Re: [PATCH v6 4/9] media: mtk-mdp: don't pm_run_time_get/put for master comp in clock_on

[houlong wei <houlong.wei@mediatek.com>]

On Mon, 2021-08-02 at 20:12 +0800, Eizan Miyamoto wrote:
> The original intent of commit 86698b9505bbc ("media: mtk-mdp: convert
> mtk_mdp_dev.comp array to list") was to create a list to track all
> the
> MDP components that needed to have their clocks enabled/disabled when
> calling mtk_mdp_comp_clock_on/off. However, there was a bug inside
> mtk_mdp_register_component where the args to a call to list_add were
> swapped. The result is that only one component was added to
> mtk_mdp_dev.comp_list because comp_list was instead being
> repeatedly added to the single element lists headed by each
> mtk_mdp_comp.
> 
> The order of the args to list_add in mtk_mdp_register_component was
> fixed in https://patchwork.kernel.org/patch/11742895/ (Fix Null
> pointer
> dereference when calling list_add).
> 
> Then, as a result of https://patchwork.kernel.org/patch/11530769/
> (mtk-mdp: use pm_runtime in MDP component driver) if all the
> components
> are added to the component list, the mdp "master" / rdma0 component
> ends up having pm_runtime_get_sync() called on it twice recursively:
> 
>     rpm_resume+0x694/0x8f8
>     __pm_runtime_resume+0x7c/0xa0 ***NOTE***
>     mtk_mdp_comp_clock_on+0x48/0x104 [mtk_mdp]
>     mtk_mdp_pm_resume+0x2c/0x44 [mtk_mdp]
>     pm_generic_runtime_resume+0x34/0x48
>     __genpd_runtime_resume+0x6c/0x80
>     genpd_runtime_resume+0x104/0x1ac
>     __rpm_callback+0x120/0x238
>     rpm_callback+0x34/0x8c
>     rpm_resume+0x7a0/0x8f8
>     __pm_runtime_resume+0x7c/0xa0 ***NOTE***
>     mtk_mdp_m2m_start_streaming+0x2c/0x3c [mtk_mdp]
> 
> (The calls to pm_runtime_get_sync are inlined and correspond to the
> calls to __pm_runtime_resume). It is not correct to have
> pm_runtime_get_sync called recursively and the second call will block
> indefinitely.
> 
> As a result of all that, this change factors
> mtk_mdp_comp_clock_on/off
> into mtk_mdp_comp_power_on/off and moves the calls to
> pm_runtime_get/put into the power_on/off functions.
> 
> This change then special-cases the master/rdma0 MDP component and
> does
> these things:
> - the master/rdma0 component is not added to mtk_mdp_dev.comp_list
> - the master/rdma0 component has its clocks (*but not power*) toggled
>   by mtk_mpd_comp_clock_on/off inside mtk_mdp_clock_on/off.
> - the other components have their clocks *and* power toggled with
>   mtk_mdp_comp_power_on/off.
> 
> This change introduces the assumption that mtk_mdp_pm_resume will
> always be called though a callback from pm_runtime_get_sync made on
> the
> master / rdma0 component.
> 
> Signed-off-by: Eizan Miyamoto <eizan@chromium.org>
> ---
> 

Reviewed-by: Houlong Wei <houlong.wei@mediatek.com>

> (no changes since v1)
> 
[...]