[PATCH 11/11] drm/amdkfd: inc counter on child ranges with xnack off

[Felix Kuehling]

Am 2021-06-29 um 2:02 p.m. schrieb Alex Sierra:
> During GPU page table invalidation with xnack off, new ranges
> split may occur concurrently in the same prange. Creating a new
> child per split. Each child should also increment its
> invalid counter, to assure GPU page table updates in these
> ranges.
>
> Signed-off-by: Alex Sierra <alex.sierra at amd.com>
> ---
>  drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 8 +++++++-
>  1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 294c9480a184..67605e4bc3c6 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1622,6 +1622,7 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
>  		unsigned long start, unsigned long last)
>  {
>  	struct svm_range_list *svms = prange->svms;
> +	struct svm_range *pchild;
>  	struct kfd_process *p;
>  	int r = 0;
>  
> @@ -1633,6 +1634,12 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
>  	if (!p->xnack_enabled) {
>  		int evicted_ranges;
>  
> +		list_for_each_entry(pchild, &prange->child_list, child_list) {
> +			mutex_lock_nested(&pchild->lock, 1);

You should check the address range of the child range and only increment
pchild->invalid if it intersects with start..last.


> +			atomic_inc(&pchild->invalid);
> +			mutex_unlock(&pchild->lock);
> +		}
> +
>  		atomic_inc(&prange->invalid);

Also add an address range check for the parent range.

Please make this the first patch in the series, or review/submit it
separately from the series. The series is about the reorganization of
VRAM management. This patch is an unrelated bug-fix.

Regards,
  Felix


>  		evicted_ranges = atomic_inc_return(&svms->evicted_ranges);
>  		if (evicted_ranges != 1)
> @@ -1650,7 +1657,6 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm,
>  		schedule_delayed_work(&svms->restore_work,
>  			msecs_to_jiffies(AMDGPU_SVM_RANGE_RESTORE_DELAY_MS));
>  	} else {
> -		struct svm_range *pchild;
>  		unsigned long s, l;
>  
>  		pr_debug("invalidate unmap svms 0x%p [0x%lx 0x%lx] from GPUs\n",