Re: [RFC PATCH] clk: core: refine disable unused clocks

[Jerome Brunet <jbrunet@baylibre.com>]

On Fri 08 Nov 2024 at 15:59, Chuan Liu <chuan.liu@amlogic.com> wrote:

> hi Jerome:
>
>     Tranks for your REF. I looked at your patch and there are some parts
> that I don't quite understand: The original intention of
> CLK_OPS_PARENT_ENABLE was to solve the issue of "parents need enable
> _during _gate/ungate, set rate and re-parent" when setting a clock. After
> setting the clock, it can still be disabled. However, from what I see in
> your patch, the handling logic seems more like "parents need _always _ gate
> during clock gate period"?

As explained in the description, the problem with CLK_IGNORE_UNUSED and
CLK_OPS_PARENT_ENABLE is that you'll get cycle of enable/disable, which
will disable any parent clock that may have a been enabled and expected
to be ignored.

IOW, the CCF changes the state of the tree while inspecting it.
This change solves that.

>
> On 10/4/2024 9:39 PM, Jerome Brunet wrote:
>> [ EXTERNAL EMAIL ]
>>
>> As it as been pointed out numerous times, flagging a clock with
>> CLK_IGNORE_UNUSED does _not_ guarantee that clock left enabled will stay
>> on. The clock will get disabled if any enable/disable cycle happens on it
>> or its parent clocks.
>>
>> Because enable/disable cycles will disable unused clocks,
>> clk_disable_unused() should not trigger such cycle. Doing so disregard
>> the flag if set for any parent clocks. This is problematic with
>> CLK_OPS_PARENT_ENABLE handling.
>>
>> To solve this, and a couple other issues, pass the parent status to the
>> child while walking the subtree, and return whether child ignored disable,
>> or not.
>>
>> * Knowing the parent status allows to safely disable clocks with
>>    CLK_OPS_PARENT_ENABLE when the parent is enabled. Otherwise it means
>>    that, while the clock is not gated it is effectively disabled. Turning on
>>    the parents to sanitize the sitation would bring back our initial
>>    problem, so just let it sanitize itself when the clock gets used.
>>
>> * If a clock is not actively used (enabled_count == 0), does not have
>>    CLK_IGNORE_UNUSED but the hw enabled all the way to the root clock, and a
>>    child ignored the disable, it should ignore the disable too. Doing so
>>    avoids disabling what is feading the children. Let the flag trickle down
>>    the tree. This has the added benefit to transfer the information to the
>>    unprepare path, so we don't unprepare the parent of a clock that ignored
>>    a disable.
>>
>> * An enabled clock must be prepared in CCF but we can't rely solely on
>>    counts at clk_disable_unused() stage. Make sure an enabled clock is
>>    considered prepared too, even if does not implement the related callback.
>>    Also make sure only disabled clocks get unprepared.
>>
>> Signed-off-by: Jerome Brunet <jbrunet@baylibre.com>
>> ---
>>
>>   This is sent as an RFC to continue the discussion started by Chuan.
>>   It is not meant to be applied as it is.
>>
>>
>>   drivers/clk/clk.c | 92 ++++++++++++++++++++++++++++++-----------------
>>   1 file changed, 60 insertions(+), 32 deletions(-)
>>
>> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
>> index d02451f951cf..41c4504a41f1 100644
>> --- a/drivers/clk/clk.c
>> +++ b/drivers/clk/clk.c
>> @@ -332,17 +332,6 @@ static bool clk_core_is_enabled(struct clk_core *core)
>>                  }
>>          }
>>
>> -       /*
>> -        * This could be called with the enable lock held, or from atomic
>> -        * context. If the parent isn't enabled already, we can't do
>> -        * anything here. We can also assume this clock isn't enabled.
>> -        */
>> -       if ((core->flags & CLK_OPS_PARENT_ENABLE) && core->parent)
>
> This judgment of CLK_OPS_PARENT_ENABLE seems redundant. According to
> normal logic, if the parent is disabled, its children will also be
> forced to disable. This seems unrelated to whether CLK_OPS_PARENT_ENABLE
> is configured.😳

It's removed.

>
>> -               if (!clk_core_is_enabled(core->parent)) {
>> -                       ret = false;
>> -                       goto done;
>> -               }
>> -
>>          ret = core->ops->is_enabled(core->hw);
>>   done:
>>          if (core->rpm_enabled)
>> @@ -1454,22 +1443,39 @@ static void clk_core_disable_unprepare(struct clk_core *core)
>>          clk_core_unprepare_lock(core);
>>   }
>>
>> -static void __init clk_unprepare_unused_subtree(struct clk_core *core)
>> +static bool __init clk_unprepare_unused_subtree(struct clk_core *core,
>> +                                               bool parent_prepared)
>>   {
>>          struct clk_core *child;
>> +       bool prepared;
>>
>>          lockdep_assert_held(&prepare_lock);
>>
>> +       /*
>> +        * Relying on count is not possible at this stage, so consider
>> +        * prepared an enabled clock, in case only .is_enabled() is
>> +        * implemented
>> +        */
>> +       if (parent_prepared)
>> +               prepared = (clk_core_is_prepared(core) ||
>> +                           clk_core_is_enabled(core));
>> +       else
>> +               prepared = false;
>> +
>>          hlist_for_each_entry(child, &core->children, child_node)
>> -               clk_unprepare_unused_subtree(child);
>> +               if (clk_unprepare_unused_subtree(child, prepared) &&
>> +                   prepared && !core->prepare_count)
>> +                       core->flags |= CLK_IGNORE_UNUSED;
>>
>> -       if (core->prepare_count)
>> -               return;
>> +       if (core->flags & CLK_IGNORE_UNUSED || core->prepare_count)
>> +               goto out;
>>
>> -       if (core->flags & CLK_IGNORE_UNUSED)
>> -               return;
>> +       if (!parent_prepared && (core->flags & CLK_OPS_PARENT_ENABLE))
>> +               goto out;
>>
>> -       if (clk_core_is_prepared(core)) {
>> +       /* Do not unprepare an enabled clock */
>> +       if (clk_core_is_prepared(core) &&
>> +           !clk_core_is_enabled(core)) {
>>                  trace_clk_unprepare(core);
>>                  if (core->ops->unprepare_unused)
>>                          core->ops->unprepare_unused(core->hw);
>> @@ -1477,27 +1483,50 @@ static void __init clk_unprepare_unused_subtree(struct clk_core *core)
>>                          core->ops->unprepare(core->hw);
>>                  trace_clk_unprepare_complete(core);
>>          }
>> +
>> +out:
>> +       return (core->flags & CLK_IGNORE_UNUSED) && prepared;
>>   }
>>
>> -static void __init clk_disable_unused_subtree(struct clk_core *core)
>> +static bool __init clk_disable_unused_subtree(struct clk_core *core,
>> +                                             bool parent_enabled)
>>   {
>>          struct clk_core *child;
>>          unsigned long flags;
>> +       bool enabled;
>>
>>          lockdep_assert_held(&prepare_lock);
>>
>> -       hlist_for_each_entry(child, &core->children, child_node)
>> -               clk_disable_unused_subtree(child);
>> +       flags = clk_enable_lock();
>>
>> -       if (core->flags & CLK_OPS_PARENT_ENABLE)
>> -               clk_core_prepare_enable(core->parent);
>> +       /* Check if the clock is enabled from root to this clock */
>> +       if (parent_enabled)
>> +               enabled = clk_core_is_enabled(core);
>> +       else
>> +               enabled = false;
>>
>> -       flags = clk_enable_lock();
>> +       hlist_for_each_entry(child, &core->children, child_node)
>> +               /*
>> +                * If any child ignored disable, this clock should too,
>> +                * unless there is, valid reason for the clock to be enabled
>> +                */
>> +               if (clk_disable_unused_subtree(child, enabled) &&
>> +                   enabled && !core->enable_count)
>> +                       core->flags |= CLK_IGNORE_UNUSED;
>>
>> -       if (core->enable_count)
>> +       if (core->flags & CLK_IGNORE_UNUSED || core->enable_count)
>>                  goto unlock_out;
>>
>> -       if (core->flags & CLK_IGNORE_UNUSED)
>> +       /*
>> +        * If the parent is disabled but the gate is open, we should sanitize
>> +        * the situation. This will avoid an unexpected enable of the clock as
>> +        * soon as the parent is enabled, without control of CCF.
>> +        *
>> +        * Doing so is not possible with a CLK_OPS_PARENT_ENABLE clock without
>> +        * forcefully enabling a whole part of the subtree.  Just let the
>> +        * situation resolve it self on the first enable of the clock
>> +        */
>> +       if (!parent_enabled && (core->flags & CLK_OPS_PARENT_ENABLE))
>>                  goto unlock_out;
>>
>>          /*
>> @@ -1516,8 +1545,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core)
>>
>>   unlock_out:
>>          clk_enable_unlock(flags);
>> -       if (core->flags & CLK_OPS_PARENT_ENABLE)
>> -               clk_core_disable_unprepare(core->parent);
>> +       return (core->flags & CLK_IGNORE_UNUSED) && enabled;
>>   }
>>
>>   static bool clk_ignore_unused __initdata;
>> @@ -1550,16 +1578,16 @@ static int __init clk_disable_unused(void)
>>          clk_prepare_lock();
>>
>>          hlist_for_each_entry(core, &clk_root_list, child_node)
>> -               clk_disable_unused_subtree(core);
>> +               clk_disable_unused_subtree(core, true);
>>
>>          hlist_for_each_entry(core, &clk_orphan_list, child_node)
>> -               clk_disable_unused_subtree(core);
>> +               clk_disable_unused_subtree(core, true);
>>
>>          hlist_for_each_entry(core, &clk_root_list, child_node)
>> -               clk_unprepare_unused_subtree(core);
>> +               clk_unprepare_unused_subtree(core, true);
>>
>>          hlist_for_each_entry(core, &clk_orphan_list, child_node)
>> -               clk_unprepare_unused_subtree(core);
>> +               clk_unprepare_unused_subtree(core, true);
>>
>>          clk_prepare_unlock();
>>
>> --
>> 2.45.2
>>

-- 
Jerome