Specifically, I verified that *before* the previous patch to the
ed25519 C code, the last bit of code would fail with a SIGSEGV when
trying to read the empty signature.
For historical reasons we have a fair bit of distinct sysroot
initialization going on. A lot of code is calling *just* the new
`ensure_writable()` API, which does basically what you'd expect...
except if we're not using a mount namespace.
Which is the case in unit tests and legacy setups.
Change this API to also ensure the sysroot is fully initialized
even in those cases. Specifically we'll have `self->sysroot_fd`.
For now, callers that need `/boot` also need to separately
call `_ensure_boot_fd()`.
It's falliable, and in one place we were actually ignoring
the error and leaving a `NULL` repo object which is just a trap
for people coming along later since it's rarely nullable.
Quite a while ago we switched to loading the repo at the same time
as the sysroot; convert callers in the library to use this infallible
accessor.
Prep for another patch which will use the repo object.
https://github.com/coreos/coreos-assembler/pull/2921 broke this
test which is intentionally causing a systemd unit to fail.
As they say, necessity is the mother of invention. They don't
say though that need always causes particularly *beautiful* things
to be invented...
The pattern this API uses in C is to allow the input parameters
pointer targets to be `NULL`, and it doesn't return values in that
case.
A further complexity here is that the API will still return `NULL`
for symbolic links.
But Rust can't express this pattern as is, so we were always
returning values but in `Option<T>` wrappers that the caller needed
to unwrap for the metadata.
(We really want an even more efficient API here that avoids the glib
objects entirely, e.g. no reason not to pass directly back a type
that lets Rust directly read from the fd for bare repos, but
that can come later)
Because the source is in a subdirectory, we lose out on cargo target
autodiscovery.
I noticed this when I edited one of the tests in a way that
should have failed, but didn't...
We recently discovered `list_objects()` is inefficient with memory.
The more efficient `list_objects_set()` API isn't yet public, but
this fsck code actually just skips over non-commit objects, and
we already have an API to list just those.
No real changes.
```
$ cargo fix --edition
note: Switching to Edition 2021 will enable the use of the version 2 feature resolver in Cargo.
This may cause some dependencies to be built with fewer features enabled than previously.
More information about the resolver changes may be found at https://doc.rust-lang.org/nightly/edition-guide/rust-2021/default-cargo-resolver.html
When building the following dependencies, the given features will no longer be used:
libc v0.2.126 removed features: extra_traits
The following differences only apply when building with dev-dependencies:
getrandom v0.2.6 removed features: std
```
which looks OK to me.
When we did the merger, it turns out cargo by default is basically
going to include all of stuff in the git repository root directory
which is "libostree". We just want the stuff in `rust-bindings/`.
I initially tried adding `include = "rust-bindings/"` but
according to
https://doc.rust-lang.org/cargo/reference/manifest.html#the-exclude-and-include-fields
specifying `include` means that `exclude` is
ignored, which is kind of annoying. Further, doing so *also*
turns off the cargo automatic rules for handling e.g. `gitignore`.
So for now I went with the approach of adding everything from the C
library stuff into `exclude/`.
I messed this up; the last release should inherit from the previous
release (N-1) and not the previous to that (N-2).
I think (hope) this isn't an ABI break...
Just noticed this when I was going to add a new symbol.
In https://github.com/ostreedev/ostree/pull/2633 I realized
that our CI only builds git of libostree or git of rust-bindings,
not git of both. And we definitely want to test the latter too,
so e.g. the Rust tests *also* become tests for changes to the C code.
In a prior change we discovered that for bad historical reasons
libostree was returning a mapping "object type+checksum" => "metadata"
but the "metadata" was redundant and pointless.
Optimize the prune API to use a (currently internal) object listing
API which returns a set, not a map. This allows `GHashTable` to
avoid allocating a separate array for the values, neatly cutting
memory usage in half (from ~13MB to ~6MB) on my test case of a
dry-run prune of a FCOS build.
There were some changes to the sys API for introspection fixes.
And add a feature for the current release, which is something
I'll add to the checklist for releases.
I was looking at https://github.com/ostreedev/ostree/pull/2632
and confused at the usage of
`GVariant *value = g_variant_new ("(b@as)", TRUE, g_variant_new_strv (NULL, 0));`
which looked strange - why the empty strv?
It turns out that this is a historical legacy of the time when
ostree had pack files. And nothing actually cares about the values
of these variants; we should have an API that returns a proper set,
and not a hash.
But...since all of these things have exactly the same value, instead
of allocating lots of redundant copies on the heap, just have
them all hold a refcount on a shared value.
This cuts the heap usage from 20MB to 13MB on a test FCOS repository
build.
It inherently depends on the individual build, and can't
really be an official stable API for introspection users.
I've noticed the value of this flip flop when doing local builds.
I'm fairly certain no one is trying to use it from a higher level
language.
It'd probably make sense to even drop from the official C API,
but I'm trying to be conservative with that.
I was looking at our `.gir` and noticed we had the cmdprivate bits
because the pattern for excluding headers is `-private.h`, which
didn't match `cmdprivate.h`.
There's a pile of new warnings in GCC 12 from F36, I dug into
them but most of them seem inscrutable - I am not seeing the bugs.
For now to unblock further work, stop using this flag.
I'm aiming to do some more work on the Rust side around `fsck`
like functionality, and this is a useful primitive. There isn't
a great Rust crate for xattrs, and I think it's better to share this
code.
It's really tempting to remove `make syntax-check`, it has very
very rarely found any real problems.
But anyways, just exclude all the binding code because it trips
up random problems we simply don't care about like mentions of
`O_NDELAY` in the `GLib-2.0.gir`.
Fix up the paths for the crates now that the Rust bindings are in
`rust/`.
We can't today include the test suite because it depends on `ostree-rs-ext`
which would make everything circular.
(Building that now requires a separate `cd tests/inst && cargo build`)
I want to write APIs that *require* the caller to have set up
an ostree transaction. It's natural to require passing a guard
to do so. But then we want an accessor for the repo.
The underlying `ostree_repo_load_file()` API has the caller pass
`NULL` for output arguments it doesn't want. This isn't sanely
bindable in Rust - what the generator does is always request
all values, but maps them all to `Option<T>`.
The main cases are where a user wants either metadata, or both
metadata and content. This API gives just metadata; it's a
bit more efficient as we don't need to open the file, and doesn't
require the caller to `unwrap()`.
Colin Walters
