[PATCH] ARM: OMAP4: PM: Keep static dep between MPUSS and ABE clockdomain

[Cousson, Benoit]

On 5/15/2012 5:00 PM, Shilimkar, Santosh wrote:
> On Tue, May 15, 2012 at 8:02 PM, Cousson, Benoit<b-cousson at ti.com>  wrote:
>> + Paul
>>
>> Hi Tarun,
>>
>> On 5/15/2012 1:42 PM, Tarun Kanti DebBarma wrote:
>>> Commit 68523f4233de5f233478dde0a63047b4efb710b8 (ARM: OMAP4:
>>> Workaround the OCP synchronisation issue with 32K synctimer)
>>> does not include GP Timers in ABE domain. Since synchronization
>>> issue is applicable to all GPTIMER[1-12], we also need to set
>>> static dependency of MPUSS with abe_clkdm and l4_per_clkdm.
>>> Dependency with l4_per_clkdm timers is already set in commit
>>> 12f27826bdaf56b01cbdfc8bdeb577ebc106dee3 (ARM: OMAP4: PM: Keep
>>> static dep between MPUSS-EMIF and MPUSS-L3/L4 and DUCATI-L3).
>>> Therefore, set static dependency of MPUSS with abe_clkdm.
>>
>> It seems to me that these various static dep workaround patches are more and more confusing and should require some further investigation / explanation.
>>
> This is a new BUG which has not made it to errata list yet. It will
> make it eventually.
>
>> If we keep doing that we will end up having every clock domains always ON each time the CPU is active. This is a very brute force approach not really acceptable for mainline and for PM point of view.
>>
> Indeed.
>
>> Here we are forcing the ABE domain to be ON each time the MPU is ON even if we do not have any timer used inside the domain.
>>
> Actually the BUG is really related to timers running on 32KHz and only
> in that case such a WA is needed. BTW, the WA is suggested by hardware
> team.

That does not mean we have to implement it that way, or there is no 
other WA. HW team tends to stop investigating as soon as one WA is found.

>> It was mostly OK to do that for the wakeup domain due to the small power impact, but doing that on the L4_PER and ABE seems a little bit too much.
>>
> L4PER and ABE should not be set default....

Indeed.

>> The fix assumes as well that the MPU is the only user of that timer. What if either DSP or IPU uses it as well?
>>
>
>> Moreover the previous 68523f4233de5f2 commit did add tons of deps that does not seems to be really justified by any HW errata AFAIK.
>> That does not mean they are not needed, but I think we should either remove them or add some more explanation.
>>
> EMIF and L3 are covered as part of the errata's. Most if these static
> deps issues never worked properly and people ended up hacking
> like disable L3 when display ON etc etc.

Yeah, that's why we have to be more explicit because some of them were 
already investigated further since last year, and as you said some are 
already deprecated.

>> +       /*
>> +        * The dynamic dependency between MPUSS ->  MEMIF and
>> +        * MPUSS ->  L4_PER/L3_* and DUCATI ->  L3_* doesn't work as
>> +        * expected. The hardware recommendation is to enable static
>> +        * dependencies for these to avoid system lock ups or random crashes.
>> +        */
>> +       mpuss_clkdm = clkdm_lookup("mpuss_clkdm");
>> +       emif_clkdm = clkdm_lookup("l3_emif_clkdm");
>> +       l3_1_clkdm = clkdm_lookup("l3_1_clkdm");
>> +       l3_2_clkdm = clkdm_lookup("l3_2_clkdm");
>> +       l4_per_clkdm = clkdm_lookup("l4_per_clkdm");
>> +       ducati_clkdm = clkdm_lookup("ducati_clkdm");
>> +       if ((!mpuss_clkdm) || (!emif_clkdm) || (!l3_1_clkdm) ||
>> +               (!l3_2_clkdm) || (!ducati_clkdm) || (!l4_per_clkdm))
>> +               goto err2;
>> +
>> +       ret = clkdm_add_wkdep(mpuss_clkdm, emif_clkdm);
>>
>> AFAIK, this one is the only one covered by an errata. It might be good to add a comment to explained the issue.
>>
>> +       ret |= clkdm_add_wkdep(mpuss_clkdm, l3_1_clkdm);
>> +       ret |= clkdm_add_wkdep(mpuss_clkdm, l3_2_clkdm);
>> +       ret |= clkdm_add_wkdep(mpuss_clkdm, l4_per_clkdm);
>> +       ret |= clkdm_add_wkdep(ducati_clkdm, l3_1_clkdm);
>> +       ret |= clkdm_add_wkdep(ducati_clkdm, l3_2_clkdm);
>>
>> Do we have errata for any of these ones?
>>
> If we forget about the latest timer issues doing the round only EMIF and L3
> deps with different initiators were needed.

Is there an errata for these ones?

> L4PER was because of UART
> idle mode issue which I fixed recently. At least with that fix, L4PER should
> be killed. Will be good to take the latest findings on the static deps issues
> and update above list.

Yes, that was my point. We have to reduce that list if possible and 
potentially add some details to understand why these deps are needed.

> These timer OCP sync issue has really created a big mess again...
> Timer is 3 domains. AON, ABE and L4pER and the WA suggested is
> static dep as if it is free. There is no other WA.

There is always other WA. Playing with global clock domain settings to 
prevent clock transition at IP level is always the easy but worst WA.

A much better approach is to play with IP local idle management. At 
least that will ensure that the WA is applied only if the module is in 
use. Another one will be to read until the timer has changed its value. 
OK, that one will generate a huge delay that might be un-acceptable for 
a timer using a 32k clock, but might be doable for a one at sys_clk, 
assuming this bug does exist in that case.

My point is that we have to refine the way we are applying this WA if it 
is really needed for all the cases. Enabling at boot time only is 
clearly not a good approach.

In any case, this WA should be module centric. If we change the clock 
domain partitioning for OMAP5, we will have to change again that fix 
since we are hard coding the clock domain in that code.
This is the timers IP that does have this issue, so it should be handled 
at that level.

Regards,
Benoit