Re: [PATCH v2] mvneta: fix "napi poll" infinite loop

[Maxime Chevallier <maxime.chevallier@bootlin.com>]

Hello Julien,

On Wed, 11 Sep 2024 13:28:46 +0200
Julien Blais <webmaster@jbsky.fr> wrote:

> In percpu mode, when there's a network load, one of the cpus can be
> solicited without having anything to process.
> If 0 is returned to napi poll, napi will ignore the next requests,
> causing an infinite loop with ISR handling.
> 
> Without this change, patches hang around fixing the queue at 0 and
> the interrupt remains stuck on the 1st CPU.
> The percpu conf is useless in this case, so we might as well remove it.
> 
> Signed-off-by: Julien Blais <webmaster@jbsky.fr>

If this patch is a fix, you must also include a "Fixes" tag stating
which commit you are fixing. You must also include in the subject
whether you are targetting the "net-next" tree or "net". In your case,
this would be the "net" tree, that gathers the bugfixes.

> ---
>  drivers/net/ethernet/marvell/mvneta.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
> index 3f124268b..b6e89b888 100644
> --- a/drivers/net/ethernet/marvell/mvneta.c
> +++ b/drivers/net/ethernet/marvell/mvneta.c
> @@ -3186,7 +3186,10 @@ static int mvneta_poll(struct napi_struct *napi, int budget)
>  
>  	if (rx_done < budget) {
>  		cause_rx_tx = 0;
> -		napi_complete_done(napi, rx_done);
> +		if (rx_done)
> +			napi_complete_done(napi, rx_done);
> +		else
> +			napi_complete(napi);

I don't quite get this patch. as napi_complete is just calling
napi_complete_done(napi, 0), so this is basically

if (rx_done != 0)
	napi_complete_done(napi, rx_done);
else
	napi_complete_done(napi, 0);

So, nothing actually changes.

Can you elaborate more on the issue you are facing ?

Thanks,

Maxime