Re: [PATCH v6 2/3] media: imx: add a driver for i.MX8MQ mipi csi rx phy and controller

[Martin Kepplinger <martin.kepplinger@puri.sm>]

Am Freitag, dem 16.07.2021 um 13:47 +0300 schrieb Laurent Pinchart:
> Hi Martin,
> 
> On Fri, Jul 16, 2021 at 10:47:14AM +0200, Martin Kepplinger wrote:
> > Am Freitag, dem 16.07.2021 um 00:52 +0300 schrieb Laurent Pinchart:
> > > On Thu, Jul 15, 2021 at 09:37:24AM +0200, Martin Kepplinger
> > > wrote:
> > > > Am Mittwoch, dem 14.07.2021 um 21:24 +0300 schrieb Laurent
> > > > Pinchart:
> > > > > On Wed, Jul 14, 2021 at 01:19:30PM +0200, Martin Kepplinger
> > > > > wrote:
> > > > > > Add a driver to support the i.MX8MQ MIPI CSI receiver. The
> > > > > > hardware side
> > > > > > is based on
> > > > > > https://source.codeaurora.org/external/imx/linux-imx/tree/drivers/media/platform/imx8/mxc-mipi-csi2_yav.c?h=imx_5.4.70_2.3.0
> > > > > > 
> > > > > > It's built as part of VIDEO_IMX7_CSI because that's
> > > > > > documented to support
> > > > > > i.MX8M platforms. This driver adds i.MX8MQ support where
> > > > > > currently only the
> > > > > > i.MX8MM platform has been supported.
> > > > > > 
> > > > > > Signed-off-by: Martin Kepplinger
> > > > > > <martin.kepplinger@puri.sm>
> > > > > > ---
> > > > > >  drivers/staging/media/imx/Makefile           |   1 +
> > > > > >  drivers/staging/media/imx/imx8mq-mipi-csi2.c | 949
> > > > > > +++++++++++++++++++
> > > > > >  2 files changed, 950 insertions(+)
> > > > > >  create mode 100644 drivers/staging/media/imx/imx8mq-mipi-
> > > > > > csi2.c
> > > > > > 
> > > > > > diff --git a/drivers/staging/media/imx/Makefile
> > > > > > b/drivers/staging/media/imx/Makefile
> > > > > > index 6ac33275cc97..19c2fc54d424 100644
> > > > > > --- a/drivers/staging/media/imx/Makefile
> > > > > > +++ b/drivers/staging/media/imx/Makefile
> > > > > > @@ -16,3 +16,4 @@ obj-$(CONFIG_VIDEO_IMX_CSI) += imx6-mipi-
> > > > > > csi2.o
> > > 
> > > [snip]
> > > 
> > > > > > +static int imx8mq_mipi_csi_calc_hs_settle(struct csi_state
> > > > > > *state)
> > > > > > +{
> > > > > > +       u32 width = state-
> > > > > > >format_mbus[MIPI_CSI2_PAD_SINK].width;
> > > > > > +       u32 height = state-
> > > > > > >format_mbus[MIPI_CSI2_PAD_SINK].height;
> > > > > > +       s64 link_freq;
> > > > > > +       u32 lane_rate;
> > > > > > +
> > > > > > +       /* Calculate the line rate from the pixel rate. */
> > > > > > +       link_freq = v4l2_get_link_freq(state->src_sd-
> > > > > > >ctrl_handler,
> > > > > > +                                      state->csi2_fmt-
> > > > > > >width,
> > > > > > +                                      state-
> > > > > > >bus.num_data_lanes * 2);
> > > > > > +       if (link_freq < 0) {
> > > > > > +               dev_err(state->dev, "Unable to obtain link
> > > > > > frequency: %d\n",
> > > > > > +                       (int)link_freq);
> > > > > > +               return link_freq;
> > > > > > +       }
> > > > > > +
> > > > > > +       lane_rate = link_freq * 2;
> > > > > > +       if (lane_rate < 80000000 || lane_rate > 1500000000)
> > > > > > {
> > > > > > +               dev_dbg(state->dev, "Out-of-bound lane rate
> > > > > > %u\n", lane_rate);
> > > > > > +               return -EINVAL;
> > > > > > +       }
> > > > > > +
> > > > > > +       /*
> > > > > > https://community.nxp.com/t5/i-MX-Processors/Explenation-for-HS-SETTLE-parameter-in-MIPI-CSI-D-PHY-registers/m-p/764275/highlight/true#M118744
> > > > > >  */
> > > > > > +       if (lane_rate < 250000000)
> > > > > > +               state->hs_settle = 0xb;
> > > > > > +       else if (lane_rate < 500000000)
> > > > > > +               state->hs_settle = 0x8;
> > > > > > +       else
> > > > > > +               state->hs_settle = 0x6;
> > > > > 
> > > > > We could possibly compute this value based on the formula
> > > > > from the table
> > > > > in that page, but maybe that's overkill ? If you want to give
> > > > > it a try,
> > > > > it would be along those lines.
> > > > > 
> > > > >         /*
> > > > >          * The D-PHY specification requires Ths-settle to be
> > > > > in the range
> > > > >          * 85ns + 6*UI to 140ns + 10*UI, with the unit
> > > > > interval UI being half
> > > > >          * the clock period.
> > > > >          *
> > > > >          * The Ths-settle value is expressed in the hardware
> > > > > as a multiple of
> > > > >          * the Esc clock period:
> > > > >          *
> > > > >          * Ths-settle = (PRG_RXHS_SETTLE + 1) * Tperiod of
> > > > > RxClkInEsc
> > > > >          *
> > > > >          * Due to the one cycle inaccuracy introduced by
> > > > > rounding, the
> > > > >          * documentation recommends picking a value away from
> > > > > the boundaries.
> > > > >          * Let's pick the average.
> > > > >          */
> > > > >         esc_clk_rate = clk_get_rate(...);
> > > > > 
> > > > >         min_ths_settle = 85 + 6 * 1000000 / (lane_rate /
> > > > > 1000);
> > > > >         max_ths_settle = 140 + 10 * 1000000 / (lane_rate /
> > > > > 1000);
> > > > >         ths_settle = (min_ths_settle + max_ths_settle) / 2;
> > > > > 
> > > > >         state->hs_settle = ths_settle * esc_clk_rate /
> > > > > 1000000000 - 1;
> > > > 
> > > > I experimented a bit but would like to leave this as a task for
> > > > later
> > > > if that's ok. it's correct and simple now. also, using
> > > > clks[i].clk
> > > > based on the name string would feel better to submit seperately
> > > > later.
> > > 
> > > That's OK with me, but I may then submit a patch on top fairly
> > > soon :-)
> > > Have you been able to test if this code works on your device ?
> > > The main
> > > reason why I think it's better is that it doesn't hardcode a
> > > specific
> > > escape clock frequency assumption, so it should be able to
> > > accommodate a
> > > wider range of use cases. If we change it later, there's always a
> > > risk
> > > of regressions, while if we do this from the start, we'll figure
> > > out
> > > quickly if it doesn't work in some cases.
> > 
> > taking your code basically as-is doesn't yet work, but it helps a
> > bit.
> 
> Thanks for testing.
> 
> > tbh I don't even know how to correctly read that table /
> > calculation:
> > what is the exact relation of the calculated Ths_settle time
> > inverval
> > to the hs_settle register bits?
> 
> The PRG_RXHS_SETTLE field stores a number of timer ticks to cover the
> Ths-settle internal. The D-PHY arms the timer when it detects the
> transition to LP-00, and ignores transitions on the lane until the
> timer
> expires. The timer is clocked by the escape clock.
> 
> What hs_settle value do you currently use, and what value does my
> code
> produce ?
> 
> > if the 2 of us can't quickly figure it out I can ask NXP via that
> > community forum issue and I created
> > https://source.puri.sm/Librem5/linux-next/-/issues/340 so I won't
> > forget about it.
> 

hi Laurent,

the below patch for hs_settle works (and calculates either the same or
+/- 1 for the hs_settle value, compared to the table), but getting the
esc clock looks really scary how I do it: how would you do that?


@@ -284,6 +285,9 @@ static int imx8mq_mipi_csi_calc_hs_settle(struct
csi_state *state)
 {
 	s64 link_freq;
 	u32 lane_rate;
+	u32 esc_clk_rate = 0;
+	u32 i, min_ths_settle, max_ths_settle, ths_settle_ns,
esc_clk_period_ns;
+	char *p;
 
 	/* Calculate the line rate from the pixel rate. */
 	link_freq = v4l2_get_link_freq(state->src_sd->ctrl_handler,
@@ -302,20 +306,44 @@ static int imx8mq_mipi_csi_calc_hs_settle(struct
csi_state *state)
 	}
 
 	/*
-	 * The following table is the source for this:
-	 *
https://community.nxp.com/t5/i-MX-Processors/Explenation-for-HS-SETTLE-parameter-in-MIPI-CSI-D-PHY-registers/m-p/764275/highlight/true#M118744
-	 * but it would be even better to calculate the value for any
-	 * given datarate.
+	 * The D-PHY specification requires Ths-settle to be in the
range
+	 * 85ns + 6*UI to 140ns + 10*UI, with the unit interval UI
being half
+	 * the clock period.
+	 *
+	 * The Ths-settle value is expressed in the hardware as a
multiple of
+	 * the Esc clock period:
+	 *
+	 * Ths-settle = (PRG_RXHS_SETTLE + 1) * Tperiod of RxClkInEsc
+	 *
+	 * Due to the one cycle inaccuracy introduced by rounding, the
+	 * documentation recommends picking a value away from the
boundaries.
+	 * Let's pick the average.
 	 */
-	if (lane_rate < 250000000)
-		state->hs_settle = 0xb;
-	else if (lane_rate < 500000000)
-		state->hs_settle = 0x8;
-	else
-		state->hs_settle = 0x6;
-
-	dev_dbg(state->dev, "start stream: lane rate %u hs_settle
%u\n",
-		lane_rate, state->hs_settle);
+	for (i = 0; i < CSI2_NUM_CLKS; i++) {
+		p = (char *)__clk_get_name(state->clks[i].clk);
+		/* we're getting csi1_esc here */
+		if (strlen(p) > 7)
+			p += 5;
+
+		dev_dbg(state->dev, "comparing: %s to esc\n", p);
+		if (!strcmp(p, "esc"))
+			esc_clk_rate = clk_get_rate(state-
>clks[i].clk);
+	}
+
+	if (!esc_clk_rate)
+		dev_err(state->dev, "Could not get esc clock rate\n");
+
+	dev_dbg(state->dev, "esc clk rate: %u\n", esc_clk_rate);
+	esc_clk_period_ns = 1000000000 / esc_clk_rate;
+
+	min_ths_settle = 85 + 6 * 1000000 / (lane_rate / 1000);
+	max_ths_settle = 140 + 10 * 1000000 / (lane_rate / 1000);
+	ths_settle_ns = (min_ths_settle + max_ths_settle) / 2;
+
+	state->hs_settle = ths_settle_ns / esc_clk_period_ns - 1;
+
+	dev_dbg(state->dev, "lane rate %u Ths_settle %u hs_settle
%u\n",
+		lane_rate, ths_settle_ns, state->hs_settle);
 
 	return 0;