Re: [PATCH v5 3/5] drm/msm/dp: set stream_pixel rate directly

[Stephen Boyd <swboyd@chromium.org>]

Quoting Dmitry Baryshkov (2022-03-03 23:58:58)
> On Fri, 4 Mar 2022 at 07:31, Stephen Boyd <swboyd@chromium.org> wrote:
> >
> > Quoting Dmitry Baryshkov (2022-03-03 20:23:06)
> > > On Fri, 4 Mar 2022 at 01:32, Stephen Boyd <swboyd@chromium.org> wrote:
> > > >
> > > > Quoting Dmitry Baryshkov (2022-02-16 21:55:27)
> > > > > The only clock for which we set the rate is the "stream_pixel". Rather
> > > > > than storing the rate and then setting it by looping over all the
> > > > > clocks, set the clock rate directly.
> > > > >
> > > > > Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org>
> > > > [...]
> > > > > diff --git a/drivers/gpu/drm/msm/dp/dp_ctrl.c b/drivers/gpu/drm/msm/dp/dp_ctrl.c
> > > > > index 07f6bf7e1acb..8e6361dedd77 100644
> > > > > --- a/drivers/gpu/drm/msm/dp/dp_ctrl.c
> > > > > +++ b/drivers/gpu/drm/msm/dp/dp_ctrl.c
> > > > > @@ -1315,7 +1315,7 @@ static void dp_ctrl_set_clock_rate(struct dp_ctrl_private *ctrl,
> > > > >         DRM_DEBUG_DP("setting rate=%lu on clk=%s\n", rate, name);
> > > > >
> > > > >         if (num)
> > > > > -               cfg->rate = rate;
> > > > > +               clk_set_rate(cfg->clk, rate);
> > > >
> > > > This looks bad. From what I can tell we set the rate of the pixel clk
> > > > after enabling the phy and configuring it. See the order of operations
> > > > in dp_ctrl_enable_mainlink_clocks() and note how dp_power_clk_enable()
> > > > is the one that eventually sets a rate through dp_power_clk_set_rate()
> > > >
> > > >         dp_ctrl_set_clock_rate(ctrl, DP_CTRL_PM, "ctrl_link",
> > > >                                         ctrl->link->link_params.rate * 1000);
> > > >
> > > >         phy_configure(phy, &dp_io->phy_opts);
> > > >         phy_power_on(phy);
> > > >
> > > >         ret = dp_power_clk_enable(ctrl->power, DP_CTRL_PM, true);
> > >
> > > This code has been changed in the previous patch.
> > >
> > > Let's get back a bit.
> > > Currently dp_ctrl_set_clock_rate() doesn't change the clock rate. It
> > > just stores the rate in the config so that later the sequence of
> > > dp_power_clk_enable() -> dp_power_clk_set_rate() ->
> > > [dp_power_clk_set_link_rate() -> dev_pm_opp_set_rate() or
> > > msm_dss_clk_set_rate() -> clk_set_rate()] will use that.
> > >
> > > There are only two users of dp_ctrl_set_clock_rate():
> > > - dp_ctrl_enable_mainlink_clocks(), which you have quoted above.
> > >   This case is handled in the patch 1 from this series. It makes
> >
> > Patch 1 form this series says DP is unaffected. Huh?
> >
> > > dp_ctrl_enable_mainlink_clocks() call dev_pm_opp_set_rate() directly
> > > without storing (!) the rate in the config, calling
> > > phy_configure()/phy_power_on() and then setting the opp via the
> > > sequence of calls specified above
>
> Note, this handles the "ctrl_link" clock.
>
> > >
> > > - dp_ctrl_enable_stream_clocks(), which calls dp_power_clk_enable()
> > > immediately afterwards. This call would set the stream_pixel rate
> > > while enabling stream clocks. As far as I can see, the stream_pixel is
> > > the only stream clock. So this patch sets the clock rate without
> > > storing in the interim configuration data.
> > >
> > > Could you please clarify, what exactly looks bad to you?
> > >
>
> Note, this handles the "stream_pixel" clock.
>
> >
> > I'm concerned about the order of operations changing between the
> > phy being powered on and the pixel clk frequency being set. From what I
> > recall the pixel clk rate operations depend on the phy frequency being
> > set (which is done through phy_configure?) so if we call clk_set_rate()
> > on the pixel clk before the phy is set then the clk frequency will be
> > calculated badly and probably be incorrect.
>
> But the order of operations is mostly unchanged. The only major change
> is that the opp point is now set before calling the
> phy_configure()/phy_power_on()

Yes that's my concern. The qmp phy driver has a couple clk_set_rate()
calls in the .configure_dp_phy callback. That is called from
phy_power_on() (see qcom_qmp_phy_power_on() and qcom_qmp_phy_dp_ops).
Looking at qcom_qmp_v3_phy_configure_dp_phy() it does

        clk_set_rate(dp_clks->dp_link_hw.clk, dp_opts->link_rate * 100000);
        clk_set_rate(dp_clks->dp_pixel_hw.clk, pixel_freq);

and I believe the child of dp_pixel_hw is find_clock("stream_pixel").
Looks like that is DISP_CC_MDSS_DP_PIXEL_CLK which is
disp_cc_mdss_dp_pixel_clk_src for the rate settable part. That has
clk_dp_ops which is clk_rcg2_dp_set_rate() for the set rate part. That
wants the parent clk frequency to be something non-zero to use in
rational_best_approximation(). If the clk_set_rate("stream_pixel") call
is made before phy_power_on() then the parent_rate in
clk_rcg2_dp_set_rate() won't be valid and the pixel clk frequency will
be wrong.

>
> For the pixel clock the driver has:
> static int dp_ctrl_enable_stream_clocks(struct dp_ctrl_private *ctrl)
> {
>         int ret = 0;
>
>         dp_ctrl_set_clock_rate(ctrl, DP_STREAM_PM, "stream_pixel",
>                                         ctrl->dp_ctrl.pixel_rate * 1000);
>
>         ret = dp_power_clk_enable(ctrl->power, DP_STREAM_PM, true);
> [skipped the error handling]
> }
>
> dp_power_clk_enable() doesn't have any special handlers for the the
> DP_STREAM_PM,
> so this code would be equivalent to the following pseudo code (given
> that there is only one stream clock):
>
> unsigned int rate = ctrl->dp_ctrl.pixel_rate * 1000;
>
> /* dp_ctrl_set_clock_rate() */
> cfg = find_clock_cfg("stream_pixel");
> cfg->rate = rate;
>
> /* dp_power_clk_enable() */
> clk = find_clock("stream_pixel")
> clk_set_rate(clk, cfg->rate);
> clk_prepare_enable(clk);
>
> The proposed patch does exactly this.
>
> Please correct me if I'm wrong.
>