RE: hugetlb demand paging patch part [2/3]

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

David Gibson wrote on Thursday, April 15, 2004 8:27 PM
> Ah!  So it's just an optimiziation - it makes a bit more sense to me
> now.  I had assumed that this case (hugepage get_user_pages()) would
> be sufficiently rare that it would not require optimization.
> Apparently not.

It's a huge deal because for *every* I/O, kernel has to do get_user_pages()
to lock the page, it's really gets in the way with the spin_lock as well.

        spin_lock(&mm->page_table_lock);
        do {
                struct page *map;
                int lookup_write = write;
                while (!(map = follow_page(mm, start, lookup_write))) {

With current state of art platform, I/O requirement pushes into 200K
per second, this become quite significant.


> Do you know where the cycles are going without this optimization?  In
> particular, could it be just the find_vma() in hugepage_vma() called
> before follow_huge_addr()?  I note that IA64 is the only arch to have
> a non-trivial hugepage_vma()/follow_huge_addr() and that its
> follow_huge_addr() doesn't actually use the vma passed in.

That's one, plus the spin lock mentioned above.


> If we could get rid of follow_hugetlb_pages() it would remove an ugly
> function from every arch, which would be nice.

I hope the goal here is not to trim code for existing prefaulting scheme.
That function has to go for demand paging, and demand paging comes with
a performance price most people don't realize.  If the goal here is to
make the code prettier, I vote against that.