[RFC PATCH 0/9] dt: dependencies (for deterministic driver initialization order based on the DT)

Stephen Warren swarren at wwwdotorg.org
Wed Aug 27 09:22:06 PDT 2014

On 08/27/2014 08:44 AM, Catalin Marinas wrote:
> On Wed, Aug 27, 2014 at 11:34:32AM +0100, Grant Likely wrote:
>> On Tue, 26 Aug 2014 11:11:07 +0100, Mark Rutland <mark.rutland at arm.com> wrote:
>>> On Tue, Aug 26, 2014 at 10:42:04AM +0100, Alexander Holler wrote:
>>>> Am 26.08.2014 10:49, schrieb Thierry Reding:
>>>>> On Tue, Aug 26, 2014 at 09:42:08AM +0100, Grant Likely wrote:
>>>>>> On Mon, 25 Aug 2014 15:37:16 +0200, Thierry Reding <thierry.reding at gmail.com> wrote:
>>>>> [...]
>>>>>>> There are somewhat standardized bindings for the above and especially
>>>>>>> for bindings of the type that clocks implement this is trivial. We can
>>>>>>> simply iterate over each (phandle, specifier) tuple and check that the
>>>>>>> corresponding clock provider can be resolved (which typically means that
>>>>>>> it's been registered with the common clock framework).
>>>>>>> For regulators (and regulator-like bindings) the problem is somewhat
>>>>>>> more difficult because they property names are not standardized. One way
>>>>>>> to solve this would be to look for property names with a -supply suffix,
>>>>>>> but that could obviously lead to false positives. One alternative that I
>>>>>>> think could eliminate this would be to explicitly list dependencies in
>>>>>>> drivers. This would allow core code to step through such a list and
>>>>>>> resolve the (phandle, specifier) tuples.
>>>>>> False positives and negatives may not actually be a problem. It is
>>>>>> suboptimal, certainly, but it shouldn't outright break the kernel.
>>>>> There could be cases where some random integer in a cell could be
>>>>> interpreted as a phandle and resolve to a struct device_node. I suppose
>>>>> it might be unlikely, but not impossible, that the device_node could
>>>>> even match a device in the correct subsystem and you'd get a wrong
>>>>> dependency. Granted, a wrong dependency may not be catastrophic in that
>>>>> it won't lead to a crash, but it could lead to various kinds of
>>>>> weirdness and hard to diagnose problems.
>>>> You need either the type information in the DTB (that's why I've add
>>>> those "dependencies" to identify phandles), or you need to know every
>>>> binding (at "dependency-resolve-time" to identify phandles.
>>> While having type information in the DTB would be fantastic, it's not
>>> something we can expect from the systems already in the wild, and I
>>> worry how it would interact with bootloaders that modify the DTB (I
>>> don't know if any modify properties with phandles).
>> Anything we do here is firmly in the realm of optimization and
>> improvement. Adding data to the tree is fine as long as we don't make
>> the kernel depend on it. Older platforms will continue to work without
>> the optimization.
> It's not just optimisation but an important feature for new arm64 SoCs.
> Given some Tegra discussions recently, in many cases the machine_desc
> use on arm is primarily to initialise devices in the right order. If we
> can solve this in a more deterministic way (other than deferred
> probing), we avoid the need for a dedicated SoC platform driver (or
> machine_desc) or workarounds like different initcall levels and explicit
> DT parsing.

A lot of the ordering is SW driver dependencies. I'm not sure how much 
of that can accurately be claimed as HW dependencies. As such, I'm not 
sure that putting dependencies into DT would be a good idea; it doesn't 
feel like HW data, and might well change if we restructure SW. It'd need 
some detailed research though.

More information about the linux-arm-kernel mailing list