Since commit 605ad7f184b60cfaacbc038aa6c55ee68dee3c89 "tcp: refine TSO
autosizing",
the TSQ limit is computed as the smaller of
sysctl_tcp_limit_output_bytes and max(2 * skb->truesize, sk->sk_pacing_rate >>
10).
For some connections this approach sets a low limit, reducing throughput
dramatically.
Add a call to skb_orphan() to sis900_start_xmit()
to speed up packets delivery from the kernel to the driver.
Test:
netperf -H remote -l -2000000 -- -s 1000000
before patch:
MIGRATED TCP STREAM TEST from 0.0.0.0 () port 0 AF_INET to remote () port 0
AF_INET : demo
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
87380 327680 327680 341.79 0.05
after patch:
MIGRATED TCP STREAM TEST from 0.0.0.0 () port 0 AF_INET to remote () port 0
AF_INET : demo
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
87380 327680 327680 1.29 12.54
Signed-off-by: Sergej Benilov <sergej.beni...@googlemail.com>
---
drivers/net/ethernet/sis/sis900.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/net/ethernet/sis/sis900.c
b/drivers/net/ethernet/sis/sis900.c
index fd812d2e..ca17b50c 100644
--- a/drivers/net/ethernet/sis/sis900.c
+++ b/drivers/net/ethernet/sis/sis900.c
@@ -1604,6 +1604,7 @@ sis900_start_xmit(struct sk_buff *skb, struct net_device
*net_dev)
unsigned int index_cur_tx, index_dirty_tx;
unsigned int count_dirty_tx;
+ skb_orphan(skb);
spin_lock_irqsave(&sis_priv->lock, flags);
/* Calculate the next Tx descriptor entry. */
--
2.17.1
