[PATCH 3/3] drm/amdkfd: Check bitmap_mapped flag to skip retry fault

[Felix Kuehling]

On 2023-09-29 10:11, Philip Yang wrote:
> Use bitmap_mapped flag to check if range already mapped to the specific
> GPU, to skip the retry fault from different page of the same range.
>
> Remove prange validate_timestamp which is not accurate for multiple
> GPUs.
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>

Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>


> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 24 ++++++++----------------
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.h |  1 -
>   2 files changed, 8 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index ac65bf25c685..5e063d902a46 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -43,10 +43,6 @@
>   
>   #define AMDGPU_SVM_RANGE_RESTORE_DELAY_MS 1
>   
> -/* Long enough to ensure no retry fault comes after svm range is restored and
> - * page table is updated.
> - */
> -#define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING	(2UL * NSEC_PER_MSEC)
>   #if IS_ENABLED(CONFIG_DYNAMIC_DEBUG)
>   #define dynamic_svm_range_dump(svms) \
>   	_dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms)
> @@ -365,7 +361,6 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start,
>   	INIT_LIST_HEAD(&prange->deferred_list);
>   	INIT_LIST_HEAD(&prange->child_list);
>   	atomic_set(&prange->invalid, 0);
> -	prange->validate_timestamp = 0;
>   	mutex_init(&prange->migrate_mutex);
>   	mutex_init(&prange->lock);
>   
> @@ -1876,8 +1871,6 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
>   	}
>   
>   	svm_range_unreserve_bos(ctx);
> -	if (!r)
> -		prange->validate_timestamp = ktime_get_boottime();
>   
>   free_ctx:
>   	kfree(ctx);
> @@ -3162,15 +3155,6 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   		goto out_unlock_range;
>   	}
>   
> -	/* skip duplicate vm fault on different pages of same range */
> -	if (ktime_before(timestamp, ktime_add_ns(prange->validate_timestamp,
> -				AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING))) {
> -		pr_debug("svms 0x%p [0x%lx %lx] already restored\n",
> -			 svms, prange->start, prange->last);
> -		r = 0;
> -		goto out_unlock_range;
> -	}
> -
>   	/* __do_munmap removed VMA, return success as we are handling stale
>   	 * retry fault.
>   	 */
> @@ -3196,6 +3180,14 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   		goto out_unlock_range;
>   	}
>   
> +	/* skip duplicate vm fault on different pages of same range */
> +	if (svm_range_partial_mapped_dev(gpuidx, prange, addr, addr)) {
> +		pr_debug("svms 0x%p [0x%lx %lx] already restored on gpu %d\n",
> +			 svms, prange->start, prange->last, gpuidx);
> +		r = 0;
> +		goto out_unlock_range;
> +	}
> +
>   	pr_debug("svms %p [0x%lx 0x%lx] best restore 0x%x, actual loc 0x%x\n",
>   		 svms, prange->start, prange->last, best_loc,
>   		 prange->actual_loc);
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> index 10c92c5e23a7..3afc33a3dd30 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> @@ -125,7 +125,6 @@ struct svm_range {
>   	uint32_t			actual_loc;
>   	uint8_t				granularity;
>   	atomic_t			invalid;
> -	ktime_t				validate_timestamp;
>   	struct mmu_interval_notifier	notifier;
>   	struct svm_work_list_item	work_item;
>   	struct list_head		deferred_list;