Re: [PATCH v2] drm/tidss: Fix synclost issue while disabling layer due to erratum i2097

[Yashas D <y-d@ti.com>]

On 07/09/26 18:45, Tomi Valkeinen wrote:
> Hi,
> 
Hi Tomi,
Thanks for the comments.

> On 01/07/2026 09:33, Yashas D wrote:
>> A gap beyond a frame duration for disabling the layer in overlay manager
> 
> I don't think that's right. The key is that the layer is disabled on 
> frame N, and the GO bit is set on frame N+1. So the duration can be less 
> than a frame, if the frame boundary is between disable and GO.
> 

Will reword it.

>> and go bit setting on the corresponding video port may lead to SYNC_LOST.
> 
> I think it's quite guaranteed, not "may" =).
> 

The errata documentation that I have attached says it "may" lead to 
SYNC_LOST. I had tried to force it and it doesnt give SYNC_LOST all the 
time.

>> Fix by moving the layer to the non-visible area and deferring the 
>> ENABLE=0
>> write, immediately before the GO bit, guaranteeing both occur within the
>> same frame window.
> 
> Maybe: "Fix by moving the layer outside the visible area, and deferring 
> the ENABLE=0 write to happen just before the GO bit write, ..."
> 

Will fix the wording.

>> Link: https://www.ti.com/lit/pdf/sprz530 # i2097
>>
>> Signed-off-by: Yashas D <y-d@ti.com>
>> ---
>>
>> changelog v1->v2:
>> - Flushed the pending layers in dispc_vp_disable to prevent hanging 
>> layers in hardware
>>    when dispc_vp_go is not called in CRTC teardown and modeset paths.
>> - Removed the AM62L switch case as its the same as the default case.
>>
>> Link to v1: https://lore.kernel.org/all/20260630103802.1853564-1-y- 
>> d@ti.com/
>>
>>   drivers/gpu/drm/tidss/tidss_dispc.c | 118 +++++++++++++++++++++++++++-
>>   drivers/gpu/drm/tidss/tidss_dispc.h |   1 +
>>   2 files changed, 118 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/tidss/tidss_dispc.c b/drivers/gpu/drm/ 
>> tidss/tidss_dispc.c
>> index 58d5eb033bdb..02a3cf50e0ce 100644
>> --- a/drivers/gpu/drm/tidss/tidss_dispc.c
>> +++ b/drivers/gpu/drm/tidss/tidss_dispc.c
>> @@ -38,6 +38,8 @@
>>   #include "tidss_dispc_regs.h"
>>   #include "tidss_scale_coefs.h"
>> +#define OVR_LAYER_MAX_POS(mask)        FIELD_MAX(mask)
> 
> Maybe just use FIELD_MAX() in the code below.
> 

Will use FIELD_MAX() in the next version.

>> +
>>   static const u16 tidss_k2g_common_regs[DISPC_COMMON_REG_TABLE_LEN] = {
>>       [DSS_REVISION_OFF] =                    0x00,
>>       [DSS_SYSCONFIG_OFF] =                   0x04,
>> @@ -467,6 +469,8 @@ struct dispc_device {
>>       struct clk *fclk;
>> +    unsigned long fclk_rate;
>> +
>>       bool is_enabled;
>>       struct dss_vp_data vp_data[TIDSS_MAX_PORTS];
>> @@ -477,6 +481,9 @@ struct dispc_device {
>>       u32 memory_bandwidth_limit;
>>       struct dispc_errata errata;
>> +
>> +    /* WA for erratum i2097: OVR Layer Disable May Cause Sync Lost. */
>> +    u32 pending_disable_layers[TIDSS_MAX_PORTS];
>>   };
>>   static void dispc_write(struct dispc_device *dispc, u16 reg, u32 val)
>> @@ -1134,6 +1141,12 @@ void dispc_vp_prepare(struct dispc_device 
>> *dispc, u32 hw_videoport,
>>       bool align, onoff, rf, ieo, ipc, ihs, ivs;
>>       u32 hsw, hfp, hbp, vsw, vfp, vbp;
>> +    /*
>> +     * WA for erratum i2097: clear any stale layer disable tracking
>> +     * state left over from the previous VP enable/disable cycle.
>> +     */
>> +    dispc->pending_disable_layers[hw_videoport] = 0;
> 
> When is such stale state left over?
> 

It was used defensively if by chance any stale state is left. But since 
we are also doing it in dispc_vp_disable, it might not be required here. 
Will remove it.

>> +
>>       fmt = dispc_vp_find_bus_fmt(dispc, hw_videoport, tstate- 
>> >bus_format,
>>                       tstate->bus_flags);
>> @@ -1210,6 +1223,23 @@ void dispc_vp_enable(struct dispc_device 
>> *dispc, u32 hw_videoport)
>>   void dispc_vp_disable(struct dispc_device *dispc, u32 hw_videoport)
>>   {
>> +    u32 layer;
>> +
>> +    /*
>> +     * WA for erratum i2097: flush any pending layer disables directly.
>> +     * dispc_vp_go() is not called in the CRTC teardown and modeset 
>> paths
> 
> I think "teardown" is what happens when e.g. one unloads the module. I 
> assume you mean crtc disable here? Or what's "crtc teardown" and 
> "modeset", exactly?
> 

Yes, I meant tidss_crtc_atomic_disable and tidss_crtc_atomic_enable. 
Will fix the wording.

>> +     * so pending_disable_layers would otherwise never be consumed, 
>> leaving
>> +     * layers enabled in hardware. Sync lost does not matter here 
>> since the
> 
> What do you mean does not matter? If it causes sync lost error prints, 
> it matters.
> 

No GO bit write follows a raw ENABLE=0 in dispc_vp_disable(), so no 
frame-boundary event that DSS could flag as SYNC_LOST is generated on 
this path. Will fix the wording.

>> +     * VP is being disabled.
>> +     */
>> +    for (layer = 0; layer < dispc->feat->num_vids; layer++) {
>> +        if (dispc->pending_disable_layers[hw_videoport] & BIT(layer))
>> +            OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                    DISPC_OVR_ATTRIBUTES(layer),
>> +                    0, DISPC_OVR_ATTRIBUTES_ENABLE_MASK);
>> +    }
>> +    dispc->pending_disable_layers[hw_videoport] = 0;
>> +
>>       VP_REG_FLD_MOD(dispc, hw_videoport, DISPC_VP_CONTROL, 0,
>>                  DISPC_VP_CONTROL_ENABLE_MASK);
>>   }
>> @@ -1233,6 +1263,38 @@ void dispc_vp_go(struct dispc_device *dispc, 
>> u32 hw_videoport)
>>   {
>>       WARN_ON(VP_REG_GET(dispc, hw_videoport, DISPC_VP_CONTROL,
>>                  DISPC_VP_CONTROL_GOBIT_MASK));
>> +
>> +    if (dispc->errata.i2097 &&
>> +        dispc->pending_disable_layers[hw_videoport]) {
>> +        u32 layer;
>> +        u32 delay_ns;
>> +
>> +        /* WA for erratum i2097: set GO bit #1 to latch position
> 
> Start a comment block with "/*" line.
> 

Will fix it.

> What do you mean with "set GO bit #1"? Do you just mean writing 1 to GO 
> bit? If so, just "set GO bit to latch..." is enough.
> 
> 

Yes I meant writing 1 to bit. Will use "set GO bit to latch...".

>> +         * changes into the DSS pipeline, wait 10 DSS functional clock
>> +         * cycles, then write ENABLE=0.
>> +         */
>> +        VP_REG_FLD_MOD(dispc, hw_videoport, DISPC_VP_CONTROL, 1,
>> +                   DISPC_VP_CONTROL_GOBIT_MASK);
>> +
>> +        if (dispc->fclk_rate)
>> +            delay_ns = DIV_ROUND_UP_ULL((u64)10 * NSEC_PER_SEC,
>> +                            dispc->fclk_rate);
>> +        else
>> +            delay_ns = 500;
> 
> When do we not have fclk_rate? Shouldn't we always have it?
> 

Added it as a defensive check. Will remove it.


>> +
>> +        ndelay(delay_ns);
>> +
>> +        for (layer = 0; layer < dispc->feat->num_vids; layer++) {
>> +            if (dispc->pending_disable_layers[hw_videoport] &
>> +                BIT(layer))
>> +                OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                        DISPC_OVR_ATTRIBUTES(layer),
>> +                        0,
>> +                        DISPC_OVR_ATTRIBUTES_ENABLE_MASK);
>> +        }
>> +        dispc->pending_disable_layers[hw_videoport] = 0;
> 
> My first thought was if this causes problems if vsync happens somewhere 
> in the middle here, and tidss_crtc does pageflip handling. But I think 
> that can't happen, as this is only called from tidss_crtc, inside a spin 
> lock. If that's the case, I think a comment here somewhere would still 
> be a good thing.

Will add the comment.

> 
>> +    }
>> +
>>       VP_REG_FLD_MOD(dispc, hw_videoport, DISPC_VP_CONTROL, 1,
>>                  DISPC_VP_CONTROL_GOBIT_MASK);
>>   }
>> @@ -1503,6 +1565,53 @@ void dispc_ovr_enable_layer(struct dispc_device 
>> *dispc,
>>       if (dispc->feat->subrev == DISPC_K2G)
>>           return;
>> +    if (dispc->errata.i2097 && !enable) {
>> +        /*
>> +         * WA for erratum i2097:
>> +         *
>> +         * Do not write ENABLE=0 directly. Instead move the layer to
>> +         * the non-visible area so it contributes no pixels.
>> +         *
>> +         * Position register layout differs per SoC:
>> +         *   J721E : DISPC_OVR_ATTRIBUTES2, X[13:0],  Y[29:16] (14-bit)
>> +         *   Others: DISPC_OVR_ATTRIBUTES,  X[17:6],  Y[30:19] (12-bit)
>> +         */
>> +        switch (dispc->feat->subrev) {
>> +        case DISPC_J721E:
>> +            OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                    DISPC_OVR_ATTRIBUTES2(layer),
>> +                    OVR_LAYER_MAX_POS(DISPC_OVR_ATTRIBUTES2_POSX_MASK),
>> +                    DISPC_OVR_ATTRIBUTES2_POSX_MASK);
>> +            OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                    DISPC_OVR_ATTRIBUTES2(layer),
>> +                    OVR_LAYER_MAX_POS(DISPC_OVR_ATTRIBUTES2_POSY_MASK),
>> +                    DISPC_OVR_ATTRIBUTES2_POSY_MASK);
>> +            break;
>> +        default:
>> +            OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                    DISPC_OVR_ATTRIBUTES(layer),
>> +                    OVR_LAYER_MAX_POS(DISPC_OVR_ATTRIBUTES_POSX_MASK),
>> +                    DISPC_OVR_ATTRIBUTES_POSX_MASK);
>> +            OVR_REG_FLD_MOD(dispc, hw_videoport,
>> +                    DISPC_OVR_ATTRIBUTES(layer),
>> +                    OVR_LAYER_MAX_POS(DISPC_OVR_ATTRIBUTES_POSY_MASK),
>> +                    DISPC_OVR_ATTRIBUTES_POSY_MASK);
>> +            break;
>> +        }
> 
> You could probably use dispc_ovr_set_plane() here. It writes one extra 
> register, but I don't see much harm in that, and gets rid of this 
> duplication.
> 

Yes, will use that instead of duplicating again.

>> +
>> +        dispc->pending_disable_layers[hw_videoport] |= BIT(layer);
>> +        return;
>> +    }
>> +
>> +    if (dispc->errata.i2097 && enable) {
>> +        /*
>> +         * Layer being re-enabled: cancel any pending disable so
>> +         * dispc_vp_go() does not write ENABLE=0 after we have
>> +         * just written ENABLE=1 here.
>> +         */
>> +        dispc->pending_disable_layers[hw_videoport] &= ~BIT(layer);
>> +    }
>> +
>>       OVR_REG_FLD_MOD(dispc, hw_videoport, DISPC_OVR_ATTRIBUTES(layer),
>>               !!enable, DISPC_OVR_ATTRIBUTES_ENABLE_MASK);
>>   }
>> @@ -2871,6 +2980,12 @@ static void dispc_init_errata(struct 
>> dispc_device *dispc)
>>           dispc->errata.i2000 = true;
>>           dev_info(dispc->dev, "WA for erratum i2000: YUV formats 
>> disabled\n");
>>       }
>> +
>> +    if (dispc->feat->subrev != DISPC_K2G) {
>> +        dispc->errata.i2097 = true;
>> +        dev_info(dispc->dev,
>> +             "WA for erratum i2097: OVR layer disable uses non- 
>> visible area\n");
> 
> I don't think an dev_info print is required. The other WA info print is 
> there as it's a rather big one, disabling YUV formats.
> 

Will remove it.

>> +    }
>>   }
>>   /*
>> @@ -3059,7 +3174,8 @@ int dispc_init(struct tidss_device *tidss)
>>               __func__, PTR_ERR(dispc->fclk));
>>           return PTR_ERR(dispc->fclk);
>>       }
>> -    dev_dbg(dev, "DSS fclk %lu Hz\n", clk_get_rate(dispc->fclk));
>> +    dispc->fclk_rate = clk_get_rate(dispc->fclk);
>> +    dev_dbg(dev, "DSS fclk %lu Hz\n", dispc->fclk_rate);
>>       of_property_read_u32(dispc->dev->of_node, "max-memory-bandwidth",
>>                    &dispc->memory_bandwidth_limit);
>> diff --git a/drivers/gpu/drm/tidss/tidss_dispc.h b/drivers/gpu/drm/ 
>> tidss/tidss_dispc.h
>> index 739d211d0018..e2ce9fdf1279 100644
>> --- a/drivers/gpu/drm/tidss/tidss_dispc.h
>> +++ b/drivers/gpu/drm/tidss/tidss_dispc.h
>> @@ -57,6 +57,7 @@ struct dispc_vid_info {
>>   struct dispc_errata {
>>       bool i2000; /* DSS Does Not Support YUV Pixel Data Formats */
>> +    bool i2097; /* OVR Layer Disable May Cause Sync Lost */
>>   };
>>   enum dispc_vp_bus_type {
> 

Let me know if any more changes are required, will send a v3 after that.
Thanks,
Yashas D