ_ONCE(x))?
> Surely you should be stashing READ_ONCE(x) into a local variable?

Not a bad idea.

> 
> And there are a _lot_ of places which access kswapd_classzone_idx
> without a lock.  Are you sure this patch is sufficient?

I am not sure, but KCSAN did not complain about other places after running extended
period of testing, so I will look into once other places show up later.