Re: CVS commit: src/sys/net

[Ryota Ozaki <ozaki-r%netbsd.org@localhost>]

On Thu, Feb 18, 2016 at 12:08 PM, Ryota Ozaki <ozaki-r%netbsd.org@localhost> wrote:
> On Thu, Feb 18, 2016 at 12:18 AM, Roy Marples <roy%marples.name@localhost> wrote:
>> On 17/02/2016 01:46, Roy Marples wrote:
>>> Patch attached to solve change from a priority array into a bit mask
>>> approach where we swallow the whole queue in a softint.
>>
>> I did the same tests as noted in PR 50602 - put the system under load by
>> building clang from src and qt5 from pkgsrc at the same time, both with
>> -j3 on my Core2 @2.4Ghz dev machine.
>> The load was measured at about 9 by top, the machine itself was pretty
>> un-useable.
>>
>> It also has a wm(4) interface, which the reporter had.
>> Also an iwn(4) interface.
>> Toggling both as fast as possible - my daughter power cycled the AP
>> while I unplugged / replugged the ethernet cable.
>>
>> Doing this for about 10 minutes resulted in no lossage reported with my
>> patch and like state transitions happened perfectly.
>
> Good.
>
>>
>> I would imagine any lossage would occur as a result of a driver or
>> hardware defect, where more important information is likely found on the
>> console as well.
>
> IIUC amd64/i386 have the fast softint facility and it tries to schedule
> a softint just after hwintr finishes, and so more than one events are
> unlikely to come before the softint starts? Please someone correct
> me if my guess is wrong.

Okay, I've changed my mind. Let's commit your patch anyway and
keep an eye on the lossage messages to know how often and how
many lossage happens. If there are issues, we can improve then.
It's not too late.

I'm sorry for disturbing you.
  ozaki-r