Re: [PATCH v2 2/6] drm/amdgpu: move devcoredump generation to a worker

["Christian König" <christian.koenig@amd.com>]

On 2/26/26 10:37, Pierre-Eric Pelloux-Prayer wrote:
> Update the way drm_coredump_printer is used based on its documentation
> and Xe's code: the main idea is to generate the final version in one go
> and then use memcpy to return the chunks requested by the caller of
> amdgpu_devcoredump_read.
> 
> The generation is moved to a separate worker thread.
> 
> This cuts the time to copy the dump from 40s to ~0s on my machine.
> 
> Signed-off-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
> Acked-by: Alex Deucher <alexander.deucher@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu.h           |  5 ++
>  .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c  | 66 +++++++++++++++++--
>  .../gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h  |  9 +++
>  3 files changed, 74 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 057c8bd2ad89..ae81a428cfb5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -1200,6 +1200,11 @@ struct amdgpu_device {
>  
>  	struct amdgpu_reset_domain	*reset_domain;
>  
> +#ifdef CONFIG_DEV_COREDUMP
> +	/* If a coredump state capture is in progress don't start a new one. */
> +	bool coredump_in_progress;
> +#endif

Don't keep that as state around, gather it directly from the work item.

> +
>  	struct mutex			benchmark_mutex;
>  
>  	bool                            scpm_enabled;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
> index 42a969512dcc..0808ca98ccd9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.c
> @@ -34,6 +34,8 @@ void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
>  }
>  #else
>  
> +#define AMDGPU_CORE_DUMP_SIZE_MAX (256 * 1024 * 1024)
> +
>  const char *hw_ip_names[MAX_HWIP] = {
>  	[GC_HWIP]		= "GC",
>  	[HDP_HWIP]		= "HDP",
> @@ -196,11 +198,9 @@ static void amdgpu_devcoredump_fw_info(struct amdgpu_device *adev,
>  }
>  
>  static ssize_t
> -amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,
> -			void *data, size_t datalen)
> +amdgpu_devcoredump_format(char *buffer, size_t count, struct amdgpu_coredump_info *coredump)
>  {
>  	struct drm_printer p;
> -	struct amdgpu_coredump_info *coredump = data;
>  	struct drm_print_iterator iter;
>  	struct amdgpu_vm_fault_info *fault_info;
>  	struct amdgpu_ip_block *ip_block;
> @@ -208,7 +208,6 @@ amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,
>  
>  	iter.data = buffer;
>  	iter.offset = 0;
> -	iter.start = offset;
>  	iter.remain = count;
>  
>  	p = drm_coredump_printer(&iter);
> @@ -323,11 +322,60 @@ amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,
>  	return count - iter.remain;
>  }
>  
> +static ssize_t
> +amdgpu_devcoredump_read(char *buffer, loff_t offset, size_t count,
> +			void *data, size_t datalen)
> +{
> +	struct amdgpu_coredump_info *coredump = data;
> +	ssize_t byte_copied;
> +
> +	if (!coredump)
> +		return -ENODEV;
> +
> +	flush_work(&coredump->work);
> +
> +	if (!coredump->formatted)
> +		return -ENODEV;
> +
> +	if (offset >= coredump->formatted_size)
> +		return 0;
> +
> +	byte_copied = count < coredump->formatted_size - offset ? count :
> +		coredump->formatted_size - offset;
> +	memcpy(buffer, coredump->formatted + offset, byte_copied);
> +
> +	return byte_copied;
> +}
> +
>  static void amdgpu_devcoredump_free(void *data)
>  {
> +	struct amdgpu_coredump_info *coredump = data;
> +
> +	cancel_work_sync(&coredump->work);
> +	coredump->adev->coredump_in_progress = false;
> +	kfree(coredump->formatted);
>  	kfree(data);
>  }
>  
> +static void amdgpu_devcoredump_deferred_work(struct work_struct *work)
> +{
> +	struct amdgpu_coredump_info *coredump = container_of(work, typeof(*coredump), work);
> +
> +	dev_coredumpm(coredump->adev->dev, THIS_MODULE, coredump, 0, GFP_NOWAIT,
> +		      amdgpu_devcoredump_read, amdgpu_devcoredump_free);
> +
> +	/* Do a one-time preparation of the coredump output because
> +	 * repeatingly calling drm_coredump_printer is very slow.
> +	 */
> +	coredump->formatted_size =
> +		amdgpu_devcoredump_format(NULL, AMDGPU_CORE_DUMP_SIZE_MAX, coredump);
> +	coredump->formatted = kvzalloc(coredump->formatted_size, GFP_KERNEL);
> +	if (!coredump->formatted)
> +		return;
> +	amdgpu_devcoredump_format(coredump->formatted, coredump->formatted_size, coredump);
> +	coredump->adev->coredump_in_progress = false;
> +}
> +
>  void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
>  		     bool vram_lost, struct amdgpu_job *job)
>  {
> @@ -335,10 +383,15 @@ void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
>  	struct amdgpu_coredump_info *coredump;
>  	struct drm_sched_job *s_job;
>  
> +	if (adev->coredump_in_progress)
> +		return;
> +
>  	coredump = kzalloc(sizeof(*coredump), GFP_NOWAIT);
>  	if (!coredump)
>  		return;
>  
> +	adev->coredump_in_progress = true;

That is racy, clearly don't do that.

Regards,
Christian.

> +
>  	coredump->skip_vram_check = skip_vram_check;
>  	coredump->reset_vram_lost = vram_lost;
>  
> @@ -361,8 +414,9 @@ void amdgpu_coredump(struct amdgpu_device *adev, bool skip_vram_check,
>  
>  	ktime_get_ts64(&coredump->reset_time);
>  
> -	dev_coredumpm(dev->dev, THIS_MODULE, coredump, 0, GFP_NOWAIT,
> -		      amdgpu_devcoredump_read, amdgpu_devcoredump_free);
> +	/* Kick off coredump formatting to a worker thread. */
> +	INIT_WORK(&coredump->work, amdgpu_devcoredump_deferred_work);
> +	queue_work(system_unbound_wq, &coredump->work);
>  
>  	drm_info(dev, "AMDGPU device coredump file has been created\n");
>  	drm_info(dev, "Check your /sys/class/drm/card%d/device/devcoredump/data\n",
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
> index ef9772c6bcc9..4c37a852b74a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dev_coredump.h
> @@ -35,9 +35,18 @@ struct amdgpu_coredump_info {
>  	struct amdgpu_device            *adev;
>  	struct amdgpu_task_info         reset_task_info;
>  	struct timespec64               reset_time;
> +
> +	struct work_struct work;
> +
>  	bool                            skip_vram_check;
>  	bool                            reset_vram_lost;
>  	struct amdgpu_ring              *ring;
> +	/* Readable form of coredevdump, generate once to speed up
> +	 * reading it (see drm_coredump_printer's documentation).
> +	 */
> +	ssize_t				formatted_size;
> +	char				*formatted;
> +
>  };
>  #endif
>