Re: [PATCH v6 08/17] drm/msm/dsi/phy: Fix reading zero as PLL rates when unprepared

[Dmitry Baryshkov]

On 18/06/2025 16:34, Krzysztof Kozlowski wrote:
> On 18/06/2025 15:07, Dmitry Baryshkov wrote:
> > On Wed, Jun 18, 2025 at 10:28:10AM +0200, Krzysztof Kozlowski wrote:
> > > On 13/06/2025 16:04, Dmitry Baryshkov wrote:
> > > > On 13/06/2025 17:02, Krzysztof Kozlowski wrote:
> > > > > On 13/06/2025 15:55, Dmitry Baryshkov wrote:
> > > > > > >    
> > > > > > > @@ -361,24 +373,47 @@ static int dsi_pll_7nm_lock_status(struct dsi_pll_7nm *pll)
> > > > > > >    
> > > > > > >    static void dsi_pll_disable_pll_bias(struct dsi_pll_7nm *pll)
> > > > > > >    {
> > > > > > > +	unsigned long flags;
> > > > > > >    	u32 data;
> > > > > > >    
> > > > > > > +	spin_lock_irqsave(&pll->pll_enable_lock, flags);
> > > > > > > +	--pll->pll_enable_cnt;
> > > > > > > +	if (pll->pll_enable_cnt < 0) {
> > > > > > > +		spin_unlock_irqrestore(&pll->pll_enable_lock, flags);
> > > > > > > +		DRM_DEV_ERROR_RATELIMITED(&pll->phy->pdev->dev,
> > > > > > > +					  "bug: imbalance in disabling PLL bias\n");
> > > > > > > +		return;
> > > > > > > +	} else if (pll->pll_enable_cnt > 0) {
> > > > > > > +		spin_unlock_irqrestore(&pll->pll_enable_lock, flags);
> > > > > > > +		return;
> > > > > > > +	} /* else: == 0 */
> > > > > > > +
> > > > > > >    	data = readl(pll->phy->base + REG_DSI_7nm_PHY_CMN_CTRL_0);
> > > > > > >    	data &= ~DSI_7nm_PHY_CMN_CTRL_0_PLL_SHUTDOWNB;
> > > > > > >    	writel(0, pll->phy->pll_base + REG_DSI_7nm_PHY_PLL_SYSTEM_MUXES);
> > > > > > >    	writel(data, pll->phy->base + REG_DSI_7nm_PHY_CMN_CTRL_0);
> > > > > > > +	spin_unlock_irqrestore(&pll->pll_enable_lock, flags);
> > > > > > >    	ndelay(250);
> > > > > >
> > > > > > What is this ndelay protecting? Is is to let the hardware to wind down
> > > > > > correctly? I'm worried about dsi_pll_disable_pll_bias() beng followed up
> > > > > > by dsi_pll_enable_pll_bias() in another thread, which would mean that
> > > > > > corresponding writes to the REG_DSI_7nm_PHY_CMN_CTRL_0 can come up
> > > > > > without any delay between them.
> > > > >
> > > > > Great question, but why do you ask me? The code was there already and
> > > > > MSM DRM drivers are not something I know and could provide context about.
> > > >
> > > > Because it's you who are changing the code as you've faced the issue
> > > > with recalc_rate.
> > > Heh, the answer is then: I don't know. I think authors of the code could
> > > know.
> >
> > The 10nm HPG documents a 250ns interval between enabling PLL bias and
> > and enabling the PLL via the CMN_PLL_CNTRL register. There is no extra
> > delay between disabling the PLL, disabling FIFO and remobing PLL bias.
> > Please adjust the code for 7nm and 10nm PHYs accordingly.
>
> I can drop this 250 ns here, if that's what you ask me. But fixing
> anything in 10nm is not relevant to this patchset. You were already
> asking me for different fixes for some different things and I find it
> not acceptable anymore. Stop blocking this patchset with every little
> existing issue.

I think that it is a common practice to ask to fix the issue in relevant 
pieces. For example, we frequently ask to fix all the DT files if there 
was an issue / workaround reported against a selected set of those.

In this case you can send a fix for 10nm separately, but please post a 
fix for that platform too.

> Or merge this code without this patch if a fix for reading PLL as zero
> anyhow is questionable for you.

I think I've asked it at some point, to split the generic code parts and 
the DSI enablement into two different patch series, so that they could 
be picked up separately.

No, the fix is not questionable. The patch causes questions though.



> Best regards,
> Krzysztof


--
With best wishes
Dmitry