Re: [Patch 3/3] prepopulate/cache cleared pages

[Andi Kleen <ak@suse.de>]

On Thursday 23 February 2006 10:29, Arjan van de Ven wrote:
> This patch adds an entry for a cleared page to the task struct. The main
> purpose of this patch is to be able to pre-allocate and clear a page in a
> pagefault scenario before taking any locks (esp mmap_sem),
> opportunistically. Allocating+clearing a page is an very expensive 
> operation that currently increases lock hold times quite bit (in a threaded 
> environment that allocates/use/frees memory on a regular basis, this leads
> to contention).
> 
> This is probably the most controversial patch of the 3, since there is
> a potential to take up 1 page per thread in this cache. In practice it's
> not as bad as it sounds (a large degree of the pagefaults are anonymous 
> and thus immediately use up the page). One could argue "let the VM reap
> these" but that has a few downsides; it increases locking needs but more,
> clearing a page is relatively expensive, if the VM reaps the page again
> in case it wasn't needed, the work was just wasted.

Looks like an incredible bad hack. What workload was that again where it helps?
And how much? I think before we can consider adding that ugly code you would a far better
rationale.

-Andi