Re: [patch 0/4] MAP_NOZERO v2 - VM_NOZERO/MAP_NOZERO early summer madness

[Andy Isaacson <adi@hexapodia.org>]

On Mon, Jul 02, 2007 at 06:55:40PM -0400, Rik van Riel wrote:
> You could easily replace the cookie with a pointer to a free
> page pool.

It just occurred to me that something like this is *required* to get the
performance benefit from MAP_NOZERO on a busy system.  With Davide's
current proposal, if there are N jobs with different "nozero cookies"
busy allocating and deallocating pages on a single-NUMA-node system,
then there's only a 1/N chance that the page returned by __alloc_pages
will have the correct cookie, so (N-1)/N percent of the time MAP_NOZERO
will have no positive effect -- 90% of the time for the case of N=10.

(Of course on NUMA systems node affinity will probably make the
situation a bit better.)

So I'm afraid that it sounds like additional complexity would be
required so that __alloc_pages could track and preferentially return
pages with the correct "nozero cookie".


In a mostly unrelated complaint, I note that having a function named
"alloc_zeroed_page_vma" which returns a potentially nonzero page is, um,
unintuitive.  It needs a name which does not claim it returns zeroed
pages.  I'm no good at names, but perhaps "alloc_available_page_vma"?

-andy