Re: 230-objrmap fixes for 2.6.3-mjb2

[Dave McCracken <dmccr@us.ibm.com>]

--On Wednesday, March 03, 2004 19:51:22 +0100 Andrea Arcangeli
<andrea@suse.de> wrote:

> ok, you used PG_locked that already exists for another purpose so it was
> not clear, another bitflag would be ok.

I forgot PG_locked already existed :)

> the main remaining issue to solve (and run at runtime) is the logic is
> to keep this flag consistent with all vmas pointing to the page having
> VM_LOCKED set. I'm not sure if it's worth it.

That's been the main stumbling block to anyone actually trying it.  It
involves per-page housekeeping at every mlock() and every time a locked vma
gets unmapped.

> what we do in 2.4 and that works pretty well, that is simply to refile
> pages into the active list if they're mlocked, so we don't waste too
> much cpu on them since we don't analyze them too often. this should work
> pretty well for everybody, or peraphs google may prefer to have a fully
> consistent PG_mlocked.

The point would be to have a way of finding and skipping the locked page
before we go look up the vmas for it.  2.4 doesn't have that problem
because it's working from the vma.

Dave McCracken