[PATCH 2/5] ARM: add SMP support for Broadcom mobile SoCs

[Stephen Boyd]

On 05/05/14 15:02, Alex Elder wrote:
> On 04/04/2014 12:55 PM, Stephen Boyd wrote:
>> On 04/03/14 19:18, Alex Elder wrote:
>>> +
>>> +/*
>>> + * Secondary startup method setup routine to extract the location of
>>> + * the secondary boot register from a "cpu" or "cpus" device tree
>>> + * node.  Only the first seen secondary boot register value is used;
>>> + * any others are ignored.  The secondary boot register value must be
>>> + * non-zero.
>>> + *
>>> + * Returns 0 if successful or an error code otherwise.
>>> + */
>>> +static int __init of_enable_method_setup(struct device_node *node)
>>> +{
>>> +	int ret;
>>> +
>>> +	/* Ignore all but the first one specified */
>>> +	if (secondary_boot)
>>> +		return 0;
>>> +
>>> +	ret = of_property_read_u32(node, OF_SECONDARY_BOOT, &secondary_boot);
>>> +	if (ret)
>>> +		pr_err("%s: missing/invalid " OF_SECONDARY_BOOT " property\n",
>>> +			node->name);
>>> +
>>> +	return ret;
>>> +}
>> I don't understand why we need this. Why can't we get the secondary boot
>> address from the /cpus node in the smp_prepare_cpus op. It isn't that
>> hard to get access to the cpus node there via of_find_node_by_path().
>> Then we don't need patch 1 at all. If it turns out to be common stuff,
>> we can always have the common function live in arm common code or maybe
>> even be a devicetree API.
> I already responded to this, but never got any response.  I
> was preparing to re-send this series and wanted to try to
> pull the added feature (patch 1) out and not be dependent on
> it.  But I think it's a bit ugly so I'm hoping to get a
> blessing to proceed with what I originally proposed.  For
> reference, here's the thread:
>     https://lkml.org/lkml/2014/4/3/421
>
> What I'm trying to do is get the value of a "secondary-boot-reg"
> property from a node known to have an "enable-method" property
> that matches the method name supplied in CPU_METHOD_OF_DECLARE().

I don't recall seeing any documentation about the secondary-boot-reg
property. Please make sure it is documented in the next patch series.

>
> Using the callback function as I originally proposed, this is
> very easy.  When arm_dt_init_cpu_maps() parses the "cpus" portion
> of the device tree it calls set_smp_ops_by_method() for a
> matching "cpu" or "cpus" node, and that function supplies
> the node to the callback function. The callback can extract
> additional property values if needed.
>
> If I hold off until smp_prepare_cpus() is called, I have to
> re-parse the device tree to find the "cpus" node (this is
> in itself trivial).  I then need to re-parse that node to
> verify the matching "enable-method" property is found before
> looking for the parameter information I need for that enable
> method.  I would really prefer not to re-do this parsing
> step.  It's imprecise and a little inefficient, and it
> duplicates (but not exactly) logic that's already performed
> by arm_dt_init_cpu_maps().

Do you have any devices where the enable-method and secondary-boot-reg
isn't the same across all CPUs? A lot of the complexity comes from
broken DTs that don't specify a secondary-boot-reg along with the
enable-method. From the description and the code it seems that we should
just always put these two properties in the cpus node to make things
simple and precise. I agree it's a minor duplication to read the DT
again to get the /cpus node and read a property out of it, but I doubt
you could even measure the difference.

>
> One more point of clarification.  This "secondary-boot-reg"
> value is *not* the secondary boot address--that is, it's
> not the address secondary cores jump to when they are
> activated.  Instead, this is the address of a register
> that's used to request the ROM code release a core from
> its ROM-implemented holding pen.  For this machine,
> control jumps at that point to secondary_startup(),
> defined in arch/arm/kernel/head.S.

Yes it wouldn't be possible to specify the entry address in DT (depends
on compile time factors). How is this different from cpu-release-addr
though? It looks like it describes something similar to what ePAPR
describes and how arm64 uses it (although those two slightly differ).
Assuming it's paired with a different enable-method than spin-table I
don't see a problem reusing the same name.

>
> So...
>
> Stephen, I'd like to hear from you whether my explanation
> is adequate, and whether you think my addition and use of
> CPU_METHOD_OF_DECLARE_SETUP() is reasonable.  (If you have
> a suggestion for a better name, I'm open.)
>
> If you still don't like it, I'll follow up with a
> new version of the patches, this time parsing the
> device tree in the smp_prepare_cpus() method as
> you suggested.  I don't want this to hold up getting
> this SMP support into the kernel.
>

This was my train of thought. It annoys me that we have smp ops at two
levels. We already have SMP ops that deal with SMP things and now we
have another level that is associated with the enable-method. If we
really need this callback why couldn't we just collapse it with the
smp_ops we already have? When we do that we realize that
smp_prepare_cpus() is a fairly similar place to do this sort of thing,
but changing that callback to take the node is annoying to propagate to
all users. But do we really need that node? It seems simple enough to
just get the cpus node and then read out the property we care about and
we can do all that in smp_prepare_cpus() one time (and keep the mapping
around forever) without requiring any core ARM changes.

Do you have any more complicated use cases for this setup hook? RIght
now I'm not convinced that we need to add all this framework for
something that looks like it could be done in a couple lines in the
prepare_cpus() step.

	cpus = of_find_node_by_path("/cpus")
	ret = of_property_read_u32(cpus, "secondary-boot-reg", &addr);


-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation