Netlink support for FreeBSD 13.2

[ydirson]

Closes #3194, to the point that Xen Guest Agent can works on FreeBSD.

Requires breaking-change #3367.

Downstream support in rust-netlink:

• netlink-packet-route: FreeBSD support available since v0.18
• rtnetlink: FreeBSD support available since v0.14.1 (from FreeBSD support rust-netlink/rtnetlink#49)

Also see #3555 for a backport to libc-0.2 branch using a feature flag to avoid the breaking change by default.

[rustbot]

Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @JohnTitor (or someone else) soon.

Please see the contribution instructions for more information. Namely, in order to ensure the minimum review times lag, PR authors and assigned reviewers should ensure that the review label (S-waiting-on-review and S-waiting-on-author) stays updated, invoking these commands when appropriate:

• @rustbot author: the review is finished, PR author should check the comments and take action accordingly
• @rustbot review: the author is ready for a review, this PR will be queued again in the reviewer's queue

[ydirson]

I can't find a description of what the failing test does; obviously it's compiling some C code (I infer it's to check that the Rust-level definition is compatible with the C definition) and was never told to use the proper C header. This seems to hint that we don't want those definitions exposed for versions which don't implement them, that is FreeBSD 13.1 and earlier.
This seems to imply we want to extend build.rs to define freebsd13_2 (and then likely other ones). A possiblity would be to change which_freebsd() to return (major, minor) instead of just major. What would be the strategy then, adding submodules of freebsd13 for minor versions, and duplicating the netlink definitions for freebsd13_2 and freebsd14 ?

For the tests themselves:

• I already get 126 errors when running the tests on FreeBSD 13.2 without any change
• if I add netlink/netlink.h to the list in test_freebsd that seems to otherwise address the problem exposed by the test, likely it should then be made conditional with a [freebsd13_2 || freebsd14] config predicate?
• about the previous item, I note an existing [freebsd13] item, does that magically includes freebsd14 or would that be a bug? Edit: found the magic ;)

This also seems to hint that more CI configs will be needed.

[ydirson]

The freebsd14 warning about NETLINK_GENERIC not being used is likely about its shadowing. Should we add a cfg restriction to the one in freensd/mod.rs? I'm not even sure it has any usefulness, and is not listed in the semver list for freebsd, so maybe we should just nuke it instead?

[JohnTitor]

The freebsd14 warning about NETLINK_GENERIC not being used is likely about its shadowing. Should we add a cfg restriction to the one in freensd/mod.rs? I'm not even sure it has any usefulness, and is not listed in the semver list for freebsd, so maybe we should just nuke it instead?

Because it's already defined:

libc/src/unix/bsd/freebsdlike/freebsd/mod.rs

Line 3342 in 0b758af

pub const NETLINK_GENERIC: ::c_int = 0;

[ydirson]

I see it's already defined. The problem is, FreeBSD has 2 defines with same name and different values, each for a different API and defined in their own header file. The separate header files allow this to work for C by creating a poor man's namespacing, but how are we to handle that in the Rust case?

Some options I can see include:

• change the definition of the other FreeBSD API providing the clashing define (breaking API for them, no clue how widespread its use is in Rust or even if it is used at all - can we know that?)
• use a different name for the FreeBSD Netlink API's const: that make the FreeBSD Rust Netlink API incompatible with that of Linux-like OS, which does not look clean and has to be handled specially in crates building on this to provide the rustified API (or really on any crate using the libc Netlink API - that raises some documentation issues to make sure this shape-changing API would be properly used)
• change the definition of the Netlink API in libc crate, possibly with a compatibility layer on linux-like for transition (which would make the Netlink API more special than the rest of libc)

What do you think?

[asomers]

Regarding NETLINK_GENERIC, ask @GuillaumeGomez . He's the one who added the definition from if_mib.h . Also, grep.app shows that his sysinfo crate is the only one that's using it as he intended. There are many more that try to use your version. Including https://github.com/bytecodealliance/rustix/blob/db143034d80dbe2dc62368b732378c89f16e7569/src/net/types.rs#L800 , which seems to be confused by the definition in libc. One version or the other will have to change. I suspect that the least overall pain to the community would come by removing or renaming the definition from if_mib.h

Also, you should move your definitions out of the freebsd13 and freebsd14 modules and into the freebsd/mod.rs file. The former two should only be used for stuff that changes between versions. Symbols that newly appear in a newer version of FreeBSD can safely be defined by libc in for all OS versions. For example, ino_t changed between versions, so it must be defined in freebsd11, freebsd12, etc.

[GuillaumeGomez]

FreeBSD and its handling of APIs is a nightmare. They don't care about compatibility, they change types/constants instead of creating new ones. So if NETLINK_GENERIC exists in two different places with different values, to prevent breaking libc's API, here's what I recommend: adding docs to the current NETLINK_GENERIC telling where to look for which values and then create both constants in two submodules (which maybe mention each others?).

[asomers]

What do you mean @GuillaumeGomez ? This problem is caused by if anything too much backwards-compatibility, not too little. The NETLINK_GENERIC mib that you're using was added in 1996, and it still works. The NETLINK_GENERIC that @ydirson wants, by contrast, is based on an RFC written in 2003. By then it was too late to change the name of the old constant, but changing the name of the new constant would break programs' cross-platform compatibility. Hence isolating them in separate headers.

Is there any precedent for adding submodules to libc?

[GuillaumeGomez]

I mean that if they wanted to change the value of NETLINK_GENERIC, they should have created a new constant (called NETLINK_GENERIC2 for example) instead of just changing the value. They often break API, the worst being in structs where they change the fields order or even remove/add some.

Is there any precedent for adding submodules to libc?

No clue but if there is a good enough reason, I think it could be ok (up to them though!).

[asomers]

I mean that if they wanted to change the value of NETLINK_GENERIC, they should have created a new constant (called NETLINK_GENERIC2 for example) instead of just changing the value. They often break API, the worst being in structs where they change the fields order or even remove/add some.

Are you talking about stuff like struct stat? That, I think, is really cool. FreeBSD uses ELF symbol versioning to change structs like that. That means zero problems for C code; old code will still compile, and old binaries will still run. And yet new code can take advantage of new features without introducing new syscalls like Linux's stat64. The only problems come when using FFI, like Rust does. Rust basically chose to be easily cross-compilable at the expense of not being able to take advantage of ELF symbol versions.

Is there any precedent for adding submodules to libc?

No clue but if there is a good enough reason, I think it could be ok (up to them though!).

I don't see why it couldn't be done. I think it's preferable to move the NETLINK_GENERIC mib into a submodule, because unlink the NETLINK_GENERIC protocol family it is platform-specific. If we do that, what do you think the submodule should be named, and what other symbols should go with it?

[GuillaumeGomez]

Are you talking about stuff like struct stat? That, I think, is really cool. FreeBSD uses ELF symbol versioning to change structs like that. That means zero problems for C code; old code will still compile, and old binaries will still run. And yet new code can take advantage of new features without introducing new syscalls like Linux's stat64. The only problems come when using FFI, like Rust does. Rust basically chose to be easily cross-compilable at the expense of not being able to take advantage of ELF symbol versions.

I don't know enough about this to comment. From my understanding, having multiple types depending on which minimum version you support is much better and simpler. Anyway, there is no point debating over whether or not FreeBSD took the right decision here.

[ydirson]

Also, you should move your definitions out of the freebsd13 and freebsd14 modules and into the freebsd/mod.rs file.

Pushed this, but:

• it won't compile anywhere before the name clash gets solved, so I stacked a temporary commit to comment the if_mib definition out to keep the branch usable
• libc-test now reports problems on freebsd12, which was the reason why I had moved the code in version-specific modules

[asomers]

* libc-test now reports problems on `freebsd12`, which was the reason why I had moved the code in version-specific modules

You will have to add an exception in libc-test. See copy_file_range for an example.

[ydirson]

Starting to introduce namespaces in libc will have an impact on those generated tests, we'll likely need someone mastering this part?

[ydirson]

Now rebased on top of #3367 to get rid of the previous hack around NETLINK_GENERIC

[bors]

☔ The latest upstream changes (presumably #3341) made this pull request unmergeable. Please resolve the merge conflicts.

[GuillaumeGomez]

This is a breaking change btw (moving a constant to a different module).

[ydirson]

Yes, this is handled separately as #3367, on which this PR is based

[ydirson]

Rebased onto 0.2.151.

Now forks for real in xen-guest-agent.
Can't wait for v0.3 really 😃

[ydirson]

Since 0.2 was branched out I cannot really rebase this branch any more to keep it usable (rustix just pushed a new version demanding libc 0.2.152), so I pushed the new rebased version to https://github.com/xcp-ng/rust-libc/tree/freebsd-netlink-0.2

[ydirson]

This last push removes the pcap/nflog part, which is indeed not directly related, and not necessary for our use case.

That removed part is at https://github.com/xcp-ng/rust-libc/tree/freebsd-pcap-nflog in case it is useful for a new PR.

[ydirson]

... and rebased to avoid failure of obsolete freebsd12 build job

[bors]

☔ The latest upstream changes (presumably #3587) made this pull request unmergeable. Please resolve the merge conflicts.