Tracking issue for libc 1.0
joshtriplett opened this issue ยท 16 comments
This issue tracks changes we want to make in libc 1.0.
(Note that this tracking issue previously referred to libc 0.3, this should now be assumed to refer to 1.0 instead)
64-bit time_t
and off_t
On some 32-bit targets, it is unclear whether time_t
should be 32-bit or 64-bit since in C this depends on what pre-processor macros are defined when the C standard library headers are included (e.g. _FILE_OFFSET_BITS
, _TIME_BITS
on glibc) or what version of the libc headers are used (e.g. musl 1.2 changed time_t
to always be 64-bit). This was achieved without breaking ABI compatibility by giving the new 64-bit API separate symbol names (clock_gettime64
) and having the header files transparently redirect the old function names to the new symbols. The old symbols remain available for programs built against old versions of the libc headers.
This makes it difficult for the libc
crate since we can only have one definition of off_t
and time_t
, and yet we need our definitions of those types to match the ones used in any C code we link against.
- musl 1.2: #3068
- Switch to 64-bit time on musl
- Deprecate all the 32-bit variants of time and file offset functions
- On musl 1.2, the unsuffixed variants of time and file offset functions are the 64-bit versions, so perhaps allow those to remain undeprecated.
- #2934 and #2935
- Decide how to handle
time64_t
andoff64_t
.- Should we provide only the 64-bit suffixed versions? Should we provide the unsuffixed versions and have them be 64-bit? Should we provide the 32-bit versions at all?
Breaking changes
Most of these are mistakes that were made early on which cannot be fixed without making a breaking change. libc 1.0 is a good opportunity to make these changes.
- Remove
iconv
on all platforms - #3131
- #2971
- #2903
- Check #3161
- Remove
vfork
: #1574 - #3367
- Review all breakage canidates in the issue list: https://github.com/rust-lang/libc/issues?q=label%3Abreakage-candidate
Policy
These don't involve any code changes, but instead require a decision for how to handle breaking changes moving forward.
- Define an MSRV policy (rust-lang/libs-team#72)
- Define the supported platform versions (e.g. glibc, musl, etc.)
Anonymous structs and unions
Some C types are defined using anonymous structs and unions, such as sigaction
or siginfo
. To properly expose these, we need support for anonymous structs and unions in Rust. This feature is part of RFC 2102, but the implementation is still a work in progress.
- We should implement rust-lang/rfcs#2102 first?
Trait implementations
RFC 2235 added implementations of Eq
, Hash
and Debug
for all libc types. However it turns out that this is not appropriate for many libc types, especially in the presence of padding, unions and extensible structs.
We would like to un-accept this RFC for libc 1.0 and completely remove Eq
and Hash
trait implementations for all types. It is still an open question whether we want to keep Debug
trait implementations, but we are tending towards removing them as well.
- Remove impls of
Eq
,Hash
, etc. rust-lang/rfcs#2235 - Either remove
Debug
or make it best-effort.
Testing
We use ctest2 to check that the struct and function definitions in libc match those in C header files. However ctest2 is based on an old fork of the Rust parser, which doesn't handle new syntax well. It might be worth rewriting it or replacing it with a more modern tool (possibly based on bindgen/cbindgen).
- Rewrite ctest2 (which is used on libc-test) to support newer edition and syntax
Implementation plan
- Fork a libc-0.2 branch from master on which development for libc 0.2 will continue.
- The master branch can take PRs that make breaking changes.
- Anyone that needs a feature in a release soon will be encouraged to make a PR to the 0.2 branch and a release can be made from that.
- Once all desired features for libc 1.0 are ready, we can made a release! ๐
On the LFS64 symbol removal point (#2935) the current PR is attempting to achieve rust crate API compatibility (by aliasing LFS64 things to their non-LFS64 equivalents), hiding the effective ABI break in musl libc (it's not a real break since those symbols never "really" existed, but they existed enough for the libc crate to wind up depending on them ๐). If 0.3 is going to be a breaking change we could do a simpler change where we simply stop all the LFS64 functions and types from the libc crate (which is arguably closer to the musl libc API).
My expectation was that we'd merge the API-compatible change in a 0.2.x (and maybe then deprecate the aliases?) and remove them "later" in a 0.3 release, but if 0.3 is imminent, maybe we just drop the items completely?
Note that the rust libc crate has to do "something" here to work with musl 1.2.4+ (@wesleywiser moved rustc's self-contained version to 1.2.3 since that's ABI-compatible with 1.1.24 but 1.2.4 breaks the crate's use of the ABI).
@bossmc Figuring out how to deal with time64 has been one of the two biggest blocking issues for this. (The other is having unnamed fields of struct and union types, for which there's an accepted RFC that still needs compiler support.)
We could just try to match the API of musl as closely as possible when building for musl, but that just pushes the problem onto whoever uses libc.
Personally, I would like to minimize the degree to which people have to write a pile of cfg
s in order to use the 64-bit functions. libc::foo64
seems preferable to #[cfg(...)] use libc::foo as foo64; #[cfg(not(...))] use libc::foo64;
(with corresponding complexity for other functions and types).
Is there some reasonable way we could generally make the same function names and types available on every platform that has 64-bit support, so that people only need cfg
if they want to support platforms that don't have 64-bit support?
So, the changes for 64 bits version of time and file offset function is just for musl
? Can't 32 bit glibc
be included in the package?
Fork a libc-0.2 branch from master on which development for libc 0.2 will continue.
Created the libc-0.2
branch. Once the musl 1.2 PR (or any other breaking changes) gets merged, I'll update the bors config and docs.
Fork a libc-0.2 branch from master on which development for libc 0.2 will continue.
Created the
libc-0.2
branch. Once the musl 1.2 PR (or any other breaking changes) gets merged, I'll update the bors config and docs.
Great to see 0.3 starting to take shape!
Is there an ETA for the breakage-candidate
s to land in main branch, and for a 0.3.0? Not asking for a firm date, my question is really "should I invest time to get a feature-guarded version of #3367 / #3201 into 0.2, or is 0.3 just right off the corner?"
I expect it will take at least a few months until 0.3 is ready. As you can see from the list at the top of this issue, there is still quite a lot of work to do.
Will PR submitted to libc-0.2
get merged into main
for 0.3? Should I make two PRs basing both libc-0.2
and main
if I want it available for both 0.2 and 0.3?
You should make 2 separate PRs.
According to both RFC 1242 and RFC 3119 and also https://blog.rust-lang.org/2017/02/06/roadmap.html , mature crates, which are part of "official Rust", should not have 0.x version numbers. So, please, release 1.0, not 0.3. Previous discussion: #547 . We should take this chance and release 1.0, not 0.3. libc is mature enough.
Later we can release 2.0, etc, if needed.
If there exist previous discussion, where you decided to release 0.3, not 1.0, then I'm sorry. The discussion is not linked from first comment
Obviously, next libc release should use semver trick ( https://github.com/dtolnay/semver-trick ), because libc types widely used in public interfaces
Also, we possibly should wait for extern types ( rust-lang/rust#43467 ). Many libc types are good candidates, for example, FILE ( https://docs.rs/libc/0.2.153/libc/enum.FILE.html )
I've updated the tracking issue to talk about libc 1.0 instead of 0.3.
No pull requests was accepted in one whole month. So, our 1.0 work is stalled
It seems the reviewer for this repository is overwhelmed: there are 53 pull requests awaiting review, and almost all of them are for @JohnTitor. So do you need help? I'm willing to help push forward.