On Sun, 9 Jun 2019, iam@xxxxxxxxxxx wrote: > While playing with eBPF i tried tcpstates.py > (https://github.com/iovisor/bcc/blob/master/tools/tcpstates.py) and > noticed very low time for TCP state transition SYN_RECV -> ESTABLISHED > > My numbers from tests: > > SYN_RECV -> ESTABLISHED 0.015 > SYN_RECV -> ESTABLISHED 0.017 > SYN_RECV -> ESTABLISHED 0.051 > > From handshake > diagram(https://user-images.githubusercontent.com/1006307/58944706-0ffdb580-878b-11e9-95d3-8e7a4f85d8b0.png) > it looks like that transition time from SYN_RECV to ESTABLISHED should > be near RTT between hosts? > > TCP Fast Open turned off on host. > > OS: Ubuntu 18.10 (GNU/Linux 4.18.0-21-generic x86_64) > > tcpdump -tttttv > > 00:00:00.000000 IP (tos 0x28, ttl 49, id 280, offset 0, flags [DF], proto TCP (6), length 60) > dst_host > src_host: Flags [S], cksum 0x46ae (correct), seq 4063608731, win 29200, options [mss 1460,sackOK,TS val 332512899 ecr 0,nop,wscale 6], length 0 > 00:00:00.000071 IP (tos 0x0, ttl 64, id 0, offset 0, flags [DF], proto TCP (6), length 60) > src_host > dst_host: Flags [S.], cksum 0x84c2 (incorrect -> 0x3c01), seq 1516815880, ack 4063608732, win 28960, options [mss 1460,sackOK,TS val 2341429130 ecr 332512899,nop,wscale 7], length 0 > 00:00:00.079183 IP (tos 0x28, ttl 49, id 281, offset 0, flags [DF], proto TCP (6), length 52) > dst_host > src_host: Flags [.], cksum 0xda11 (correct), ack 1, win 457, options [nop,nop,TS val 332512918 ecr 2341429130], length 0 > 00:00:00.093311 IP (tos 0x0, ttl 64, id 60333, offset 0, flags [DF], proto TCP (6), length 93) > src_host > dst_host: Flags [P.], cksum 0x84e3 (incorrect -> 0xb341), seq 1:42, ack 1, win 227, options [nop,nop,TS val 2341429223 ecr 332512918], length 41 > such low numbers are between syn_recv and syn+ack send, but between > syn_recv and established it should be 00:00:00.079183 > > ping between hosts in this tcpdump is around 83ms > > Looks like nf_conntrack is doing this: > > https://github.com/torvalds/linux/blob/80f232121b69cc69a31ccb2b38c1665d770b0710/net/netfilter/nf_conntrack_proto_tcp.c#L1033-L1037 > > Is it bug or i'm wrong? That part deals with simultaneous open. Please explain what do you mean? Conntrack must deal with slow/long delay network connections, therefore the internal timeouts are conservative. Best regards, Jozsef - E-mail : kadlec@xxxxxxxxxxxxxxxxx, kadlecsik.jozsef@xxxxxxxxxxxxx PGP key : http://www.kfki.hu/~kadlec/pgp_public_key.txt Address : Wigner Research Centre for Physics, Hungarian Academy of Sciences H-1525 Budapest 114, POB. 49, Hungary
