Re: [PATCH v1] arm64: dts: imx8mm: fix missing pgc_vpu_* power domain parent

[Lucas Stach <l.stach@pengutronix.de>]

Hi Vitor,

Am Dienstag, dem 16.04.2024 um 17:08 +0100 schrieb Vitor Soares:
> On Tue, 2024-04-16 at 11:53 +0100, Vitor Soares wrote:
> > ++ Peng Fan <peng.fan@nxp.com>
> > 
> > Greetings,
> > 
> > 
> > On Wed, 2024-04-10 at 12:01 +0100, Vitor Soares wrote:
> > > Hi Lucas,
> > > 
> > > On Tue, 2024-04-09 at 17:44 +0100, Vitor Soares wrote:
> > > > On Tue, 2024-04-09 at 16:36 +0200, Lucas Stach wrote:
> > > > > Am Dienstag, dem 09.04.2024 um 14:22 +0100 schrieb Vitor Soares:
> > > > > > Hi Lucas,
> > > > > > 
> > > > > > Thanks for your feedback.
> > > > > > 
> > > > > > On Tue, 2024-04-09 at 11:13 +0200, Lucas Stach wrote:
> > > > > > > Hi Vitor,
> > > > > > > 
> > > > > > > Am Dienstag, dem 09.04.2024 um 09:58 +0100 schrieb Vitor
> > > > > > > Soares:
> > > > > > > > From: Vitor Soares <vitor.soares@toradex.com>
> > > > > > > > 
> > > > > > > > The pgc_vpu_* nodes miss the reference to the power domain
> > > > > > > > parent,
> > > > > > > > leading the system to hang during the resume.
> > > > > > > > 
> > > > > > > This change is not correct. The vpumix domain is controlled
> > > > > > > through
> > > > > > > the
> > > > > > > imx8mm-vpu-blk-ctrl and must not be directly triggered by the
> > > > > > > child
> > > > > > > domains in order to guarantee proper power sequencing.
> > > > > > > 
> > > > > > > If the sequencing is incorrect for resume, it needs to be
> > > > > > > fixed
> > > > > > > in
> > > > > > > the
> > > > > > > blk-ctrl driver. I'll happily assist if you have any
> > > > > > > questions
> > > > > > > about
> > > > > > > this intricate mix between GPC and blk-ctrl hardware/drivers.
> > > > > >  
> > > > > > I'm new into the topic, so I tried to follow same approach as
> > > > > > in
> > > > > > imx8mp
> > > > > > DT.
> > > > > > 
> > > > > That's a good hint, the 8MP VPU GPC node additions missed my
> > > > > radar.
> > > > > The
> > > > > direct dependency there between the GPC domains is equally wrong.
> > > > > 
> > > > > > I also checked the imx8mq DT and it only have one domain for
> > > > > > the
> > > > > > VPU in the GPC. It seem blk-ctrl also dependes on pgc_vpu_* to
> > > > > > work
> > > > > > properly.
> > > > > > 
> > > > > > The blk-ctrl driver hangs on imx8m_blk_ctrl_power_on() when
> > > > > > access
> > > > > > the
> > > > > > ip registers for the soft reset. I tried to power-up the before
> > > > > > the
> > > > > > soft reset, but it didn't work.
> > > > > > 
> > > > > The runtime_pm_get_sync() at the start of that function should
> > > > > ensure
> > > > > that bus GPC domain aka vpumix is powered up. Can you check if
> > > > > that
> > > > > is
> > > > > happening?
> > > > 
> > > > I checked bc->bus_power_dev->power.runtime_status and it is
> > > > RPM_ACTIVE.
> > > > 
> > > > Am I looking to on the right thing? It is RPM_ACTIVE event before
> > > > runtime_pm_get_sync().
> > > 
> > > During the probe I can see that
> > > bus_power_dev->power.runtime_status = RPM_SUSPENDED and then vpumix
> > > is
> > > powered up on GPC driver.
> > > 
> > > On resume routine I can't see this flow. bus_power_dev-
> > > > power.runtime_status = RPM_ACTIVE and vpumix end up not being
> > > > powered-
> > > up.
> > > 
> > > I checked the suspend flow and the GPC tries to poweroff vpumix.
> > > 
> > > 
> > 
> > My understanding is that when resuming the 38310000.video-codec, the
> > vpumix isn't powered up. It happens because runtime_status and
> > runtime_last_status = RPM_ACTIVE. 
> > 
> > I tried to change blk-ctrl suspend routine to force the runtime_status
> > = RPM_SUSPENDED, but the system ended up hanging on another device.
> > 
> > From the comment in blk-ctrl suspend, we rely on PM_SLEEP code that
> > iterates over dpm_list for suspend/resume.
> > I did look at the dpm_list, and it changes the order on every boot. 
> > 
> > With all the tests, I also found that the system randomly hangs on
> > dispblk-lcdif suspend. I have confirmed this device is in a different
> > place in the dpm_list (not sure if it is the root cause). 
> > I haven't understood how blk-ctrl ensures the correct order there yet. 
> > 
Random order of the DPM list seems like a good find to investigate
further.

> > Taking the following dpm_list excerpt:
> > idx - device
> > ------------------------------
> > ...                                                                  
> > 191 - imx-pgc-domain.7                                               
> > 192 - imx-pgc-domain.8                                               
> > 193 - imx-pgc-domain.9                                               
> > 194 - 38330000.blk-ctrl                                              
> > 195 - 38310000.video-codec                                           
> > 196 - 38300000.video-codec                                           
> > ...
> > 205 - genpd:0:38330000.blk-ctrl
> > 206 - genpd:1:38330000.blk-ctrl
> > 207 - genpd:2:38330000.blk-ctrl
> > 208 - genpd:3:38330000.blk-ctrl
> > ------------------------------
> > 
> > Shouldn't genpd devices be before 38330000.blk-ctrl?
> > As their power domain is GPC and the blk-ctrl power domain is genpd.
> > 
> 
> I did the following change to have genpd device before 38330000.blk-ctrl
> on dpm_list and it did work.
> 
> diff --git a/drivers/pmdomain/imx/imx8m-blk-ctrl.c b/drivers/pmdomain/imx/imx8m-blk-ctrl.c
> index ca942d7929c2..0f1471dcd4e8 100644
> --- a/drivers/pmdomain/imx/imx8m-blk-ctrl.c
> +++ b/drivers/pmdomain/imx/imx8m-blk-ctrl.c
> @@ -220,6 +220,7 @@ static int imx8m_blk_ctrl_probe(struct platform_device *pdev)
>                         return dev_err_probe(dev, PTR_ERR(bc->bus_power_dev),
>                                              "failed to attach power domain \"bus\"\n");
>         }
> +       device_move(dev, bc->bus_power_dev, DPM_ORDER_PARENT_BEFORE_DEV);
>  
>         for (i = 0; i < bc_data->num_domains; i++) {
>                 const struct imx8m_blk_ctrl_domain_data *data = &bc_data->domains[i];
> @@ -268,6 +269,7 @@ static int imx8m_blk_ctrl_probe(struct platform_device *pdev)
>                                       data->gpc_name);
>                         goto cleanup_pds;
>                 }
> +               device_move(dev, domain->power_dev, DPM_ORDER_PARENT_BEFORE_DEV);
>  
>                 domain->genpd.name = data->name;
>                 domain->genpd.power_on = imx8m_blk_ctrl_power_on;
> 
> any concern about this approach?
> 
I'm a bit uncomfortable with calling such a low-level function from
this driver. Also we don't really want to move the device to a new
parent, but just want to ensure proper order on the dpm list. Adding a
device_link between the devices seems like the better way to do so.

Regards,
Lucas

> Best regards,
> Vitor Soares
> > 
> > > 
> > > > 
> > > > 
> > > > > 
> > > > > Regards,
> > > > > Lucas
> > > > > 
> > > > > > Do you have an idea how we can address this within blk-ctrl?
> > > > > > 
> > > > > > Best regards,
> > > > > > Vitor
> > > > 
> > > 
> > 
>