On 7/11/19 9:28 AM, Christoph Paasch wrote:
>
>
>> On Jul 10, 2019, at 9:26 PM, Eric Dumazet <eric.duma...@gmail.com> wrote:
>>
>>
>>
>> On 7/10/19 8:53 PM, Prout, Andrew - LLSC - MITLL wrote:
>>>
>>> Our initial rollout was v4.14.130, but I reproduced it with v4.14.132 as
>>> well, reliably for the samba test and once (not reliably) with synthetic
>>> test I was trying. A patched v4.14.132 with this patch partially reverted
>>> (just the four lines from tcp_fragment deleted) passed the samba test.
>>>
>>> The synthetic test was a pair of simple send/recv test programs under the
>>> following conditions:
>>> -The send socket was non-blocking
>>> -SO_SNDBUF set to 128KiB
>>> -The receiver NIC was being flooded with traffic from multiple hosts (to
>>> induce packet loss/retransmits)
>>> -Load was on both systems: a while(1) program spinning on each CPU core
>>> -The receiver was on an older unaffected kernel
>>>
>>
>> SO_SNDBUF to 128KB does not permit to recover from heavy losses,
>> since skbs needs to be allocated for retransmits.
>
> Would it make sense to always allow the alloc in tcp_fragment when coming
> from __tcp_retransmit_skb() through the retransmit-timer ?
4.15+ kernels have :
if (unlikely((sk->sk_wmem_queued >> 1) > sk->sk_sndbuf &&
tcp_queue != TCP_FRAG_IN_WRITE_QUEUE)) {
Meaning that things like TLP will succeed.
Anything we add in TCP stack to overcome the SO_SNDBUF by twice the limit
_will_ be exploited at scale.
I am not sure we want to continue to support small SO_SNDBUF values, as this
makes no sense today.
We use 64 MB auto tuning limit, and /proc/sys/net/ipv4/tcp_notsent_lowat to 1
MB.
I would rather work (when net-next reopens) on better collapsing at rtx to
allow reduction of the overhead.
Something like :
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index
f6a9c95a48edb234e4d4e21bf585744fbaf9a0a7..d5c85986209cd162cf39edb787b1385cb2c8b630
100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -2860,7 +2860,7 @@ static int __net_init tcp_sk_init(struct net *net)
net->ipv4.sysctl_tcp_early_retrans = 3;
net->ipv4.sysctl_tcp_recovery = TCP_RACK_LOSS_DETECTION;
net->ipv4.sysctl_tcp_slow_start_after_idle = 1; /* By default, RFC2861
behavior. */
- net->ipv4.sysctl_tcp_retrans_collapse = 1;
+ net->ipv4.sysctl_tcp_retrans_collapse = 3;
net->ipv4.sysctl_tcp_max_reordering = 300;
net->ipv4.sysctl_tcp_dsack = 1;
net->ipv4.sysctl_tcp_app_win = 31;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index
d61264cf89ef66b229ecf797c1abfb7fcdab009f..05cd264f98b084f62eaf2ef9d6e14a392670d02c
100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -3015,8 +3015,6 @@ static bool tcp_collapse_retrans(struct sock *sk, struct
sk_buff *skb)
next_skb_size = next_skb->len;
- BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
-
if (next_skb_size) {
if (next_skb_size <= skb_availroom(skb))
skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
@@ -3054,8 +3052,6 @@ static bool tcp_collapse_retrans(struct sock *sk, struct
sk_buff *skb)
/* Check if coalescing SKBs is legal. */
static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
{
- if (tcp_skb_pcount(skb) > 1)
- return false;
if (skb_cloned(skb))
return false;
/* Some heuristics for collapsing over SACK'd could be invented */
@@ -3114,7 +3110,7 @@ int __tcp_retransmit_skb(struct sock *sk, struct sk_buff
*skb, int segs)
struct inet_connection_sock *icsk = inet_csk(sk);
struct tcp_sock *tp = tcp_sk(sk);
unsigned int cur_mss;
- int diff, len, err;
+ int diff, len, maxlen, err;
/* Inconclusive MTU probe */
@@ -3165,12 +3161,13 @@ int __tcp_retransmit_skb(struct sock *sk, struct
sk_buff *skb, int segs)
return -ENOMEM;
diff = tcp_skb_pcount(skb);
+ maxlen = (sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse & 2) ?
len : cur_mss;
+ if (skb->len < maxlen)
+ tcp_retrans_try_collapse(sk, skb, maxlen);
tcp_set_skb_tso_segs(skb, cur_mss);
diff -= tcp_skb_pcount(skb);
if (diff)
tcp_adjust_pcount(sk, skb, diff);
- if (skb->len < cur_mss)
- tcp_retrans_try_collapse(sk, skb, cur_mss);
}
/* RFC3168, section 6.1.1.1. ECN fallback */
