Skip to contentSkip to conversation

drm/amdkfd: avoid svm dump when dynamic debug disabled

PATCH In review Alex Sierra 2 participants 2 messages Jul 8, 2023 – Jul 19, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdkfd: avoid svm dump when dynamic debug disabled

svm_range_debug_dump should not be called at all when dynamic debug is disabled to avoid iterating over SVM lists. This could drop performance, specially with big number of SVM ranges.

---
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +4 -2
index 479c4f66afa7..4fb427fc5942 100644
@@ -821,7 +821,7 @@ svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
  *
  * Context: The caller must hold svms->lock
  */
-static void svm_range_debug_dump(struct svm_range_list *svms)
+static int svm_range_debug_dump(struct svm_range_list *svms)
 {
 	struct interval_tree_node *node;
 	struct svm_range *prange;
@@ -847,6 +847,8 @@ static void svm_range_debug_dump(struct svm_range_list *svms)
 			 prange->actual_loc);
 		node = interval_tree_iter_next(node, 0, ~0ULL);
 	}
+
+	return 0;
 }
 
 static int
@@ -3563,7 +3565,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
 			break;
 	}
 
-	svm_range_debug_dump(svms);
+	pr_debug("%d", svm_range_debug_dump(svms));
 
 	mutex_unlock(&svms->lock);
 	mmap_read_unlock(mm);
Signature · 1 line
2.32.0
maintainer
#1 lore raw

Am 2023-07-08 um 12:57 schrieb Alex Sierra:

Quoted 33 lines

svm_range_debug_dump should not be called at all when dynamic debug is disabled to avoid iterating over SVM lists. This could drop performance, specially with big number of SVM ranges.

---
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3 -1
index 479c4f66afa7..4fb427fc5942 100644
@@ -821,7 +821,7 @@ svm_range_is_same_attrs(struct kfd_process *p, struct svm_range *prange,
   *
   * Context: The caller must hold svms->lock
   */
-static void svm_range_debug_dump(struct svm_range_list *svms)
+static int svm_range_debug_dump(struct svm_range_list *svms)
  {
  	struct interval_tree_node *node;
  	struct svm_range *prange;
@@ -847,6 +847,8 @@ static void svm_range_debug_dump(struct svm_range_list *svms)
  			 prange->actual_loc);
  		node = interval_tree_iter_next(node, 0, ~0ULL);
  	}
+
+	return 0;
  }
 
  static int

drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, · line 3568

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +1 -1
@@ -3563,7 +3565,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
  			break;
  	}
 
-	svm_range_debug_dump(svms);
+	pr_debug("%d", svm_range_debug_dump(svms));

This is a bit hacky. I would use the way dynamic_hex_dump is defined as an example for how to do this without the dummy pr_debug and without returning a dummy result from svm_range_debug_dump:

#define dynamic_svm_range_dump(svms) \
_dynamic_func_call_no_desc("svm_range_dump", svm_range_debug_dump, svms)

Then instead of calling svm_range_debug_dump directly, call dynamic_svm_range_dump(svms).

Regards,
  Felix

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0 -0
  	mutex_unlock(&svms->lock);
  	mmap_read_unlock(mm);