linuxcncrsh: Allow IPv4 as fallback when IPv6 is disabled - #4403
Conversation
| return -1; | ||
| } | ||
| } | ||
| setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &optval, sizeof(optval)); |
There was a problem hiding this comment.
Dual-stack here relies on net.ipv6.bindv6only defaulting to 0. With that sysctl
set to 1, socket(AF_INET6, ...) still succeeds, so the fallback never fires and
IPv4 clients get ECONNREFUSED. Measured:
v6only=0 IPv4 client: connected OK
v6only=1 IPv4 client: connect FAILED (Connection refused)
Worth pinning the assumption rather than inheriting it, on the IPv6 branch only?
int zero = 0;
setsockopt(sockfd, IPPROTO_IPV6, IPV6_V6ONLY, &zero, sizeof(zero));
| addr.addr6.sin6_port = htons(port); | ||
| socklen_t slen = sizeof(addr.addr6); | ||
|
|
||
| sockfd = socket(AF_INET6, SOCK_STREAM, 0); |
There was a problem hiding this comment.
Is EAFNOSUPPORT the only way the IPv6 listener can be unavailable in practice?
A restricted netns or an LSM/seccomp policy returns EACCES or EPERM here, and
there are setups where socket() succeeds but bind() to :: is what fails. Since
IPv4 is a strict fallback, any downside to taking it on any IPv6 failure?
| struct sockaddr_in6 csa6; | ||
| struct sockaddr_in csa4; | ||
| } csa; | ||
| socklen_t csal = isipv4 ? sizeof(csa.csa4) : sizeof(csa.csa6); |
There was a problem hiding this comment.
accept() writes the real length back, so does the family need to be known here?
sizeof(csa) is correct for both.
Broader: would sockaddr_storage plus getnameinfo(..., NI_NUMERICHOST |
NI_NUMERICSERV) let the union, isipv4, and the duplicated logging branch below
all disappear?
| { | ||
| if (ctx.sock < 0) | ||
| return; | ||
| if(ctx.cmdtimeout > 0.0) { |
There was a problem hiding this comment.
The latch clears, but the NML command is still in flight, so its completion or
error lands on whichever client polls next. Intended trade-off, or would
recording emcCommandSerialNumber at close time and discarding status up to it
be cheap enough to close the desync?
2 participants
The new linuxcncrsh server used v4-mapped-on-v6 to have one socket accept both IPv6 and IPv4 connections. However, if IPv6 is disabled at the kernel command line (or not compiled in), then it would fail to create a listening socket. Now there is a IPv4-only fallback in case IPv6 fails. This PR also fixes a lockup in the processing of SET commands when a client disconnects while a SET command is in flight.
Fixes #4400
Fixes #4401