Re: [PATCH v2 2/5] ARM: OMAP2+: DRA7: Add hwmod entries for PWMSS

[Franklin S Cooper Jr.]

Hi Paul

On 02/18/2016 12:58 AM, Paul Walmsley wrote:
> Hi Franklin
>
> On Wed, 17 Feb 2016, Franklin S Cooper Jr. wrote:
>
>> On 08/31/2015 10:51 AM, Paul Walmsley wrote:
>>> On Thu, 16 Jul 2015, R, Vignesh wrote:
>>>> On 07/16/2015 03:24 AM, Paul Walmsley wrote:
>>>>> On Wed, 3 Jun 2015, Vignesh R wrote:
>>>>>
>>>>>> Add hwmod entries for the PWMSS on DRA7.
>>>>>>
>>>>>> Set l4_root_clk_div as the main_clk of PWMSS. It is fixed-factored clock
>>>>>> equal to L4PER2_L3_GICLK/2(l3_iclk_div/2).
>>>>>> As per AM57x TRM SPRUHZ6[1], October 2014, Section 29.1.3 Table 29-4,
>>>>>> clock source to PWMSS is L4PER2_L3_GICLK. But it is actually
>>>>>> L4PER2_L3_GICLK/2. The TRM does not show the division by 2.
>>>>> Is the divide-by-two coming from PWMSS_EPWM.EPWM_TBCTL[HSPCLKDIV]?  Or is 
>>>>> HSPCLKDIV a separate divider after the divide-by-2 you mention above?
>>>> No, it not related to HSPCLKDIV. The TRM wrongly states L4PER2_L3_GICLK
>>>> as clock input for PWMSS. But actually  L4PER2_L4_GICLK(=L3_GICLK/2) is
>>>> the clock input for PWMSS. This will be updated in TRM soon.
>>> OK
>>>
>>>>>> --- a/arch/arm/mach-omap2/omap_hwmod_7xx_data.c
>>>>>> +++ b/arch/arm/mach-omap2/omap_hwmod_7xx_data.c
>>>>>> @@ -362,6 +362,149 @@ static struct omap_hwmod dra7xx_dcan2_hwmod = {
>>>>>>  	},
>>>>>>  };
>>>>>>  
>>>>>> +/* pwmss  */
>>>>>> +static struct omap_hwmod_class_sysconfig dra7xx_epwmss_sysc = {
>>>>>> +	.rev_offs	= 0x0,
>>>>>> +	.sysc_offs	= 0x4,
>>>>>> +	.sysc_flags	= SYSC_HAS_SIDLEMODE | SYSC_HAS_RESET_STATUS,
>>>>> This doesn't match SPRUHZ6 Table 29-13 "PWMSS_SYSCONFIG".  There's no 
>>>>> RESETSTATUS bit.  There is a SOFTRESET bit. 
>>>>>
>>>>> Could you please confirm whether this is intentional?
>>>> sorry my bad... I will change this in v3.
>>> OK
>>>
>>>>>> +/* ecap0 */
>>>>>> +struct omap_hwmod dra7xx_ecap0_hwmod = {
>>>>>> +	.name		= "ecap0",
>>>>>> +	.class		= &dra7xx_ecap_hwmod_class,
>>>>>> +	.clkdm_name	= "l4per2_clkdm",
>>>>>> +	.main_clk	= "l4_root_clk_div",
>>>>> Looking at SPRUHZ6 Section 29.1.4.2 "PWMSS Modules Local Clock Gating", 
>>>>> there appears to be a local "mini-PRCM" for the PWMSS which implements 
>>>>> clock gating and reports back on the status of what I'd guess is the local 
>>>>> clock gating FSM.
>>>>>
>>>>> So from that point of view, you should probably create a clock driver that 
>>>>> manages both the clock gate request bit and the FSM status bit.  It should 
>>>>> be something that can be reused for the other PWMSS IP blocks.  Then 
>>>>> you'd create per-IP block clock nodes and set the main_clk to point to 
>>>>> that node.
>>>> Since, this register is within the config space of PWMSS, the individual
>>>> gating and reporting for the modules within PWMSS
>>>> (PWMSS_CLKCONFIG) is currently being taken care by pwm-tipwmss.c(almost
>>>> the sole function this driver is doing). It has been the same since
>>>> am335x. Adding new clock nodes will result in driver changes and also
>>>> changes to am335x, am437x (and other platforms) hwmod files. It also
>>>> involves adding new nodes to clocks.dtsi and it will be difficult to
>>>> maintain backward compatibility for older platforms. Is it not better to
>>>> keep this as is, in order to maintain consistency (with am335x, am437x
>>>> etc) and also that these clock bits are within IP's config space?
>>> It's certainly possible that we as maintainers didn't look closely enough 
>>> at the AM33xx data for the PWMSS when we merged it.  But if that's 
>>> incorrect too, then now is the time to fix this.  Otherwise it will never 
>>> get fixed, since each new group of people patching this code will keep 
>>> punting it off to the indeterminate future.
>>>
>>> In terms of hwmod data: based on the register maps in sections 29.4.3, 
>>> 29.3.3, and 29.2.3 of SPRUHZ6, none of these subdevices are hwmod devices.  
>>> They don't support the Highlander OCP registers, they have no individual 
>>> PRCM registers or register bitfields, and all of the idle and status 
>>> reporting is to the PWMSS top-level IP block itself.  So it looks to me 
>>> like the eCAP, eQEP, and ePWM modules should be registered via DT, rather 
>>> than via hwmod.  It looks like you can get away with using the 
>>> "simple-bus" abstraction, but you might ultimately have to define 
>>> something custom here.  However, the PWMSS top level subsystem, described 
>>> in section 29.1, does have the OCP registers, sideband signals, etc., and 
>>> thus should remain a hwmod-registered device (via DT).
>>>
>>> In terms of the clock data: based on section 29.1.4, section 29.1.5.2, 
>>> figure 29-3, and table 29-4, there are several clock gating control bits.  
>>> These should be modeled as clock nodes in the Linux common clock 
>>> framework.  Furthermore these registers are located inside the PWMSS 
>>> subsystem itself, and are only accessible when the PWMSS IP block is 
>>> functional in a PM runtime point of view: i.e., when the block is mapped 
>>> into memory space, clocked and out of reset, etc.  So the clock types for 
>>> the PWMSS_CFG IP block should most likely be implemented in the PWMSS 
>>> driver.  I think you'll still be able to define the clock node data itself 
>>> in DT, but this will probably need a closer look.  Ideally, since the 
>>> clock node register address data is the same for all three subsystem 
>>> instances, one would be able to simply include a DT include file three 
>>> times; but the DT binding data format may ultimately make this 
>>> impractical.
>>>
>>> In terms of the transition from the old approach to this approach, it ooks 
>>> to me like the first thing to do would be to convert
>>> drivers/pwm/pwm-tipwmss.c to define some clk_ops and register some clock
>>> nodes with the CCF.  You've got the meat of the clock gating control code
>>> there already.  Then the next thing to do would be to to get rid of
>>> pwmss_submodule_state_change() and use
>>> clk_{prepare,enable,disable,unprepare}*() in the drivers/pwm/pwm-ti*.c
>>> subdrivers instead.  All of that looks like it should be possible to  
>>> implement in a backwards-compatible way.  Then you'd convert the eCAP, 
>>> eQEP, and ePWM code to probe as platform_devices, driven from a simple-bus 
>>> or pwmss-bus or whatever in the DT data.  Naturally, the AM33xx/43xx data 
>>> should be fixed also.  
>> I am working on updating this patchset on behalf of Vignesh
>> based on your suggestion. During the process of converting
>> the pwm-tipwmss.c to be a clock provider I discovered that
>> there is an issue with the PWMSS local clock gates. For
>> example on AM437x if you run the below commands you will get
>> a " Custom Error: MASTER M2 (64-bit) TARGET L4_PER_0 (Read):
>> Data Access in User mode during Functional access" error.
>>
>> rmmod pwm_tiecap
>> rmmod pwm_bl
>> rmmod pwm_tiecap
>> modprobe pwm_tiecap
>> modprobe pwm_bl
>>
>> Full Log: http://pastebin.com/sEyy52HV
>>
>> On device powerup, the various PWMSS local clock gates are
>> "unlocked". The remove call in the ecap driver gates the
>> local PWMSS clock gate for the ecap. The probe of the ecap
>> driver "unlocks" the clock gate for the ecap. However, once
>> you try to access any of the ecap registers which is
>> indirectly accessed by the pwm_bl driver you will get an
>> error. This is because the ecap clock is still being gated.
> OK I'm not sure I understand what's going on, particularly the part about 
> locking and unlocking.  Are you saying that the pwm_bl driver calls into 
> the pwm_tiecap module to write to the PWMSS_CLKCONFIG registers to ungate 
> the ECAP clock, and then the hardware silently ignores the write?  If 
> that's the case, shouldn't we be seeing some warning messages from a 
> failure to ungate the clock from a subsequent PWMSS_CLKSTATUS poll?  Or am 
> I misunderstanding what's going on here?
Pwm-tipwmss.c exports a function
pwmss_submodule_state_change which interacts with
PWMSS_CLKCONFIG and PWMSS_CLKSTATUS registers. Both
pwm-tiehrpwm.c and pwm-tiecap.c calls this exported function
to unlock the gate at probe time and lock the gate when the
driver is removed.

So when the gate fails to be unlocked after it previously is
locked there isn't a way to know this via the
PWMSS_CLKSTATUS registers. So the driver "believes" that the
gate is unlocked and when pwm_bl runs it calls the
set_polarity function. Set_polarity in the case of AM437x gp
evm maps to the ecap's  ecap_pwm_set_polarity function call.
This call then attempts to write to the ecap registers which
results in the external abort since the clock to the ecap is
still gated.

When the ecap and tihrpwm driver request to unlock the clock
gate it already checks the XXX_CLK_EN_ACK bitfields within
CLKSTATUS and it shows that the clocks "should" be unlocked.
So there is an issue with the IP.

>> This has been verified by hardware folks on our side. Its
>> been determined these registers are not needed since they
>> were designed for devices that didn't have a PRCM.
>> Therefore, the current plan is to update the trms of
>> AM335x,AM437x,DRA7 and AM57x to make it clear that these
>> registers shouldn't be touched.
> OK 
>
>> I plan on sending a patch for the pwm-tipwmss.c driver
>> essentially removing anything that touches the
>> PWMSS_CLKCONFIG and PWMSS_CLKSTATUS registers. I will also
>> be sending a v3 version of this patch since I believe the
>> struct omap_hwmod dra7xx_ecap0_hwmod and similar entries you
>> had comments about before are ok to leave as is based on the
>> above findings.
> Including the comments regarding dra7xx_epwmss_sysc ?  

Yes, I plan on addressing the change Vignesh acknowledged
regarding changing RESETSTATUS to SOFTRESET. For your other
comments I believe he already explained why it has to be set
in that particular way. If there is anything that I missed
or something that isn't clear please let me know.
>
>
> - Paul