On 6/7/19 10:31 AM, Sergej Benilov wrote:
> On Fri, 7 Jun 2019 at 19:26, Sergej Benilov
> <sergej.beni...@googlemail.com> wrote:
>>
>> 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
>>
>
> Thanks to Eric Dumazet for suggesting this patch
Note that this suggests the driver is not performing TX completion fast enough.
Looking at the driver, I do not see anything requesting interrupt mitigation,
so this might also be caused by a race in the driver (some skbs being not TX
completed until
another unrelated xmit is requested)
