Re: [PATCH net] tcp: refine memory limit test in tcp_fragment()

[Eric Dumazet <eric.dumazet@gmail.com>]

On 7/2/19 8:27 PM, Tony Lu wrote:
> Hello Eric,
> 
> 	We have applied that commit e358f4af19db ("tcp: tcp_fragment() should apply sane memory limits")
> 	as a hotpatch in production environment. We found that it will make
> 	tcp long connection reset during sending out packet when applying
> 	that commit. 
> 	
> 	Our applications which in A/B test have suffered that
> 	and made them retransmit large data, and then caused retransmission
> 	storm and lower the performance and increase RT.
> 
> 	Therefore we discontinued to apply this hotpatch in A/B test.
> 
> 	After invesgation, we found this patch already fix this issue in
> 	stable. Before applying this patch, we have some questions:
> 

Which stable version are you referring to exactly ?

> 	1. This commit in stable hard coded a magic number 0x20000. I am
> 	wondering this value and if there any better solution.

0x20000 is two times 64KB, please read the changelog for the rationale.

> 	2. Is there any known or unknown side effect? If any, we could test
> 	it in some suspicious scenarios before testing in prod env.

No known side effect.

Honestly, applications setting small SO_SNDBUF values can not expect good TCP performance anyway.


> 
> 	Thanks.
> 
> Cheers,
> Tony Lu
> 
> On Fri, Jun 21, 2019 at 06:09:55AM -0700, Eric Dumazet wrote:
>> tcp_fragment() might be called for skbs in the write queue.
>>
>> Memory limits might have been exceeded because tcp_sendmsg() only
>> checks limits at full skb (64KB) boundaries.
>>
>> Therefore, we need to make sure tcp_fragment() wont punish applications
>> that might have setup very low SO_SNDBUF values.
>>
>> Fixes: f070ef2ac667 ("tcp: tcp_fragment() should apply sane memory limits")
>> Signed-off-by: Eric Dumazet <edumazet@google.com>
>> Reported-by: Christoph Paasch <cpaasch@apple.com>
>> ---
>>  net/ipv4/tcp_output.c | 3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
>> index 00c01a01b547ec67c971dc25a74c9258563cf871..0ebc33d1c9e5099d163a234930e213ee35e9fbd1 100644
>> --- a/net/ipv4/tcp_output.c
>> +++ b/net/ipv4/tcp_output.c
>> @@ -1296,7 +1296,8 @@ int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
>>  	if (nsize < 0)
>>  		nsize = 0;
>>  
>> -	if (unlikely((sk->sk_wmem_queued >> 1) > sk->sk_sndbuf)) {
>> +	if (unlikely((sk->sk_wmem_queued >> 1) > sk->sk_sndbuf &&
>> +		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE)) {
>>  		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
>>  		return -ENOMEM;
>>  	}
>> -- 
>> 2.22.0.410.gd8fdbe21b5-goog