Re: [PATCH V2 net 10/20] net/ena: remove redundant logic in napi callback for busy poll mode

[Eric Dumazet]

On Sun, 2016-12-04 at 15:19 +0200, Netanel Belgazal wrote:
> sk_busy_loop can call the napi callback few million times a sec.
> For each call there is unmask interrupt.
> We want to reduce the number of unmasks.
> 
> Add an atomic variable that will tell the napi handler if
> it was called from irq context or not.
> Unmask the interrupt only from irq context.
> 
> A schenario where the driver left with missed unmask isn't feasible.
> when ena_intr_msix_io is called the driver have 2 options:
> 1)Before napi completes and call napi_complete_done
> 2)After calling napi_complete_done
> 
> In the former case the napi will unmask the interrupt as needed.
> In the latter case napi_complete_done will remove napi from the schedule
> list so napi will be rescheduled (by ena_intr_msix_io) and interrupt
> will be unmasked as desire in the 2nd napi call.
> 
> Signed-off-by: Netanel Belgazal <netanel@xxxxxxxxxxxxxxxxx>
> ---


This looks very complicated to me.

I guess you missed the recent patches that happened on net-next ?

2e713283751f494596655d9125c168aeb913f71d net/mlx4_en: use napi_complete_done() return value
364b6055738b4c752c30ccaaf25c624e69d76195 net: busy-poll: return busypolling status to drivers
21cb84c48ca0619181106f0f44f3802a989de024 net: busy-poll: remove need_resched() from sk_can_busy_loop()
217f6974368188fd8bd7804bf5a036aa5762c5e4 net: busy-poll: allow preemption in sk_busy_loop()

napi_complete_done() return code can be used by a driver,
no need to add yet another atomic operation in fast path.

Anyway, this looks wrong :

@@ -1186,6 +1201,7 @@ static irqreturn_t ena_intr_msix_io(int irq, void *data)
 {
        struct ena_napi *ena_napi = data;
 
+       atomic_set(&ena_napi->unmask_interrupt, 1);
        napi_schedule(&ena_napi->napi);
 
You probably wanted :

if (napi_schedule_prep(n)) {
	atomic_set(&ena_napi->unmask_interrupt, 1);
	__napi_schedule(n);
}



Please rework this napi poll using core infrastructure.

busypoll logic should be centralized, not reimplemented in different ways in a driver.

Thanks.