Re: [PATCH 4.19 625/639] packet: fix data-race in fanout_flow_is_huge()

[Pavel Machek]

Hi!

> From: Eric Dumazet <edumazet@xxxxxxxxxx>
> 
> [ Upstream commit b756ad928d98e5ef0b74af7546a6a31a8dadde00 ]
> 
> KCSAN reported the following data-race [1]
> 
> Adding a couple of READ_ONCE()/WRITE_ONCE() should silence it.
> 
> Since the report hinted about multiple cpus using the history
> concurrently, I added a test avoiding writing on it if the
> victim slot already contains the desired value.

>  static bool fanout_flow_is_huge(struct packet_sock *po, struct sk_buff *skb)
>  {
> -	u32 rxhash;
> +	u32 *history = po->rollover->history;
> +	u32 victim, rxhash;
>  	int i, count = 0;
>  
>  	rxhash = skb_get_hash(skb);
>  	for (i = 0; i < ROLLOVER_HLEN; i++)
> -		if (po->rollover->history[i] == rxhash)
> +		if (READ_ONCE(history[i]) == rxhash)
>  			count++;
>  
> -	po->rollover->history[prandom_u32() % ROLLOVER_HLEN] = rxhash;
> +	victim = prandom_u32() % ROLLOVER_HLEN;
> +
> +	/* Avoid dirtying the cache line if possible */
> +	if (READ_ONCE(history[victim]) != rxhash)
> +		WRITE_ONCE(history[victim], rxhash);
> +

Replacing simple asignment with if() is ... not nice and with all the
"volatile" magic in _ONCE macros may not be win for
everyone. [Actually, I don't think this is win here. This is not
exactly hot path, is it? Is it likely that array aready contains
required value?]

If this is going to get more common, should we get
WRITE_ONCE_NONDIRTY() macro hiding the uglyness?

Best regards,
								Pavel

-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Attachment: signature.asc
Description: Digital signature