This reverts commit d97e7b1eb8afd7a404466533b0bc192351b760c7.
Needed for the next revert patch.
Signed-off-by: Philip Yang <Philip.Yang@xxxxxxx>
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 60 ++++++++++++++++++++++++++++
drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 3 ++
2 files changed, 63 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 4d000c63cde8..3422eee8d0d0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1145,6 +1145,66 @@ svm_range_add_child(struct svm_range *prange, struct mm_struct *mm,
list_add_tail(&pchild->child_list, &prange->child_list);
}
+/**
+ * svm_range_split_by_granularity - collect ranges within granularity boundary
+ *
+ * @p: the process with svms list
+ * @mm: mm structure
+ * @addr: the vm fault address in pages, to split the prange
+ * @parent: parent range if prange is from child list
+ * @prange: prange to split
+ *
+ * Trims @prange to be a single aligned block of prange->granularity if
+ * possible. The head and tail are added to the child_list in @parent.
+ *
+ * Context: caller must hold mmap_read_lock and prange->lock
+ *
+ * Return:
+ * 0 - OK, otherwise error code
+ */
+int
+svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
+ unsigned long addr, struct svm_range *parent,
+ struct svm_range *prange)
+{
+ struct svm_range *head, *tail;
+ unsigned long start, last, size;
+ int r;
+
+ /* Align splited range start and size to granularity size, then a single
+ * PTE will be used for whole range, this reduces the number of PTE
+ * updated and the L1 TLB space used for translation.
+ */
+ size = 1UL << prange->granularity;
+ start = ALIGN_DOWN(addr, size);
+ last = ALIGN(addr + 1, size) - 1;
+
+ pr_debug("svms 0x%p split [0x%lx 0x%lx] to [0x%lx 0x%lx] size 0x%lx\n",
+ prange->svms, prange->start, prange->last, start, last, size);
+
+ if (start > prange->start) {
+ r = svm_range_split(prange, start, prange->last, &head);
+ if (r)
+ return r;
+ svm_range_add_child(parent, mm, head, SVM_OP_ADD_RANGE);
+ }
+
+ if (last < prange->last) {
+ r = svm_range_split(prange, prange->start, last, &tail);
+ if (r)
+ return r;
+ svm_range_add_child(parent, mm, tail, SVM_OP_ADD_RANGE);
+ }
+
+ /* xnack on, update mapping on GPUs with ACCESS_IN_PLACE */
+ if (p->xnack_enabled && prange->work_item.op == SVM_OP_ADD_RANGE) {
+ prange->work_item.op = SVM_OP_ADD_RANGE_AND_MAP;
+ pr_debug("change prange 0x%p [0x%lx 0x%lx] op %d\n",
+ prange, prange->start, prange->last,
+ SVM_OP_ADD_RANGE_AND_MAP);
+ }
+ return 0;
+}
static bool
svm_nodes_in_same_hive(struct kfd_node *node_a, struct kfd_node *node_b)
{
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
index 026863a0abcd..be11ba0c4289 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
@@ -172,6 +172,9 @@ struct kfd_node *svm_range_get_node_by_id(struct svm_range *prange,
int svm_range_vram_node_new(struct kfd_node *node, struct svm_range *prange,
bool clear);
void svm_range_vram_node_free(struct svm_range *prange);
+int svm_range_split_by_granularity(struct kfd_process *p, struct mm_struct *mm,
+ unsigned long addr, struct svm_range *parent,
+ struct svm_range *prange);
int svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
uint32_t vmid, uint32_t node_id, uint64_t addr,
bool write_fault);
--
2.35.1
