Re: [PATCH v2 6/7] i2c: pxa: support i2c controller from DT

From: Grant Likely <grant.likely-s3s/WqlpOiPyB63q8FvJNQ@public.gmane.org>
To: Russell King - ARM Linux <linux-lFZ/pmaqli7XmaaqVzeoHQ@public.gmane.org>
Cc: eric.y.miao-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org,
	devicetree-discuss-uLR06cmDAlY/bJ5BZ2RsiQ@public.gmane.org,
	Haojian Zhuang
	<haojian.zhuang-eYqpPyKDWXRBDgjK7y7TUQ@public.gmane.org>,
	linux-arm-kernel-IAPFreCvJWM7uuMidbF8XUB+6BGkLq7r@public.gmane.org,
	alan-VuQAYsv1563Yd54FQh9/CA@public.gmane.org
Subject: Re: [PATCH v2 6/7] i2c: pxa: support i2c controller from DT
Date: Sat, 30 Jul 2011 18:38:46 -0600	[thread overview]
Message-ID: <20110731003846.GA21433@ponder.secretlab.ca> (raw)
In-Reply-To: <20110730153728.GA17570-l+eeeJia6m9vn6HldHNs0ANdhmdF6hFW@public.gmane.org>

On Sat, Jul 30, 2011 at 04:37:28PM +0100, Russell King - ARM Linux wrote:
> On Sat, Jul 30, 2011 at 04:29:50AM -1000, Mitch Bradley wrote:
> > On 7/29/2011 6:55 AM, Russell King - ARM Linux wrote:
> >> On Fri, Jul 29, 2011 at 10:52:22AM -0600, Grant Likely wrote:
> >>> On Thu, Jul 28, 2011 at 02:41:32PM +0800, Haojian Zhuang wrote:
> >>>> -	/*
> >>>> -	 * If "dev->id" is negative we consider it as zero.
> >>>> -	 * The reason to do so is to avoid sysfs names that only make
> >>>> -	 * sense when there are multiple adapters.
> >>>> -	 */
> >>>> -	i2c->adap.nr = dev->id;
> >>>> -	snprintf(i2c->adap.name, sizeof(i2c->adap.name), "pxa_i2c-i2c.%u",
> >>>> -		 i2c->adap.nr);
> >>>>
> >>>> -	i2c->clk = clk_get(&dev->dev, NULL);
> >>>> +	if (np) {
> >>>> +		i2c->adap.nr = idx++;
> >>>
> >>> Use this so that a bus number gets dynamically assigned:
> >>> 		i2c->adap.nr = -1;
> >>>
> >>>> +		snprintf(i2c->adap.name, sizeof(i2c->adap.name),
> >>>> +			"pxa2xx-i2c.%u", i2c->adap.nr);
> >>>> +		i2c->clk = clk_get_sys(i2c->adap.name, NULL);
> >>>
> >>> Missing i2c->adap.dev.of_node = dev->dev.of_node;
> >>
> >> And here we go again.  Is it really the case that this DT stuff doesn't
> >> have stable device names?

The "full_name" is indeed stable for a device_node, so that isn't
actually the problem, but the full_name is the entire path from the
root of the tree to the device.  The problem is that it is far too
long for to be used as (struct device*)->kobj.name, and it has '/'
characters in it which aren't particularly friendly.  The short node
name is unique among its siblings, but not globally.

Currently the kernel constructs a 'short' name when creating a device
with a heuristic that tries to use the physical address of the device
and the node name.  If that isn't available (ie. non-memory mapped
devices), then it uses a globally incremented integer to assign each
device a unique name so that sysfs doesn't freak out.

See of_device_make_bus_id() in drivers/of/platform.c

The /real/ problem is that I don't much like the heuristic; but I
haven't been able to think of anything better.  There is enough
uncertainty (very tiny, but uncertainty never the less) on device
names, and the fact that I'm hoping to improve the method of creating
device names, that I don't want to rely on them for matching up
resources.

Instead, the dev_resdata array that is passed into
of_platform_populate() has a field to override the heuristic generated
name which neatly solves the problem entirely without tracking down
all the references that need to be added, and has the added
advantage of DT and non-DT platforms using the same names.

The clk_get() changes in this patch are absolutely the wrong solution.

> > Device tree names are completely stable, based on hardware addresses  
> > that don't change from boot to boot.  Even for buses where access  
> > addresses are dynamically assigned, the device tree "reg property"  
> > address is based on a stable addressing form.  For example, PCI devices  
> > use the (stable) configuration address as the reg property and USB  
> > devices use the (stable) hub port number.
> >
> > People's tendency to want to assign sequential small integers in Linux  
> > has nothing to do with the device tree.  I suspect that it's a carryover  
> > from the historical Unix major/minor device numbering model, but in any  
> > case, there's nothing unstable about the device tree naming model. Quite 
> > the opposite - stable naming was a fundamental criterion when I designed 
> > Open Firmware.
> 
> Which means - if Grant is accepting the conversion of ARM to DT and
> upstreaming it, he needs to keep an eye on this madness and reject
> stuff which tries to do as per this patch.

Correct, I will nack/reject any patches messing about with clock
attachment or similar changes.  I was distracted by the other issues,
so I didn't comment on it in this patch.

g.