control page: avoid "minor" page faults

By default, even private writable pages are mapped
with the RW bit disabled in the PTE. This causes a
"minor" page fault when the page is first written
to. To avoid this, make sure that vm_inert_page()
uses the proper page protection bits and mark the
VMA as VM_IO to keep the rest of the VM code out.

1 files changed, 10 insertions, 3 deletions

diff --git a/litmus/ctrldev.c b/litmus/ctrldev.c
index ec0fc4b1ab55..9969ab17c190 100644
--- a/litmus/ctrldev.c
+++ b/litmus/ctrldev.c
@@ -95,9 +95,16 @@ static int litmus_ctrl_mmap(struct file* filp, struct vm_area_struct* vma)
                 return -EINVAL;
 
         vma->vm_ops = &litmus_ctrl_vm_ops;
-        /* this mapping should not be kept across forks,
-         * and cannot be expanded */
-        vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
+        /* This mapping should not be kept across forks,
+         * cannot be expanded, and is not a "normal" page. */
+        vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND | VM_IO;
+
+        /* We don't want the first write access to trigger a "minor" page fault
+         * to mark the page as dirty.  This is transient, private memory, we
+         * don't care if it was touched or not. __S011 means RW access, but not
+         * execute, and avoids copy-on-write behavior.
+         * See protection_map in mmap.c.  */
+        vma->vm_page_prot = __S011;
 
         err = alloc_ctrl_page(current);
         if (!err)