Re: [RFC PATCH 1/2] mac80211: make max_network_latency notifier atomic safe

[Johannes Berg <johannes@sipsolutions.net>]

On Wed, 2010-06-09 at 12:20 +0200, Florian Mickler wrote:
> On Wed, 09 Jun 2010 11:38:07 +0200
> Johannes Berg <johannes@sipsolutions.net> wrote:
> 
> > On Wed, 2010-06-09 at 11:15 +0200, florian@mickler.org wrote:
> > > In order to have the pm_qos framework be callable from interrupt
> > > context, all listeners have to also be callable in that context.
> > 
> > That makes no sense at all. Why add work structs _everywhere_ in the
> > callees and make the API harder to use and easy to get wrong completely,
> > instead of just adding a single work struct that will be queued from the
> > caller and dealing with the locking complexity etc. just once.


> There are only two listeners at the moment. I suspect that most future
> uses of the framework need to be atomic, as the driver that
> requests a specific quality of service probably doesn't want to get into
> races with the provider of that service(listener). So i suspected the
> network listener to be the special case.  

Well even if it doesn't _want_ to race with it, a lot of drivers like
USB drivers etc. can't really do anything without deferring to a
workqueue.

And what's the race anyway? You get one update, defer the work, and if
another update happens inbetween you just read the new value when the
work finally runs -- and you end up doing it only once instead of twice.
That doesn't seem like a problem.

> The race between service-provider and qos-requester for non-atomic
> contextes is already there, isn't it? so, locking complexity shouldn't
> be worse than before.

I have no idea how it works now? I thought you can't request an update
from an atomic context.

However, if you request a QoS value, it is fundamentally that -- a
request. There's no guarantee as to when or how it will be honoured.

> But my first approach to this is seen here:
> https://lists.linux-foundation.org/pipermail/linux-pm/2010-June/026902.html

Icky too.

> A third possibility would be to make it dependent on the
> type of the constraint, if blocking notifiers are allowed or not. 
> But that would sacrifice API consistency (update_request for one
> constraint is allowed to be called in interrupt context and
> update_request for another would be not).

I don't see what's wrong with the fourth possibility: Allow calling
pm_qos_update_request() from atomic context, but change _it_ to schedule
off a work that calls the blocking notifier chain. That avoids the
complexity in notify-API users since they have process context, and also
in request-API users since they can call it from any context.

johannes