[PATCH 1/6 v9] ARM: Add basic architecture support for VIA/WonderMedia 85xx SoC's

[Alexey Charkov]

2010/12/21 Ryan Mallon <ryan at bluewatersys.com>:
> On 12/21/2010 12:00 PM, Alexey Charkov wrote:
>> 2010/12/21 Ryan Mallon <ryan at bluewatersys.com>:
>>> On 12/21/2010 10:48 AM, Alexey Charkov wrote:
>>>> 2010/12/20 Ryan Mallon <ryan at bluewatersys.com>:
>>>>> On 12/21/2010 08:54 AM, Alexey Charkov wrote:
>>>>>> This adds support for the family of Systems-on-Chip produced initially
>>>>>> by VIA and now its subsidiary WonderMedia that have recently become
>>>>>> widespread in lower-end Chinese ARM-based tablets and netbooks.
>>>>>>
>>>>>> Support is included for both VT8500 and WM8505, selectable by a
>>>>>> configuration switch at kernel build time.
>>>>>>
>>>>>> Included are basic machine initialization files, register and
>>>>>> interrupt definitions, support for the on-chip interrupt controller,
>>>>>> high-precision OS timer, GPIO lines, necessary macros for early debug,
>>>>>> pulse-width-modulated outputs control, as well as platform device
>>>>>> configurations for the specific drivers implemented elsewhere.
>>>>>>
>>>>>> Signed-off-by: Alexey Charkov <alchark at gmail.com>
>>>>>
>>>>> Hi Alexey,
>>>>>
>>>>> Quick review below.
>>>
>>>> <snip>
>>>>>> +void __init wmt_set_resources(void)
>>>>>> +{
>>>>>> +     resources_lcdc[0].start = wmt_current_regs->lcdc;
>>>>>> +     resources_lcdc[0].end = wmt_current_regs->lcdc + SZ_1K - 1;
>>>>>> +     resources_lcdc[1].start = wmt_current_irqs->lcdc;
>>>>>> +     resources_lcdc[1].end = wmt_current_irqs->lcdc;
>>>>>
>>>>> Ah, this makes more sense. But why have all the indirection? The
>>>>> wmt_regmaps table could just be replaced with #defines and then have
>>>>> separate device files for the VT8500 and the WM8505. This would also
>>>>> make clearer which variants have which peripherals.
>>>>>
>>>>
>>>> This was the way I implemented it originally. However, Arnd made quite
>>>> a valid suggestion to allow runtime selection of the chip variant,
>>>> thus registers and interrupts need to be held in an indexed data type
>>>> instead of just compile-time macros. In addition, there is now some
>>>> overall movement towards unification of binary kernel images for
>>>> different ARM variants (as far as I can see), so this would be
>>>> required in any case.
>>>>
>>>> Furthermore, as with many unbranded Chinese products, it's somewhat
>>>> difficult to reliably determine the exact chip version used in your
>>>> netbook without disassembling it. Reading a hardware register for
>>>> identification is easier :)
>>>
>>> Okay, that makes sense. I still think there must be a better way than
>>> having a massive indirect table with all the values. Why not detect the
>>> variant in the core code and then have something like:
>>>
>>> int init_devices(void)
>>> {
>>>        int board_type = detect_board_type();
>>>
>>>        switch (board_type) {
>>>        case BOARD_TYPE_VT8500:
>>>                return vt8500_init_devices();
>>>
>>>        case BOARD_TYPE_WM8505:
>>>                return wm8500_init_devices();
>>>        }
>>>
>>>        pr_err("Unknown board type\n");
>>>        BUG(); /* panic()? */
>>>        while (1)
>>>                ;
>>> }
>>>
>>> Then you can have the peripheral setup for each of the variants in their
>>> own files and use #defines. It may get tricky in a couple of places if
>>> you need to be able to access some value directly which is different on
>>> each of the variants, but that can be handled on a case by case basis.
>>> The majority of the numbers will be passed into drivers via the resource
>>> structs.
>>>
>>
>> This is more or less what I'm doing right now - except for the
>> separation between different files. I tried to avoid duplication of
>> similar things here. Is the indirect table really a big issue? I'm a
>> bit reluctant to copy about the whole devices.c for each chip variant,
>> which would be otherwise required. Further, it would add more
>> complexity to the timer, irq, gpio, i8042 and probably some other
>> places.
>
> Yeah, agreed about the duplication. My approach would require the common
> peripherals to be defined for each variant. I'm not entirely against the
> indirect table (and am even starting to think it may be the best overall
> solution), it's just that it can be a bit difficult to follow because to
> add a device you need to do the following:
>
>  - Add a partially empty resource/platform_device struct
>  - Add resource entries to the resource table definition
>  - Add resource values to the resource table
>  - Add assignment of resource values to device init code
>

That's actually only one step more than what machines with static
resource definitions require (the last one). Flexibility does come at
a cost, so there should be a mathematical limit to optimization of
this thing :)

> The indirection also makes it harder to quickly determine the IRQ number
> of memory address of a peripheral.
>
> The current solution using the indirect table is okay, but it would be
> nice to find a solution which reduces some of this effort.
>
>>>>>> +}
>>>>>> +
>>>>>> +int pwm_config(struct pwm_device *pwm, int duty_ns, int period_ns)
>>>>>> +{
>>>>>> +     unsigned long long c;
>>>>>> +     unsigned long period_cycles, prescale, pv, dc;
>>>>>> +
>>>>>> +     if (pwm == NULL || period_ns == 0 || duty_ns > period_ns)
>>>>>> +             return -EINVAL;
>>>>>> +
>>>>>> +     c = 25000000/2; /* wild guess --- need to implement clocks */
>>>>>> +     c = c * period_ns;
>>>>>> +     do_div(c, 1000000000);
>>>>>> +     period_cycles = c;
>>>>>
>>>>> This looks like it could be reworked to remove the do_div call.
>>>>>
>>>>
>>>> I just followed PXA implementation in this regard. Are there any
>>>> specific suggestions? Note that c should not be a complie-time
>>>> constant eventually, as this is the guessed PWM base frequency (should
>>>> be read from the hardware, but the code for clocks is not yet in).
>>>
>>> I didn't have a particular solution in mind, but often by changing the
>>> units used and rearranging the math a bit you can often avoid having to
>>> do horrible multiplies and divides.
>>>
>>> For now at least you could avoid the do_div by assigning period_cycles
>>> directly.
>>>
>>
>> It depends on period_ns, which is passed in as an argument from
>> whatever uses PWM, so I'm not sure it can be assigned directly.
>
> Ah. How big a number is period_ns? Can you do something like this instead?
>
>        period_cycles = ((250 / 2) * period_ns) / 10000;
>
> and still safely avoid overflows?
>

The only current in-kernel user of PWM is the backlight, and that
currently uses period_ns = 250000. At this value it does not overflow.
However, in a general case the base frequency will also be returned as
a large number (like 12500000) as per CLK infrastructure conventions
(once that part is implemented). Further, I can't see any built-in
reasons for period_ns to be bounded by anything below sizeof(int). The
hardware will work with up to 4096*1024*1000000000/base_frequency (see
the code for constraints), so it can in principle overflow with 32 bit
arithmetics.

Best regards,
Alexey