Fix UDP ENOBUFS handling and multicast interface quoting (#5486)

[pepone]

Addresses #5486 (items 9, 10). One of a set of PRs splitting the grouped transport audit.

• Item 9 (C++/Unix): a transient ENOBUFS on a UDP send is routine on macOS/BSD during bursts — a transient local condition, not a connection failure. It was treated as fatal and closed the connection; since UDP is best-effort, UdpTransceiver::write now drops the datagram instead of throwing. Limited to non-Windows, because noBuffers() also matches WSAEFAULT there, which must not be silently dropped.
• Item 10 (C++, C#, Java): a multicast --interface value was quoted only when it contained :, so a Windows adapter friendly name with a space (e.g. "Ethernet 2") was emitted unquoted and failed to re-parse. It is now quoted whenever it contains a separator the parser splits on (whitespace or :). The same bug existed in all three mappings; JS has no UdpEndpointI (N/A).

Item 10 has a red→green round-trip test in Ice/proxy for C++, C#, and Java; Ice/udp passes with no regression.

No changelog: with the catastrophic server-adapter case (bidirectional UDP, slated for removal in #5362) out of scope, item 9's remaining impact — a transient ENOBUFS burst on UDP sends no longer closing the connection — is minor, consistent with how the #5679 fixes were treated. Codex-audited.

[Copilot]

Pull request overview

This PR addresses two transport-layer correctness issues in the C++ implementation related to UDP behavior and proxy string round-tripping, as tracked in #5486 (items 9 and 10).

Changes:

• Treat transient UDP ENOBUFS (non-Windows) as a best-effort drop rather than a fatal error that closes the UDP connection.
• Quote --interface in UDP endpoint stringification when it contains parser separators (e.g., whitespace), fixing proxy ice_toString() round-trip for Windows-friendly adapter names.
• Add a proxy round-trip test covering an interface name containing a space.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
cpp/test/Ice/proxy/AllTests.cpp	Adds a round-trip test ensuring --interface "Ethernet 2" survives ice_toString() re-parse.
cpp/src/Ice/UdpTransceiver.cpp	Drops a datagram on non-Windows ENOBUFS instead of throwing/closing the UDP connection.
cpp/src/Ice/UdpEndpointI.cpp	Updates quoting logic for --interface to include whitespace and other separators.

[pepone]

It is not clear why Windows include WSAEFAULT in the noBuffers, seems unrelated to the use we have in StreamSocket for it.,

[pepone]

recv doesn't return WSAENOBUFS — only WSAEFAULT (send returns both). So in noBuffers() (which drives the StreamSocket read/write halve-and-retry), WSAEFAULT is the recv-side half of the pair. But retrying it isn't actually required: halving the length can't validate a bad pointer, and we always pass a correctly-sized buffer. Looks defensive/legacy (no rationale in git, predates 2012), and it's harmless only because it degrades to a throw. Unrelated to this UDP send path, which is why the drop stays non-Windows for now.

[bernardnormier]

The two fixes are correct. A few requests before approval:

1. The ENOBUFS comment is confusing. It's in write() (sending), but justifies the change via a server's UDP adapter receiving datagrams. The only way a server reaches write() is bidirectional UDP (the sendto(_peerAddr) reply path) over the socket it also receives on — and that feature is undocumented, half-implemented, and slated for removal (#5362). So the rationale shouldn't lean on it. The best-effort argument stands on its own (and covers the common client-send case too). Suggestion inline.

2. PR description — same issue: please drop the "destroying the never-recreated shared incoming-datagram connection" framing (that's the bidir/server-adapter story being removed in #5362) and justify item 9 on the client best-effort send path instead.

3. No changelog needed. With the catastrophic server-adapter case going away, item 9's remaining impact (a transient ENOBUFS burst on UDP sends no longer closing the connection) is minor, consistent with how we treated the #5679 fixes.

4. Cross-mapping: item 10 is duplicated in C# and Java — both quote the multicast interface on : only and have the identical round-trip bug for a name with a space:

• csharp/src/Ice/Internal/UdpEndpointI.cs: _mcastInterface.Contains(':', ...)
• java/.../com/zeroc/Ice/UdpEndpointI.java: _mcastInterface.indexOf(':') != -1

Please fix both in this same PR (JS has no UdpEndpointI, so it's N/A). Item 9 is C++/Unix-specific and doesn't need a parallel change here.

5. Test nit — drop the (issue #5486 item 10) tag; tests should stand on their own. Suggestion inline.

[bernardnormier]

All review points addressed: the ENOBUFS comment is now justified on best-effort UDP grounds (no bidir/server-adapter narrative), the description is reworked, item 10 is fixed in C# and Java alongside C++ (matching separator sets, with round-trip tests in all three), the issue-number test tag is gone, and the no-changelog rationale is sound. Verified the C#/Java quoting matches the C++ set. LGTM.