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: 6ad3122a08e3 ("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>

diff --git a/net/core/flow.c b/net/core/flow.c
index 3937b1b68d5bc7ad50691716ac1612332a5dc997..18e8893d4be59905ba1053fffea7cc16d0e054af 100644 (file)
--- a/net/core/flow.c
+++ b/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;
                }