flowcache: Increase threshold for refusing new allocations

The threshold for OOM protection is too small for systems with large
number of CPUs. Applications report ENOBUFs on connect() every 10
minutes.

The problem is that the variable net->xfrm.flow_cache_gc_count is a
global counter while the variable fc->high_watermark is a per-CPU
constant. Take the number of CPUs into account as well.

Fixes: 6ad3122a ("flowcache: Avoid OOM condition under preasure")
Reported-by: Lukáš Koldrt <lk@excello.cz>
Tested-by: Jan Hejl <jh@excello.cz>
Signed-off-by: Miroslav Urbanek <mu@miroslavurbanek.com>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>

net/core/flow.c

@@ -95,7 +95,6 @@ static void flow_cache_gc_task(struct work_struct *work)
 	list_for_each_entry_safe(fce, n, &gc_list, u.gc_list) {
 		flow_entry_kill(fce, xfrm);
 		atomic_dec(&xfrm->flow_cache_gc_count);
-		WARN_ON(atomic_read(&xfrm->flow_cache_gc_count) < 0);
 	}
 }
 
@@ -236,9 +235,8 @@ flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
 		if (fcp->hash_count > fc->high_watermark)
 			flow_cache_shrink(fc, fcp);
 
-		if (fcp->hash_count > 2 * fc->high_watermark ||
-		    atomic_read(&net->xfrm.flow_cache_gc_count) > fc->high_watermark) {
-			atomic_inc(&net->xfrm.flow_cache_genid);
+		if (atomic_read(&net->xfrm.flow_cache_gc_count) >
+		    2 * num_online_cpus() * fc->high_watermark) {
 			flo = ERR_PTR(-ENOBUFS);
 			goto ret_object;
 		}