On Mon, Jun 21, 2010 at 7:16 PM, Russell King - ARM Linux
<linux@xxxxxxxxxxxxxxxx> wrote:
> On Mon, Jun 21, 2010 at 06:39:10PM +0800, Eric Miao wrote:
>> On Mon, Jun 21, 2010 at 5:19 PM, Russell King - ARM Linux
>> <linux@xxxxxxxxxxxxxxxx> wrote:
>> > On Mon, Jun 21, 2010 at 05:05:34PM +0800, Eric Miao wrote:
>> >> On Mon, Jun 21, 2010 at 2:26 PM, Joonyoung Shim <jy0922.shim@xxxxxxxxxxx> wrote:
>> >> > +void __init samsung_keypad_set_platdata(struct samsung_keypad_platdata *pd)
>> >> > +{
>> >> > + struct samsung_keypad_platdata *npd;
>> >> > +
>> >> > + if (!pd) {
>> >> > + printk(KERN_ERR "%s: no platform data\n", __func__);
>> >> > + return;
>> >> > + }
>> >> > +
>> >> > + npd = kmemdup(pd, sizeof(struct samsung_keypad_platdata), GFP_KERNEL);
>> >> > + if (!npd)
>> >> > + printk(KERN_ERR "%s: no memory for platform data\n", __func__);
>> >>
>> >> This part of the code is actually duplicated again and again and again
>> >> for each device, PXA and other legacy platforms are bad references for
>> >> this. In arch/arm/mach-mmp/, it might be a bit cleaner, there are three
>> >> major points:
>> >>
>> >> 1. A minimum 'struct pxa_device_desc' for a simple description of a
>> >> device (more than 90% of the devices can be described that way),
>> >> and avoid using a comparatively heavier weight platform_device,
>> >> which can be generated at run-time
>> >>
>> >> 2. pxa_register_device() to allocate and register the platform_device
>> >> at run-time, along with the platform data
>> >
>> > It's a bad idea to make platform data be run-time discardable like this:
>> >
>> >> > +struct samsung_keypad_platdata {
>> >> > + const struct matrix_keymap_data *keymap_data;
>> >
>> > What you end up with is some platform data structures which must be kept
>> > (those which have pointers to them from the platform data), and others
>> > (the platform data itself) which can be discarded at runtime.
>> >
>> > We know that the __initdata attributations cause lots of problems -
>> > they're frequently wrong. Just see the constant hastle with __devinit
>> > et.al. The same issue happens with __initdata as well.
>> >
>> > So why make things more complicated by allowing some platform data
>> > structures to be discardable and others not to be? Is their small
>> > size (maybe 6 words for this one) really worth the hastle of getting
>> > __initdata attributations wrong (eg, on the keymap data?)
>> >
>>
>> Russell,
>>
>> The benefit I see is when multiple boards are compiled in, those
>> data not used can be automatically discarded.
>
> Yes, but only some of the data can be discarded. Continuing with the
> example in hand, while you can discard the six words which represent
> samsung_keypad_platdata, but the keymap_data can't be because that won't
> be re-allocated, which is probably a much larger data structure.
>
> So, samsung_keypad_platdata can be marked with __initdata, but the keymap
> data must not be - and this is where the confusion about what can be
> marked with __initdata starts - and eventually leads to the keymap data
> also being mistakenly marked __initdata.
Indeed. One ideal solution would be to make all the board code into
a module, yet since some of the code should be executed really
early and need to stay as built-in, looks like we need a way to
separate these two parts.
>
> Is saving six words (or so) worth the effort to track down stuff
> inappropriately marked with __initdata ? I'm sure there's bigger savings
> to be had in other parts of the kernel (such as shrinking the size of
> tcp hash tables, reducing the kernel message ring buffer, etc) if size is
> really a concern.
>
That's true, however, I think the size will become a bit more significant
if more and more board code is compiled into a single kernel.
--
To unsubscribe from this list: send the line "unsubscribe linux-input" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
