[PATCH] drm/amdkfd: Change error handling at prange update in svm_range_set_attr
Xiaogang.Chen
xiaogang.chen at amd.com
Fri Jan 31 16:58:39 UTC 2025
From: Xiaogang Chen <xiaogang.chen at amd.com>
When register a vm range at svm the added vm range may be split into multiple
subranges and/or existing pranges got spitted. The new pranges need validated
and mapped. This patch changes error handling for pranges that fail updating:
1: free prange resources and remove it from svms if its updating fails as it
will not be used.
2: return -EAGAIN when all pranges at update_list need redo valid/map,
otherwise return no -EAGAIN error to user space to indicate failure. That
removes unnecessary retries.
Signed-off-by: Xiaogang Chen <xiaogang.chen at amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 27 +++++++++++++++++++++++----
1 file changed, 23 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index e32e19196f6b..455cb98bf16a 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -3716,8 +3716,19 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
out_unlock_range:
mutex_unlock(&prange->migrate_mutex);
- if (r)
- ret = r;
+ /* this prange cannot be migraed, valid or map */
+ if (r) {
+ /* free this prange resources, remove it from svms */
+ svm_range_unlink(prange);
+ svm_range_remove_notifier(prange);
+ svm_range_free(prange, false);
+
+ /* ret got update when any r != -EAGAIN;
+ * return -EAGAIN when all pranges at update_list
+ * need redo valid/map */
+ if (r != -EAGAIN || !ret)
+ ret = r;
+ }
}
list_for_each_entry(prange, &remap_list, update_list) {
@@ -3729,8 +3740,16 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
if (r)
pr_debug("failed %d on remap svm range\n", r);
mutex_unlock(&prange->migrate_mutex);
- if (r)
- ret = r;
+
+ if (r) {
+ /* remove this prange */
+ svm_range_unlink(prange);
+ svm_range_remove_notifier(prange);
+ svm_range_free(prange, false);
+
+ if (r != -EAGAIN || !ret)
+ ret = r;
+ }
}
dynamic_svm_range_dump(svms);
--
2.25.1
More information about the amd-gfx
mailing list