Message-Id: <20121115161020.3205E3E194B@localhost>
Date:	Thu, 15 Nov 2012 16:10:20 +0000
From:	Grant Likely <grant.likely@...retlab.ca>
To:	Stepan Moskovchenko <stepanm@...eaurora.org>,
	Stepan Moskovchenko <stepanm@...eaurora.org>
Cc:	Rob Herring <robherring2@...il.com>,
	devicetree-discuss@...ts.ozlabs.org,
	David Brown <davidb@...eaurora.org>,
	Daniel Walker <dwalker@...o99.com>,
	Bryan Huntsman <bryanh@...eaurora.org>,
	Russell King <linux@....linux.org.uk>,
	linux-arm-msm@...r.kernel.org,
	linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC] dt/platform: Use cell-index for device naming if available

On Mon, 12 Nov 2012 18:48:43 -0800, Stepan Moskovchenko <stepanm@...eaurora.org> wrote:
> On 11/11/2012 5:45 PM, Stepan Moskovchenko wrote:
> >
> >> On Sun, Nov 11, 2012 at 2:32 AM, Rob Herring <robherring2@...il.com>
> >> wrote:
> >>> On 11/09/2012 06:48 PM, Stepan Moskovchenko wrote:
> >>>> Use the cell-index property to construct names for platform
> >>>> devices, falling back on the existing scheme of using the
> >>>> device register address if cell-index is not specified.
> >>>>
> >>>> The cell-index property is a more useful device identifier,
> >>>> especially in systems containing several numbered instances
> >>>> of a particular hardware block, since it more easily
> >>>> illustrates how devices relate to each other.
> >>>>
> >>>> Additionally, userspace software may rely on the classic
> >>>> <name>.<id> naming scheme to access device attributes in
> >>>> sysfs, without having to know the physical addresses of
> >>>> that device on every platform the userspace software may
> >>>> support. Using cell-index for device naming allows the
> >>>> device addresses to be hidden from userspace and to be
> >>>> exposed by logical device number without having to rely on
> >>>> auxdata to perform name overrides. This allows userspace to
> >>>> make assumptions about which sysfs nodes map to which
> >>>> logical instance of a specific hardware block.
> >>>>
> >>>> Signed-off-by: Stepan Moskovchenko <stepanm@...eaurora.org>
> >>>> ---
> >>>> I had also considered using something like the linux,label property to
> >>>> allow
> >>>> custom names for platform devices without resorting to auxdata, but the
> >>>> cell-index approach seems more in line with what cell-index was
> >>>> intended for
> >>>> and with what the pre-DT platform device naming scheme used to be.
> >>>> Please let
> >>>> me know if you think there is a better way to accomplish this.
> >>>>
> >>>> This is just being sent out as an RFC for now. If there are no
> >>>> objections, I
> >>>> will send this out as an official patch, along with (or combined with)
> >>>> a patch
> >>>> to fix up the device names in things like clock tables of any affected
> >>>> platforms.
> >>>
> >>> cell-index is basically deprecated. This has been discussed multiple
> >>> times in the past. You can use auxdata if you really need to have the
> >>> old name.
> >>
> >> Actually, I think it would be fine to use an /aliases entry to set the
> >> device name. That's the place to put global namespace information.
> >>
> >> g.
> >>
> >
> > Ah, thank you. I would prefer to stay away from auxdata, since it involves
> > placing more platform-specific data into the kernel, and it is my
> > understanding that auxdata is intended as a temporary measure. The
> > /aliases approach looks interesting, and I'll see what I can do with it -
> > hopefully I can have an RFC / patch soon. It looks like we would want an
> > "inverse" alias lookup- that is, we would need to know which alias
> > corresponds to a given node. Is it possible for a node to have multiple
> > aliases?

yes

> > If so, which shall we use to create the device name? Anyway, I
> > will further look into how these aliases work.

Well, why exactly do you want to control the names of devices? Is it so
that devices match up with what they are, or is it to make things match
up with things like clocks and regulators. If it is the latter, then no,
don't do this. Use auxdata. When the kernel requires a specific name for
a device it is very much a kernel *internal* detail. It does not make
sense to encode that into the device tree when it isn't something part
of the binding.


> >
> > Steve
> 
> Hi Grant,
> 
> Looking through the alias code, I see that the stem and the alias ID are 
> stored and parsed separately. For the current way of using aliases, this 
> makes sense. However, can you please clarify what you meant by using an 
> /aliases entry to set the device name?
> 
> The first and most straightforward approach would be to use the entire 
> alias name as the device name, making no distinction between the alias 
> stem and ID. However, since it is possible to have multiple aliases to 
> the same device, which of the aliases shall we use to construct the 
> device name? Additionally, this may cause possible problems for legacy 
> software that expects names in the format of <name>.<ID>, since '.' is 
> not a valid character for alias names as defined by the DT spec, 
> although strictly speaking this approach would successfully solve the 
> problem of giving devices predictable and controllable names.
> 
> Another way an /aliases entry could be used to set the device name is to 
> have a <name>.<ID> naming scheme, where the name comes from node->name 
> (as is done in of_device_make_bus_id) and the ID gets queried using 
> of_alias_get_id(). We would need to create a new alias stem for this 
> purpose, and suppose that something like "platform" would work. The 
> name-setting code would then roughly look as follows:
> 
> +	alias_id = of_alias_get_id(node, "platform");
> +	if (alias_id != -ENODEV) {
> +		dev_set_name(dev, "%s.%d", node->name, alias_id);
> +		return;
> +	}
> 
> The downside to this approach is that it imposes the restriction that 
> device ID numbers now have to be unique throughout the system, whereas 
> before only the <name>.<ID> combinations had to be unique. This is the 
> result of only the ID number being present in the alias table, with each 
> such ID number having the "platform" stem, and the restriction that node 
> properties shall have unique names.

Again, this all looks like trying to manipulate names to keep the kernel
happy. Since the kernel has the restrictions on naming, that is where
the fixups should be made. Either by devres or by changing the expected
name in the clk/regulator tables.

> A third possible solution is to use an alias stem prefix for defining 
> the device name. That is, the alias to set the device name would have 
> some prefix (such as "platform-" for example) and the aliases would look 
> something like platform-<name><ID>. The code to assign device names 
> would find the matching alias containing the "platform-" prefix, strip 
> the prefix, and use the resulting name and ID to construct the device 
> name. This approach would make it more obvious as to which of several 
> aliases is used to set the device name, but it imposes additional 
> structure on the stem names and causes any aliases starting with 
> "platform-" to become magical, which bothers me slightly.

And specific to the current Linux implementation details.

> Do any of these describe what you intended when you suggested using the 
> /aliases node to set device names, or is there another approach that I 
> have missed? Can you please elaborate further?

Really, all I was thinking about was allowing the device that has an
alias "eth0" to be given a name with 'eth0' in it somewhere. Since names
like that are a global namespace, /aliases is the place to get them
because there is no chance of colision with that approach.

g.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists