[PATCH 2/3] drm/amdkfd: Avoid thrashing of stack and heap

[Felix Kuehling]

Stack and heap pages tend to be shared by many small allocations.
Concurrent access by CPU and GPU is therefore likely, which can lead to
thrashing. Avoid this by setting the preferred location to system memory.

Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
Signed-off-by: Philip Yang <yangp at amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 22 +++++++++++++++++-----
 1 file changed, 17 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 8d37ee755b34..854735dbb96e 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -2310,7 +2310,8 @@ svm_range_best_restore_location(struct svm_range *prange,
 
 static int
 svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
-				unsigned long *start, unsigned long *last)
+			       unsigned long *start, unsigned long *last,
+			       bool *is_heap_stack)
 {
 	struct vm_area_struct *vma;
 	struct interval_tree_node *node;
@@ -2321,6 +2322,12 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
 		pr_debug("VMA does not exist in address [0x%llx]\n", addr);
 		return -EFAULT;
 	}
+
+	*is_heap_stack = (vma->vm_start <= vma->vm_mm->brk &&
+			  vma->vm_end >= vma->vm_mm->start_brk) ||
+			 (vma->vm_start <= vma->vm_mm->start_stack &&
+			  vma->vm_end >= vma->vm_mm->start_stack);
+
 	start_limit = max(vma->vm_start >> PAGE_SHIFT,
 		      (unsigned long)ALIGN_DOWN(addr, 2UL << 8));
 	end_limit = min(vma->vm_end >> PAGE_SHIFT,
@@ -2350,9 +2357,9 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr,
 	*start = start_limit;
 	*last = end_limit - 1;
 
-	pr_debug("vma start: 0x%lx start: 0x%lx vma end: 0x%lx last: 0x%lx\n",
-		  vma->vm_start >> PAGE_SHIFT, *start,
-		  vma->vm_end >> PAGE_SHIFT, *last);
+	pr_debug("vma [0x%lx 0x%lx] range [0x%lx 0x%lx] is_heap_stack %d\n",
+		 vma->vm_start >> PAGE_SHIFT, vma->vm_end >> PAGE_SHIFT,
+		 *start, *last, *is_heap_stack);
 
 	return 0;
 }
@@ -2417,11 +2424,13 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
 	struct svm_range *prange = NULL;
 	unsigned long start, last;
 	uint32_t gpuid, gpuidx;
+	bool is_heap_stack;
 	uint64_t bo_s = 0;
 	uint64_t bo_l = 0;
 	int r;
 
-	if (svm_range_get_range_boundaries(p, addr, &start, &last))
+	if (svm_range_get_range_boundaries(p, addr, &start, &last,
+					   &is_heap_stack))
 		return NULL;
 
 	r = svm_range_check_vm(p, start, last, &bo_s, &bo_l);
@@ -2448,6 +2457,9 @@ svm_range *svm_range_create_unregistered_range(struct amdgpu_device *adev,
 		return NULL;
 	}
 
+	if (is_heap_stack)
+		prange->preferred_loc = KFD_IOCTL_SVM_LOCATION_SYSMEM;
+
 	svm_range_add_to_svms(prange);
 	svm_range_add_notifier_locked(mm, prange);
 
-- 
2.32.0