tun: Add support for RFS on tun flows
authorTom Herbert <therbert@google.com>
Sun, 22 Dec 2013 10:54:32 +0000 (18:54 +0800)
committerDavid S. Miller <davem@davemloft.net>
Tue, 31 Dec 2013 18:31:34 +0000 (13:31 -0500)
This patch adds support so that the rps_flow_tables (RFS) can be
programmed using the tun flows which are already set up to track flows
for the purposes of queue selection.

On the receive path (corresponding to select_queue and tun_net_xmit) the
rxhash is saved in the flow_entry.  The original code only does flow
lookup in select_queue, so this patch adds a flow lookup in tun_net_xmit
if num_queues == 1 (select_queue is not called from
dev_queue_xmit->netdev_pick_tx in that case).

The flow is recorded (processing CPU) in tun_flow_update (TX path), and
reset when flow is deleted.

Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/tun.c

index a17a701..3cf0457 100644 (file)
@@ -152,6 +152,7 @@ struct tun_flow_entry {
        struct tun_struct *tun;
 
        u32 rxhash;
+       u32 rps_rxhash;
        int queue_index;
        unsigned long updated;
 };
@@ -220,6 +221,7 @@ static struct tun_flow_entry *tun_flow_create(struct tun_struct *tun,
                          rxhash, queue_index);
                e->updated = jiffies;
                e->rxhash = rxhash;
+               e->rps_rxhash = 0;
                e->queue_index = queue_index;
                e->tun = tun;
                hlist_add_head_rcu(&e->hash_link, head);
@@ -232,6 +234,7 @@ static void tun_flow_delete(struct tun_struct *tun, struct tun_flow_entry *e)
 {
        tun_debug(KERN_INFO, tun, "delete flow: hash %u index %u\n",
                  e->rxhash, e->queue_index);
+       sock_rps_reset_flow_hash(e->rps_rxhash);
        hlist_del_rcu(&e->hash_link);
        kfree_rcu(e, rcu);
        --tun->flow_count;
@@ -325,6 +328,7 @@ static void tun_flow_update(struct tun_struct *tun, u32 rxhash,
                /* TODO: keep queueing to old queue until it's empty? */
                e->queue_index = queue_index;
                e->updated = jiffies;
+               sock_rps_record_flow_hash(e->rps_rxhash);
        } else {
                spin_lock_bh(&tun->lock);
                if (!tun_flow_find(head, rxhash) &&
@@ -341,6 +345,18 @@ unlock:
        rcu_read_unlock();
 }
 
+/**
+ * Save the hash received in the stack receive path and update the
+ * flow_hash table accordingly.
+ */
+static inline void tun_flow_save_rps_rxhash(struct tun_flow_entry *e, u32 hash)
+{
+       if (unlikely(e->rps_rxhash != hash)) {
+               sock_rps_reset_flow_hash(e->rps_rxhash);
+               e->rps_rxhash = hash;
+       }
+}
+
 /* We try to identify a flow through its rxhash first. The reason that
  * we do not check rxq no. is because some cards(e.g 82599), chooses
  * the rxq based on the txq where the last packet of the flow comes. As
@@ -361,9 +377,10 @@ static u16 tun_select_queue(struct net_device *dev, struct sk_buff *skb)
        txq = skb_get_hash(skb);
        if (txq) {
                e = tun_flow_find(&tun->flows[tun_hashfn(txq)], txq);
-               if (e)
+               if (e) {
                        txq = e->queue_index;
-               else
+                       tun_flow_save_rps_rxhash(e, txq);
+               } else
                        /* use multiply and shift instead of expensive divide */
                        txq = ((u64)txq * numqueues) >> 32;
        } else if (likely(skb_rx_queue_recorded(skb))) {
@@ -728,6 +745,22 @@ static netdev_tx_t tun_net_xmit(struct sk_buff *skb, struct net_device *dev)
        if (txq >= tun->numqueues)
                goto drop;
 
+       if (tun->numqueues == 1) {
+               /* Select queue was not called for the skbuff, so we extract the
+                * RPS hash and save it into the flow_table here.
+                */
+               __u32 rxhash;
+
+               rxhash = skb_get_hash(skb);
+               if (rxhash) {
+                       struct tun_flow_entry *e;
+                       e = tun_flow_find(&tun->flows[tun_hashfn(rxhash)],
+                                       rxhash);
+                       if (e)
+                               tun_flow_save_rps_rxhash(e, rxhash);
+               }
+       }
+
        tun_debug(KERN_INFO, tun, "tun_net_xmit %d\n", skb->len);
 
        BUG_ON(!tfile);