[PATCHv3 4/9] ARM: OMAP2+: AM33XX: Reserve memory to comply with EMIF spec

[khilman@linaro.org (Kevin Hilman)]

Santosh Shilimkar <santosh.shilimkar@ti.com> writes:

> On Thursday 08 August 2013 04:05 PM, Kevin Hilman wrote:
>> Santosh Shilimkar <santosh.shilimkar@ti.com> writes:
>> 
>>> On Thursday 08 August 2013 02:16 PM, Kevin Hilman wrote:
>>>> Dave Gerlach <d-gerlach@ti.com> writes:
>>>>
>>>>> From: Vaibhav Bedia <vaibhav.bedia@ti.com>
>>>>>
>>>>> SDRAM controller on AM33XX requires that a modification of certain
>>>>> bit-fields in PWR_MGMT_CTRL register (ref. section 7.3.5.13 in
>>>>> AM335x-Rev H) is followed by a dummy read access to SDRAM. This
>>>>> scenario arises when entering a low power state like DeepSleep.
>>>>> To ensure that the read is not from a cached region we reserve
>>>>> some memory during bootup using the arm_memblock_steal() API.
>>>>
>>>> Hmm, sounds to me an awful lot like the existing omap_bus_sync() ?
>>>>
>>> All the credit of that awful omap_bus_sync() goes to me since 
>>> I introduced it. And I keep beating the hardware guys
>>> who have not left a choice but to introduce the ugly work
>>> around in software. ;-)
>> 
>> Agreed, but what's even more awful than the current version is
>> duplicating it in a slightly different way using yet another whole page
>> mapping for a single read/write location.
>> 
> The real issue is limitation of the kernel memory steal(memblock) API which
> won't let you still less than 1 MB. It would have been ok for page allocation
> because that is any way what you will get minimum on standard non-cached
> allocations.

All the more reason that the omap_bus_sync() should be refactored
slightly in a way that would be reusable for AM33xx.

Kevin