Re: [PATCH] pm_qos: Add system bus performance parameter

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Tuesday, September 07, 2010, mark gross wrote:
> On Thu, Sep 02, 2010 at 10:09:20PM +0200, Rafael J. Wysocki wrote:
> > On Thursday, September 02, 2010, mark gross wrote:
> > > On Fri, Aug 27, 2010 at 07:31:46AM -0700, Kevin Hilman wrote:
> > > > Saravana Kannan <skannan@codeaurora.org> writes:
> > > > 
> > > > > Some drivers/devices might need some minimum system bus performance to
> > > > > provide acceptable service. Provide a PM QoS parameter to send these requests
> > > > > to.
> > > > >
> > > > > The new parameter is named "system bus performance" since it is generic enough
> > > > > for the unit of the request to be frequency, bandwidth or something else that
> > > > > might be appropriate. It's up to each implementation of the QoS provider to
> > > > > define what the unit of the request would be.
> > > > >
> > > > > Signed-off-by: Saravana Kannan <skannan@codeaurora.org>
> > > > 
> > > > With this current design, only one system-wide bus would be managed.
> > > > What if a platform has more than one independently scalable bus?
> > > > 
> > > > I think the only scalable way to handle this kind of thing is to have
> > > > per-device QoS constraints that can then be combined/aggregated by parent
> > > > devices/busses.
> > > > 
> > > > At LPC this year, I've proposed per-device QoS constraints[1] as a topic
> > > > for the PM mini-conf.  I hope some folks from the MSM camp can be there
> > > > for these discussions.
> > > > 
> > > > Kevin
> > > > 
> > > > [1] http://www.linuxplumbersconf.org/2010/ocw/proposals/819
> > > 
> > > I thought a pm_qos like thing per bus would be a patch or you where
> > > going to put up to the driver model. ;)
> > > 
> > > The current pm_qos would stick around for higher level pm_qos things.
> > > So making the system bus and changing to a summation aggregation would
> > > be  temporary thing.  
> > > 
> > > Or are you you saying we shouldn't put system_bus into pm_qos at all and
> > > instead we should put effort into adding it to the driver model for
> > > buses?
> > 
> > Hmm, well, what's system_bus?
> > 
> > Rafael
> 
> My understanding is that system_bus is a somewhat generic concept that
> has meaning only in platform specific hardware configurations pm_qos
> request class for device specific buses.  Memory, SDIO, SPI, i2c etc.
> 
> I'm not sure if it should be exposed up to user mode as an ABI.  Its
> pretty abstract and its meaning is perhaps a bit mutable across target 
> architectures at this time.

No, it shouldn't.

Thanks,
Rafael