kernel: add a RPS balancer

By default the RPS delegation will happen by masking the last few bits of
skb->hash. This patch adds an inermediate hash bucket that maps the masked
hash to a RPS core. This makes RPS results much more deterministic on SMP
systems.

Signed-off-by: John Crispin <john@phrozen.org>

diff --git a/target/linux/generic/hack-4.14/600-net-core-add-RPS-balancer.patch b/target/linux/generic/hack-4.14/600-net-core-add-RPS-balancer.patch
new file mode 100644 (file)
index 0000000..6fc52ee
--- /dev/null
+++ b/target/linux/generic/hack-4.14/600-net-core-add-RPS-balancer.patch
@@ -0,0 +1,93 @@
+From 3e969c9695b45e1a052d43b367096ec99f2f0aac Mon Sep 17 00:00:00 2001
+From: John Crispin <john@phrozen.org>
+Date: Thu, 10 Aug 2017 15:58:29 +0200
+Subject: [PATCH] net: core: add RPS balancer
+
+This patch adds a hash bucket based rps hash balancer.
+
+Signed-off-by: John Crispin <john@phrozen.org>
+---
+ net/core/dev.c | 57 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
+ 1 file changed, 56 insertions(+), 1 deletion(-)
+
+--- a/net/core/dev.c
++++ b/net/core/dev.c
+@@ -3626,6 +3626,58 @@ set_rps_cpu(struct net_device *dev, stru
+       return rflow;
+ }
+ 
++#define RPS_TBL_SIZE_SHIFT    10
++#define RPS_TBL_SIZE          (1 << RPS_TBL_SIZE_SHIFT)
++struct rps_table {
++      int                     core;
++      struct timer_list       expire;
++};
++static struct rps_table rps_table[RPS_TBL_SIZE];
++static int rps_table_last_core;
++
++static void rps_table_expire(unsigned long data)
++{
++      struct rps_table *entry = (struct rps_table *) data;
++
++      entry->core = -1;
++}
++
++static int rps_table_core(struct rps_map *map)
++{
++      int i;
++
++      for (i = 0; i < map->len; i++) {
++              int cpu = map->cpus[(rps_table_last_core + i + 1) % map->len];
++              if (cpu_online(cpu)) {
++                      rps_table_last_core = cpu;
++                      return cpu;
++              }
++      }
++      return map->cpus[0];
++}
++
++static int rps_table_lookup(struct rps_map *map, u32 hash)
++{
++      int bucket = hash & 0x3ff;
++
++      if (rps_table[bucket].core < 0)
++              rps_table[bucket].core = rps_table_core(map);
++      mod_timer(&rps_table[bucket].expire, jiffies + HZ);
++
++      return rps_table[bucket].core;
++}
++
++static void rps_table_init(void)
++{
++      int i;
++
++      for (i = 0; i < RPS_TBL_SIZE; i++) {
++              rps_table[i].core = -1;
++              setup_timer(&rps_table[i].expire, rps_table_expire,
++                          (unsigned long) &rps_table[i]);
++      }
++}
++
+ /*
+  * get_rps_cpu is called from netif_receive_skb and returns the target
+  * CPU from the RPS map of the receiving queue for a given skb.
+@@ -3715,7 +3767,7 @@ static int get_rps_cpu(struct net_device
+ try_rps:
+ 
+       if (map) {
+-              tcpu = map->cpus[reciprocal_scale(hash, map->len)];
++              tcpu = rps_table_lookup(map, hash);
+               if (cpu_online(tcpu)) {
+                       cpu = tcpu;
+                       goto done;
+@@ -8800,6 +8852,10 @@ static int __init net_dev_init(void)
+               sd->backlog.weight = weight_p;
+       }
+ 
++#ifdef CONFIG_RPS
++              rps_table_init();
++#endif
++
+       dev_boot_phase = 0;
+ 
+       /* The loopback device is special if any other network devices