Message-ID: <1a5ed43e-914e-079d-96bf-c9e3912a9473@quicinc.com>
Date:   Tue, 4 Oct 2022 15:31:10 -0700
From:   Abhinav Kumar <quic_abhinavk@...cinc.com>
To:     Marijn Suijten <marijn.suijten@...ainline.org>,
        <phone-devel@...r.kernel.org>, Rob Clark <robdclark@...il.com>,
        Dmitry Baryshkov <dmitry.baryshkov@...aro.org>,
        Vinod Koul <vkoul@...nel.org>,
        <~postmarketos/upstreaming@...ts.sr.ht>,
        AngeloGioacchino Del Regno 
        <angelogioacchino.delregno@...ainline.org>,
        Konrad Dybcio <konrad.dybcio@...ainline.org>,
        Martin Botka <martin.botka@...ainline.org>,
        Jami Kettunen <jami.kettunen@...ainline.org>,
        David Airlie <airlied@...ux.ie>,
        Daniel Vetter <daniel@...ll.ch>, Sean Paul <sean@...rly.run>,
        Thomas Zimmermann <tzimmermann@...e.de>,
        Javier Martinez Canillas <javierm@...hat.com>,
        Alex Deucher <alexander.deucher@....com>,
        Douglas Anderson <dianders@...omium.org>,
        Vladimir Lypak <vladimir.lypak@...il.com>,
        <dri-devel@...ts.freedesktop.org>, <linux-kernel@...r.kernel.org>,
        <linux-arm-msm@...r.kernel.org>, <freedreno@...ts.freedesktop.org>,
        Lyude Paul <lyude@...hat.com>
Subject: Re: [PATCH 5/5] drm/dsc: Prevent negative BPG offsets from shadowing
 adjacent bitfields



On 10/4/2022 2:57 PM, Marijn Suijten wrote:
> On 2022-10-04 13:22:25, Abhinav Kumar wrote:
>>
>> On 10/1/2022 12:08 PM, Marijn Suijten wrote:
>>> msm's dsi_host specifies negative BPG offsets which fill the full 8 bits
>>> of a char thanks to two's complement: this however results in those bits
>>> bleeding into the next parameter when the field is only expected to
>>> contain 6-bit wide values.
>>> As a consequence random slices appear corrupted on-screen (tested on a
>>> Sony Tama Akatsuki device with sdm845).
>>>
>>> Use AND operators to limit all values that constitute the RC Range
>>> parameter fields to their expected size.
>>>
>>> Fixes: b9080324d6ca ("drm/msm/dsi: add support for dsc data")
>>> Signed-off-by: Marijn Suijten <marijn.suijten@...ainline.org>
>>> ---
>>>    drivers/gpu/drm/display/drm_dsc_helper.c | 6 +++---
>>>    1 file changed, 3 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/display/drm_dsc_helper.c b/drivers/gpu/drm/display/drm_dsc_helper.c
>>> index c869c6e51e2b..2e7ef242685d 100644
>>> --- a/drivers/gpu/drm/display/drm_dsc_helper.c
>>> +++ b/drivers/gpu/drm/display/drm_dsc_helper.c
>>> @@ -243,11 +243,11 @@ void drm_dsc_pps_payload_pack(struct drm_dsc_picture_parameter_set *pps_payload,
>>>    	 */
>>>    	for (i = 0; i < DSC_NUM_BUF_RANGES; i++) {
>>>    		pps_payload->rc_range_parameters[i] =
>>> -			cpu_to_be16((dsc_cfg->rc_range_params[i].range_min_qp <<
>>> +			cpu_to_be16(((dsc_cfg->rc_range_params[i].range_min_qp & 0x1f) <<
>>>    				     DSC_PPS_RC_RANGE_MINQP_SHIFT) |
>>> -				    (dsc_cfg->rc_range_params[i].range_max_qp <<
>>> +				    ((dsc_cfg->rc_range_params[i].range_max_qp & 0x1f) <<
>>>    				     DSC_PPS_RC_RANGE_MAXQP_SHIFT) |
>>> -				    (dsc_cfg->rc_range_params[i].range_bpg_offset));
>>> +				    (dsc_cfg->rc_range_params[i].range_bpg_offset & 0x3f));
>>>    	}
>>>    
>>
>> Looking at some examples of this for other vendors, they have managed to
>> limit the value to 6 bits in their drivers:
>>
>> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/i915/display/intel_vdsc.c#L532
>>
>> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/amd/display/dc/dsc/rc_calc_dpi.c#L87
>>
>> Perhaps, msm should do the same thing instead of the helper change.
> 
> Thanks, I should have done my due-diligence and look up how other
> drivers dealt with this, but wasn't immediately expecting negative
> values elsewhere.
> 
> Alas, as explained in the cover letter I opted to perform the masking in
> the PPS packing code as the DSC block code also reads these values, and
> would suddenly write 6-bit intead of 8-bit values to the
> DSC_RANGE_BPG_OFFSET registers.  Quick testing on the mentioned sdm845
> platform shows no regressions, but I'm not sure if that's safe to rely
> on?

I looked up the MDP_DSC_0_RANGE_BPG_OFFSET_* registers.
They take only a 6-bit value according to the SW documentation ( bits 5:0 )

It was always expecting only a 6-bit value and not 8.

So this change is safe.

> 
>> If you want to move to helper, other drivers need to be changed too to
>> remove duplicate & 0x3f.
> 
> Sure, we only have to confirm whether those drivers also read back the
> value(s) in rc_range_params, and expect / allow this to be 8 instead of
> 6 bits.
> 
>> FWIW, this too has already been fixed in the latest downstream driver too.
> 
> What is this supposed to mean?  Is there a downstream DPU project that
> has pending patches needing to be upstreamed?  Or is the downstream SDE,
> techpack/display, or whatever it is called nowadays, slowly using more
> DRM structs like drm_dsc_config and this drm_dsc_pps_payload_pack()
> helper function as pointed out in an earlier mail?
> 

No, what I meant was, the version of downstream driver based on which 
the upstream DSC was made seems to be an older version. Downstream 
drivers keep getting updated and we always keep trying to align with 
upstream structs.

This is true not just for DSC but even other blocks.

So as part of that effort, we started using struct drm_dsc_config . That 
change was made on newer chipsets. But the downstream SW on sdm845 based 
on which the DSC was upstreamed seems like didnt have that. Hence all 
this redundant math happened.

So this comment was more of a explanation about why this issue happened 
even though latest downstream didnt have this issue.

> Offtopic: are SDE and DPU growing closer together, hopefully achieving
> feature parity allowing the SDE project to be dropped in favour of a
> fully upstreamed DPU driver for day-one out-of-the-box mainline support
> for new SoCs (as long as work is published and on its way upstream)?
> 

There is still a lot of gap between SDE and DPU drivers at this point. 
We keep trying to upstream as many features as possible to minimize the 
gap but there is still a lot of work to do.


> - Marijn

Powered by blists - more mailing lists