Re: [PATCH v4 02/11] memory: emif: Move EMIF register defines to include/linux/

[Dave Gerlach <d-gerlach@ti.com>]

Tony,
On 07/15/2014 01:38 AM, Tony Lindgren wrote:
> * Dave Gerlach <d-gerlach@ti.com> [140714 10:44]:
>> On 07/14/2014 06:12 AM, Tony Lindgren wrote:
>>> * Dave Gerlach <d-gerlach@ti.com> [140710 19:59]:
>>>> OMAP4 and AM33XX share the same EMIF controller IP. Although there
>>>> are significant differences in the IP integration due to which
>>>> AM33XX can't reuse the EMIF driver DVFS similar to OMAP4,
>>>> it can definitely benefit by reusing the EMIF related macros
>>>> defined in drivers/memory/emif.h.
>>>>
>>>> In the current OMAP PM framework the PM code resides under
>>>> arch/arm/mach-omap2/. To enable reuse of the register defines move
>>>> the register defines in the emif header file to include/linux so that
>>>> both the EMIF driver and the AM33XX PM code can benefit.
>>>>
>>>> Signed-off-by: Dave Gerlach <d-gerlach@ti.com>
>>>> Reviewed-by: Russ Dill <russ.dill@ti.com>
>>>> Acked-by: Santosh Shililmar <santosh.shilimkar@ti.com>
>>>> Acked-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>>>> ---
>>>> v3->v4:
>>>> patch unchanged from original:
>>>> 	http://www.spinics.net/lists/linux-omap/msg95314.html
>>>>
>>>>   drivers/memory/emif.h   | 543 +---------------------------------------------
>>>>   include/linux/ti_emif.h | 558 ++++++++++++++++++++++++++++++++++++++++++++++++
>>>
>>> So far we've seen that exposing hardware registers like this
>>> will lead into various drivers misusing them. I think a better
>>> solution is to implement few targeted functions that allow
>>> sharing code between the platform idle code and memory driver.
>>>
>>> Maybe you can have the shared functions in in something like
>>> drivers/memory/ti-emif.c that's always built in? The idle
>>> code won't need any of that early on.
>>
>> Well the reason it was done this way was to utilize all of the addresses of
>> EMIF register in the ASM sleep code to do relative addressing from the EMIF
>> base address. The ASM sleep code (patch 9) needs to save and restore emif
>> context and set and unset self refresh in emif. The issues will come from
>> the ASM being copied to and running from SRAM without the ability to access
>> code in DDR (because we are shutting the EMIF off), so we would need to copy
>> these functions as well and have to worry about any issues we introduce by
>> relocating c code. Is it worth the added maintenance burden?
>
> Ah right it needs to run in SRAM. There were some relocatable
> c code patches posted a while back, so it might be worth
> revisiting that.
>
> I think it can also be done with assembly with something like
> this:
>
> 1. Make am335x idle code depends on TI_EMIF && WKUP_M3_RPROC
>
> 2. Add the memory save and restore assembly functions into
>     drivers/memory/ti-emif-sram.S
>
> 3. Allocate the SRAM preferrably with drivers/misc/sram.c
>     instead of the legacy mach-omap2/sram.c
>

This I can do, I will just need to make a change somewhere to make generic sram 
driver provide sram allocations marked for exec.

> 4. Map the idle assembly code and EMIF save and restore
>     functions into SRAM
>
> 5. Call the EMIF save and restore functions from the idle
>     assembly code at the SRAM locations and pass the save and
>     restore area in a register
>
> So basically we need to figure out a generic way to do driver
> hooks in the PM idle code even very late and early in the
> assembly code so we can keep most of the code in drivers.
> Eventually also the idle assembly code should be in the drivers
> too..

I did not consider this earlier but the cpuidle code will use the same path in 
the assembly code. The cpuidle configures the suspend path to make the emif 
actions optional (save and restore with shut off, and self refresh), so a 
generic solution probably isn't possible here as we (will) need a certain level 
of granularity of control over the emif actions, and that will be difficult to 
maintain while keeping the pm functionality out of the EMIF code.

Regards,
Dave

>
> Regards,
>
> Tony
>