Re: Always passing mm and vma down (was: [RFC][PATCH] Convert do_no_page() to a hook to avoid DFS race)

[Paul E. McKenney]

On Sat, May 31, 2003 at 10:46:18AM +0200, Ingo Oeser wrote:
> On Fri, May 30, 2003 at 04:41:50PM -0700, Paul E. McKenney wrote:
> > -struct page *
> > -ia32_install_shared_page (struct vm_area_struct *vma, unsigned long address, int no_share)
> > +int
> > +ia32_install_shared_page (struct mm_struct *mm, struct vm_area_struct *vma, unsigned long address, int write_access, pmd_t *pmd)
> 
> Why do we always pass mm and vma down, even if vma->vm_mm
> contains the mm, where the vma belongs to? Is the connection
> between a vma and its mm also protected by the mmap_sem?
> 
> Is this really necessary or an oversight and we waste a lot of
> stack in a lot of places?
> 
> If we just need it for accounting: We need current->mm, if we
> need it to locate the next vma relatively to this vma, vma->vm_mm
> is the one.

Interesting point.  The original do_no_page() API does this
as well:

	static int
	do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
		   unsigned long address, int write_access, pte_t *page_table, pmd_t *pmd)

As does do_anonymous_page().  I assumed that there were corner
cases where this one-to-one correspondence did not exist, but
must confess that I did not go looking for them.

Or is this a performance issue, avoiding a dereference and
possible cache miss?

					Thanx, Paul