Re: RFC: PERCPU_LIST to fix PR kern/52515

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

On Fri, Dec 8, 2017 at 6:51 PM, Kengo NAKAHARA <k-nakahara%iij.ad.jp@localhost> wrote:
> Hi,
>
> On 2017/12/07 14:21, Taylor R Campbell wrote:
>> I dropped this thread on the floor a while ago and I forget what the
>> status was.  I've had a patch sitting in my tree for a while which I
>> brushed off to put the list update logic in separate functions, as I
>> think chuq requested a while ago, but still keep it all isolated to
>> subr_psref.c and avoid defining any new macros.
>>
>> If you've measured that SLIST works better -- which would make sense
>> because the typical bracketed psref_acquire/release nesting makes a
>> nice LIFO stack of the things so that SLIST_REMOVE should usually be
>> done in the first iteration -- then I'm happy to replace it by SLIST.
>> We should just make sure to fix this bug before netbsd-8 goes out!
>>
>> Thoughts?
>
> I measure IPv4 forwarding performance your patch(PSREF_LIST) version
> and SLIST version. At first, the result is quite affected by
> optimization option "-falign-functions"... Based on this, it seems
> there is almost no difference between PSREF_LIST and SLIST.
>
> However, it seems your patch has large diff... From the point of code
> stability, smaller diff SLIST version would be better for netbsd-8 branch
> to fix the bug. Because your patch causes some new ATF failures such as
> ldp_regen and route_change_ifp (reported by ozaki-r@n.o). We can probably
> fix them at once but guaranteeing its stability would take more time.

Hmm, the same failures also happen with the SLIST version. IIRC there
was no regression with it two months ago and so I didn't test it enough
days ago. Anyway I guess the failures are not because of your patch.
I'm sorry for that.

I have no idea why the failures happen. The patches may dig out a
latent bug in -current.

  ozaki-r