[v1] drm/panthor: dummy read to discard the stale first GPU_TIMESTAMP read

[PATCH] drm/panthor: dummy read to discard the stale first GPU_TIMESTAMP read

Posted by ryan.zhang@oss.nxp.com 3 days, 8 hours ago

From: Ryan Zhang <ryan.zhang@oss.nxp.com>

The first read of GPU_TIMESTAMP after the GPU has been idle can return a
value that lags the real counter, while a read issued immediately after is
accurate.

Userspace correlates this timestamp with the host clock, so a stale value
shows up as skew between the device and host time domains and makes
dEQP-VK.pipeline.monolithic.timestamp.calibrated.calibration_test fail
whenever a query happens to hit a large first-read error.

Signed-off-by: Ryan Zhang <ryan.zhang@oss.nxp.com>
---
 drivers/gpu/drm/panthor/panthor_drv.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/drivers/gpu/drm/panthor/panthor_drv.c b/drivers/gpu/drm/panthor/panthor_drv.c
index 720bf889096d..956a147dc282 100644
--- a/drivers/gpu/drm/panthor/panthor_drv.c
+++ b/drivers/gpu/drm/panthor/panthor_drv.c
@@ -2,6 +2,7 @@
 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */
 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */
 /* Copyright 2019 Collabora ltd. */
+/* Copyright 2026 NXP */
 
 #ifdef CONFIG_ARM_ARCH_TIMER
 #include <asm/arch_timer.h>
@@ -772,6 +773,10 @@ static int panthor_query_timestamp_info(struct panthor_device *ptdev,
 #else
 	arg->timestamp_frequency = 0;
 #endif
+
+	/* Dummy read to avoid stale values. */
+	gpu_read64_counter(ptdev, GPU_TIMESTAMP);
+
 	arg->current_timestamp = gpu_read64_counter(ptdev, GPU_TIMESTAMP);
 	arg->timestamp_offset = gpu_read64(ptdev, GPU_TIMESTAMP_OFFSET);
 
-- 
2.43.0

Re: [PATCH] drm/panthor: dummy read to discard the stale first GPU_TIMESTAMP read

Posted by Liviu Dudau 3 days, 2 hours ago

On Mon, Sep 21, 2026 at 06:29:57PM +0800, ryan.zhang@oss.nxp.com wrote:
> From: Ryan Zhang <ryan.zhang@oss.nxp.com>
> 
> The first read of GPU_TIMESTAMP after the GPU has been idle can return a
> value that lags the real counter, while a read issued immediately after is
> accurate.
> 
> Userspace correlates this timestamp with the host clock, so a stale value
> shows up as skew between the device and host time domains and makes
> dEQP-VK.pipeline.monolithic.timestamp.calibrated.calibration_test fail
> whenever a query happens to hit a large first-read error.
> 
> Signed-off-by: Ryan Zhang <ryan.zhang@oss.nxp.com>

Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>

Best regards,
Liviu

> ---
>  drivers/gpu/drm/panthor/panthor_drv.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/drivers/gpu/drm/panthor/panthor_drv.c b/drivers/gpu/drm/panthor/panthor_drv.c
> index 720bf889096d..956a147dc282 100644
> --- a/drivers/gpu/drm/panthor/panthor_drv.c
> +++ b/drivers/gpu/drm/panthor/panthor_drv.c
> @@ -2,6 +2,7 @@
>  /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */
>  /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */
>  /* Copyright 2019 Collabora ltd. */
> +/* Copyright 2026 NXP */
>  
>  #ifdef CONFIG_ARM_ARCH_TIMER
>  #include <asm/arch_timer.h>
> @@ -772,6 +773,10 @@ static int panthor_query_timestamp_info(struct panthor_device *ptdev,
>  #else
>  	arg->timestamp_frequency = 0;
>  #endif
> +
> +	/* Dummy read to avoid stale values. */
> +	gpu_read64_counter(ptdev, GPU_TIMESTAMP);
> +
>  	arg->current_timestamp = gpu_read64_counter(ptdev, GPU_TIMESTAMP);
>  	arg->timestamp_offset = gpu_read64(ptdev, GPU_TIMESTAMP_OFFSET);
>  
> -- 
> 2.43.0
> 

-- 
====================
| I would like to |
| fix the world,  |
| but they're not |
| giving me the   |
 \ source code!  /
  ---------------
    ¯\_(ツ)_/¯

[PATCH v2] drm/panthor: dummy read to discard the stale first GPU_TIMESTAMP read

Posted by ryan.zhang@oss.nxp.com 1 day, 8 hours ago

From: Ryan Zhang <ryan.zhang@oss.nxp.com>

The first read of GPU_TIMESTAMP after the GPU has been idle can return a
value that lags the real counter, while a read issued immediately after is
accurate.

Userspace correlates this timestamp with the host clock, so a stale value
shows up as skew between the device and host time domains and makes
dEQP-VK.pipeline.monolithic.timestamp.calibrated.calibration_test fail
whenever a query happens to hit a large first-read error.

Signed-off-by: Ryan Zhang <ryan.zhang@oss.nxp.com>
Reviewed-by: Liviu Dudau <liviu.dudau@arm.com>
---
v2:
 - Added the r-b in commit msg.
 - Rebased on drm-misc-next to resolve conflict.
 - Moved the dummy read into panthor_gpu_get_timestamp() to align with
   the refactored timestamp query code on drm-misc-next.
Note: The original v1 patch was based on v6.18 and applied cleanly there.
Could we also port v1 into 6.18?

 drivers/gpu/drm/panthor/panthor_gpu.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/gpu/drm/panthor/panthor_gpu.c b/drivers/gpu/drm/panthor/panthor_gpu.c
index a383b04f101e..b8df5663f04e 100644
--- a/drivers/gpu/drm/panthor/panthor_gpu.c
+++ b/drivers/gpu/drm/panthor/panthor_gpu.c
@@ -2,6 +2,7 @@
 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */
 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */
 /* Copyright 2019 Collabora ltd. */
+/* Copyright 2026 NXP */
 
 #include <linux/bitfield.h>
 #include <linux/bitmap.h>
@@ -470,6 +471,9 @@ void panthor_gpu_resume(struct panthor_device *ptdev)
 
 u64 panthor_gpu_get_timestamp(struct panthor_device *ptdev)
 {
+	/* Dummy read to avoid stale values. */
+	gpu_read64_counter(ptdev->gpu->iomem, GPU_TIMESTAMP);
+
 	return gpu_read64_counter(ptdev->gpu->iomem, GPU_TIMESTAMP);
 }
 
-- 
2.43.0