Hi James!
On Mon, 14 Jun 2010 16:46:40 +0200
[email protected] wrote:
> With this patch we only schedule the work when in interrupt context.
>
> Before update_request was callable from interrupt-context there was a
> 1:1 relation between a change in the request-value and a notification.
> This patch restores that behaviour for all constraints that have update_request
> never called from interrupt context.
>
> The notifier mutex serializes calls to blocking_notifier_call_chain, so
> that we are serialized against any pending or currently executing notification.
>
> Signed-off-by: Florian Mickler <[email protected]>
> ---
> kernel/pm_qos_params.c | 10 +++++++---
> 1 files changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/kernel/pm_qos_params.c b/kernel/pm_qos_params.c
> index 9346906..c06cae9 100644
> --- a/kernel/pm_qos_params.c
> +++ b/kernel/pm_qos_params.c
> @@ -152,11 +152,15 @@ static s32 min_compare(s32 v1, s32 v2)
> static void pm_qos_call_notifiers(struct pm_qos_object *o,
> unsigned long curr_value)
> {
> - schedule_work(&o->notify);
> -
> if (o->atomic_notifiers)
> atomic_notifier_call_chain(o->atomic_notifiers,
> - curr_value, NULL);
> + (unsigned long) curr_value, NULL);
> +
> + if (in_interrupt())
> + schedule_work(&o->notify);
> + else
> + blocking_notifier_call_chain(o->blocking_notifiers,
> + (unsigned long) curr_value, NULL);
> }
>
> static void update_notify(struct work_struct *work)
What about this? Is this ok? I don't know if it is benign to use
in_interrupt() here. I took this idea from the
execute_in_process_context() implementation.
If this is ok, should I rebase them on your two pm_qos patches (plists
and the kzalloc removal)?
Did you already thought about some debugging stuff that would suffice
the android needs? I kind of thought about either registerieng some
notifier callback or using the perf/tracing infrastructure.
But I have not looked into it yet.
Cheers,
Flo
On Tue, 2010-06-15 at 19:23 +0200, Florian Mickler wrote:
> Hi James!
>
> On Mon, 14 Jun 2010 16:46:40 +0200
> [email protected] wrote:
>
> > With this patch we only schedule the work when in interrupt context.
> >
> > Before update_request was callable from interrupt-context there was a
> > 1:1 relation between a change in the request-value and a notification.
> > This patch restores that behaviour for all constraints that have update_request
> > never called from interrupt context.
> >
> > The notifier mutex serializes calls to blocking_notifier_call_chain, so
> > that we are serialized against any pending or currently executing notification.
> >
> > Signed-off-by: Florian Mickler <[email protected]>
> > ---
> > kernel/pm_qos_params.c | 10 +++++++---
> > 1 files changed, 7 insertions(+), 3 deletions(-)
> >
> > diff --git a/kernel/pm_qos_params.c b/kernel/pm_qos_params.c
> > index 9346906..c06cae9 100644
> > --- a/kernel/pm_qos_params.c
> > +++ b/kernel/pm_qos_params.c
> > @@ -152,11 +152,15 @@ static s32 min_compare(s32 v1, s32 v2)
> > static void pm_qos_call_notifiers(struct pm_qos_object *o,
> > unsigned long curr_value)
> > {
> > - schedule_work(&o->notify);
> > -
> > if (o->atomic_notifiers)
> > atomic_notifier_call_chain(o->atomic_notifiers,
> > - curr_value, NULL);
> > + (unsigned long) curr_value, NULL);
> > +
> > + if (in_interrupt())
> > + schedule_work(&o->notify);
> > + else
> > + blocking_notifier_call_chain(o->blocking_notifiers,
> > + (unsigned long) curr_value, NULL);
> > }
> >
> > static void update_notify(struct work_struct *work)
>
> What about this? Is this ok? I don't know if it is benign to use
> in_interrupt() here. I took this idea from the
> execute_in_process_context() implementation.
I think it will work ... but I still think it's over complex given the
listed requirements (android seems to only want atomic notifiers from
atomic contexts).
> If this is ok, should I rebase them on your two pm_qos patches (plists
> and the kzalloc removal)?
Well, I would say yes. However, for more impartial advice, I'd wait and
see what the pm maintainers want.
> Did you already thought about some debugging stuff that would suffice
> the android needs? I kind of thought about either registerieng some
> notifier callback or using the perf/tracing infrastructure.
> But I have not looked into it yet.
I was just going to try the conversion when the wakelocks stuff was
finally in and see if it worked in an android kernel.
James