net: bulk free SKBs that were delay free'ed due to IRQ context
authorJesper Dangaard Brouer <brouer@redhat.com>
Mon, 8 Feb 2016 12:15:04 +0000 (13:15 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 11 Feb 2016 16:59:09 +0000 (11:59 -0500)
The network stack defers SKBs free, in-case free happens in IRQ or
when IRQs are disabled. This happens in __dev_kfree_skb_irq() that
writes SKBs that were free'ed during IRQ to the softirq completion
queue (softnet_data.completion_queue).

These SKBs are naturally delayed, and cleaned up during NET_TX_SOFTIRQ
in function net_tx_action().  Take advantage of this a use the skb
defer and flush API, as we are already in softirq context.

For modern drivers this rarely happens. Although most drivers do call
dev_kfree_skb_any(), which detects the situation and calls
__dev_kfree_skb_irq() when needed.  This due to netpoll can call from
IRQ context.

Signed-off-by: Alexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/core/dev.c
net/core/skbuff.c

index b56c0103fa15e64ee85dcd07dee5519f275fa59c..6ec86f1a2ed96d809e4a6a4579c0325886e77bb5 100644 (file)
@@ -2407,6 +2407,7 @@ static inline struct sk_buff *napi_alloc_skb(struct napi_struct *napi,
 void napi_consume_skb(struct sk_buff *skb, int budget);
 
 void __kfree_skb_flush(void);
+void __kfree_skb_defer(struct sk_buff *skb);
 
 /**
  * __dev_alloc_pages - allocate page for network Rx
index 9b2c7a999e71442e3354104436bf005cae1bc716..3f4071a84a03fb018fb97d881d55eee68a015326 100644 (file)
@@ -3829,8 +3829,14 @@ static void net_tx_action(struct softirq_action *h)
                                trace_consume_skb(skb);
                        else
                                trace_kfree_skb(skb, net_tx_action);
-                       __kfree_skb(skb);
+
+                       if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
+                               __kfree_skb(skb);
+                       else
+                               __kfree_skb_defer(skb);
                }
+
+               __kfree_skb_flush();
        }
 
        if (sd->output_queue) {
index b64187b8777386cd98a84b82aa6316a41e96c94c..a5bd067ec1a3b85b59249129e8e4d9813e16ca7a 100644 (file)
@@ -767,7 +767,7 @@ void __kfree_skb_flush(void)
        }
 }
 
-static void __kfree_skb_defer(struct sk_buff *skb)
+static inline void _kfree_skb_defer(struct sk_buff *skb)
 {
        struct napi_alloc_cache *nc = this_cpu_ptr(&napi_alloc_cache);
 
@@ -789,6 +789,10 @@ static void __kfree_skb_defer(struct sk_buff *skb)
                nc->skb_count = 0;
        }
 }
+void __kfree_skb_defer(struct sk_buff *skb)
+{
+       _kfree_skb_defer(skb);
+}
 
 void napi_consume_skb(struct sk_buff *skb, int budget)
 {
@@ -814,7 +818,7 @@ void napi_consume_skb(struct sk_buff *skb, int budget)
                return;
        }
 
-       __kfree_skb_defer(skb);
+       _kfree_skb_defer(skb);
 }
 EXPORT_SYMBOL(napi_consume_skb);