[PATCH 3/3] drm/amdkfd: Split giant svm range

[Felix Kuehling]

Am 2022-07-25 um 08:23 schrieb Philip Yang:
> Giant svm range split to smaller ranges, align the range start address
> to max svm range pages to improve MMU TLB usage.
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 52 +++++++++++++++++++---------
>   1 file changed, 36 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index cf9565ddddf8..044bb99f88ea 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -1885,6 +1885,37 @@ __init void svm_range_set_max_pages(struct amdgpu_device *adev)
>   	max_svm_range_pages = ALIGN(max_svm_range_pages, 1ULL << 9);
>   }
>   
> +static int
> +__svm_range_add(struct svm_range_list *svms, uint64_t start, uint64_t last,
> +	       struct list_head *insert_list, struct list_head *update_list)

It would be nice to find a better name for this. Maybe 
svm_range_split_new. Maybe make the max size a parameter of the function 
for better clarity.


> +{
> +	struct svm_range *prange;
> +	uint64_t l;
> +
> +	pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n",
> +		 max_svm_range_pages, start, last);
> +
> +	while (last >= start) {
> +		if (last - start + 1 > max_svm_range_pages) {

Use a single READ_ONCE in this function to read max_svm_range_pages into 
a local variable. This should avoid race conditions with GPU hotplug. If 
you make the max size a parameter of this function, that also works if 
the caller uses READ_ONCE.


> +			if (start % max_svm_range_pages)
> +				l = ALIGN(start, max_svm_range_pages) - 1;
> +			else
> +				l = start + max_svm_range_pages - 1;
> +		} else {
> +			l = last;

I think this whole if block could be written as

     l = min(last, ALIGN_DOWN(start + max_svm_range_pages, 
max_svm_range_pages) - 1);

Regards,
   Felix


> +		}
> +
> +		prange = svm_range_new(svms, start, l);
> +		if (!prange)
> +			return -ENOMEM;
> +		list_add(&prange->list, insert_list);
> +		list_add(&prange->update_list, update_list);
> +
> +		start = l + 1;
> +	}
> +	return 0;
> +}
> +
>   /**
>    * svm_range_add - add svm range and handle overlap
>    * @p: the range add to this process svms
> @@ -1987,14 +2018,10 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
>   
>   		/* insert a new node if needed */
>   		if (node->start > start) {
> -			prange = svm_range_new(svms, start, node->start - 1);
> -			if (!prange) {
> -				r = -ENOMEM;
> +			r = __svm_range_add(svms, start, node->start - 1,
> +					    insert_list, update_list);
> +			if (r)
>   				goto out;
> -			}
> -
> -			list_add(&prange->list, insert_list);
> -			list_add(&prange->update_list, update_list);
>   		}
>   
>   		node = next;
> @@ -2002,15 +2029,8 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size,
>   	}
>   
>   	/* add a final range at the end if needed */
> -	if (start <= last) {
> -		prange = svm_range_new(svms, start, last);
> -		if (!prange) {
> -			r = -ENOMEM;
> -			goto out;
> -		}
> -		list_add(&prange->list, insert_list);
> -		list_add(&prange->update_list, update_list);
> -	}
> +	if (start <= last)
> +		r = __svm_range_add(svms, start, last, insert_list, update_list);
>   
>   out:
>   	if (r)