Re: [PATCH v2] mm: remove MIGRATE_ISOLATE check in hotpath

[Minchan Kim]

On Mon, Feb 25, 2013 at 02:50:11PM -0800, Andrew Morton wrote:
> On Mon, 25 Feb 2013 11:13:08 +0900
> Minchan Kim <minchan@xxxxxxxxxx> wrote:
> 
> > > 
> > > >
> > > > ...
> > > >
> > > > @@ -683,7 +683,7 @@ static void free_one_page(struct zone *zone, struct page *page, int order,
> > > >  	zone->pages_scanned = 0;
> > > >  
> > > >  	__free_one_page(page, zone, order, migratetype);
> > > > -	if (unlikely(migratetype != MIGRATE_ISOLATE))
> > > > +	if (unlikely(!is_migrate_isolate(migratetype)))
> > > >  		__mod_zone_freepage_state(zone, 1 << order, migratetype);
> > > >  	spin_unlock(&zone->lock);
> > > >  }
> > > 
> > > The code both before and after this patch is assuming that the
> > > migratetype in free_one_page is likely to be MIGRATE_ISOLATE.  Seems
> > > wrong.  If CONFIG_MEMORY_ISOLATION=n this ends up doing
> > > if(unlikely(true)) which is harmless-but-amusing.
> > 
> > >From the beginning of [2139cbe627, cma: fix counting of isolated pages],
> > it was wrong. We can't make sure it's very likely.
> > If it is called by order-0 page free path, it is but if it is called by
> > high order page free path, we can't.
> > So I think it would be better to remove unlikley.
> 
> Order-0 pages surely preponderate, so I'd say that "likely" is the way
> to go.

Okay then, let's rule out high order allocation.
Firstly, let's look CONFIG_MEMORY_ISOLATION=y case.
In case of order-0, free_hot_cold_page calls free_one_page very unlikely.

void free_hot_cold_page ()
{
        ...
        if (migratetype >= MIGRATE_PCPTYPES) {
                if (unlikely(is_migrate_isolate(migratetype))) {
                        free_one_page(zone, page, 0, migratetype);
                        goto out;
                }
        ...
}

So, if free_one_page is called for order-0 page, it's for only MIGRATE_ISOLATE.
So unlikely(!is_migrate_isolate(migratetype)) in free_one_page does make sense
to me.

In case of CONFIG_MEMORY_ISOLATION=n case, below is_migrate_isolate is always
false so it could be compiled out so free_one_page is called only
for high order page free path. So if you don't mind high order free path
hitting on likely/unlikely, I think current code doesn't have any problem.

 if (migratetype >= MIGRATE_PCPTYPES) {
                if (unlikely(is_migrate_isolate(migratetype))) { ==> always false
                        free_one_page(zone, page, 0, migratetype);
                        goto out;
                }

In summary, if you don't care of high order free path, there is no problem.

> 
> I don't recall anyone ever demonstrating that likely/unlikely actually
> does anything useful.  It would be interesting to have a play around,
> see if it actually does good things to the code generation.

Yes. especially about page alloc/free path. 

> 
> I think someone (perhaps in or near Dave Jones?) once had a patch which
> added counters to likely/unlikely, so the kernel can accumulate and
> then report upon the hit/miss ratio at each site.  iirc, an alarmingly
> large number of the sites were deoptimisations!

It seems you mean "Branch Profiling (Trace likely/unlikely profiler)" made by
Steven Rostedt. Anyway, it's a rather troublesome job and needs many workload
but worthy. Will queue it up to my future TODO. :)

> 
> > They are trivial patch so send it now or send it after you release
> > first mmotm after finishing merge window?
> 
> It's in mainline now.

I will send fix about only undo_isolate_page_range if you don't object my
above opinion.

Thanks.

> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>

-- 
Kind regards,
Minchan Kim
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/