Re: [PATCH] ARM: tegra: Use PMC scratch register 40 for tegra_resume() location store

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

On 01/09/2015 03:29 AM, Peter De Schrijver wrote:
> On Fri, Jan 09, 2015 at 10:51:35AM +0100, Thierry Reding wrote:
>> * PGP Signed by an unknown key
>>
>> On Thu, Jan 08, 2015 at 02:37:09PM +0200, Peter De Schrijver wrote:
>>> On Thu, Jan 08, 2015 at 11:57:43AM +0100, Thierry Reding wrote:
>>>>> Old Signed by an unknown key
>>>>
>>>> On Mon, Dec 22, 2014 at 11:00:16AM -0700, Stephen Warren wrote:
>>>>> On 12/22/2014 10:27 AM, Dmitry Osipenko wrote:
>>>>>> 22.12.2014 19:17, Stephen Warren пишет:
>>>>>>> On 12/21/2014 03:52 PM, Dmitry Osipenko wrote:
>>>>>>>> Commit 7232398abc6a ("ARM: tegra: Convert PMC to a driver") changed
>>>>>>>> tegra_resume()
>>>>>>>> location storing from late to early and as result broke suspend on tegra20.
>>>>>>>> PMC scratch register 41 was used by tegra lp1 suspend core code for storing
>>>>>>>> physical memory address of common resume function and in the same time used by
>>>>>>>> tegra20 cpuidle driver for storing cpu1 "resettable" status, so it implied
>>>>>>>> strict order of scratch register use. Fix it by using scratch 40 instead of 41
>>>>>>>> for tegra_resume() location store.
>>>>>>>
>>>>>>> You likely can't simply change the PMC scratch register usage arbitrarily;
>>>>>>> specific registers are designated for specific purposes, and code outside the
>>>>>>> Linux kernel (bootloaders, LP0 resume code, secure monitors, etc.) may depend on
>>>>>>> those specific values being in those registers. Without significant research,
>>>>>>> I'd suggest not changing the PMC scratch register usage.
>>>>>>
>>>>>> Sure, that's why I asked to verify if scratch register 40 is in use in the
>>>>>> comment after commit message.
>>>>>
>>>>> Sorry, I didn't notice that.
>>>>>
>>>>>> I've checked that u-boot doesn't use it (since
>>>>>> upstream kernel doesn't care about any other bootloader), but no idea about
>>>>>> secure monitor. It's definitely safer to avoid changing scratch regs usage, I
>>>>>> thought that proposed solution would be best from the pure code point of view.
>>>>>> So, I'm considering your answer as a rejection of the patch (please, let me know
>>>>>> if I'm wrong) and will prepare another one. Btw, it would be nice to have
>>>>>> scratch registers usage publicly documented somewhere (on "Tegra Public
>>>>>> Application Notes" webpage for example), if it's possible, of course.
>>>>>
>>>>> At this stage in Tegra20 development, I think it'd be best to avoid changing
>>>>> any scratch register usage if at all possible.
>>>>
>>>> Sorry, I had completely missed this discussion. When looking at the code
>>>> it doesn't look like this particular "resettable" status needs to be
>>>> stored in a PMC scratch register. It can't be stored in RAM because that
>>>> goes into self-refresh as part of LP1, but how about just putting it
>>>> into IRAM? That stays on in both LP1 and LP2, so should be suitable for
>>>> this use-case. It would make the code slightly more complex but using a
>>>> single scratch register for multiple purposes sounds brittle and easy to
>>>> break (as evidenced by the offending commit).
>>>>
>>>> Otherwise it would seem that PMC_SCRATCH40 is only used to store EMC
>>>> configuration data across LP0 suspend/resume, so I wouldn't think it'd
>>>> cause problems if we used that instead of PMC_SCRATCH41 to store the
>>>> "resettable" state.
>>>>
>>>
>>> No. Usually the scratch registers for EMC config data are setup once by the
>>> bootloader and never touched by the kernel after that. So I would not
>>> recommend reusing those registers for different purposes.
>>
>> Right, I misread the code in the downstream kernel. Though it's not the
>> bootloader that does it (at least on Tegra20), but some early code in
>> the kernel.
>>
>> IRAM sounds like a good candidate still. Or do you know of anything that
>> would exclude IRAM as storage location for this data?
>
> No. I can't think of a reason this flag could not be in IRAM.

The only thing you might want to watch out for is whether something else 
is using IRAM. For example, our product SW stacks use the AVP as a media 
co-processor and that runs at least some of its code from IRAM. To 
support something similar, you'd need to make sure to save/restore the 
IRAM content when using it for other purposes rather than just blindly 
over-writing it (and of course synchronize with any driver for the AVP 
execution, to ensure it was shut down first and brought back up last 
after any power saving event). Of course, we don't actually support 
loading code onto the AVP upstream at the moment, so perhaps we can 
defer handling that for now.