Re: RFC: softint-based if_input

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

On Wed, Feb 3, 2016 at 5:13 PM, Nick Hudson <skrll%netbsd.org@localhost> wrote:
> On 02/03/16 04:44, Ryota Ozaki wrote:
>>
>> On Tue, Feb 2, 2016 at 8:51 AM, Taylor R Campbell
>> <campbell+netbsd-tech-kern%mumble.net@localhost> wrote:
>
> [snip]
>>>
>>> In this case, it's not a priori clear to me why the caller must be in
>>> hardintr context.  Perhaps you want to discourage callers from using
>>> this if they're already in softint context.  But even if the caller is
>>> in softint context, maybe it's higher-priority softint context than
>>> softnet (e.g., I think skrll@ plans to make all USB softints run at
>>> softserial instead of softnet), and maybe you don't want ifp->if_input
>>> to run at >softnet priority.
>
>
> I'm still debating what to do here, but it's clear to me that all USB
> callbacks running at splsoftnet
> has issues - see
>
>         http://gnats.netbsd.org/50491
>
> and somewhat related
>
>         http://gnats.netbsd.org/49065

Hmm, it looks that the issue isn't due to splsofnet but rather due to
bad use of softnet_lock (for example it is held longer than it should be,
it can be held on socket operations that even don't run in softnet,
etc.).

Once Mp-safe work is done, softnet_lock will disappear (or at least we
stop holding it on calling a driver), and the issue will be resolved.

BTW can we replace softnet_lock with KERNEL_LOCK until getting
MP-safe work done? That may be a backdown though, it may be less
problematic.

>
>> Oh, that's a case I didn't take into account. Running if->if_input
>> at >softnet priority would break the assumption that we rely on.
>
>
> What assumption are you making?

If networking codes run in softint, they're in softnet (or lower) ipl.
So we can use splsoftnet for mutual exclusion.

>
>> We may be able to change the assert to
>>    KASSERT(cpu_intr_p() || (cpu_softintr_p() && level != SOFTINT_NET))
>> though, I think we don't need to be paranoid for the case.
>> (And I don't know how to get level. Is there API?).
>
>
> There isn't an API afaik.

I see. Thanks.

  ozaki-r