On Tue, 03 Apr 2007 15:47:05 +1000 Rusty Russell <rusty@xxxxxxxxxxxxxxx> wrote:
> On Mon, 2007-04-02 at 21:57 -0700, Andrew Morton wrote:
> > On Tue, 03 Apr 2007 14:45:02 +1000 Rusty Russell <rusty@xxxxxxxxxxxxxxx>
> > wrote:
> > > Does that mean the to function correctly every user needs some internal
> > > cursor so it doesn't end up scanning the first N entries over and over?
> > >
> >
> > If it wants to be well-behaved, and to behave as the VM expects, yes.
> >
> > There's an expectation that the callback will be performing some scan-based
> > aging operation and of course to do LRU (or whatever) aging, the callback
> > will need to remember where it was up to last time it was called.
> >
> > But it's just a guideline - callbacks could do something different but
> > in-the-spirit, I guess.
>
> Hmm, actually the callers I looked at (nfs, dcache, mbcache) seem to use
> an LRU list and just walk the first "nr_to_scan" entries, and nr_to_scan
> is always 128.
That's just because of the batching logic up in shrink_slab(). And iirc we
only break the scanning into lumps of 128 items so we can add a
cond_resched() into it.
> Someone who keeps a cursor will be disadvantaged: the other shrinkers
> could well get less effective on repeated calls, but we won't. Someone
> who picks entries at random might have the same issue.
To examine the balancing one would need to examine the value of total_scan
in shrink_slab(), rather than looking at the value which shrink_slab()
passes into the callback.
> I think it is clearest to describe how we expect everyone to work, and
> let whoever is getting creative worry about it themselves.
>
> How's this:
> ==
> Cleanup and kernelify shrinker registration.
hm, well, six-of-one, VI of the other. We save maybe four kmallocs across
the entire uptime at the cost of exposing stuff kernel-side which doesn't
need to be exposed.
But I think we need to weed that crappiness out of XFS first.
|