[PATCH] drm/xe/xe_ggtt: No need to use xe_pm_runtime_get_noresume

[Rodrigo Vivi]

On Mon, Apr 29, 2024 at 09:29:15AM -0700, Ashutosh Dixit wrote:
> Switching from xe_device_mem_access_get/put to xe_pm_runtime_get/put
> results in the following WARNING in xe_oa:
> 
> [11614.356168] xe 0000:00:02.0: Missing outer runtime PM protection
> [11614.356187] WARNING: CPU: 1 PID: 13075 at drivers/gpu/drm/xe/xe_pm.c:549 xe_pm_runtime_get_noresume+0x60/0x80 [xe]
> ...
> [11614.356377] Call Trace:
> [11614.356379]  <TASK>
> [11614.356381]  ? __warn+0x7e/0x180
> [11614.356387]  ? xe_pm_runtime_get_noresume+0x60/0x80 [xe]
> [11614.356507]  xe_ggtt_remove_node+0x22/0x80 [xe]
> [11614.356546]  xe_ttm_bo_destroy+0xea/0xf0 [xe]
> [11614.356579]  xe_oa_stream_destroy+0xf7/0x120 [xe]
> [11614.356627]  xe_oa_release+0x35/0xc0 [xe]
> [11614.356673]  __fput+0xa1/0x2d0
> [11614.356679]  __x64_sys_close+0x37/0x80
> [11614.356697]  do_syscall_64+0x6d/0x140
> [11614.356700]  entry_SYSCALL_64_after_hwframe+0x71/0x79
> [11614.356702] RIP: 0033:0x7f2b37314f67
> 
> There seems to be no reason to use xe_pm_runtime_get_noresume in xe_ggtt
> functions. Just use xe_pm_runtime_get.
> 
> Signed-off-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
> ---
>  drivers/gpu/drm/xe/xe_ggtt.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
> index 0d541f55b4fc..8548a2eb3b32 100644
> --- a/drivers/gpu/drm/xe/xe_ggtt.c
> +++ b/drivers/gpu/drm/xe/xe_ggtt.c
> @@ -404,7 +404,7 @@ static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
>  	if (err)
>  		return err;
>  
> -	xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile));
> +	xe_pm_runtime_get(tile_to_xe(ggtt->tile));
>  	mutex_lock(&ggtt->lock);
>  	err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node, bo->size,
>  					  alignment, 0, start, end, 0);
> @@ -433,7 +433,7 @@ int xe_ggtt_insert_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
>  void xe_ggtt_remove_node(struct xe_ggtt *ggtt, struct drm_mm_node *node,
>  			 bool invalidate)
>  {
> -	xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile));
> +	xe_pm_runtime_get(tile_to_xe(ggtt->tile));

we cannot do this as this place gets called from locked places.
This is a deadlock risk.
We need to ensure to have an outer caller of the xe_pm_runtime_get that will
ensure to get the device waked first, then then we continue with the _noresume
variant here that only ensures that we have an extra reference.

These warnings are indeed poping up in multiple places, and this is a good
thing since we killed the mem_access... at least now we know and have a
backtrace of the places that are putting our device at risk of deadlock
and can use this information to now find the right outer place protections.

https://gitlab.freedesktop.org/drm/xe/kernel/issues/1705

>  
>  	mutex_lock(&ggtt->lock);
>  	xe_ggtt_clear(ggtt, node->start, node->size);
> -- 
> 2.41.0
>