net/dst: use a smaller percpu_counter batch for dst entries accounting

percpu_counter_add() uses a default batch size which is quite big
on platforms with 256 cpus. (2*256 -> 512)

This means dst_entries_get_fast() can be off by +/- 2*(nr_cpus^2)
(131072 on servers with 256 cpus)

Reduce the batch size to something more reasonable, and
add logic to ip6_dst_gc() to call dst_entries_get_slow()
before calling the _very_ expensive fib6_run_gc() function.

Signed-off-by: Eric Dumazet <[email protected]>
Signed-off-by: Jakub Kicinski <[email protected]>

Author: edumazet

include/net/dst_ops.h

@@ -53,9 +53,11 @@ static inline int dst_entries_get_slow(struct dst_ops *dst)
 	return percpu_counter_sum_positive(&dst->pcpuc_entries);
 }
 
+#define DST_PERCPU_COUNTER_BATCH 32
 static inline void dst_entries_add(struct dst_ops *dst, int val)
 {
-	percpu_counter_add(&dst->pcpuc_entries, val);
+	percpu_counter_add_batch(&dst->pcpuc_entries, val,
+				 DST_PERCPU_COUNTER_BATCH);
 }
 
 static inline int dst_entries_init(struct dst_ops *dst)

net/core/dst.c

@@ -81,11 +81,11 @@ void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
 {
 	struct dst_entry *dst;
 
-	if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
+	if (ops->gc &&
+	    !(flags & DST_NOCOUNT) &&
+	    dst_entries_get_fast(ops) > ops->gc_thresh) {
 		if (ops->gc(ops)) {
-			printk_ratelimited(KERN_NOTICE "Route cache is full: "
-					   "consider increasing sysctl "
-					   "net.ipv[4|6].route.max_size.\n");
+			pr_notice_ratelimited("Route cache is full: consider increasing sysctl net.ipv6.route.max_size.\n");
 			return NULL;
 		}
 	}

net/ipv6/route.c

@@ -3195,6 +3195,9 @@ static int ip6_dst_gc(struct dst_ops *ops)
 	int entries;
 
 	entries = dst_entries_get_fast(ops);
+	if (entries > rt_max_size)
+		entries = dst_entries_get_slow(ops);
+
 	if (time_after(rt_last_gc + rt_min_interval, jiffies) &&
 	    entries <= rt_max_size)
 		goto out;