[Intel-xe] [PATCH v2 1/2] Revert "drm/xe: Pad GGTT mapping with an extra page pointing to scratch"

[Niranjana Vishwanathapura]

This reverts commit 70e94be2729e5c725c53b1aa2d10c073566bb1ca.

Resolve minor conflict by retaining XE_BO_CREATE_VRAM_MASK flag.

Coalescing GGTT invalidations is causing a bunch of hangs during
driver load and in user space. Benefit is rather small too, so
revert it to stabilize the stack.

Tested-by: Matt Roper <matthew.d.roper at intel.com>  # ADL-P
Reviewed-by: Lucas De Marchi <lucas.demarchi at intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura at intel.com>
---
 drivers/gpu/drm/xe/xe_bo.c       |  1 -
 drivers/gpu/drm/xe/xe_bo.h       |  1 +
 drivers/gpu/drm/xe/xe_bo_types.h |  4 +---
 drivers/gpu/drm/xe/xe_ggtt.c     | 30 +++++++++---------------------
 4 files changed, 11 insertions(+), 25 deletions(-)

diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
index 5460e6fe3c1f..c7f175a149be 100644
--- a/drivers/gpu/drm/xe/xe_bo.c
+++ b/drivers/gpu/drm/xe/xe_bo.c
@@ -975,7 +975,6 @@ struct xe_bo *__xe_bo_create_locked(struct xe_device *xe, struct xe_bo *bo,
 			return bo;
 	}
 
-	bo->requested_size = size;
 	if (flags & (XE_BO_CREATE_VRAM_MASK | XE_BO_CREATE_STOLEN_BIT) &&
 	    !(flags & XE_BO_CREATE_IGNORE_MIN_PAGE_SIZE_BIT) &&
 	    xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K) {
diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h
index 8f5a7ad10d09..dd58edcb9398 100644
--- a/drivers/gpu/drm/xe/xe_bo.h
+++ b/drivers/gpu/drm/xe/xe_bo.h
@@ -213,6 +213,7 @@ xe_bo_main_addr(struct xe_bo *bo, size_t page_size)
 static inline u32
 xe_bo_ggtt_addr(struct xe_bo *bo)
 {
+	XE_BUG_ON(bo->ggtt_node.size > bo->size);
 	XE_BUG_ON(bo->ggtt_node.start + bo->ggtt_node.size > (1ull << 32));
 	return bo->ggtt_node.start;
 }
diff --git a/drivers/gpu/drm/xe/xe_bo_types.h b/drivers/gpu/drm/xe/xe_bo_types.h
index db217500a970..06de3330211d 100644
--- a/drivers/gpu/drm/xe/xe_bo_types.h
+++ b/drivers/gpu/drm/xe/xe_bo_types.h
@@ -23,9 +23,7 @@ struct xe_vm;
 struct xe_bo {
 	/** @ttm: TTM base buffer object */
 	struct ttm_buffer_object ttm;
-	/** @requested_size: Requested size of this buffer object */
-	size_t requested_size;
-	/** @size: Size of this buffer object after alignment adjusting */
+	/** @size: Size of this buffer object */
 	size_t size;
 	/** @flags: flags for this buffer object */
 	u32 flags;
diff --git a/drivers/gpu/drm/xe/xe_ggtt.c b/drivers/gpu/drm/xe/xe_ggtt.c
index a430d1568890..2e31b1ce3e4c 100644
--- a/drivers/gpu/drm/xe/xe_ggtt.c
+++ b/drivers/gpu/drm/xe/xe_ggtt.c
@@ -263,6 +263,7 @@ int xe_ggtt_insert_special_node(struct xe_ggtt *ggtt, struct drm_mm_node *node,
 
 void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
 {
+	struct xe_device *xe = gt_to_xe(ggtt->gt);
 	u64 start = bo->ggtt_node.start;
 	u64 offset, pte;
 
@@ -273,12 +274,8 @@ void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
 		xe_ggtt_set_pte(ggtt, start + offset, pte);
 	}
 
-	if (bo->size == bo->requested_size) {
-		pte = xe_ggtt_pte_encode(ggtt->scratch ?: bo, 0);
-		xe_ggtt_set_pte(ggtt, start + bo->size, pte);
-	}
-
-	if (ggtt->invalidate) {
+	/* XXX: Without doing this everytime on integrated driver load fails */
+	if (ggtt->invalidate || !IS_DGFX(xe)) {
 		xe_ggtt_invalidate(ggtt->gt);
 		ggtt->invalidate = false;
 	}
@@ -287,7 +284,6 @@ void xe_ggtt_map_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
 static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
 				  u64 start, u64 end)
 {
-	u64 size = bo->size;
 	int err;
 	u64 alignment = GEN8_PAGE_SIZE;
 
@@ -295,6 +291,8 @@ static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
 		alignment = SZ_64K;
 
 	if (XE_WARN_ON(bo->ggtt_node.size)) {
+		/* Someone's already inserted this BO in the GGTT */
+		XE_BUG_ON(bo->ggtt_node.size != bo->size);
 		return 0;
 	}
 
@@ -302,21 +300,8 @@ static int __xe_ggtt_insert_bo_at(struct xe_ggtt *ggtt, struct xe_bo *bo,
 	if (err)
 		return err;
 
-	/*
-	 * We add an extra page when mapping a BO in the GGTT so we can coalesce
-	 * GGTT invalidations. Without this extra page GGTT prefetches can leave
-	 * entries in the TLB pointing to an invalidation GGTT entry when in
-	 * fact we have programmed this GGTT entry to a valid entry. BO aligned
-	 * to 64k already have padding so no need to add an extra page.
-	 */
-	if (bo->size == bo->requested_size) {
-		size += SZ_4K;
-		if (end != U64_MAX)
-			end += SZ_4K;
-	}
-
 	mutex_lock(&ggtt->lock);
-	err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node, size,
+	err = drm_mm_insert_node_in_range(&ggtt->mm, &bo->ggtt_node, bo->size,
 					  alignment, 0, start, end, 0);
 	if (!err)
 		xe_ggtt_map_bo(ggtt, bo);
@@ -354,6 +339,9 @@ void xe_ggtt_remove_bo(struct xe_ggtt *ggtt, struct xe_bo *bo)
 	if (XE_WARN_ON(!bo->ggtt_node.size))
 		return;
 
+	/* This BO is not currently in the GGTT */
+	XE_BUG_ON(bo->ggtt_node.size != bo->size);
+
 	xe_ggtt_remove_node(ggtt, &bo->ggtt_node);
 }
 
-- 
2.21.0.rc0.32.g243a4c7e27