Part of what xe_bo_restore_kernel does, is restore BO's GGTT mappings
which may have been lost during a power state change. Missing is
restoring the GGTT entries without BO mappings to a known state (e.g.,
scratch pages). Update xe_bo_restore_kernel to clear the entire GGTT
before restoring BO's GGTT mappings.
v2:
- Include missing local change of tile and id variable (CI)
v3:
- Fixed kernel doc (CI)
v4:
- Only clear holes (CI)
Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Cc: Lucas De Marchi <lucas.demar...@intel.com>
Cc: Matthew Auld <matthew.a...@intel.com>
Signed-off-by: Matthew Brost <matthew.br...@intel.com>
Cc: <sta...@vger.kernel.org> # v6.8+
---
drivers/gpu/drm/xe/xe_bo_evict.c | 8 +++++++-
drivers/gpu/drm/xe/xe_ggtt.c | 19 ++++++++++++++++---
drivers/gpu/drm/xe/xe_ggtt.h | 2 ++
3 files changed, 25 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/xe/xe_bo_evict.c b/drivers/gpu/drm/xe/xe_bo_evict.c
index 8fb2be061003..d7bb3dbb41d6 100644
--- a/drivers/gpu/drm/xe/xe_bo_evict.c
+++ b/drivers/gpu/drm/xe/xe_bo_evict.c
@@ -123,7 +123,8 @@ int xe_bo_evict_all(struct xe_device *xe)
* @xe: xe device
*
* Move kernel BOs from temporary (typically system) memory to VRAM via CPU.
All
- * moves done via TTM calls.
+ * moves done via TTM calls. All GGTT are restored too, first by clearing GGTT
+ * to known state and then restoring individual BO's GGTT mappings.
*
* This function should be called early, before trying to init the GT, on
device
* resume.
@@ -131,8 +132,13 @@ int xe_bo_evict_all(struct xe_device *xe)
int xe_bo_restore_kernel(struct xe_device *xe)
{
struct xe_bo *bo;
+ struct xe_tile *tile;
+ u8 id;
int ret;
+ for_each_tile(tile, xe, id)
+ xe_ggtt_clear(tile->mem.ggtt);
+
spin_lock(&xe->pinned.lock);
for (;;) {
bo = list_first_entry_or_null(&xe->pinned.evicted,
diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
index 558fac8bb6fb..2fc498b89878 100644
--- a/drivers/gpu/drm/xe/xe_ggtt.c
+++ b/drivers/gpu/drm/xe/xe_ggtt.c
@@ -140,7 +140,7 @@ static void xe_ggtt_set_pte_and_flush(struct xe_ggtt *ggtt,
u64 addr, u64 pte)
ggtt_update_access_counter(ggtt);
}
-static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
+static void __xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start, u64 size)
{
u16 pat_index = tile_to_xe(ggtt->tile)->pat.idx[XE_CACHE_WB];
u64 end = start + size - 1;
@@ -160,6 +160,19 @@ static void xe_ggtt_clear(struct xe_ggtt *ggtt, u64 start,
u64 size)
}
}
+static void xe_ggtt_initial_clear(struct xe_ggtt *ggtt);
+
+/**
+ * xe_ggtt_clear() - GGTT clear
+ * @ggtt: the &xe_ggtt to be cleared
+ *
+ * Clear all GGTT to a known state
+ */
+void xe_ggtt_clear(struct xe_ggtt *ggtt)
+{
+ xe_ggtt_initial_clear(ggtt);
+}
+
static void ggtt_fini_early(struct drm_device *drm, void *arg)
{
struct xe_ggtt *ggtt = arg;
@@ -277,7 +290,7 @@ static void xe_ggtt_initial_clear(struct xe_ggtt *ggtt)
/* Display may have allocated inside ggtt, so be careful with clearing
here */
mutex_lock(&ggtt->lock);
drm_mm_for_each_hole(hole, &ggtt->mm, start, end)
- xe_ggtt_clear(ggtt, start, end - start);
+ __xe_ggtt_clear(ggtt, start, end - start);
xe_ggtt_invalidate(ggtt);
mutex_unlock(&ggtt->lock);
@@ -294,7 +307,7 @@ static void ggtt_node_remove(struct xe_ggtt_node *node)
mutex_lock(&ggtt->lock);
if (bound)
- xe_ggtt_clear(ggtt, node->base.start, node->base.size);
+ __xe_ggtt_clear(ggtt, node->base.start, node->base.size);
drm_mm_remove_node(&node->base);
node->base.size = 0;
mutex_unlock(&ggtt->lock);
diff --git a/drivers/gpu/drm/xe/xe_ggtt.h b/drivers/gpu/drm/xe/xe_ggtt.h
index 27e7d67de004..b7ae440cdebf 100644
--- a/drivers/gpu/drm/xe/xe_ggtt.h
+++ b/drivers/gpu/drm/xe/xe_ggtt.h
@@ -13,6 +13,8 @@ struct drm_printer;
int xe_ggtt_init_early(struct xe_ggtt *ggtt);
int xe_ggtt_init(struct xe_ggtt *ggtt);
+void xe_ggtt_clear(struct xe_ggtt *ggtt);
+
struct xe_ggtt_node *xe_ggtt_node_init(struct xe_ggtt *ggtt);
void xe_ggtt_node_fini(struct xe_ggtt_node *node);
int xe_ggtt_node_insert_balloon(struct xe_ggtt_node *node,
--
2.34.1
