Re: [libgpiod][PATCH 0/2] bindings: rust: feature gate unreleased features

["Erik Schilling" <erik.schilling@xxxxxxxxxx>]

On Mon Oct 9, 2023 at 2:43 PM CEST, Bartosz Golaszewski wrote:
> On Mon, Oct 9, 2023 at 1:38 PM Erik Schilling <erik.schilling@xxxxxxxxxx> wrote:
> >
> > On Mon Oct 9, 2023 at 10:58 AM CEST, Bartosz Golaszewski wrote:
> > > On Fri, Oct 6, 2023 at 9:24 AM Erik Schilling <erik.schilling@xxxxxxxxxx> wrote:
> > > >
> > > > When releasing the 0.2.0 version of the libgpiod crate, I did not
> > > > realize that there were C lib features that were not released yet.
> > > > Helpfully, vhost-device's CI went up in flames and revealed this [1].
> > > >
> > > > This suggests a way to handle that and sketches how further updates can
> > > > be handled.
> > > >
> > > > I acknowledge that this may be very strange to C developers...
> > > > Traditionally, one would just use whatever your distro provided and the
> > > > distro would make sure that dependencies update in lock-step.
> > > >
> > > > However, in Rust the default way to consume libraries is to pull them
> > > > from crates.io. This is a balancing act for -sys crates which link to
> > > > distro-provided libraries on the system. Since crates.io does not wait
> > > > for distros to update their libraries, crates will need to support a
> > > > wider range of system libraries.
> > > >
> > > > This sets up / sketches the infrastructure for that.
> > > >
> > > > Only the first commit is intended to be merged. The second one just
> > > > sketches how a release will look like once it happens.
> > > >
> > > > [1] https://buildkite.com/rust-vmm/vhost-device-ci/builds/1746#018b0110-b9d3-468a-973c-c3bbc27cd479
> > > >
> > > > To: Bartosz Golaszewski <brgl@xxxxxxxx>
> > > > To: Linux-GPIO <linux-gpio@xxxxxxxxxxxxxxx>
> > > > Cc: Viresh Kumar <viresh.kumar@xxxxxxxxxx>
> > > > Cc: Manos Pitsidianakis <manos.pitsidianakis@xxxxxxxxxx>
> > > > Cc: Kent Gibson <warthog618@xxxxxxxxx>
> > > >
> > > > Signed-off-by: Erik Schilling <erik.schilling@xxxxxxxxxx>
> > > > ---
> > > > Erik Schilling (2):
> > > >       bindings: rust: feature gate unreleased features
> > > >       DONOTMERGE: bindings: rust: simulate v2.1 release
> > > >
> > > >  bindings/rust/libgpiod-sys/Cargo.toml        |  9 +++++++--
> > > >  bindings/rust/libgpiod/Cargo.toml            |  4 ++++
> > > >  bindings/rust/libgpiod/Makefile.am           |  2 +-
> > > >  bindings/rust/libgpiod/README.md             | 14 ++++++++++++++
> > > >  bindings/rust/libgpiod/src/line_request.rs   |  2 ++
> > > >  bindings/rust/libgpiod/tests/line_request.rs |  1 +
> > > >  6 files changed, 29 insertions(+), 3 deletions(-)
> > > > ---
> > > > base-commit: e7b02c2259d97c77107c77b68e3bc1664e6703c1
> > > > change-id: 20231006-b4-bindings-old-version-fix-789973703b77
> > > >
> > > > Best regards,
> > > > --
> > > > Erik Schilling <erik.schilling@xxxxxxxxxx>
> > > >
> > >
> > > I'm currently discussing a similar problem with a developer who
> > > offered to help make Python bindings released on PyPi more
> > > user-friendly. His suggestion for python bindings is what many PyPi
> > > packages that provide bindings to C libraries do: just compile the
> > > library statically and make it part of the C extension module bundled
> > > together with python code.
> > >
> > > Given that rust programs are statically linked, maybe this is a
> > > solution we could use?
> >
> > It may be one option that we could provide, but I still see need for
> > linking against system libs.
> >
>
> Sure, this is what the python bindings will get. A switch to link
> against the system libgpiod if desired.

How does this switch work with python? Will it be some environment
variable or does pip have something similar to the Rust --feature flags?
>
> > Rust is already a bit special - at least compared to traditional C
> > projects - in that it pulls all the code of all dependencies and builds
> > the entire bunch into a static lib. This is still upsetting some people
> > who prefer each dependency to be its own self-contained dynamic library
> > that can be updated without updating the rest of it. Rust can make
> > that work by being a mono-culture in term of build tools. Everything on
> > crates.io builds through cargo in the same way, so tooling can assist
> > with auditing and updating crates.
> >
> > While distros - to various degrees - are starting to accept that this
> > may be a viable approach, I imagine that they would still feel uneasy
> > about some Rust code statically linking a C lib. They spend a lot of
> > effort making sure that the C lib is built according to their packaging
> > guidelines and do not want to repeat that effort for each Rust binary
> > that ends up including libgpiod.
> >
> > Hence, I think that we would still need to provide a way to build and
> > link against the system libs.
> >
> > However, that does not necessarily means
> > that we need to feature gate pending features like suggested here.
> > We _could_ just tie the Rust bindings to the libgpiod version 1:1 and
> > make bindings only support one version of the C lib.
> >
> > That would come with some implications for the release  though. We would
> > effectively need to manage bugfixes to the bindings through libgpiod's
> > stable branches. For the version numbers to be non-confusing, we would
> > probably also need to always do the releases together. So there may
> > be needs to do a new patch release for libgpiod even if only a fix was
> > needed in the bindings.
> >
> > Then, we may end up in a situation where SemVer changes may be
> > conflicting between bindings and the core lib. With the Rust bindings
> > being new, my recent fix of a soundness issue required changes to types
> > that required a bump of the major version. The C version was entirely
> > unaffected. Should we have bumped its major for this too?
> >
>
> Absolutely not! I expect rust bindings to still be quite volatile in
> terms of API stability but libgpiod v2 is pretty mature. Distros
> already seem to be slow adopting the new release so I definitely don't
> want to confuse them even more by bumping libgpiod major version
> because of a change in rust bindings.
>
> We decided to fully decouple rust from libgpiod and let's keep it that way.

Sure. I brought it up as rethorical question to demonstrate the kind of
problems. I fully agree that separate version numbers are the way to go.

>
> > I outlined these kind of issues in my initial thread around publishing
> > to crates.io [1]. We could of course just document somewhere which
> > version of the Rust crate pulls in which version of the C lib.
> >
> > But is it worth it? We would still need to maintain the nested build
> > next to building against system libs. libgpiod is of course fairly
> > boring to build. But we would still need to tell users to install the
> > few build dependencies, C tools may update and the latest crates.io
> > release may fail building.
>
> The core libgpiod library requires a sane toolchain and standard
> library with GNU extensions period. That will not change and I'm quite
> convinced that if an environment has cargo and rustcc and it will have
> a C compiler as well.

Yeah. I think making sure people got the dependencies installed will be
an easy part. That said... yesterday I had to extend my aarch64 SDK
with autoconf-archive for doing a build :).

Anyway, I am mostly concerned about being "odd" and entering less well
tested paths. We are already a "special" Rust crate since we link to a
C lib outside of the Rust ecosystem. But at least we can rely on
somebody already having done the work of providing the library. Now, I
wonder in which ways an autoconf, configure and compilation automation
from a build tool that is optimized to build Rust code can fail.

I am not aware of any popular Rust crate that does this kind of static
linking to C libs by default. So I am not really counting on all the
tooling around this being very well tested across all environments and
targets.

But maybe I am just overly pessimistic :->.

Do you dislike the conditional compilations here? Alternatively, we
could also just backport the fixes that I did to the stable branch and
do a bindings release from there. Then the offending access will be gone
and we could just bump the required C lib version for the Rust bindings
on new releases from the master branch.

That would not lock us into any future and would give us more time to
consider whats best.

- Erik

>
> Bart
>
> >
> > I totally see that static builds may be helpful in some cases. But I
> > would assume that the maintenance overhead of nesting the build is less
> > controllable compared to what I suggest in this patch. The C lib already
> > follows the good practices around versioning and API compatibility. So
> > I would assume that maintaining the conditional compiles will be fairly
> > simple.
> >
> > Once another major bump happens in the C lib, we can just follow the
> > bump by removing all existing conditionals and bumping the required
> > floor version for the Rust bindings.
> >
> > All in all, I only see minor advantages for users while seeing a lot of
> > maintenance burden.
> >
> > [1] https://lore.kernel.org/all/CT0CRWOTJIEO.20BGIDMLFM0E8@fedora/
> >
> > - Erik
> >