Re: [RFD PATCH 3/5] sched/cpufreq_schedutil: make worker kthread be SCHED_DEADLINE
From: Peter Zijlstra
Date: Tue Mar 28 2017 - 06:21:12 EST
On Tue, Mar 28, 2017 at 11:29:31AM +0200, Vincent Guittot wrote:
> On 27 March 2017 at 18:50, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Fri, Mar 24, 2017 at 02:08:58PM +0000, Juri Lelli wrote:
> >> Worker kthread needs to be able to change frequency for all other
> >> threads.
> >>
> >> Make it special, just under STOP class.
> >
> > *yuck* ;-)
> >
> > So imagine our I2C/SPI bus is 'busy' and its mutex taken, then this
> > 'soecial' task will need to boost it. Now add BWI to your thinking and
> > shudder.
> >
> >
> > On IRC broonie mentioned that:
> >
> > - most PMIC operations are fire and forget (no need to wait for a
> > response).
> > - PMIC 'packets' are 'small'.
> > - SPI has the possibility to push stuff on the queue.
> >
> > Taken together this seems to suggest we can rework cpufreq drivers to
> > function in-context, either directly push the packet on the bus if
> > available, or queue it and let whoever owns it sort it without blocking.
> >
> > It might be possible to rework/augment I2C to also support pushing stuff
> > on a queue.
>
> But sending new voltage value to PMIC is only part of the sequence.
> When cpufreq set a new opp, it does
>
> -set new voltage
> -wait for the voltage to settle down.
> -set the new clock frequency
>
> you can even have to switch to an intermediate clock source.
>
> When such sequence is managed by the kernel, we can't easily git ride
> of a kthread
That stinks :-(
The whole blocking and bandwidth inheritance stuff gives me nightmares.
Makes all this stuff almost impossible to analyse.
Esp. if the bus (SPI/I2C) is shared with some other 'expensive' device
like a DSP or MMC flash crud.