Re: [PATCH v3 RESEND 07/11] pwm: imx: Provide atomic PWM support for i.MX PWMv2

[Boris Brezillon <boris.brezillon@free-electrons.com>]

On Tue, 3 Jan 2017 23:01:11 +0100
Lukasz Majewski <lukma@denx.de> wrote:

> Hi Boris, Stefan,
> 
> > On Tue, 03 Jan 2017 09:29:40 -0800
> > Stefan Agner <stefan@agner.ch> wrote:
> >   
> > > On 2017-01-03 04:46, Boris Brezillon wrote:
> > > <snip>  
> > > >> > Well, regarding the imx_pwm_apply_v2() suggested by Stefan, I
> > > >> > think we both agreed that most of the code was unneeded when
> > > >> > all we want to do is disable the PWM.    
> > > >>
> > > >> So for the PATCH 7/11 we fix the issue with recalculating clocks
> > > >> when we want to disable PWM.
> > > >>
> > > >> if (state->enabled) {
> > > >> 		c = clk_get_rate(imx->clk_per);
> > > >> 		c *= state->period;
> > > >>
> > > >> 		do_div(c, 1000000000);
> > > >> 		period_cycles = c;
> > > >>
> > > >> 		prescale = period_cycles / 0x10000 + 1;
> > > >>
> > > >> 		period_cycles /= prescale;
> > > >> 		c = (unsigned long long)period_cycles *
> > > >> 		    state->duty_cycle;
> > > >> 		do_div(c, state->period);
> > > >> 		duty_cycles = c;
> > > >>
> > > >> 		/*
> > > >> 		 * According to imx pwm RM, the real period value
> > > >> 		 * should be PERIOD value in PWMPR plus 2.
> > > >> 		 */
> > > >> 		if (period_cycles > 2)
> > > >> 			period_cycles -= 2;
> > > >> 		else
> > > >> 			period_cycles = 0;
> > > >>
> > > >> 		/*
> > > >> 		 * Enable the clock if the PWM is not already
> > > >> 		 * enabled.
> > > >> 		 */
> > > >> 		if (!cstate.enabled) {
> > > >> 			ret = clk_prepare_enable(imx->clk_per);
> > > >> 			if (ret)
> > > >> 			return ret;
> > > >> 		}
> > > >>
> > > >> 		/*
> > > >> 		 * Wait for a free FIFO slot if the PWM is
> > > >> already
> > > >> 		 * enabled, and flush the FIFO if the PWM was
> > > >> disabled
> > > >> 		 * and is about to be enabled.
> > > >> 		 */
> > > >> 		if (cstate.enabled)
> > > >> 			imx_pwm_wait_fifo_slot(chip, pwm);
> > > >> 		else
> > > >> 			imx_pwm_sw_reset(chip);
> > > >>
> > > >> 		writel(duty_cycles, imx->mmio_base + MX3_PWMSAR);
> > > >> 		writel(period_cycles, imx->mmio_base +
> > > >> MX3_PWMPR);
> > > >>
> > > >> 		writel(MX3_PWMCR_PRESCALER(prescale) |
> > > >> 		       MX3_PWMCR_DOZEEN | MX3_PWMCR_WAITEN |
> > > >> 		       MX3_PWMCR_DBGEN |
> > > >> MX3_PWMCR_CLKSRC_IPG_HIGH | MX3_PWMCR_EN,
> > > >> 		       imx->mmio_base + MX3_PWMCR);
> > > >> 	} else {
> > > >>
> > > >> 		writel(0, imx->mmio_base + MX3_PWMCR);
> > > >>
> > > >> 		/* Disable the clock if the PWM is currently
> > > >> enabled. */ if (cstate.enabled)
> > > >> 			clk_disable_unprepare(imx->clk_per);
> > > >> 	}
> > > >>
> > > >>    
> > > > 
> > > > Yep.
> > > >     
> > > 
> > > This looks like a good transformation of the current Patch 7, but
> > > once you merge my patch, it will look slightly different...  
> > 
> > Yes. I think we should just unconditionally enable/disable the per_clk
> > at function entry/exit. The prepare_enable() call is almost free
> > when the clk is already enabled, so it's not like we're adding a huge
> > overhead by doing that.  
> 
> So in the above snippet we should replace:
> 
> if (!cstate.enabled) {
> 	ret = clk_prepare_enable(imx->clk_per);
> 	if (ret)
> 		return ret;
> }
> 
> with
> 	ret = clk_prepare_enable(imx->clk_per);
> 	if (ret)
> 		return ret;
> 
> And 
> 
> if (cstate.enabled)
> 	clk_disable_unprepare(imx->clk_per);
> 
> with 
> 	clk_disable_unprepare(imx->clk_per);

That's what I had in mind.

> 
> >   
> > >   
> > > >>    
> > > >> >
> > > >> > My concern was more about the way PWM changes are applied
> > > >> > (->apply() returns before the change is actually applied), but
> > > >> > I agreed that it could be fixed later on (if other people
> > > >> > think it's really needed), since the existing code already
> > > >> > handles it this way.    
> > > >>
> > > >> This is the issue with FIFO setting - but for now we do not deal
> > > >> with it.    
> > > > 
> > > > Exactly.
> > > >     
> > > >>    
> > > >> >    
> > > >> > > No clear decision what to change until today when Stefan
> > > >> > > prepared separate (concise) patch (now I see what is the
> > > >> > > problem).   
> > > >> >
> > > >> > The patch proposed by Stefan is addressing a different
> > > >> > problem: the periph clock has to be enabled before accessing
> > > >> > registers.    
> > > >>
> > > >> So for this reason Stefan's patch [1] always enable the clock no
> > > >> matter if PWM clock is generated or not.    
> > > > 
> > > > Yes.
> > > >     
> > > >>    
> > > >> >    
> > > >> > >    
> > > >> > > >
> > > >> > > > Same goes for the regression introduced in patch 2: I
> > > >> > > > think it's better to keep things bisectable on all
> > > >> > > > platforms (even if it appeared to work by chance on imx7,
> > > >> > > > it did work before this change).    
> > > >> > >
> > > >> > > Could you be more specific about your idea to solve this
> > > >> > > problem?    
> > > >> >
> > > >> > Stefan already provided a patch, I just think it should be
> > > >> > fixed before patch 2 to avoid breaking bisectibility.    
> > > >>
> > > >> My idea is as follows:
> > > >>
> > > >> I will drop patch v2 (prepared by Sasha) and then squash
> > > >> Stefan's patch [1] to patch 7/11. The "old" ipg enable code will
> > > >> be removed with other not needed code during conversion.    
> > > > 
> > > > How about keeping patch 2 but enabling/disabling the periph clk
> > > > in imx_pwm_config() instead of completely dropping the
> > > > enable/disable clk sequence.
> > > > 
> > > > In patch 7 you just add the logic we talked about earlier:
> > > > unconditionally enable the periph clk when entering the
> > > > imx_pwm_apply_v2() function and disable it before leaving the
> > > > function.
> > > > 
> > > > This way you can preserve bisectibility and still get rid of the
> > > > ipg clk.
> > > > 
> > > > Stefan, what's your opinion?    
> > > 
> > > We will get rid of the ipg clocks anyway in patch 8 (which removes
> > > those functions completely).
> > > 
> > > So I think Lukasz approach should be fine, just drop patch 2 and
> > > squash my patch into patch 7.  
> > 
> > Well, the end result will be same (ipg_clk will be gone after patch
> > 8), but then it's hard to track why this clock suddenly disappeared.
> > I still think it's worth adding an extra commit explaining that
> > enabling the per_clk before accessing IP registers is needed on some
> > platforms (imx7), and that IPG clk is actually not required until we
> > start using it as a source for the PWM signal generation.
> > 
> > Maybe I'm the only one to think so. In this case, feel free to drop
> > patch 2.  
> 
> If you feel really bad about this issue, then we can drop patch 2 and:
> 
> reorganize patch 7/11 to 
>  - keep code, which adds imx_pwm_apply_v2() function code (just moves it
> as is) 
>  - remove .apply = imx_pwm_apply_v2 entry from pwm_ops structure.
> 
> 
> On top of it add patch to enable/disable unconditionally the
> imx->clk_per clock to avoid problems on imx7 (and state them in commit
> message).
> 
> Then we add separate patch with 
> .apply = imx_pwm_apply_v2 to pwm_ops structure to enable "new" atomic
> approach.
> 
> And at last we apply patch 8/11, which removes the code for old (non
> atomic) behaviour.
> 
> All the issues are documented in this way on the cost of having
> "dead" (I mean not used) imx_pwm_apply_v2() for two commits.
> 

This looks even more complicated.
Sorry, but I don't see the problem with modifying patch 2 to enable
per_clk instead of ipg_clk. Can you explain what's bothering you?

If you really want to do the change after patch 7, fine, but in this
case, keep the existing logic: enable/disable ipg_clk in
imx_pwm_apply_v2() until you drop the ipg_clk and replace the ipg_clk
enable/disable sequence by the equivalent enable/disable per_clk one.