Re: [RFC PATCH net-next 2/2] net: napi: Skip poll when arming GRO timer in busy poll

["Björn Töpel" <bjorn@kernel.org>]

Dragos Tatulea <dtatulea@nvidia.com> writes:

> From: Martin Karsten <mkarsten@uwaterloo.ca>
>
> As referenced in the previous patch, having the GRO timer scheduled
> while poll is running can lead to issues.

Outline the issues here as well, so we don't need to cross-reference the
kernel logs.

> Skip the extra call to napi->poll() when the GRO timer is armed. This
> removes the need for having a separate __busy_poll_stop routine and its
> code is moved directly into the relevant places in busy_poll_stop.

Nice!

> Signed-off-by: Martin Karsten <mkarsten@uwaterloo.ca>
> Co-developed-by: Dragos Tatulea <dtatulea@nvidia.com>
> Signed-off-by: Dragos Tatulea <dtatulea@nvidia.com>
> ---
>  net/core/dev.c | 61 +++++++++++++++++++++++++-------------------------
>  1 file changed, 30 insertions(+), 31 deletions(-)
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 1487d4946dcf..d4829c2484f5 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -6869,26 +6869,6 @@ static void skb_defer_free_flush(void)
>  
>  #if defined(CONFIG_NET_RX_BUSY_POLL)
>  
> -static void __busy_poll_stop(struct napi_struct *napi, unsigned long timeout)
> -{
> -	unsigned long flags;
> -
> -	if (!timeout) {
> -		gro_normal_list(&napi->gro);
> -		__napi_schedule(napi);
> -		return;
> -	}
> -
> -	/* Flush too old packets. If HZ < 1000, flush all packets */
> -	gro_flush_normal(&napi->gro, HZ >= 1000);
> -
> -	local_irq_save(flags);
> -	hrtimer_start(&napi->timer, ns_to_ktime(timeout),
> -		      HRTIMER_MODE_REL_PINNED);
> -	clear_bit(NAPI_STATE_SCHED, &napi->state);
> -	local_irq_restore(flags);
> -}
> -
>  enum {
>  	NAPI_F_PREFER_BUSY_POLL	= 1,
>  	NAPI_F_END_ON_RESCHED	= 2,
> @@ -6898,14 +6878,14 @@ static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock,
>  			   unsigned flags, u16 budget)
>  {
>  	struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx;
> -	unsigned long timeout = 0;
> +	unsigned long timeout;
>  	int rc;
>  
>  	/* Busy polling means there is a high chance device driver hard irq
>  	 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
>  	 * set in napi_schedule_prep().
> -	 * Since we are about to call napi->poll() once more, we can safely
> -	 * clear NAPI_STATE_MISSED.
> +	 * Since we either call napi->poll() once more or start the timer,
> +	 * we can safely clear NAPI_STATE_MISSED.
>  	 *
>  	 * Note: x86 could use a single "lock and ..." instruction
>  	 * to perform these two clear_bit()
> @@ -6918,13 +6898,28 @@ static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock,
>  
>  	if (flags & NAPI_F_PREFER_BUSY_POLL) {
>  		napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi);
> -		if (napi->defer_hard_irqs_count) {
> -			/* Timer will be scheduled after napi poll to avoid
> -			 * firing during a slow poll which could cause the
> -			 * queue to get stuck with interrupts disabled and no
> -			 * scheduled timer.
> +		timeout = napi_get_gro_flush_timeout(napi);
> +		if (napi->defer_hard_irqs_count && timeout) {
> +			unsigned long flags;
> +
> +			/* Drop prefer-busy state as in napi_complete_done(). */
> +			clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
> +			netpoll_poll_unlock(have_poll_lock);
> +
> +			/* Flush too old packets. If HZ < 1000, flush all
> +			 * packets.
> +			 */
> +			gro_flush_normal(&napi->gro, HZ >= 1000);
> +			local_irq_save(flags);
> +			hrtimer_start(&napi->timer, ns_to_ktime(timeout),
> +				      HRTIMER_MODE_REL_PINNED);
> +			clear_bit(NAPI_STATE_SCHED, &napi->state);
> +			local_irq_restore(flags);

(Same swap arm timer/clear as patch 1.)

> +
> +			/* Timer started, so need for another call to
                                            ^
                                       "so *NO* need for another..."



Björn