lkml.org 
[lkml]   [2022]   [May]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH 2/8] iommu: mtk_iommu: Lookup phandle to retrieve syscon to infracfg
From
On 2022-05-18 09:29, AngeloGioacchino Del Regno wrote:
> Il 17/05/22 16:12, Robin Murphy ha scritto:
>> On 2022-05-17 14:21, AngeloGioacchino Del Regno wrote:
>>> This driver will get support for more SoCs and the list of infracfg
>>> compatibles is expected to grow: in order to prevent getting this
>>> situation out of control and see a long list of compatible strings,
>>> add support to retrieve a handle to infracfg's regmap through a
>>> new "mediatek,infracfg" phandle.
>>>
>>> In order to keep retrocompatibility with older devicetrees, the old
>>> way is kept in place, but also a dev_warn() was added to advertise
>>> this change in hope that the user will see it and eventually update
>>> the devicetree if this is possible.
>>>
>>> Signed-off-by: AngeloGioacchino Del Regno
>>> <angelogioacchino.delregno@collabora.com>
>>> ---
>>>   drivers/iommu/mtk_iommu.c | 40 +++++++++++++++++++++++++--------------
>>>   1 file changed, 26 insertions(+), 14 deletions(-)
>>>
>>> diff --git a/drivers/iommu/mtk_iommu.c b/drivers/iommu/mtk_iommu.c
>>> index 71b2ace74cd6..cfaaa98d2b50 100644
>>> --- a/drivers/iommu/mtk_iommu.c
>>> +++ b/drivers/iommu/mtk_iommu.c
>>> @@ -1134,22 +1134,34 @@ static int mtk_iommu_probe(struct
>>> platform_device *pdev)
>>>       data->protect_base = ALIGN(virt_to_phys(protect),
>>> MTK_PROTECT_PA_ALIGN);
>>>       if (MTK_IOMMU_HAS_FLAG(data->plat_data, HAS_4GB_MODE)) {
>>> -        switch (data->plat_data->m4u_plat) {
>>> -        case M4U_MT2712:
>>> -            p = "mediatek,mt2712-infracfg";
>>> -            break;
>>> -        case M4U_MT8173:
>>> -            p = "mediatek,mt8173-infracfg";
>>> -            break;
>>> -        default:
>>> -            p = NULL;
>>> +        infracfg = syscon_regmap_lookup_by_phandle(dev->of_node,
>>> "mediatek,infracfg");
>>> +        if (IS_ERR(infracfg)) {
>>> +            dev_warn(dev, "Cannot find phandle to mediatek,infracfg:"
>>> +                      " Please update your devicetree.\n");
>>
>> Is this really a dev_warn-level problem? There's no functional impact,
>> given that we can't stop supporting the original binding any time
>> soon, if ever, so I suspect this is more likely to just annoy users
>> and CI systems than effect any significant change.
>>
>
> The upstream devicetrees were updated to use the new handle and this is
> a way to
> warn about having outdated DTs... besides, I believe that CIs will
> always get the
> devicetree from the same tree that the kernel was compiled from (hence
> no message
> will be thrown).
>
> In any case, if you think that a dev_info would be more appropriate, I
> can change
> that no problem.

If there's some functional impact from using the old binding vs. the new
one then it's reasonable to inform the user of that (as we do in
arm-smmu, for example).

However if you change an established binding for non-functional reasons,
then you get to support both bindings, and it's not the end user's
problem at all. There seems to be zero reason to update an existing DTB
for this difference alone, so TBH I don't think it deserves a message at
all.

>>> +            /*
>>> +             * Legacy devicetrees will not specify a phandle to
>>> +             * mediatek,infracfg: in that case, we use the older
>>> +             * way to retrieve a syscon to infra.
>>> +             *
>>> +             * This is for retrocompatibility purposes only, hence
>>> +             * no more compatibles shall be added to this.
>>> +             */
>>> +            switch (data->plat_data->m4u_plat) {
>>> +            case M4U_MT2712:
>>> +                p = "mediatek,mt2712-infracfg";
>>> +                break;
>>> +            case M4U_MT8173:
>>> +                p = "mediatek,mt8173-infracfg";
>>> +                break;
>>> +            default:
>>> +                p = NULL;
>>> +            }
>>> +
>>> +            infracfg = syscon_regmap_lookup_by_compatible(p);
>>
>> Would it not make sense to punt this over to the same mechanism as for
>> pericfg, such that it simplifies down to something like:
>>
>>      if (IS_ERR(infracfg) && plat_data->infracfg) {
>>          infracfg =
>> syscon_regmap_lookup_by_compatible(plat_data->infracfg);
>>          ...
>>      }
>>
>> ?
>>
>> TBH if we're still going to have a load of per-SoC data in the driver
>> anyway then I don't see that we really gain much by delegating one
>> aspect of it to DT, but meh. I would note that with the phandle
>> approach, you still need some *other* flag in the driver to know
>> whether a phandle is expected to be present or not, whereas a NULL vs.
>> non-NULL string is at least neatly self-describing.
>>
>
> That would be possible but, as Yong also pointed out, we should try to
> reduce the
> per-SoC data in the driver by commonizing as much as possible, because
> this driver
> supports a very long list of SoCs (even though they're not all
> upstreamed yet),
> and the list is going to grow even more with time: this is also why I
> have changed
> the MT8195 pericfg regmap lookup with a phandle like I've done for infra.

That's fair enough, but it's not what the commit message says. The "long
list of compatible strings" complaint could be addressed at face value
by refactoring without changing the DT binding at all.

However, I didn't think I'd have to point out why that argument doesn't
apply to existing SoCs which we have to support with the original
binding too; take another look at the switch statement above and have a
think...

If we have to maintain infracfg compatible data *somewhere* in the
driver, which we do, then it seems more logical to keep it with the rest
of the data rather than scattered through the code, that's the main
point I wanted to make here.

> There would also be another way, which would imply adding a generic
> compatible
> "mediatek,infracfg" to the infra syscon node, but I really don't like
> that for
> more than one reason, one of which is that this poses an issue, for
> which it's
> not guaranteed that the registers are in infracfg and not infracfg_ao (even
> though the offsets are the same), so then we would be back to ground zero.

No, we still have to support the existing binding in the context of the
existing binding. There *are* some reasonable arguments for moving
future SoCs to the phandle binding, they just haven't been presented in
these patches ;)

Thanks,
Robin.

\
 
 \ /
  Last update: 2022-05-18 13:08    [W:0.087 / U:0.404 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site