Re: [PATCH 02/22] Do not sanity check order in the fast path

[Dave Hansen]

On Thu, 2009-04-23 at 10:58 +0100, Mel Gorman wrote:
> > How about this:  I'll go and audit the use of order in page_alloc.c to
> > make sure that having an order>MAX_ORDER-1 floating around is OK and
> > won't break anything. 
> 
> Great. Right now, I think it's ok but I haven't audited for this
> explicily and a second set of eyes never hurts.

OK, after looking through this, I have a couple of ideas.  One is that
we do the MAX_ORDER check in __alloc_pages_internal(), but *after* the
first call to get_page_from_freelist().  That's because I'm worried if
we ever got into the reclaim code with a >MAX_ORDER 'order'.  Such as:

void wakeup_kswapd(struct zone *zone, int order)
{
...
        if (pgdat->kswapd_max_order < order)
                pgdat->kswapd_max_order = order;
        if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
                return;
        if (!waitqueue_active(&pgdat->kswapd_wait))
                return;
        wake_up_interruptible(&pgdat->kswapd_wait);
}

unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
                                gfp_t gfp_mask, nodemask_t *nodemask)
{
        struct scan_control sc = {
...
                .order = order,
                .mem_cgroup = NULL,
                .isolate_pages = isolate_pages_global,
                .nodemask = nodemask,
        };

        return do_try_to_free_pages(zonelist, &sc);
}

This will keep us only checking 'order' once for each
alloc_pages_internal() call.  It is an extra branch, but it is out of
the really, really hot path since we're about to start reclaim here
anyway.

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e2f2699..1e3a01e 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1498,6 +1498,13 @@ restart:
 			zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET);
 	if (page)
 		goto got_pg;
+	/*
+	 * We're out of the rocket-hot area above, so do a quick sanity
+	 * check.  We do this here to avoid ever trying to do any reclaim
+	 * of >=MAX_ORDER areas which can never succeed, of course.
+	 */
+	if (order >= MAX_ORDER)
+		goto nopage;
 
 	/*
 	 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and


-- Dave

--
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/