Re: [PATCH 2.6.14-rc2] fix incorrect mm->hiwater_vm and mm->hiwater_rss

[Hugh Dickins <hugh@veritas.com>]

On Wed, 21 Sep 2005, Christoph Lameter wrote:
> On Wed, 21 Sep 2005, Hugh Dickins wrote:
> 
> Looks like this is staying off the critical code paths. So ok.

Great, thanks.

> > I'm also rearranging the rss,anon_rss accounting.  Maybe come back
> > to the hiwaters later on?
> 
> I'd be very interested to see the rearrangement.

Sorry, you'll be imagining something much grander than I meant.
And sorry, I have to work my way through splitting it out from
the total and then posting, can't cut it out for you right now.

But all I was meaning was, as part of the preparatory patches for
the narrowed and split page_table_lock, not included in last month's
prototype, I'm (a) switching to file_rss plus anon_rss rather than rss
including anon_rss, as we discussed earlier (you believe the SGI boxes
won't need that, but in general it seems silly to be doing two atomic
updates instead of one); and (b) moving the "freed" rss updating up
from inside the "tlb" operations (the prototype patch had the wrong
locking for that); and (c) batching the updates (i.e. one add or sub
instead of many incs or decs) per-page_table-page in copy_pte_range
as in zap_pte_range.

I don't expect you to disagree with it; but nor do I expect you
to be anything but disappointed, realizing that's all I meant.

Hugh