TCP connection locally isn't possible

[zephyrbot]

Reported by Paul Sokolovsky:

Even with GH-3411 fixed and local UDP communication working OK, TCP connection from one local context to another is not possible. What happens is that connect() call seemingly succeeds, but listening socket never has accept callback called. Enabling debug logging, following can be seen:

[net/ctx] [DBG] net_context_bind: (0x0013e050): Context 0x00136ce0 binding to TCP 192.0.2.1:5555 iface 0x00136760
[net/tcp] [DBG] net_tcp_change_state: (0x0013e050): state@0x00136fa0 CLOSED (0) => LISTEN (1)
[net/conn] [DBG] net_conn_register: (0x0013e050): [0/2/6/0x11] remote 0x00000000/-/0 local 0x0013ee90/192.0.2.1/5555 cb 0x00109dba ud 0x00136ce0
=============== lsock: <socket 136ce0 type=2>
[net/conn] [DBG] net_conn_register: (0x0013e050): [1/65363/6/0x23] remote 0x0013eeb0/127.0.0.1/5555 local 0x00000000/-/48205 cb 0x0010945c ud 0x00136d38
[net/tcp] [DBG] net_tcp_change_state: (0x0013e050): state@0x00137038 CLOSED (0) => SYN_SENT (2)
[net/tcp] [DBG] net_tcp_trace: (0x0013e050): buf 0x0013d6c0 src 48205 dst 5555 seq 0x3db82d47 ack 0x0000 (0) flags uaprSf win 1280 chk 0x2ce0
src addr: 10200c0
net_send_data(13d6c0): 1
[net/conn] [DBG] net_conn_input: (0x0013e050): Check TCP listener for buf 0x0013d6c0 src port 48205 dst port 5555 family 2 chksum 0x2ce0
[net/conn] [DBG] net_conn_input: (0x0013e050): [0] match found cb 0x00109dba ud 0x00136ce0 rank 0x11
[net/ctx] [DBG] _tcp_syn_rcvd: (0x0013e050): SYN received from 127.0.0.1 port 48205
[net/tcp] [DBG] net_tcp_change_state: (0x0013e050): state@0x00136fa0 LISTEN (1) => SYN_RCVD (3)
[net/tcp] [DBG] net_tcp_trace: (0x0013e050): buf 0x0013d708 src 5555 dst 48205 seq 0x3d389dfa ack 0x3db82d48 (0) flags uAprSf win 1280 chk 0x519c
src addr: 10200c0
net_send_data(13d708): 1
[net/conn] [DBG] net_conn_input: (0x0013e050): Check TCP listener for buf 0x0013d708 src port 5555 dst port 48205 family 2 chksum 0x519c
[net/conn] [DBG] net_conn_input: (0x0013e050): [1] match found cb 0x0010945c ud 0x00136d38 rank 0x23
*** remove temp conn, replace w/establ
[net/conn] [DBG] net_conn_unregister: (0x0013e050): [0] connection handler 0x001373bc removed
[net/conn] [DBG] net_conn_register: (0x0013e050): [1/2/6/0x33] remote 0x0013eba0/127.0.0.1/5555 local 0x0013eba8/192.0.2.1/48205 cb 0x00109778 ud 0x00136d38
[net/tcp] [DBG] net_tcp_change_state: (0x0013e050): state@0x00137038 SYN_SENT (2) => ESTABLISHED (4)
[net/tcp] [DBG] prepare_segment: (0x0013e050): Protocol family 0 not supported
[net/ctx] [DBG] send_control_segment: (0x0013e050): SYN_ACK sent to 0.0.0.0 port 0
[net/ctx] [DBG] send_control_segment: (0x0013e050): SYN sent to 0.0.0.0 port 0
Setting recv cb after connect()
========== client <socket 136d38 type=2>
before accept
net_send_data(13d6c0): 0
[net/ctx] [DBG] ack_timeout: (0x0013f050): Did not receive ACK in 1000ms
[net/tcp] [DBG] net_tcp_trace: (0x0013f050): buf 0x0013d6c0 src 5555 dst 0 seq 0x0000 ack 0x3db82d48 (0) flags uApRsf win 0 chk 0x6d1c
[net/ctx] [DBG] send_reset: (0x0013f050): RST sent to 0.0.0.0 port 0
net_send_data(13d6c0): -125
[net/tcp] [DBG] net_tcp_change_state: (0x0013f050): state@0x00136fa0 SYN_RCVD (3) => LISTEN (1)
net_send_data(13d6c0): 0

There're at least 2 problem spots:

• The fact that "net_conn_unregister: (0x0013e050): [0]" called, this doesn't happen when connecting from outside
• SYN ACK, instead of being sent back to local connecting socket, is sent via an empty TCP segment: "send_control_segment: (0x0013e050): SYN_ACK sent to 0.0.0.0 port 0"

I thought 1 is the route cause of 2, but even commenting out net_conn_unregister, "SYN_ACK sent to 0.0.0.0 port 0" still happens.

Overall, I spent an hour trying different hacks and workarounds, and can't get thru to what's wrong exactly. Connection handling is definitely not good, but there's a big problem beyond.

I guess this is not high priority, but it shows that some things in the stack work not general enough, so it would be nice to get to the root cause of this.

(Imported from Jira ZEP-2001)

[zephyrbot]

by Jukka Rissanen:

There are two patches applied already to net that should help with loopback TCP connections:

net: context: Set the local port correctly in accept
net: context: Connect callback was set too late

I also created a test application (echo-loopback) that tests the UDP and TCP loopback connectivity. It is currently still in DRAFT as there is a memory leak somewhere and I have not yet found it. https://gerrit.zephyrproject.org/r/#/c/13077/

[zephyrbot]

by Vakul Garg:

• loopback.tgz (size: 1929)

[zephyrbot]

by Vakul Garg:

Attached a test case. It creates a TCP server socket listening on loopback address in a thread.
From main function, a client socket connected to TCP server socket and starts sending data.

For each data sent, following errors appear:

[net/tcp] [ERR] net_tcp_get_hdr: {assert: 'frag' failed}
[net/tcp] [ERR] net_tcp_ack_received: pkt 0x0040d524 has no TCP header

To compile the test case, unpack the attached tarball in zephyr/samples/net.

[nashif]

@pfalcon @jukkar is this still an issue, please update status.

[pfalcon]

Will need to recheck, yeah.

[pfalcon]

Didn't have chance to retest, would target for 1.14 re-release testing.

[jukkar]

TCP should work if CONFIG_NET_LOOPBACK is enabled. Probably does not work if that is missing.

[jukkar]

This is old one and should work if CONFIG_NET_LOOPBACK is enabled -> closing for now.