On Thu, 25 Sep 2014, Lina Iyer wrote:
> > How would a general "keep track of the targets of all interrupts in
> > the system" mechanism make use of this?
> Sorry, I do not understand your question.
> PM QoS is only interested in the IRQs specified in the QoS request. If
> there are no requests that need to be associated with an IRQ, then PM
> QoS will not register for an affinity change notification.
Right, and I really hate the whole per irq notifier. It's a rats nest
of life time issues and other problems.
It also does not tell you whether an irq is disabled, reenabled or
removed, which will change the qos constraints as well unless you
plaster all drivers with updates to qos for those cases.
So what about adding a qos field to irq_data itself, have a function
to update it and let the irq core keep track of the per cpu irq
relevant qos constraints and provide an evaluation function or a
notifier for the PM/idle code?
That's going to need some serious thought as well, but it should avoid
most of the nasty notifier and lifetime issue which the per irq
notifiers provide.
Thoughts?
tglx
On Wed, Oct 08 2014 at 09:03 -0600, Thomas Gleixner wrote:
>On Thu, 25 Sep 2014, Lina Iyer wrote:
>> > How would a general "keep track of the targets of all interrupts in
>> > the system" mechanism make use of this?
>> Sorry, I do not understand your question.
>> PM QoS is only interested in the IRQs specified in the QoS request. If
>> there are no requests that need to be associated with an IRQ, then PM
>> QoS will not register for an affinity change notification.
>
>Right, and I really hate the whole per irq notifier. It's a rats nest
>of life time issues and other problems.
>
>It also does not tell you whether an irq is disabled, reenabled or
>removed, which will change the qos constraints as well unless you
>plaster all drivers with updates to qos for those cases.
>
>So what about adding a qos field to irq_data itself, have a function
>to update it and let the irq core keep track of the per cpu irq
>relevant qos constraints and provide an evaluation function or a
>notifier for the PM/idle code?
If that isnt intrusive in the IRQ core, then we can make it work for PM
QoS. The issue that I am concerned is that, it might result in back and
forth between IRQ and PM QoS frameworks. If that doesnt happen, then we
are good with this approach.
>
>That's going to need some serious thought as well, but it should avoid
>most of the nasty notifier and lifetime issue which the per irq
>notifiers provide.
Sure. I will look into this.
>
>Thoughts?
Thank you.
Lina
>
>
>
>
On Fri, 10 Oct 2014, Lina Iyer wrote:
> On Wed, Oct 08 2014 at 09:03 -0600, Thomas Gleixner wrote:
> > On Thu, 25 Sep 2014, Lina Iyer wrote:
> > > > How would a general "keep track of the targets of all interrupts in
> > > > the system" mechanism make use of this?
> > > Sorry, I do not understand your question.
> > > PM QoS is only interested in the IRQs specified in the QoS request. If
> > > there are no requests that need to be associated with an IRQ, then PM
> > > QoS will not register for an affinity change notification.
> >
> > Right, and I really hate the whole per irq notifier. It's a rats nest
> > of life time issues and other problems.
> >
> > It also does not tell you whether an irq is disabled, reenabled or
> > removed, which will change the qos constraints as well unless you
> > plaster all drivers with updates to qos for those cases.
> >
> > So what about adding a qos field to irq_data itself, have a function
> > to update it and let the irq core keep track of the per cpu irq
> > relevant qos constraints and provide an evaluation function or a
> > notifier for the PM/idle code?
>
> If that isnt intrusive in the IRQ core, then we can make it work for PM
> QoS. The issue that I am concerned is that, it might result in back and
> forth between IRQ and PM QoS frameworks. If that doesnt happen, then we
> are good with this approach.
I can't tell that upfront, but I think it's worth to explore it.
Thanks,
tglx
Hi Thomas,
On Fri, Oct 17 2014 at 01:29 -0600, Thomas Gleixner wrote:
>On Fri, 10 Oct 2014, Lina Iyer wrote:
>> On Wed, Oct 08 2014 at 09:03 -0600, Thomas Gleixner wrote:
>> > On Thu, 25 Sep 2014, Lina Iyer wrote:
>> > > > How would a general "keep track of the targets of all interrupts in
>> > > > the system" mechanism make use of this?
>> > > Sorry, I do not understand your question.
>> > > PM QoS is only interested in the IRQs specified in the QoS request. If
>> > > there are no requests that need to be associated with an IRQ, then PM
>> > > QoS will not register for an affinity change notification.
>> >
>> > Right, and I really hate the whole per irq notifier. It's a rats nest
>> > of life time issues and other problems.
>> >
>> > It also does not tell you whether an irq is disabled, reenabled or
>> > removed, which will change the qos constraints as well unless you
>> > plaster all drivers with updates to qos for those cases.
>> >
>> > So what about adding a qos field to irq_data itself, have a function
>> > to update it and let the irq core keep track of the per cpu irq
>> > relevant qos constraints and provide an evaluation function or a
>> > notifier for the PM/idle code?
>>
>> If that isnt intrusive in the IRQ core, then we can make it work for PM
>> QoS. The issue that I am concerned is that, it might result in back and
>> forth between IRQ and PM QoS frameworks. If that doesnt happen, then we
>> are good with this approach.
>
>I can't tell that upfront, but I think it's worth to explore it.
>
I was able to review the options and I attempted a few methods, but
off-loading the QoS onto the IRQ framework, made it quite complex to
manage it. QoS values for each of the four constraints and the
constraints could be one of 3 types - min, max or sum, makes it a whole
lot of mess handling it in IRQ code.
I was able to get QoS to be notified of IRQ affinity changes without
using notifiers, but, I still am yet to find a way to modify QoS
requests on enable/disable of IRQ.
I will send the RFC of the new patchset, would be interested in taking a
look and let me know what you think.
Thanks,
Lina