[PATCH v4] drm/i915/display: Optimize panel power-on wait time

Jani Nikula jani.nikula at intel.com
Tue Jul 29 13:27:26 UTC 2025


On Tue, 29 Jul 2025, Dibin Moolakadan Subrahmanian <dibin.moolakadan.subrahmanian at intel.com> wrote:
> On 29-07-2025 13:47, Jani Nikula wrote:
>> On Mon, 28 Jul 2025, Dibin Moolakadan Subrahmanian<dibin.moolakadan.subrahmanian at intel.com> wrote:
>>>   The current wait_panel_status() uses intel_de_wait() with a long timeout
>>>   (5000ms), which is suboptimal on Xe platforms where the underlying
>>>   xe_mmio_wait32() employs an exponential backoff strategy. This leads
>>>   to unnecessary delays during resume or power-on  when the panel becomes
>>>   ready earlier than the full timeout.
>> It's not about the timeout, it's about the exponentially increasing poll
>> delay.
>>
>>>   This patch replaces intel_de_wait() with read_poll_timeout() +
>>>   intel_de_read() to actively poll the register at given interval and exit
>>>   early when panel is ready, improving resume latency
>> Please do not say "this patch" in commit messages. Just use the
>> imperative "Replace ...".
>>
>> The commit messages is unnecessarily indented with a space.
>
> will correct this.
>
>>
>>> Changes in v2:
>>>   Replaced  two-phase intel_de_wait() with  read_poll_timeout()
>>>   + intel_de_read()
>>>
>>> Changes in v3:
>>>   - Add poll_interval_ms argument  'wait_panel_status' function.
>>>   - Modify 'wait_panel_status' callers with proper poll interval
>>>
>>> Changes in v4:
>>>   - Change 'wait_panel_off' poll interval to 10ms
>>>
>>> Signed-off-by: Dibin Moolakadan Subrahmanian<dibin.moolakadan.subrahmanian at intel.com>
>>> ---
>>>   drivers/gpu/drm/i915/display/intel_pps.c | 41 +++++++++++++++++-------
>>>   1 file changed, 30 insertions(+), 11 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/display/intel_pps.c b/drivers/gpu/drm/i915/display/intel_pps.c
>>> index b64d0b30f5b1..56ef835fc2eb 100644
>>> --- a/drivers/gpu/drm/i915/display/intel_pps.c
>>> +++ b/drivers/gpu/drm/i915/display/intel_pps.c
>>> @@ -22,6 +22,7 @@
>>>   #include "intel_pps.h"
>>>   #include "intel_pps_regs.h"
>>>   #include "intel_quirks.h"
>>> +#include <linux/iopoll.h>
>> Please look at how includes are ordered in every single file in i915.
>
> will correct this.
>
>>
>>>   static void vlv_steal_power_sequencer(struct intel_display *display,
>>>   				      enum pipe pipe);
>>> @@ -600,14 +601,18 @@ void intel_pps_check_power_unlocked(struct intel_dp *intel_dp)
>>>   #define IDLE_CYCLE_MASK		(PP_ON | PP_SEQUENCE_MASK | PP_CYCLE_DELAY_ACTIVE | PP_SEQUENCE_STATE_MASK)
>>>   #define IDLE_CYCLE_VALUE	(0     | PP_SEQUENCE_NONE | 0                     | PP_SEQUENCE_STATE_OFF_IDLE)
>>>   
>>> +#define PANEL_MAXIMUM_ON_TIME_MS		(5000)
>> The name of the macro is misleading. For single-use things, maybe better
>> to just keep the value inline as it were.
>>
>> Side note, the parenthesis are superfluous here.
>
> will correct this.
>
>>> +
>>>   static void intel_pps_verify_state(struct intel_dp *intel_dp);
>>>   
>>>   static void wait_panel_status(struct intel_dp *intel_dp,
>>> -			      u32 mask, u32 value)
>>> +			      u32 mask, u32 value, int poll_interval_ms)
>> Can we not add the extra parameter please? Can we have a meaningful
>> default instead? 10 ms? Is the 1 ms poll interval really required?
>
> Motive behind adding new parameter is to adjust the poll time based on case.
> Currently each call is taking different time to complete as below
> for panel power off time   - 82 ms
> for panel power cycle      - 0.074 ms
> for panel power on         -  327 ms
>
> Making default poll interval 10ms will  increase panel power cycle time  to 10ms

It's still opening a can of worms. The 5 second timeout is relevant for
all cases either, so you'd have to pass *both*. And it's not so simple
anymore.

BR,
Jani.

>
>
>>
>>>   {
>>>   	struct intel_display *display = to_intel_display(intel_dp);
>>>   	struct intel_digital_port *dig_port = dp_to_dig_port(intel_dp);
>>>   	i915_reg_t pp_stat_reg, pp_ctrl_reg;
>>> +	int ret;
>>> +	u32 reg_val;
>> Nitpick, usually just "val".
> will correct this.
>>
>>>   	lockdep_assert_held(&display->pps.mutex);
>>>   
>>> @@ -624,14 +629,27 @@ static void wait_panel_status(struct intel_dp *intel_dp,
>>>   		    intel_de_read(display, pp_stat_reg),
>>>   		    intel_de_read(display, pp_ctrl_reg));
>>>   
>>> -	if (intel_de_wait(display, pp_stat_reg, mask, value, 5000))
>>> -		drm_err(display->drm,
>>> -			"[ENCODER:%d:%s] %s panel status timeout: PP_STATUS: 0x%08x PP_CONTROL: 0x%08x\n",
>>> -			dig_port->base.base.base.id, dig_port->base.base.name,
>>> -			pps_name(intel_dp),
>>> -			intel_de_read(display, pp_stat_reg),
>>> -			intel_de_read(display, pp_ctrl_reg));
>>> +	if (poll_interval_ms <= 0)
>>> +		poll_interval_ms = 1; //if <0 is passed go with 1ms
>> Without the parameter, we could get rid of checks like this.
>>
>> The comment just duplicates what the code already says.
>>
>> Also, we don't use // comments.
> will correct this.
>>
>>> +
>>> +	ret = read_poll_timeout(intel_de_read, reg_val,
>>> +				((reg_val & mask) == value),
>>> +				(poll_interval_ms * 1000),  // poll intervell
>>> +				(PANEL_MAXIMUM_ON_TIME_MS * 1000),  // total timeout (us)
>>> +				true,
>>> +				display, pp_stat_reg);
>> The outer parenthesis in the parameters are superfluous.
>>
>> The comments are useless (and have a typo too).
>>
>>> +
>>> +	if (ret == 0)
>>> +		goto panel_wait_complete;
>> We do use goto in kernel, but primarily for error handling. Please use
>>
>> 	if (ret)
>>
>> here, and the whole drm_err() thing remains unchanged, and doesn't
>> become part of the patch...
>>
> will correct this.
>
>>>   
>>> +	drm_err(display->drm,
>>> +		"dibin [ENCODER:%d:%s] %s panel status timeout: PP_STATUS: 0x%08x PP_CONTROL: 0x%08x\n",
>> ...and it'll be easier to notice you've left your name in the debug
>> logs. Oops.
>>> +		dig_port->base.base.base.id, dig_port->base.base.name,
>>> +		pps_name(intel_dp),
>>> +		intel_de_read(display, pp_stat_reg),
>>> +		intel_de_read(display, pp_ctrl_reg));
>>> +
>>> +panel_wait_complete:
>>>   	drm_dbg_kms(display->drm, "Wait complete\n");
>>>   }
>>>   
>>> @@ -644,7 +662,8 @@ static void wait_panel_on(struct intel_dp *intel_dp)
>>>   		    "[ENCODER:%d:%s] %s wait for panel power on\n",
>>>   		    dig_port->base.base.base.id, dig_port->base.base.name,
>>>   		    pps_name(intel_dp));
>>> -	wait_panel_status(intel_dp, IDLE_ON_MASK, IDLE_ON_VALUE);
>>> +
>>> +	wait_panel_status(intel_dp, IDLE_ON_MASK, IDLE_ON_VALUE, 20);
>>>   }
>>>   
>>>   static void wait_panel_off(struct intel_dp *intel_dp)
>>> @@ -656,7 +675,7 @@ static void wait_panel_off(struct intel_dp *intel_dp)
>>>   		    "[ENCODER:%d:%s] %s wait for panel power off time\n",
>>>   		    dig_port->base.base.base.id, dig_port->base.base.name,
>>>   		    pps_name(intel_dp));
>>> -	wait_panel_status(intel_dp, IDLE_OFF_MASK, IDLE_OFF_VALUE);
>>> +	wait_panel_status(intel_dp, IDLE_OFF_MASK, IDLE_OFF_VALUE, 10);
>>>   }
>>>   
>>>   static void wait_panel_power_cycle(struct intel_dp *intel_dp)
>>> @@ -683,7 +702,7 @@ static void wait_panel_power_cycle(struct intel_dp *intel_dp)
>>>   	if (remaining)
>>>   		wait_remaining_ms_from_jiffies(jiffies, remaining);
>>>   
>>> -	wait_panel_status(intel_dp, IDLE_CYCLE_MASK, IDLE_CYCLE_VALUE);
>>> +	wait_panel_status(intel_dp, IDLE_CYCLE_MASK, IDLE_CYCLE_VALUE, 1);
>>>   }
>>>   
>>>   void intel_pps_wait_power_cycle(struct intel_dp *intel_dp)

-- 
Jani Nikula, Intel


More information about the Intel-gfx mailing list