Re: Softirq priority inversion from "softirq: reduce latencies"

From: Peter Hurley
Date: Mon Feb 29 2016 - 13:53:35 EST


On 02/29/2016 10:24 AM, Eric Dumazet wrote:
> On lun., 2016-02-29 at 10:05 -0800, Peter Hurley wrote:
>
>> While I appreciate the attempt, that's not the problem.
>>
>> Just to be clear
>>
>> if (time_before(jiffies, end) && !need_resched() &&
>> --max_restart)
>> goto restart;
>>
>> aborts softirq *even if 0ns have elapsed*, if NET_RX has woken a process.
>
>
> Sure, now remove the 1st and 2nd condition.

Well just removing the 2nd condition has everything working fine,
because that fixes the priority inversion.


> You would still 'abort' (ie wakeup ksoftirqd really) when --max_restart
> becomes 0

Sure. Which would mean there's contended heavy i/o load so the driver
has to fallback to non-DMA. That's an acceptable outcome.


> So, instead of some subtle load dependent bug, you know have a reliable
> trigger.

There's no "subtle load dependent bug" here.

The driver has a fallback mode of operation that it relies on without
DMA. Of course, as I already wrote, this has consequences.

If system resources are _actually contended_, then naturally, fighting
for cpu and i/o time is fine, and I'm happy to do that in ksoftirqd.

However, when system resources are _not_ contended, it makes no
sense to be forced to revert to ksoftirqd resolution, which is strictly
intended as fallback.

Or flipping your argument on its head, why not just _always_ execute
softirq in ksoftirqd?