Re: Google's mm problems

[Christian Ehrhardt]

On Wed, Oct 31, 2001 at 12:07:17AM +0100, Daniel Phillips wrote:
> Hi, I've been taking a look on a mm problem that Ben Smith of Google posted 
> a couple of weeks ago.  As it stands, Google can't use 2.4 yet because all
> known flavors of 2.4 mm fall down in one way or another.  This is not good.

Andrea suggested that this might be a mlock bug and someone else
pointed out that madvise instead of mlock exhibits similar behaviour.
So I looked at this code and this patch looks obviously correct:

--- mlock.c.old Sat Nov  3 19:53:43 2001
+++ mlock.c     Sat Nov  3 19:55:18 2001
@@ -90,7 +90,6 @@
        left->vm_end = start;
        right->vm_start = end;
        right->vm_pgoff += (right->vm_start - left->vm_start) >> PAGE_SHIFT;
-       vma->vm_flags = newflags;
        left->vm_raend = 0;
        right->vm_raend = 0;
        if (vma->vm_file)

This assignment is redundant and it is not protected by any locks.
As far as I can see vma->vm_mm->page_table_lock is supposed to protect
modifications of vma->vm_flags. If this is true we probably need this
as well:

--- filemap.c.old       Sat Nov  3 21:44:12 2001
+++ filemap.c   Sat Nov  3 21:46:36 2001
@@ -2178,7 +2178,9 @@

        if (start == vma->vm_start) {
                if (end == vma->vm_end) {
+                       spin_lock(&vma->vm_mm->page_table_lock);
                        setup_read_behavior(vma, behavior);
+                       spin_unlock(&vma->vm_mm->page_table_lock);
                        vma->vm_raend = 0;
                } else
                        error = madvise_fixup_start(vma, end, behavior);

I doubt that this is the reason for the google memory problems, but
who knows?

    regards  Christian

-- 
THAT'S ALL FOLKS!