Re: MP-safe IPsec SPD

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

On Tue, Aug 8, 2017 at 3:07 AM, Taylor R Campbell
<campbell+netbsd-tech-kern%mumble.net@localhost> wrote:
>> Date: Mon, 7 Aug 2017 12:40:11 +0900
>> From: Ryota Ozaki <ozaki-r%netbsd.org@localhost>
>>
>> Thank you for your reviewing! I think I handled all your suggestions.
>> Let me know if I missed or misunderstood something.
>
> Thanks!  Looks good.
>
> Couple tiny nits:
>
> - `pserialize critical section' ---> `pserialize read section'?
>   (Although we sometimes say `critical section' in conversation, the
>   functions are called pserialize_read_enter/exit, and it may not be
>   immediately clear whether `pserialize critical section' means a
>   nonexclusive reader or an exclusive writer.)
>
> - __read_mostly for key_psz?

Thanks. Changed so.

>
> While skimming the patch, I noticed an odd global variable acq_seq
> that is used only under certain #ifdefs and seems to modified without
> any synchronization.  I wonder what it's there for and whether it is
> OK for updates to be lost.

I think we can just remove IPSEC_NONBLOCK_ACQUIRE because, I guess,
nobody uses the option for long years.

>
>> BTW can we have an API to assert that we are in a pserialize critical
>> section? Sometimes I want to write something like:
>>   KASSERT(pserialize_in_read_critical_section());
>> rather than writing the constraint in a comment.
>
> That would be nice to have.  It would require for each machine some
> way to ask whether the current IPL is at least some specified IPL,
> which as far as I know we don't currently have.  (Nor do we have any
> MI way to compare IPLs (both numerical orders are used, so IPL_VM <
> IPL_HIGH may be true or false), nor any MI way to query the current
> IPL.)

Hmm, IPL things are so MD. Please someone work on it :-|

  ozaki-r