linux: Allow connect() on DGRAM sockets with port 0

[obiwac]

Summary

Linux allows connect() with destination port 0, but FreeBSD returns
EADDRNOTAVAIL in in_pcbconnect() and in6_pcbladdr(). Some Linux programs
(e.g. the new Android CLI tool) rely on this for routability checks,
however.

Add an SV_CONN_PORT0 flag to sv_flags which is set in the Linux sysvecs,
allowing the port 0 guard in the connect paths to be bypassed.

Test Plan

I have compiled a program that tries to connect to the zero port on FreeBSD and on Linux. Running the FreeBSD binary on FreeBSD, connect() fails with EADDRNOTAVAIL, which is the normal behaviour on FreeBSD. Running the Linux binary on FreeBSD, connect() succeeds.

The Android CLI also now works on FreeBSD.

[pouria]

I also faced this problem yesterday.

sys/netinet/in_pcb.c

1128

Maybe update this line to something like "unless INP_CONNECT_PORT_ZERO specified"

1154

sys/netinet6/in6_pcb.c

403

[obiwac]

respond to review comments

[obiwac]

sys/netinet/in_pcb.c

1128

well you do still have to set port = 0 in that situation

[glebius]

Usually in the kernel when we do Linux compatibility checks we check SV_PROC_ABI(p) == SV_ABI_LINUX. The sockets may outlive the process, so for a worst case we would put the flag on the socket. However in a connect(2) syscall the process is always present. Other important moment is that sockets may also be shared between processes. Weird to share one between a Linux process and FreeBSD process, but possible.

The correct patch is to modify only the in_pcbconnect() and in6_pcbconnect() to check if SV_PROC_ABI(curthread->td_proc) == SV_ABI_LINUX.

[kib]

In D60659#1391787, @glebius wrote:

Usually in the kernel when we do Linux compatibility checks we check SV_PROC_ABI(p) == SV_ABI_LINUX.

No, we prefer to have flags in sv_sysent to qualify modifications to the common syscall code paths, See the suggestion in D60701 and flags like SV_SIG_DISCIGN

[glebius]

In D60659#1391791, @kib wrote:

No, we prefer to have flags in sv_sysent to qualify modifications to the common syscall code paths, See the suggestion in D60701 and flags like SV_SIG_DISCIGN

Yes, that would be even better, cause we may have yet another ABI that wants this behavior.

[obiwac]

make INP_CONNECT_PORT_ZERO a sysvec flag (SV_CONN_PORT0) instead

[obiwac]

remove useless whitespace change

[glebius]

Thanks! LGTM Let's get ack from @kib and give some time to emulation people to review.

[glebius]

Note: commit message singles out the datagram sockets, but the change applies to all sockets. How does it work on Linux? We should emulate their bug completely.

[obiwac]

In D60659#1392229, @glebius wrote:

Note: commit message singles out the datagram sockets, but the change applies to all sockets. How does it work on Linux? We should emulate their bug completely.

yeah, linux allows ports 0 on all sockets. it's because the previous diff was adding the flag just to dgram sockets. commit message amended