Re: Note describing poor dcache utilization under high memory pressure

[Daniel Phillips]

On January 29, 2002 10:00 pm, Oliver Xymoron wrote:
> On Tue, 29 Jan 2002, Daniel Phillips wrote:
> 
> > On January 29, 2002 06:25 pm, Rik van Riel wrote:
> > > On Tue, 29 Jan 2002, Oliver Xymoron wrote:
> > >
> > > > Daniel's approach seems to be workable (once he's spelled out all the
> > > > details) but it misses the big performance win for fork/exec, which is
> > > > surely the common case. Given that exec will be throwing away all these
> > > > mappings, we can safely assume that we will not be inheriting many shared
> > > > mappings from parents of parents so Daniel's approach also still ends up
> > > > marking most of the pages RO still.
> > >
> > > It gets worse.  His approach also needs to adjust the reference
> > > counts on all pages (and swap pages).
> >
> > Well, Rik, time to present your algorithm.  I assume it won't reference
> > counts on pages, and will do some kind of traversal of the mm tree.  Note
> > however, that I did investigate the class of algorithm you are interested in,
> > and found only nasty, complex solutions there, with challenging locking
> > problems.  (I also looked at a number of possible improvements to virtual
> > scanning, as you know, and likewise only found ugly or inadequate solutions.)
> 
> I think it goes something like this:
> 
> fork:
>   detach page tables from parent
>   retain pointer to "backing page tables" in parent and child
>   update use count in page tables
>   "prefault" tables for current stack and instruction pages in both parent
>     and child
> 
> page fault:
>   if faulted on page table:
>     look up backing page tables
>     if use count > 1: copy, dec use count
>     else: take ownership
> 
> > Before you sink a lot of time into it though, you might add up the actual
> > overhead you're worried about above, and see if it moves the needle in a real
> > system.
> 
> I'm pretty sure something like the above does signficantly less work in
> the fork/exec case, which is the important one.

With fork/exec, for each page table there are two cases:

  - The parent instantiated the page table.  In this case the extra work to
    set the ptes RO (only for CoW pages) is insignificant.

  - The parent is still sharing the page table with its parent and so the
    ptes are still set RO.

I don't see how there is a whole lot of fat to cut here. 

-- 
Daniel