drm/vmwgfx: Use VM_PFNMAP instead of VM_MIXEDMAP when possible

Commit Message

Thomas Hellström (Intel) Jan. 15, 2020, 2:15 p.m. UTC

From: Thomas Hellstrom <[e-mail]>

For shared, and read-only private mappings of graphics memory, use
VM_PFNMAP instead of VM_MIXEDMAP. This means less accounting overhead
when inserting and removing page-table entries. TTM doesn't do this
by default, since there was a performance problem with book-keeping of
write-combined mappings. Since vmwgfx solely uses cached mappings, that's
not a problem and now that the TTM vm has largely been turned into
helpers, we can use VM_PFNMAP on a per-driver basis

Signed-off-by: Thomas Hellstrom <[e-mail]>
Reviewed-by: Roland Scheidegger <[e-mail]>
---
 drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c | 4 ++++
 1 file changed, 4 insertions(+)

Patch

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c
index ce288756531b..aa7e50f63b94 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c
@@ -45,6 +45,10 @@  int vmw_mmap(struct file *filp, struct vm_area_struct *vma)
 
 	vma->vm_ops = &vmw_vm_ops;
 
+	/* Use VM_PFNMAP rather than VM_MIXEDMAP if not a COW mapping */
+	if ((vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) != VM_MAYWRITE)
+		vma->vm_flags = (vma->vm_flags & ~VM_MIXEDMAP) | VM_PFNMAP;
+
 	return 0;
 }