On 2018/12/6 下午4:08, Li RongQing wrote:
tun flow entry 'updated' fields are written when receive
every packet. Thus if a flow is receiving packets from a
particular flow entry, it'll cause false-sharing with
all the other who has looked it up, so move it in its own
cache line
and update 'queue_index' and 'update' field only when
they are changed to reduce the cache false-sharing.
Signed-off-by: Zhang Yu <zhangy...@baidu.com>
Signed-off-by: Wang Li <wangl...@baidu.com>
Signed-off-by: Li RongQing <lirongq...@baidu.com>
---
drivers/net/tun.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/net/tun.c b/drivers/net/tun.c
index 835c73f42ae7..d0745dc81976 100644
--- a/drivers/net/tun.c
+++ b/drivers/net/tun.c
@@ -201,7 +201,7 @@ struct tun_flow_entry {
u32 rxhash;
u32 rps_rxhash;
int queue_index;
- unsigned long updated;
+ unsigned long updated ____cacheline_aligned_in_smp;
};
#define TUN_NUM_FLOW_ENTRIES 1024
@@ -539,8 +539,10 @@ static void tun_flow_update(struct tun_struct *tun, u32
rxhash,
e = tun_flow_find(head, rxhash);
if (likely(e)) {
/* TODO: keep queueing to old queue until it's empty? */
- e->queue_index = queue_index;
- e->updated = jiffies;
+ if (e->queue_index != queue_index)
+ e->queue_index = queue_index;
+ if (e->updated != jiffies)
+ e->updated = jiffies;
Acked-by: Jason Wang <jasow...@redhat.com>
Btw, do you have perf numbers for this?
Thanks
sock_rps_record_flow_hash(e->rps_rxhash);
} else {
spin_lock_bh(&tun->lock);