Re: [PATCH V2] i2c-designware: Add Intel Baytrail PMIC I2C bus support

[Maxime Ripard]

Hi David,

On Tue, Oct 07, 2014 at 12:14:20PM -0700, David E. Box wrote:
> Hi Maxime,
> 
> On Thu, Sep 25, 2014 at 11:47:52AM +0200, Maxime Ripard wrote:
> > Hi David,
> > 
> > On Tue, Sep 23, 2014 at 12:58:54PM -0700, David E. Box wrote:
> > > Hi Maxime,
> > > 
> > > On Tue, Sep 23, 2014 at 09:00:57PM +0200, Maxime Ripard wrote:
> > > > Hi David,
> > > > 
> > > > On Tue, Sep 23, 2014 at 11:40:26AM -0700, David E. Box wrote:
> > > > > This patch implements an I2C bus sharing mechanism between the host and platform
> > > > > hardware on select Intel BayTrail SoC platforms using the X-Powers AXP288 PMIC.
> > > > > 
> > > > > On these platforms access to the PMIC must be shared with platform hardware. The
> > > > > hardware unit assumes full control of the I2C bus and the host must request
> > > > > access through a special semaphore. Hardware control of the bus also makes it
> > > > > necessary to disable runtime pm to avoid interfering with hardware transactions.
> > > > > 
> > > > > Signed-off-by: David E. Box <david.e.box@xxxxxxxxxxxxxxx>
> > > > 
> > > > Sorry for stepping in like this without really knowing your platform,
> > > > but wouldn't using the hwspinlock framework make more sense than
> > > > hardcoding your own internal functions here?
> > > 
> > > I looked into this but didn't see a clear way on our platform to identify the
> > > semaphore seperately from doing it in the designware platform driver. The way
> > > we can find it now is through evaluating an ACPI _SEM object on every i2c device
> > > that gets probed by the dw driver since at probe time we can get the acpi handle.
> > 
> > And you have no way to turn it around and identify which semaphore is
> > associated to which i2c bus?
> > 
> > If so, there is probably some way to associate a given instance of the
> > i2c driver to one semaphore.
> > 
> > > Without this handle however there isn't a clear way of evaluating the _SEM
> > > object which would be needed to register a hwspinlock in separate code.
> > > 
> > > Plus it would still require changes to the designware i2c core, though admittedly
> > > having a generic hwspinlock pointer added to the struct is cleaner.
> > 
> > Not only cleaner, but that could also be used by other platforms that
> > are using this I2C driver (and since it's a designware IP, there must
> > be quite a lot) together with hardware locking.
> > 
> 
> After again considering a way to make this work I don't think this api can fit
> well with our platform. Acquisition of this semaphore is through a mailbox
> sequence where we set one register and then poll another for a value that
> confirms we have the lock. For best performance we need to be able to
> periodically sleep while waiting for that confirmation. This time can vary
> widely as it's dependent on the component we are requesting the semaphore from
> which is itself a user of that bus.
> 
> While we could simply fail after a short time, reattempts would still need
> to happen in the i2c-designware driver and the timing would be completely
> dependent on our hardware, making it less clean for reuse. In addition,
> if we timed out, we would have to immediately call unlock to cancel the
> mailbox transaction. This may not fit well with reuse either.

Ok, if Wolfram is ok with it, and if it makes your life much easier,
I'm ok :)

Thanks!
Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

Attachment: signature.asc
Description: Digital signature