summaryrefslogtreecommitdiffstats
path: root/net/ipv4/udp.c
diff options
context:
space:
mode:
authorEric Dumazet <edumazet@google.com>2016-12-07 09:19:33 -0800
committerDavid S. Miller <davem@davemloft.net>2016-12-08 13:25:07 -0500
commitc8c8b127091b758f5768f906bcdeeb88bc9951ca (patch)
tree6721fe5d6de0ca0ddd61b4356539d1621538c2bd /net/ipv4/udp.c
parent2408022eeada9b0d96cb6a40bccf0ec2aa280bab (diff)
downloadop-kernel-dev-c8c8b127091b758f5768f906bcdeeb88bc9951ca.zip
op-kernel-dev-c8c8b127091b758f5768f906bcdeeb88bc9951ca.tar.gz
udp: under rx pressure, try to condense skbs
Under UDP flood, many softirq producers try to add packets to UDP receive queue, and one user thread is burning one cpu trying to dequeue packets as fast as possible. Two parts of the per packet cost are : - copying payload from kernel space to user space, - freeing memory pieces associated with skb. If socket is under pressure, softirq handler(s) can try to pull in skb->head the payload of the packet if it fits. Meaning the softirq handler(s) can free/reuse the page fragment immediately, instead of letting udp_recvmsg() do this hundreds of usec later, possibly from another node. Additional gains : - We reduce skb->truesize and thus can store more packets per SO_RCVBUF - We avoid cache line misses at copyout() time and consume_skb() time, and avoid one put_page() with potential alien freeing on NUMA hosts. This comes at the cost of a copy, bounded to available tail room, which is usually small. (We might have to fix GRO_MAX_HEAD which looks bigger than necessary) This patch gave me about 5 % increase in throughput in my tests. skb_condense() helper could probably used in other contexts. Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Paolo Abeni <pabeni@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net/ipv4/udp.c')
-rw-r--r--net/ipv4/udp.c12
1 files changed, 11 insertions, 1 deletions
diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 16d88ba..f5628ad 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c
@@ -1199,7 +1199,7 @@ int __udp_enqueue_schedule_skb(struct sock *sk, struct sk_buff *skb)
{
struct sk_buff_head *list = &sk->sk_receive_queue;
int rmem, delta, amt, err = -ENOMEM;
- int size = skb->truesize;
+ int size;
/* try to avoid the costly atomic add/sub pair when the receive
* queue is full; always allow at least a packet
@@ -1208,6 +1208,16 @@ int __udp_enqueue_schedule_skb(struct sock *sk, struct sk_buff *skb)
if (rmem > sk->sk_rcvbuf)
goto drop;
+ /* Under mem pressure, it might be helpful to help udp_recvmsg()
+ * having linear skbs :
+ * - Reduce memory overhead and thus increase receive queue capacity
+ * - Less cache line misses at copyout() time
+ * - Less work at consume_skb() (less alien page frag freeing)
+ */
+ if (rmem > (sk->sk_rcvbuf >> 1))
+ skb_condense(skb);
+ size = skb->truesize;
+
/* we drop only if the receive buf is full and the receive
* queue contains some other skb
*/
OpenPOWER on IntegriCloud