Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <f6493f87-479a-7755-914d-5bb7d1560ee3@quicinc.com>
Date:   Tue, 16 Aug 2022 09:31:29 -0700
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Subject: Re: [PATCH v3] drm/msm/dp: check hpd_state before push idle pattern
 at dp_bridge_disable()
Content-Language: en-US
To:     Stephen Boyd <swboyd@chromium.org>,
        Abhinav Kumar <quic_abhinavk@quicinc.com>, <agross@kernel.org>,
        <airlied@linux.ie>, <bjorn.andersson@linaro.org>,
        <daniel@ffwll.ch>, <dianders@chromium.org>,
        <dmitry.baryshkov@linaro.org>, <robdclark@gmail.com>,
        <sean@poorly.run>, <vkoul@kernel.org>
CC:     <quic_aravindh@quicinc.com>, <quic_sbillaka@quicinc.com>,
        <freedreno@lists.freedesktop.org>,
        <dri-devel@lists.freedesktop.org>, <linux-arm-msm@vger.kernel.org>,
        <linux-kernel@vger.kernel.org>
References: <1660159551-13828-1-git-send-email-quic_khsieh@quicinc.com>
 <CAE-0n533SUb3Bg=pR8Fhwo-M5qLWiti4nzLR-rSGVAsrXgEYNQ@mail.gmail.com>
 <dbda8bce-2890-e5e3-4052-073a52eb06a6@quicinc.com>
 <CAE-0n51NyrP8CikcK_3wj4EEsurmmSZ4RY3pLhJJmkY2_8wNZw@mail.gmail.com>
 <0641a116-5b58-4305-bf2d-f53dcb747276@quicinc.com>
 <1e792f49-febf-43bf-d828-8ecf99cbeba3@quicinc.com>
 <CAE-0n50QXiJs=k78Tmd7om28MgWChypwC8LPRzF2jx_qB5+0FQ@mail.gmail.com>
From:   Kuogee Hsieh <quic_khsieh@quicinc.com>
In-Reply-To: <CAE-0n50QXiJs=k78Tmd7om28MgWChypwC8LPRzF2jx_qB5+0FQ@mail.gmail.com>
Content-Type: text/plain; charset="UTF-8"; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


On 8/15/2022 10:08 AM, Stephen Boyd wrote:
> Quoting Kuogee Hsieh (2022-08-11 08:20:01)
>> On 8/10/2022 6:00 PM, Abhinav Kumar wrote:
>>> Even then, you do have a valid point. DRM framework should not have
>>> caused the disable path to happen without an enable.
>>>
>>> I went through the stack mentioned in the issue.
>>>
>>> Lets see this part of the stack:
>>>
>>> dp_ctrl_push_idle+0x40/0x88
>>>   dp_bridge_disable+0x24/0x30
>>>   drm_atomic_bridge_chain_disable+0x90/0xbc
>>>   drm_atomic_helper_commit_modeset_disables+0x198/0x444
>>>   msm_atomic_commit_tail+0x1d0/0x374
>>>
>>> In drm_atomic_helper_commit_modeset_disables(), we call
>>> disable_outputs().
>>>
>>> AFAICT, this is the only place which has a protection to not call the
>>> disable() flow if it was not enabled here:
>>>
>>> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/drm_atomic_helper.c#L1063
>>>
>>>
>>> But that function is only checking crtc_state->active. Thats set by
>>> the usermode:
>>>
>>> https://gitlab.freedesktop.org/drm/msm/-/blob/msm-next/drivers/gpu/drm/drm_atomic_uapi.c#L407
>>>
>>>
>>> Now, if usermode sets that to true and then crashed it can bypass this
>>> check and we will crash in the location kuogee is trying to fix.
> That seems bad, no? We don't want userspace to be able to crash and then
> be able to call the disable path when enable never succeeded.
>
>>>  From the issue mentioned in
>>> https://gitlab.freedesktop.org/drm/msm/-/issues/17, the reporter did
>>> mention the usermode crashed.
>>>
>>> So this is my tentative analysis of whats happening here.
>>>
>>> Ideally yes, we should have been protected by the location mentioned
>>> above in disable_outputs() but looks to me due to the above hypothesis
>>> its getting bypassed.
> Can you fix the problem there? Not fixing it means that every driver out
> there has to develop the same "fix", when it could be fixed in the core
> one time.

The reporter is running GNOME Display Manager (gdm3) instead of chrome.

We can not duplicate this problem locally. Hence we can not confirm this 
is the root cause of this bug or not.

Do you know who is more proper to investigate more into this?

> Ideally drivers are simple. They configure the hardware for what the
> function pointer is asking for. State management and things like that
> should be pushed into the core framework so that we don't have to
> duplicate that multiple times.
>
>>> Thanks
>>>
>>> Abhinav
>>>
>>>
>> Ii sound likes that there is a hole either at user space or drm.
>>
>> But that should not cause dp_bridge_disable() at dp driver to crash.
> Agreed.
>
>> Therefore it is properly to check hdp_state condition at
>> dp_bridge_disable() to prevent it from crashing.
>>
> Disagree. Userspace shouldn't be able to get drm into a wedged state.

not sure for this.

I agree if this is simple driver such as i2c which does not need to 
maintain any states during operation.

but mdp/dp is far more complexity. we really do not want to have any 
crash happen at mdp/dp in the filed.

would you please reconsider our point to add this hdp_state checking 
here to prevent any crash happen.