Re: [PATCH] ath9k: break out of irq handler after 5 jiffies

[Ben Greear <greearb@candelatech.com>]

On 02/26/2018 02:08 PM, Arend van Spriel wrote:
> On 2/26/2018 10:39 PM, Ben Greear wrote:
>> On 02/07/2018 07:39 AM, Ben Greear wrote:
>>>
>>>
>>> On 02/07/2018 02:55 AM, Johannes Berg wrote:
>>>> On Wed, 2018-02-07 at 10:16 +0100, Felix Fietkau wrote:
>>>>> On 2018-02-07 00:05, greearb@candelatech.com wrote:
>>>>>> From: Ben Greear <greearb@candelatech.com>
>>>>>>
>>>>>> In case where the system is sluggish, we should probably break out
>>>>>> early.  Maybe this will fix issues where the OS thinks the IRQ handler
>>>>>> is not responding and disables the IRQ because 'nobody cared'
>>>>>>
>>>>>> Signed-off-by: Ben Greear <greearb@candelatech.com>
>>>>>
>>>>> 5 jiffies as a hardcoded value is a bad idea, since it produces
>>>>> different behavior based on CONFIG_HZ.
>>>
>>> I figured that was a benefit since it would run shorter duration on
>>> systems with
>>> a faster HZ clock.
>>>
>>>>
>>>> Also, err, NAPI? Or is something else is going on here?
>>>
>>> I don't really know, but part of my test was running traffic while
>>> creating
>>> 1200 stations, so likely there were lots of higher-level lock
>>> contention that
>>> slowed down sending pkts up the stack.
>>>
>>> I got a bunch of errors about IRQs being ignored because nobody
>>> cared.  I noticed
>>> that the ath9k loop could handle up to 500 or so frames, and that
>>> seemed like too
>>> many for my particular test case.
>>>
>>> Once I put in this patch, I did not see the 'nobody cared' error again.
>>>
>>> There could easily be a better fix.  If you all want me to use a fixed
>>> time instead
>>> of HZ, then please suggest a value.  I was testing with HZ of 1000, btw.
>>
>> Hello,
>>
>> I don't mind changing this patch, but I could use some guidance as to what
>> values you all want me to use.
>>
>> Should I use a millisecond based clock instead of jiffies?
>>
>> What time duration do you want if 5 Jiffies (or 5ms) is not desired?
>
> Hi Ben,
>
> Instead of using some time unit you could consider breaking out after handing 'x' number of frames and make 'x' configurable through debugfs.

I don't see why you would care about number of pkts...it is just a proxy for time, right?

So, in that case, then using jiffies (or some other fast timer) seems the most useful.

Thanks,
Ben

>
> Regards,
> Arend
>


-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com