[PATCH 13/13] driver-core: probe dependencies before probing

[Tomeu Vizoso]

On 17 June 2015 at 20:13, Mark Brown <broonie at kernel.org> wrote:
> On Wed, Jun 17, 2015 at 03:42:23PM +0200, Tomeu Vizoso wrote:
>> Before actually probing a device, find out what dependencies it has and
>> do our best to ensure that they are available at this point.
>
>> This is accomplished by finding out what platform devices need to be
>> probed so the dependencies are available.
>
> ...and then trying to probe them first.
>
>> If any dependencies are still unavailable after that (most probably a
>> missing driver or an error in the HW description from the firmware), we
>> print a nice error message so that people don't have to add a zillion of
>> printks to find out why a device asked for its probe to be deferred.
>
> So, I think I like this approach though I've not done a full pass
> through and I'm not sure how expensive it gets (there's definitely room
> for optimisation as the patch notes).

Have measured it and the overhead doesn't seem to be much, in the
version that I'm close to send.

> I'm not 100% sure I see what
> prints this error message you're referring to (I'm just seeing debug
> prints).

Right, so far I have left them as debug messages because I have so far
tested the series on just one platform and I'm not sure if there
wouldn't be lots of noise in others.

>> +static struct fwnode_handle *get_enclosing_platform_dev(
>> +                                             struct fwnode_handle *fwnode)
>
> Only platform devices?

Yes, this code assumes that devices on other buses will be registered
and probed when their enclosing platform devices are.

>> +static void check_dependencies_per_class(struct class *class, void *data)
>> +{
>> +     struct fwnode_handle *fwnode = data;
>> +     struct list_head *deps;
>> +     struct fwnode_dependency *dep, *tmp;
>> +
>> +     if (!class->get_dependencies)
>> +             return;
>> +
>> +     deps = class->get_dependencies(fwnode);
>> +     if (!deps)
>> +             return;
>> +
>> +     list_for_each_entry_safe(dep, tmp, deps, dependency) {
>> +             if (!check_dependency(dep->fwnode))
>> +                     pr_debug("Dependency '%s' not available\n",
>> +                              fwnode_get_name(dep->fwnode));
>> +
>> +             list_del(&dep->dependency);
>> +             kfree(dep);
>> +     }
>> +
>> +     kfree(deps);
>
> OK, so the caller is responsible for freeing everything and the class
> must allocate - this definitely suggests that
>
> I'm not sure there's any benefit in having deps be dynamically allocated
> here, just put it on the stack and iterate through the list - the
> iteration is going to be cheap if we get nothing back (probably the
> common case) and probably cheaper than the alloc/free.

Have done this and I like it more.

> One thing here is that I was under the impression classes were supposed
> to be going away...

Actually, while looking at more firmware node properties to parse for
dependencies, I found a rather common case in which the bindings are
implemented by individual drivers and not subsystems. Some examples
are nvidia,dpaux, nvidia,panel and nvidia,ddc-i2c-bus.

So in my next version I have dropped class callbacks and have gone
with a way for classes, drivers, whatever to just register a function
to extract dependencies.

Thanks,

Tomeu