Re: [PATCH V2 6/6] pinctrl: tegra: Add complete device tree support

[Stephen Warren <swarren-3lzwWm7+Weoh9ZMKESR00Q@public.gmane.org>]

On 03/21/2012 10:07 PM, Dong Aisheng wrote:
> On Thu, Mar 22, 2012 at 12:07:27AM +0800, Stephen Warren wrote:
>> On 03/21/2012 03:35 AM, Dong Aisheng wrote:
>>> On Wed, Mar 21, 2012 at 01:44:39AM +0800, Stephen Warren wrote:
>>>> Implement pinctrl_ops dt_node_to_map() and dt_free_map(). These allow
>>>> complete specification of the desired pinmux configuration using device
>>>> tree.
...
>>>> +	for_each_child_of_node(np_config, np) {
>>>> +		ret = of_property_read_string(np, "nvidia,function", &function);
>>>> +		if (ret < 0)
>>>> +			function = NULL;
>>>> +
>>>> +		for (i = 0; i < ARRAY_SIZE(cfg_params); i++) {
>>>> +			ret = of_property_read_u32(np, cfg_params[i].property,
>>>> +						   &val);
>>>> +			if (!ret) {
>>>> +				config = TEGRA_PINCONF_PACK(
>>>> +						cfg_params[i].param, val);
>>>> +				ret = add_config(&configs, &num_configs,
>>>> +						 config);
>>>> +				if (ret < 0)
>>>> +					goto error;
>>>> +			}
>>>> +		}
>>>> +
>>>> +		of_property_for_each_string(np, "nvidia,pins", prop, group) {
>>>
>>> If we calculate out the strings count and allocate corresponding size array, we may not
>>> need to keep krealloc the maps and configs array size for each entry.
>>> And this may be a little higher efficient.
>>
>> That's true. However, it'd require the code to loop once to determine
>> how many properties are present and how many entries there are in the
>> pin list. Then, loop again to actually construct the mapping table
>> array. This is all added complexity that doesn't affect correctness. I'd
>> rather get the simple code going first, and then refine it later if
>> there turns out to be a performance issue.
>>
> Can we use of_property_count_strings?

It'd be possible to avoid some of the reallocs this way. We could
realloc once per node rather than once per (node, pin). It does make the
code a bit more complex though, since you have to reserve space up-front
during the one realloc and so have to store separate num_maps and
num_maps_allocated. I'll see how bad it gets, and maybe include it in v3
if it isn't horrible.