Re: [PATCH v5 1/2] drm/i915/xe3lpd: Power request asserting/deasserting

[Jani Nikula <jani.nikula@linux.intel.com>]

On Tue, 05 Nov 2024, Mika Kahola <mika.kahola@intel.com> wrote:
> There is a HW issue that arises when there are race conditions
> between TCSS entering/exiting TC7 or TC10 states while the
> driver is asserting/deasserting TCSS power request. As a
> workaround, Display driver will implement a mailbox sequence
> to ensure that the TCSS is in TC0 when TCSS power request is
> asserted/deasserted.
>
> The sequence is the following
>
> 1. Read mailbox command status and wait until run/busy bit is
>    clear
> 2. Write mailbox data value '1' for power request asserting
>    and '0' for power request deasserting
> 3. Write mailbox command run/busy bit and command value with 0x1
> 4. Read mailbox command and wait until run/busy bit is clear
>    before continuing power request.
>
> v2: Rename WA function (Gustavo)
>     Limit WA only for PTL platform with a TODO note (Gustavo)
>     Add TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY for clarity when writing
>     register data (Gustavo)
>     Move register defs from i915_reg.h to intel_cx0_phy_regs.h (Gustavo)
> v3: Use "struct intel_display" instead of "struct drm_i915_private" (Jani)
>     Move defs above C10 definitions in the
>     intel_cx0_phy_regs.h file (Gustavo)
>     Move drm_WARN_ON() inside WA function (Gustavo)
>     Rename workaround function as wa_14020908590() (Gustvo)
>     Use boolean enable instead of if-else structure (Raag)
> v4: Drop drm_WARN_ON() (Raag)
>     Fix function definition to fit into a single line (Raag)
>
> Reviewed-by: Raag Jadav <raag.jadav@intel.com>
> Signed-off-by: Mika Kahola <mika.kahola@intel.com>
> ---
>  .../gpu/drm/i915/display/intel_cx0_phy_regs.h |  8 +++++
>  drivers/gpu/drm/i915/display/intel_tc.c       | 31 +++++++++++++++++++
>  2 files changed, 39 insertions(+)
>
> diff --git a/drivers/gpu/drm/i915/display/intel_cx0_phy_regs.h b/drivers/gpu/drm/i915/display/intel_cx0_phy_regs.h
> index f0e5c196eae4..5a0b55cca4a3 100644
> --- a/drivers/gpu/drm/i915/display/intel_cx0_phy_regs.h
> +++ b/drivers/gpu/drm/i915/display/intel_cx0_phy_regs.h
> @@ -200,6 +200,14 @@
>  #define   XELPDP_SSC_ENABLE_PLLA			REG_BIT(1)
>  #define   XELPDP_SSC_ENABLE_PLLB			REG_BIT(0)
>  
> +#define TCSS_DISP_MAILBOX_IN_CMD		_MMIO(0x161300)
> +#define   TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY	REG_BIT(31)
> +#define   TCSS_DISP_MAILBOX_IN_CMD_CMD_MASK	REG_GENMASK(7, 0)
> +#define   TCSS_DISP_MAILBOX_IN_CMD_DATA(val)	(TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY | \

Why does this contain TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY? You set it
separately anyway (and that's how it should be).

> +						REG_FIELD_PREP(TCSS_DISP_MAILBOX_IN_CMD_CMD_MASK, val))
> +
> +#define TCSS_DISP_MAILBOX_IN_DATA		_MMIO(0x161304)
> +
>  /* C10 Vendor Registers */
>  #define PHY_C10_VDR_PLL(idx)		(0xC00 + (idx))
>  #define   C10_PLL0_FRACEN		REG_BIT8(4)
> diff --git a/drivers/gpu/drm/i915/display/intel_tc.c b/drivers/gpu/drm/i915/display/intel_tc.c
> index b16c4d2d4077..e40d55f4c0c4 100644
> --- a/drivers/gpu/drm/i915/display/intel_tc.c
> +++ b/drivers/gpu/drm/i915/display/intel_tc.c
> @@ -1013,6 +1013,30 @@ xelpdp_tc_phy_wait_for_tcss_power(struct intel_tc_port *tc, bool enabled)
>  	return true;
>  }
>  
> +static void wa_14020908590(struct intel_display *display, bool enable)

Yeah I still don't like functions named wa_14020908590. It's
meaningless. What does it do?

> +{
> +	/* check if mailbox is running busy */
> +	if (intel_de_wait_for_clear(display, TCSS_DISP_MAILBOX_IN_CMD,
> +				    TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY, 10)) {
> +		drm_dbg_kms(display->drm,
> +			    "Timeout waiting for TCSS mailbox run/busy bit to clear\n");
> +		return;
> +	}
> +
> +	intel_de_write(display, TCSS_DISP_MAILBOX_IN_DATA, enable);

Not a fan of bool -> u32 implicit conversion here, with the register
contents not described.

> +	intel_de_write(display, TCSS_DISP_MAILBOX_IN_CMD,
> +		       TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY |
> +		       TCSS_DISP_MAILBOX_IN_CMD_DATA(0x1));
> +
> +	/* wait to clear mailbox running busy bit before continuing */
> +	if (intel_de_wait_for_clear(display, TCSS_DISP_MAILBOX_IN_CMD,
> +				    TCSS_DISP_MAILBOX_IN_CMD_RUN_BUSY, 10)) {
> +		drm_dbg_kms(display->drm,
> +			    "Timeout after writing data to mailbox. Mailbox run/busy bit did not clear\n");
> +		return;
> +	}
> +}
> +
>  static void __xelpdp_tc_phy_enable_tcss_power(struct intel_tc_port *tc, bool enable)
>  {
>  	struct drm_i915_private *i915 = tc_to_i915(tc);
> @@ -1022,6 +1046,13 @@ static void __xelpdp_tc_phy_enable_tcss_power(struct intel_tc_port *tc, bool ena
>  
>  	assert_tc_cold_blocked(tc);
>  
> +	/*
> +	 * Gfx driver WA 14020908590 for PTL tcss_rxdetect_clkswb_req/ack
> +	 * handshake violation when pwwreq= 0->1 during TC7/10 entry
> +	 */
> +	if (DISPLAY_VER(i915) == 30)
> +		wa_14020908590(&i915->display, enable);

You should add

	struct intel_display *display = &i915->display;

local variable already in this patch, so the next patch doesn't have to
modify the above line again. You can do the subsequent conversions in
the follow-up.

BR,
Jani.


> +
>  	val = intel_de_read(i915, reg);
>  	if (enable)
>  		val |= XELPDP_TCSS_POWER_REQUEST;

-- 
Jani Nikula, Intel