[PATCH 2/2] of/fdt: skip unflattening of disabled nodes

[Frank Rowand]

On 10/09/17 11:59, Rob Herring wrote:
> On Sun, Oct 8, 2017 at 5:57 PM, Frank Rowand <frowand.list at gmail.com> wrote:
>> On 10/03/17 09:18, Rob Herring wrote:
>>> For static DT usecases, we don't need the disabled nodes and can skip
>>> unflattening. This saves a significant amount of RAM in memory constrained
>>> cases. In one example on STM32F469, the RAM usage goes from 118K to 26K.
>>>
>>> There are a few cases in the kernel that modify the status property
>>> dynamically. These all are changes from enabled to disabled, depend on
>>> OF_DYNAMIC or are not FDT based (PDT based).
>>>
>>> Cc: Nicolas Pitre <nicolas.pitre at linaro.org>
>>> Cc: Frank Rowand <frowand.list at gmail.com>
>>> Signed-off-by: Rob Herring <robh at kernel.org>
>>> ---
>>> For more background, see this presentation from Nico:
>>>
>>> https://connect.linaro.org/resource/sfo17/sfo17-100/
>>>
>>>  drivers/of/fdt.c | 4 ++++
>>>  1 file changed, 4 insertions(+)
>>>
>>> diff --git a/drivers/of/fdt.c b/drivers/of/fdt.c
>>> index f8c39705418b..efe91c6856a0 100644
>>> --- a/drivers/of/fdt.c
>>> +++ b/drivers/of/fdt.c
>>> @@ -396,6 +396,10 @@ static int unflatten_dt_nodes(const void *blob,
>>>               if (WARN_ON_ONCE(depth >= FDT_MAX_DEPTH))
>>>                       continue;
>>>
>>> +             if (!IS_ENABLED(CONFIG_OF_DYNAMIC) &&
>>> +                 !of_fdt_device_is_available(blob, offset))
>>> +                     continue;
>>> +
>>>               if (!populate_node(blob, offset, &mem, nps[depth],
>>>                                  &nps[depth+1], dryrun))
>>>                       return mem - base;
>>>
>>
>> Hi Rob,
>>
>> I strongly support the idea of this patch, but there may be an
>> issue we have to resolve.  I'm pretty sure we had talked about
>> the issue a long time ago, and it has been sitting on my todo
>> list.
>>
>> We have two sets of node traversal macros and functions.  One
>> set honors the status property, and the other ignores it.  If
>> I recall our previous discussion properly, we want the normal
>> usage to honor the status property and only a tiny (or maybe
>> non-existent) set of locations to be allowed to ignore the
>> status property.
> 
> Ignoring status is a bug for a static DT. There could be places that
> expect the node to be present, but disabled. Those may be bugs too.
> 
>> A rough sense of how often the status property is honored or
>> not is:
>>
>>    $ git grep for_each_child_of_node | wc -l
>>    293
>>    $ git grep of_get_next_child | wc -l
>>    103
>>
>>    $ git grep for_each_available_child_of_node | wc -l
>>    106
>>    $ git grep of_get_next_available_child | wc -l
>>    20
>>
>> Many of the cases where the status flag is ignored will not
>> actually encounter a node that is not available, so many of
>> the cases where status is not checked could currently be
>> checking status.
> 
> For many nodes, status simply makes no sense or at least is undefined.
> 
>> And just for completeness, there are a number of standalone
>> checks for whether a node is available:
>>
>>    $ git grep of_device_is_available | wc -l
>>    128
> 
> I'm surprised it's that many. It's a low-level detail that the core
> should handle. We'd also need to make things like of_find_node_by_name
> honor status.
> 
>> It will be a pain to manually check all of the sites that
>> ignore the status property, but that task should be done.
>>
>> In the mean time, maybe we could flush out the few cases
>> that currently depend on ignoring the status property by
>>
>>    - making for_each_child_of_node() and of_get_next_child()
>>      actually check for valid status
>>
>>    - provide a temporary (one or two kernel release)
>>      CONFIG option to allow the old behavior for
>>      for_each_child_of_node() and of_get_next_child()
>>      just in case we miss any locations that need to
>>      be fixed
>>
>>    - fix up the few places in core device tree code that
>>      actually need to ignore status (if such places exist)
>>
>> In the end, the *_available_*() interfaces should be
>> removed, because the normal behavior of node traversal
>> should be to only traverse nodes that are available.
> 
> I'm not sure this is really something we want or need to fix.
> 
> I could just make this depend on OF_KOBJ instead. Then practically no
> one would see any change as almost everyone enables sysfs (and in turn
> /proc/device-tree).

Yes, depending on OF_KOBJ should reduce the risk significantly.  Good
enough for me (you can take my reviewed-by for that).

I'm still leaving a clean up of checking or ignoring status on my
long term todo list.

-Frank

> 
> Rob
>