add -Zgitoxide=fetch feature toggle and implementation.

This allows to use `gitoxide` for all fetch operations, boosting performance
for fetching the `crates.io` index by a factor of 2.2x, while being consistent
on all platforms.

For trying it, nightly builds of `cargo` can specify `-Zgitoxide=fetch`.
It's also possible to set the `__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2=1` environment
variable (value matters), which is when `-Zgitoxide=none` can be used
to use `git2` instead.

Limitations
-----------
Note that what follows are current shortcomings that will be addressed in future PRs.

- it's likely that authentication around the `ssh` protocol will work differently in practice
  as it uses the `ssh` program.
- clones from `file://` based crates indices will need the `git` binary to serve the locatl repository.
- the progress bar shown when fetching doesn't work like the orgiinal, but should already feel 'faster'.
This commit is contained in:
Sebastian Thiel 2022-12-01 13:45:25 +01:00
parent c61b0f0680
commit cfffda9ae5
No known key found for this signature in database
GPG key ID: 9CB5EE7895E8268B
17 changed files with 866 additions and 118 deletions

View file

@ -74,7 +74,7 @@ jobs:
os: windows-latest
rust: stable-msvc
other: i686-pc-windows-msvc
- name: Windows x86_64 gnu nightly
- name: Windows x86_64 gnu nightly # runs out of space while trying to link the test suite
os: windows-latest
rust: nightly-gnu
other: i686-pc-windows-gnu
@ -102,7 +102,16 @@ jobs:
if: "!contains(matrix.rust, 'stable')"
- run: cargo test
# The testsuite generates a huge amount of data, and fetch-smoke-test was
- name: Clear intermediate test output
run: |
df -h
rm -rf target/tmp
df -h
- name: gitoxide tests (all git-related tests)
run: cargo test git
env:
__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2: 1
# The testsuite generates a huge amount of data, and fetch-smoke-test was
# running out of disk space.
- name: Clear test output
run: |
@ -156,6 +165,18 @@ jobs:
- run: rustup update stable && rustup default stable
- run: cargo test --manifest-path crates/resolver-tests/Cargo.toml
test_gitoxide:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v3
- run: rustup update --no-self-update stable && rustup default stable
- run: rustup target add i686-unknown-linux-gnu
- run: sudo apt update -y && sudo apt install gcc-multilib libsecret-1-0 libsecret-1-dev -y
- run: rustup component add rustfmt || echo "rustfmt not available"
- run: cargo test
env:
__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2: 1
build_std:
runs-on: ubuntu-latest
env:
@ -196,7 +217,7 @@ jobs:
permissions:
contents: none
name: bors build finished
needs: [docs, rustfmt, test, resolver, build_std]
needs: [docs, rustfmt, test, resolver, build_std, test_gitoxide]
runs-on: ubuntu-latest
if: "success() && github.event_name == 'push' && github.ref == 'refs/heads/auto-cargo'"
steps:

View file

@ -30,6 +30,8 @@ filetime = "0.2.9"
flate2 = { version = "1.0.3", default-features = false, features = ["zlib"] }
git2 = "0.16.0"
git2-curl = "0.17.0"
gix = { version = "0.38.0", default-features = false, features = ["blocking-http-transport-curl", "progress-tree"] }
gix-features-for-configuration-only = { version = "0.27.0", package = "gix-features", features = [ "parallel" ] }
glob = "0.3.0"
hex = "0.4"
hmac = "0.12.1"

View file

@ -247,3 +247,10 @@ pub fn tag(repo: &git2::Repository, name: &str) {
false
));
}
/// Returns true if gitoxide is globally activated.
///
/// That way, tests that normally use `git2` can transparently use `gitoxide`.
pub fn cargo_uses_gitoxide() -> bool {
std::env::var_os("__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2").map_or(false, |value| value == "1")
}

View file

@ -716,6 +716,8 @@ unstable_cli_options!(
doctest_xcompile: bool = ("Compile and run doctests for non-host target using runner config"),
dual_proc_macros: bool = ("Build proc-macros for both the host and the target"),
features: Option<Vec<String>> = (HIDDEN),
gitoxide: Option<GitoxideFeatures> = ("Use gitoxide for the given git interactions, or all of them if no argument is given"),
jobserver_per_rustc: bool = (HIDDEN),
minimal_versions: bool = ("Resolve minimal dependency versions instead of maximum"),
direct_minimal_versions: bool = ("Resolve minimal dependency versions instead of maximum (direct dependencies only)"),
mtime_on_use: bool = ("Configure Cargo to update the mtime of used files"),
@ -827,6 +829,74 @@ where
parse_check_cfg(crates.into_iter()).map_err(D::Error::custom)
}
#[derive(Debug, Copy, Clone, Default, Deserialize)]
pub struct GitoxideFeatures {
/// All fetches are done with `gitoxide`, which includes git dependencies as well as the crates index.
pub fetch: bool,
/// When cloning the index, perform a shallow clone. Maintain shallowness upon subsequent fetches.
pub shallow_index: bool,
/// When cloning git dependencies, perform a shallow clone and maintain shallowness on subsequent fetches.
pub shallow_deps: bool,
/// Checkout git dependencies using `gitoxide` (submodules are still handled by git2 ATM, and filters
/// like linefeed conversions are unsupported).
pub checkout: bool,
/// A feature flag which doesn't have any meaning except for preventing
/// `__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2=1` builds to enable all safe `gitoxide` features.
/// That way, `gitoxide` isn't actually used even though it's enabled.
pub internal_use_git2: bool,
}
impl GitoxideFeatures {
fn all() -> Self {
GitoxideFeatures {
fetch: true,
shallow_index: true,
checkout: true,
shallow_deps: true,
internal_use_git2: false,
}
}
/// Features we deem safe for everyday use - typically true when all tests pass with them
/// AND they are backwards compatible.
fn safe() -> Self {
GitoxideFeatures {
fetch: true,
shallow_index: false,
checkout: true,
shallow_deps: false,
internal_use_git2: false,
}
}
}
fn parse_gitoxide(
it: impl Iterator<Item = impl AsRef<str>>,
) -> CargoResult<Option<GitoxideFeatures>> {
let mut out = GitoxideFeatures::default();
let GitoxideFeatures {
fetch,
shallow_index,
checkout,
shallow_deps,
internal_use_git2,
} = &mut out;
for e in it {
match e.as_ref() {
"fetch" => *fetch = true,
"shallow-index" => *shallow_index = true,
"shallow-deps" => *shallow_deps = true,
"checkout" => *checkout = true,
"internal-use-git2" => *internal_use_git2 = true,
_ => {
bail!("unstable 'gitoxide' only takes `fetch`, 'shallow-index', 'shallow-deps' and 'checkout' as valid inputs")
}
}
}
Ok(Some(out))
}
fn parse_check_cfg(
it: impl Iterator<Item = impl AsRef<str>>,
) -> CargoResult<Option<(bool, bool, bool, bool)>> {
@ -879,6 +949,13 @@ impl CliUnstable {
for flag in flags {
self.add(flag, &mut warnings)?;
}
if self.gitoxide.is_none()
&& std::env::var_os("__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2")
.map_or(false, |value| value == "1")
{
self.gitoxide = GitoxideFeatures::safe().into();
}
Ok(warnings)
}
@ -967,6 +1044,13 @@ impl CliUnstable {
"doctest-xcompile" => self.doctest_xcompile = parse_empty(k, v)?,
"doctest-in-workspace" => self.doctest_in_workspace = parse_empty(k, v)?,
"panic-abort-tests" => self.panic_abort_tests = parse_empty(k, v)?,
"jobserver-per-rustc" => self.jobserver_per_rustc = parse_empty(k, v)?,
"gitoxide" => {
self.gitoxide = v.map_or_else(
|| Ok(Some(GitoxideFeatures::all())),
|v| parse_gitoxide(v.split(',')),
)?
}
"host-config" => self.host_config = parse_empty(k, v)?,
"target-applies-to-host" => self.target_applies_to_host = parse_empty(k, v)?,
"publish-timeout" => self.publish_timeout = parse_empty(k, v)?,

View file

@ -52,7 +52,7 @@ mod cargo_uninstall;
mod common_for_install_and_uninstall;
mod fix;
mod lockfile;
mod registry;
pub(crate) mod registry;
mod resolve;
pub mod tree;
mod vendor;

View file

@ -619,9 +619,6 @@ pub fn configure_http_handle(config: &Config, handle: &mut Easy) -> CargoResult<
handle.useragent(&format!("cargo {}", version()))?;
}
// Empty string accept encoding expands to the encodings supported by the current libcurl.
handle.accept_encoding("")?;
fn to_ssl_version(s: &str) -> CargoResult<SslVersion> {
let version = match s {
"default" => SslVersion::Default,
@ -631,13 +628,15 @@ pub fn configure_http_handle(config: &Config, handle: &mut Easy) -> CargoResult<
"tlsv1.2" => SslVersion::Tlsv12,
"tlsv1.3" => SslVersion::Tlsv13,
_ => bail!(
"Invalid ssl version `{}`,\
choose from 'default', 'tlsv1', 'tlsv1.0', 'tlsv1.1', 'tlsv1.2', 'tlsv1.3'.",
s
"Invalid ssl version `{s}`,\
choose from 'default', 'tlsv1', 'tlsv1.0', 'tlsv1.1', 'tlsv1.2', 'tlsv1.3'."
),
};
Ok(version)
}
// Empty string accept encoding expands to the encodings supported by the current libcurl.
handle.accept_encoding("")?;
if let Some(ssl_version) = &http.ssl_version {
match ssl_version {
SslVersionConfig::Single(s) => {

View file

@ -1,5 +1,10 @@
pub use self::source::GitSource;
pub use self::utils::{fetch, GitCheckout, GitDatabase, GitRemote};
mod known_hosts;
mod oxide;
mod source;
mod utils;
pub mod fetch {
pub type Error = gix::env::collate::fetch::Error<gix::refspec::parse::Error>;
}

View file

@ -0,0 +1,338 @@
//! This module contains all code sporting `gitoxide` for operations on `git` repositories and it mirrors
//! `utils` closely for now. One day it can be renamed into `utils` once `git2` isn't required anymore.
use crate::ops::HttpTimeout;
use crate::util::{human_readable_bytes, network, MetricsCounter, Progress};
use crate::{CargoResult, Config};
use cargo_util::paths;
use gix::bstr::{BString, ByteSlice};
use log::debug;
use std::cell::RefCell;
use std::path::Path;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, Weak};
use std::time::{Duration, Instant};
/// For the time being, `repo_path` makes it easy to instantiate a gitoxide repo just for fetching.
/// In future this may change to be the gitoxide repository itself.
pub fn with_retry_and_progress(
repo_path: &std::path::Path,
config: &Config,
cb: &(dyn Fn(
&std::path::Path,
&AtomicBool,
&mut gix::progress::tree::Item,
&mut dyn FnMut(&gix::bstr::BStr),
) -> Result<(), crate::sources::git::fetch::Error>
+ Send
+ Sync),
) -> CargoResult<()> {
std::thread::scope(|s| {
let mut progress_bar = Progress::new("Fetch", config);
network::with_retry(config, || {
let progress_root: Arc<gix::progress::tree::Root> =
gix::progress::tree::root::Options {
initial_capacity: 10,
message_buffer_capacity: 10,
}
.into();
let root = Arc::downgrade(&progress_root);
let thread = s.spawn(move || {
let mut progress = progress_root.add_child("operation");
let mut urls = RefCell::new(Default::default());
let res = cb(
&repo_path,
&AtomicBool::default(),
&mut progress,
&mut |url| {
*urls.borrow_mut() = Some(url.to_owned());
},
);
amend_authentication_hints(res, urls.get_mut().take())
});
translate_progress_to_bar(&mut progress_bar, root)?;
thread.join().expect("no panic in scoped thread")
})
})
}
fn translate_progress_to_bar(
progress_bar: &mut Progress<'_>,
root: Weak<gix::progress::tree::Root>,
) -> CargoResult<()> {
let read_pack_bytes: gix::progress::Id =
gix::odb::pack::bundle::write::ProgressId::ReadPackBytes.into();
let delta_index_objects: gix::progress::Id =
gix::odb::pack::index::write::ProgressId::IndexObjects.into();
let resolve_objects: gix::progress::Id =
gix::odb::pack::index::write::ProgressId::ResolveObjects.into();
// We choose `N=10` here to make a `300ms * 10slots ~= 3000ms`
// sliding window for tracking the data transfer rate (in bytes/s).
let mut last_update = Instant::now();
let mut counter = MetricsCounter::<10>::new(0, last_update);
let mut tasks = Vec::with_capacity(10);
let update_interval = std::time::Duration::from_millis(300);
let short_check_interval = Duration::from_millis(50);
while let Some(root) = root.upgrade() {
let not_yet = last_update.elapsed() < update_interval;
if not_yet {
std::thread::sleep(short_check_interval);
}
root.sorted_snapshot(&mut tasks);
fn progress_by_id(
id: gix::progress::Id,
task: &gix::progress::Task,
) -> Option<&gix::progress::Value> {
(task.id == id).then(|| task.progress.as_ref()).flatten()
}
fn find_in<K>(
tasks: &[(K, gix::progress::Task)],
cb: impl Fn(&gix::progress::Task) -> Option<&gix::progress::Value>,
) -> Option<&gix::progress::Value> {
tasks.iter().find_map(|(_, t)| cb(t))
}
if let Some(objs) = find_in(&tasks, |t| progress_by_id(resolve_objects, t)) {
// Resolving deltas.
let objects = objs.step.load(Ordering::Relaxed);
let total_objects = objs.done_at.expect("known amount of objects");
let msg = format!(", ({objects}/{total_objects}) resolving deltas");
progress_bar.tick(objects, total_objects, &msg)?;
} else if let Some((objs, read_pack)) =
find_in(&tasks, |t| progress_by_id(read_pack_bytes, t)).and_then(|read| {
find_in(&tasks, |t| progress_by_id(delta_index_objects, t))
.map(|delta| (delta, read))
})
{
// Receiving objects.
let objects = objs.step.load(Ordering::Relaxed);
let total_objects = objs.done_at.expect("known amount of objects");
let received_bytes = read_pack.step.load(Ordering::Relaxed);
let now = Instant::now();
if !not_yet {
counter.add(received_bytes, now);
last_update = now;
}
let (rate, unit) = human_readable_bytes(counter.rate() as u64);
let msg = format!(", {rate:.2}{unit}/s");
progress_bar.tick(objects, total_objects, &msg)?;
}
}
Ok(())
}
fn amend_authentication_hints(
res: Result<(), crate::sources::git::fetch::Error>,
last_url_for_authentication: Option<gix::bstr::BString>,
) -> CargoResult<()> {
let Err(err) = res else { return Ok(()) };
let e = match &err {
crate::sources::git::fetch::Error::PrepareFetch(
gix::remote::fetch::prepare::Error::RefMap(gix::remote::ref_map::Error::Handshake(err)),
) => Some(err),
_ => None,
};
if let Some(e) = e {
use anyhow::Context;
let auth_message = match e {
gix::protocol::handshake::Error::Credentials(_) => {
"\n* attempted to find username/password via \
git's `credential.helper` support, but failed"
.into()
}
gix::protocol::handshake::Error::InvalidCredentials { .. } => {
"\n* attempted to find username/password via \
`credential.helper`, but maybe the found \
credentials were incorrect"
.into()
}
gix::protocol::handshake::Error::Transport(_) => {
let msg = concat!(
"network failure seems to have happened\n",
"if a proxy or similar is necessary `net.git-fetch-with-cli` may help here\n",
"https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli"
);
return Err(anyhow::Error::from(err)).context(msg);
}
_ => None,
};
if let Some(auth_message) = auth_message {
let mut msg = "failed to authenticate when downloading \
repository"
.to_string();
if let Some(url) = last_url_for_authentication {
msg.push_str(": ");
msg.push_str(url.to_str_lossy().as_ref());
}
msg.push('\n');
msg.push_str(auth_message);
msg.push_str("\n\n");
msg.push_str("if the git CLI succeeds then `net.git-fetch-with-cli` may help here\n");
msg.push_str(
"https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli",
);
return Err(anyhow::Error::from(err)).context(msg);
}
}
Err(err.into())
}
/// The reason we are opening a git repository.
///
/// This can affect the way we open it and the cost associated with it.
pub enum OpenMode {
/// We need `git_binary` configuration as well for being able to see credential helpers
/// that are configured with the `git` installation itself.
/// However, this is slow on windows (~150ms) and most people won't need it as they use the
/// standard index which won't ever need authentication, so we only enable this when needed.
ForFetch,
}
impl OpenMode {
/// Sometimes we don't need to pay for figuring out the system's git installation, and this tells
/// us if that is the case.
pub fn needs_git_binary_config(&self) -> bool {
match self {
OpenMode::ForFetch => true,
}
}
}
/// Produce a repository with everything pre-configured according to `config`. Most notably this includes
/// transport configuration. Knowing its `purpose` helps to optimize the way we open the repository.
/// Use `config_overrides` to configure the new repository.
pub fn open_repo(
repo_path: &std::path::Path,
config_overrides: Vec<BString>,
purpose: OpenMode,
) -> Result<gix::Repository, gix::open::Error> {
gix::open_opts(repo_path, {
let mut opts = gix::open::Options::default();
opts.permissions.config = gix::permissions::Config::all();
opts.permissions.config.git_binary = purpose.needs_git_binary_config();
opts.with(gix::sec::Trust::Full)
.config_overrides(config_overrides)
})
}
/// Convert `git` related cargo configuration into the respective `git` configuration which can be
/// used when opening new repositories.
pub fn cargo_config_to_gitoxide_overrides(config: &Config) -> CargoResult<Vec<BString>> {
use gix::config::tree::{gitoxide, Core, Http, Key};
let timeout = HttpTimeout::new(config)?;
let http = config.http_config()?;
let mut values = vec![
gitoxide::Http::CONNECT_TIMEOUT.validated_assignment_fmt(&timeout.dur.as_millis())?,
Http::LOW_SPEED_LIMIT.validated_assignment_fmt(&timeout.low_speed_limit)?,
Http::LOW_SPEED_TIME.validated_assignment_fmt(&timeout.dur.as_secs())?,
// Assure we are not depending on committer information when updating refs after cloning.
Core::LOG_ALL_REF_UPDATES.validated_assignment_fmt(&false)?,
];
if let Some(proxy) = &http.proxy {
values.push(Http::PROXY.validated_assignment_fmt(proxy)?);
}
if let Some(check_revoke) = http.check_revoke {
values.push(Http::SCHANNEL_CHECK_REVOKE.validated_assignment_fmt(&check_revoke)?);
}
if let Some(cainfo) = &http.cainfo {
values.push(
Http::SSL_CA_INFO.validated_assignment_fmt(&cainfo.resolve_path(config).display())?,
);
}
values.push(if let Some(user_agent) = &http.user_agent {
Http::USER_AGENT.validated_assignment_fmt(user_agent)
} else {
Http::USER_AGENT.validated_assignment_fmt(&format!("cargo {}", crate::version()))
}?);
if let Some(ssl_version) = &http.ssl_version {
use crate::util::config::SslVersionConfig;
match ssl_version {
SslVersionConfig::Single(version) => {
values.push(Http::SSL_VERSION.validated_assignment_fmt(&version)?);
}
SslVersionConfig::Range(range) => {
values.push(
gitoxide::Http::SSL_VERSION_MIN
.validated_assignment_fmt(&range.min.as_deref().unwrap_or("default"))?,
);
values.push(
gitoxide::Http::SSL_VERSION_MAX
.validated_assignment_fmt(&range.max.as_deref().unwrap_or("default"))?,
);
}
}
} else if cfg!(windows) {
// This text is copied from https://github.com/rust-lang/cargo/blob/39c13e67a5962466cc7253d41bc1099bbcb224c3/src/cargo/ops/registry.rs#L658-L674 .
// This is a temporary workaround for some bugs with libcurl and
// schannel and TLS 1.3.
//
// Our libcurl on Windows is usually built with schannel.
// On Windows 11 (or Windows Server 2022), libcurl recently (late
// 2022) gained support for TLS 1.3 with schannel, and it now defaults
// to 1.3. Unfortunately there have been some bugs with this.
// https://github.com/curl/curl/issues/9431 is the most recent. Once
// that has been fixed, and some time has passed where we can be more
// confident that the 1.3 support won't cause issues, this can be
// removed.
//
// Windows 10 is unaffected. libcurl does not support TLS 1.3 on
// Windows 10. (Windows 10 sorta had support, but it required enabling
// an advanced option in the registry which was buggy, and libcurl
// does runtime checks to prevent it.)
values.push(gitoxide::Http::SSL_VERSION_MIN.validated_assignment_fmt(&"default")?);
values.push(gitoxide::Http::SSL_VERSION_MAX.validated_assignment_fmt(&"tlsv1.2")?);
}
if let Some(debug) = http.debug {
values.push(gitoxide::Http::VERBOSE.validated_assignment_fmt(&debug)?);
}
if let Some(multiplexing) = http.multiplexing {
let http_version = multiplexing.then(|| "HTTP/2").unwrap_or("HTTP/1.1");
// Note that failing to set the HTTP version in `gix-transport` isn't fatal,
// which is why we don't have to try to figure out if HTTP V2 is supported in the
// currently linked version (see `try_old_curl!()`)
values.push(Http::VERSION.validated_assignment_fmt(&http_version)?);
}
Ok(values)
}
pub fn reinitialize(git_dir: &Path) -> CargoResult<()> {
fn init(path: &Path, bare: bool) -> CargoResult<()> {
let mut opts = git2::RepositoryInitOptions::new();
// Skip anything related to templates, they just call all sorts of issues as
// we really don't want to use them yet they insist on being used. See #6240
// for an example issue that comes up.
opts.external_template(false);
opts.bare(bare);
git2::Repository::init_opts(&path, &opts)?;
Ok(())
}
// Here we want to drop the current repository object pointed to by `repo`,
// so we initialize temporary repository in a sub-folder, blow away the
// existing git folder, and then recreate the git repo. Finally we blow away
// the `tmp` folder we allocated.
debug!("reinitializing git repo at {:?}", git_dir);
let tmp = git_dir.join("tmp");
let bare = !git_dir.ends_with(".git");
init(&tmp, false)?;
for entry in git_dir.read_dir()? {
let entry = entry?;
if entry.file_name().to_str() == Some("tmp") {
continue;
}
let path = entry.path();
drop(paths::remove_file(&path).or_else(|_| paths::remove_dir_all(&path)));
}
init(git_dir, bare)?;
paths::remove_dir_all(&tmp)?;
Ok(())
}

View file

@ -2,6 +2,8 @@
//! authentication/cloning.
use crate::core::{GitReference, Verbosity};
use crate::sources::git::oxide;
use crate::sources::git::oxide::cargo_config_to_gitoxide_overrides;
use crate::util::errors::CargoResult;
use crate::util::{human_readable_bytes, network, Config, IntoUrl, MetricsCounter, Progress};
use anyhow::{anyhow, Context as _};
@ -16,6 +18,7 @@ use std::fmt;
use std::path::{Path, PathBuf};
use std::process::Command;
use std::str;
use std::sync::atomic::{AtomicBool, Ordering};
use std::time::{Duration, Instant};
use url::Url;
@ -802,7 +805,7 @@ pub fn with_fetch_options(
pub fn fetch(
repo: &mut git2::Repository,
url: &str,
orig_url: &str,
reference: &GitReference,
config: &Config,
) -> CargoResult<()> {
@ -818,7 +821,7 @@ pub fn fetch(
// If we're fetching from GitHub, attempt GitHub's special fast path for
// testing if we've already got an up-to-date copy of the repository
let oid_to_fetch = match github_fast_path(repo, url, reference, config) {
let oid_to_fetch = match github_fast_path(repo, orig_url, reference, config) {
Ok(FastPathRev::UpToDate) => return Ok(()),
Ok(FastPathRev::NeedsFetch(rev)) => Some(rev),
Ok(FastPathRev::Indeterminate) => None,
@ -880,53 +883,157 @@ pub fn fetch(
// flavors of authentication possible while also still giving us all the
// speed and portability of using `libgit2`.
if let Some(true) = config.net_config()?.git_fetch_with_cli {
return fetch_with_cli(repo, url, &refspecs, tags, config);
return fetch_with_cli(repo, orig_url, &refspecs, tags, config);
}
if config
.cli_unstable()
.gitoxide
.map_or(false, |git| git.fetch)
{
let git2_repo = repo;
let config_overrides = cargo_config_to_gitoxide_overrides(config)?;
let repo_reinitialized = AtomicBool::default();
let res = oxide::with_retry_and_progress(
&git2_repo.path().to_owned(),
config,
&|repo_path,
should_interrupt,
mut progress,
url_for_authentication: &mut dyn FnMut(&gix::bstr::BStr)| {
// The `fetch` operation here may fail spuriously due to a corrupt
// repository. It could also fail, however, for a whole slew of other
// reasons (aka network related reasons). We want Cargo to automatically
// recover from corrupt repositories, but we don't want Cargo to stomp
// over other legitimate errors.
//
// Consequently we save off the error of the `fetch` operation and if it
// looks like a "corrupt repo" error then we blow away the repo and try
// again. If it looks like any other kind of error, or if we've already
// blown away the repository, then we want to return the error as-is.
loop {
let res = oxide::open_repo(
repo_path,
config_overrides.clone(),
oxide::OpenMode::ForFetch,
)
.map_err(crate::sources::git::fetch::Error::from)
.and_then(|repo| {
debug!("initiating fetch of {:?} from {}", refspecs, orig_url);
let url_for_authentication = &mut *url_for_authentication;
let remote = repo
.remote_at(orig_url)?
.with_fetch_tags(if tags {
gix::remote::fetch::Tags::All
} else {
gix::remote::fetch::Tags::Included
})
.with_refspecs(
refspecs.iter().map(|s| s.as_str()),
gix::remote::Direction::Fetch,
)
.map_err(crate::sources::git::fetch::Error::Other)?;
let url = remote
.url(gix::remote::Direction::Fetch)
.expect("set at init")
.to_owned();
let connection =
remote.connect(gix::remote::Direction::Fetch, &mut progress)?;
let mut authenticate = connection.configured_credentials(url)?;
let connection = connection.with_credentials(
move |action: gix::protocol::credentials::helper::Action| {
if let Some(url) = action
.context()
.and_then(|ctx| ctx.url.as_ref().filter(|url| *url != orig_url))
{
url_for_authentication(url.as_ref());
}
authenticate(action)
},
);
let outcome = connection
.prepare_fetch(gix::remote::ref_map::Options::default())?
.receive(should_interrupt)?;
Ok(outcome)
});
let err = match res {
Ok(_) => break,
Err(e) => e,
};
debug!("fetch failed: {}", err);
debug!("doing a fetch for {}", url);
let git_config = git2::Config::open_default()?;
with_fetch_options(&git_config, url, config, &mut |mut opts| {
if tags {
opts.download_tags(git2::AutotagOption::All);
}
// The `fetch` operation here may fail spuriously due to a corrupt
// repository. It could also fail, however, for a whole slew of other
// reasons (aka network related reasons). We want Cargo to automatically
// recover from corrupt repositories, but we don't want Cargo to stomp
// over other legitimate errors.
//
// Consequently we save off the error of the `fetch` operation and if it
// looks like a "corrupt repo" error then we blow away the repo and try
// again. If it looks like any other kind of error, or if we've already
// blown away the repository, then we want to return the error as-is.
let mut repo_reinitialized = false;
loop {
debug!("initiating fetch of {:?} from {}", refspecs, url);
let res = repo
.remote_anonymous(url)?
.fetch(&refspecs, Some(&mut opts), None);
let err = match res {
Ok(()) => break,
Err(e) => e,
};
debug!("fetch failed: {}", err);
if !repo_reinitialized && matches!(err.class(), ErrorClass::Reference | ErrorClass::Odb)
{
repo_reinitialized = true;
debug!(
"looks like this is a corrupt repository, reinitializing \
if !repo_reinitialized.load(Ordering::Relaxed)
// We check for errors that could occour if the configuration, refs or odb files are corrupted.
// We don't check for errors related to writing as `gitoxide` is expected to create missing leading
// folder before writing files into it, or else not even open a directory as git repository (which is
// also handled here).
&& err.is_corrupted()
{
repo_reinitialized.store(true, Ordering::Relaxed);
debug!(
"looks like this is a corrupt repository, reinitializing \
and trying again"
);
if reinitialize(repo).is_ok() {
continue;
}
}
);
if oxide::reinitialize(repo_path).is_ok() {
continue;
}
}
return Err(err.into());
return Err(err.into());
}
Ok(())
},
);
if repo_reinitialized.load(Ordering::Relaxed) {
*git2_repo = git2::Repository::open(git2_repo.path())?;
}
Ok(())
})
res
} else {
debug!("doing a fetch for {}", orig_url);
let git_config = git2::Config::open_default()?;
with_fetch_options(&git_config, orig_url, config, &mut |mut opts| {
if tags {
opts.download_tags(git2::AutotagOption::All);
}
// The `fetch` operation here may fail spuriously due to a corrupt
// repository. It could also fail, however, for a whole slew of other
// reasons (aka network related reasons). We want Cargo to automatically
// recover from corrupt repositories, but we don't want Cargo to stomp
// over other legitimate errors.
//
// Consequently we save off the error of the `fetch` operation and if it
// looks like a "corrupt repo" error then we blow away the repo and try
// again. If it looks like any other kind of error, or if we've already
// blown away the repository, then we want to return the error as-is.
let mut repo_reinitialized = false;
loop {
debug!("initiating fetch of {:?} from {}", refspecs, orig_url);
let res = repo
.remote_anonymous(orig_url)?
.fetch(&refspecs, Some(&mut opts), None);
let err = match res {
Ok(()) => break,
Err(e) => e,
};
debug!("fetch failed: {}", err);
if !repo_reinitialized
&& matches!(err.class(), ErrorClass::Reference | ErrorClass::Odb)
{
repo_reinitialized = true;
debug!(
"looks like this is a corrupt repository, reinitializing \
and trying again"
);
if reinitialize(repo).is_ok() {
continue;
}
}
return Err(err.into());
}
Ok(())
})
}
}
fn fetch_with_cli(

View file

@ -79,6 +79,15 @@ fn maybe_spurious(err: &Error) -> bool {
return true;
}
}
use gix::protocol::transport::IsSpuriousError;
if let Some(err) = err.downcast_ref::<crate::sources::git::fetch::Error>() {
if err.is_spurious() {
return true;
}
}
false
}

View file

@ -40,6 +40,16 @@ the `CARGO_RUN_BUILD_STD_TESTS=1` environment variable and running `cargo test
`rust-src` component installed with `rustup component add rust-src
--toolchain=nightly`.
## Running with `gitoxide` as default git backend in tests
By default, the `git2` backend is used for most git operations. As tests need to explicitly
opt-in to use nightly features and feature flags, adjusting all tests to run with nightly
and `-Zgitoxide` is unfeasible.
This is why the private environment variable named `__CARGO_USE_GITOXIDE_INSTEAD_OF_GIT2` can be
set while running tests to automatically enable the `-Zgitoxide` flag implicitly, allowing to
test `gitoxide` for the entire cargo test suite.
## Running public network tests
Some (very rare) tests involve connecting to the public internet.

View file

@ -100,6 +100,8 @@ Each new feature described below should explain how to use it.
* [`cargo logout`](#cargo-logout) --- Adds the `logout` command to remove the currently saved registry token.
* [publish-timeout](#publish-timeout) --- Controls the timeout between uploading the crate and being available in the index
* [registry-auth](#registry-auth) --- Adds support for authenticated registries, and generate registry authentication tokens using asymmetric cryptography.
* Other
* [gitoxide](#gitoxide) --- Use `gitoxide` instead of `git2` for a set of operations.
### allow-features
@ -1277,6 +1279,21 @@ codegen-backend = true
codegen-backend = "cranelift"
```
### gitoxide
With the 'gitoxide' unstable feature, all or the the specified git operations will be performed by
the `gitoxide` crate instead of `git2`.
While `-Zgitoxide` enables all currently implemented features, one can individually select git operations
to run with `gitoxide` with the `-Zgitoxide=operation[,operationN]` syntax.
Valid operations are the following:
* `fetch` - All fetches are done with `gitoxide`, which includes git dependencies as well as the crates index.
* `shallow-index` *(planned)* - perform a shallow clone of the index.
* `shallow-deps` *(planned)* - perform a shallow clone of git dependencies.
* `checkout` *(planned)* - checkout the worktree, with support for filters and submodules.
## Stabilized and removed features
### Compile progress

View file

@ -1,5 +1,6 @@
//! Tests for some invalid .cargo/config files.
use cargo_test_support::git::cargo_uses_gitoxide;
use cargo_test_support::registry::{self, Package};
use cargo_test_support::{basic_manifest, project, rustc_host};
@ -334,23 +335,45 @@ fn bad_git_dependency() {
let p = project()
.file(
"Cargo.toml",
r#"
&format!(
r#"
[package]
name = "foo"
version = "0.0.0"
authors = []
[dependencies]
foo = { git = "file:.." }
foo = {{ git = "{url}" }}
"#,
url = if cargo_uses_gitoxide() {
"git://host.xz"
} else {
"file:.."
}
),
)
.file("src/lib.rs", "")
.build();
p.cargo("check -v")
.with_status(101)
.with_stderr(
"\
let expected_stderr = if cargo_uses_gitoxide() {
"\
[UPDATING] git repository `git://host.xz`
[ERROR] failed to get `foo` as a dependency of package `foo v0.0.0 [..]`
Caused by:
failed to load source for dependency `foo`
Caused by:
Unable to update git://host.xz
Caused by:
failed to clone into: [..]
Caused by:
URLs need to specify the path to the repository
"
} else {
"\
[UPDATING] git repository `file:///`
[ERROR] failed to get `foo` as a dependency of package `foo v0.0.0 [..]`
@ -365,8 +388,11 @@ Caused by:
Caused by:
[..]'file:///' is not a valid local file URI[..]
",
)
"
};
p.cargo("check -v")
.with_status(101)
.with_stderr(expected_stderr)
.run();
}

View file

@ -9,6 +9,7 @@ use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;
use std::thread;
use cargo_test_support::git::cargo_uses_gitoxide;
use cargo_test_support::paths::{self, CargoPathExt};
use cargo_test_support::registry::Package;
use cargo_test_support::{basic_lib_manifest, basic_manifest, git, main_file, path2url, project};
@ -1827,6 +1828,51 @@ fn fetch_downloads() {
p.cargo("fetch").with_stdout("").run();
}
#[cargo_test]
fn fetch_downloads_with_git2_first_then_with_gitoxide_and_vice_versa() {
let bar = git::new("bar", |project| {
project
.file("Cargo.toml", &basic_manifest("bar", "0.5.0"))
.file("src/lib.rs", "pub fn bar() -> i32 { 1 }")
});
let feature_configuration = if cargo_uses_gitoxide() {
// When we are always using `gitoxide` by default, create the registry with git2 as well as the download…
"-Zgitoxide=internal-use-git2"
} else {
// …otherwise create the registry and the git download with `gitoxide`.
"-Zgitoxide=fetch"
};
let p = project()
.file(
"Cargo.toml",
&format!(
r#"
[package]
name = "foo"
version = "0.5.0"
authors = []
[dependencies.bar]
git = '{url}'
"#,
url = bar.url()
),
)
.file("src/main.rs", "fn main() {}")
.build();
p.cargo("fetch")
.arg(feature_configuration)
.masquerade_as_nightly_cargo(&["unstable features must be available for -Z gitoxide"])
.with_stderr(&format!(
"[UPDATING] git repository `{url}`",
url = bar.url()
))
.run();
Package::new("bar", "1.0.0").publish(); // trigger a crates-index change.
p.cargo("fetch").with_stdout("").run();
}
#[cargo_test]
fn warnings_in_git_dep() {
let bar = git::new("bar", |project| {

View file

@ -8,6 +8,7 @@ use std::sync::atomic::{AtomicUsize, Ordering::SeqCst};
use std::sync::Arc;
use std::thread::{self, JoinHandle};
use cargo_test_support::git::cargo_uses_gitoxide;
use cargo_test_support::paths;
use cargo_test_support::{basic_manifest, project};
@ -157,8 +158,7 @@ Caused by:
https://[..]
Caused by:
",
addr = addr
"
))
.run();
@ -206,15 +206,16 @@ fn https_something_happens() {
p.cargo("check -v")
.with_status(101)
.with_stderr_contains(&format!(
"[UPDATING] git repository `https://{addr}/foo/bar`",
addr = addr
"[UPDATING] git repository `https://{addr}/foo/bar`"
))
.with_stderr_contains(&format!(
"\
Caused by:
{errmsg}
",
errmsg = if cfg!(windows) {
errmsg = if cargo_uses_gitoxide() {
"[..]SSL connect error [..]"
} else if cfg!(windows) {
"[..]failed to send request: [..]"
} else if cfg!(target_os = "macos") {
// macOS is difficult to tests as some builds may use Security.framework,
@ -258,18 +259,40 @@ fn ssh_something_happens() {
.file("src/main.rs", "")
.build();
p.cargo("check -v")
.with_status(101)
.with_stderr_contains(&format!(
"[UPDATING] git repository `ssh://{addr}/foo/bar`",
addr = addr
))
.with_stderr_contains(
let (expected_ssh_message, expected_update) = if cargo_uses_gitoxide() {
// Due to the usage of `ssh` and `ssh.exe` respectively, the messages change.
// This will be adjusted to use `ssh2` to get rid of this dependency and have uniform messaging.
let message = if cfg!(windows) {
// The order of multiple possible messages isn't deterministic within `ssh`, and `gitoxide` detects both
// but gets to report only the first. Thus this test can flip-flop from one version of the error to the other
// and we can't test for that.
// We'd want to test for:
// "[..]ssh: connect to host 127.0.0.1 [..]"
// ssh: connect to host example.org port 22: No route to host
// "[..]banner exchange: Connection to 127.0.0.1 [..]"
// banner exchange: Connection to 127.0.0.1 port 62250: Software caused connection abort
// But since there is no common meaningful sequence or word, we can only match a small telling sequence of characters.
"[..]onnect[..]"
} else {
"[..]Connection [..] by [..]"
};
(
message,
format!("[..]Unable to update ssh://{addr}/foo/bar"),
)
} else {
(
"\
Caused by:
[..]failed to start SSH session: Failed getting banner[..]
",
format!("[UPDATING] git repository `ssh://{addr}/foo/bar`"),
)
};
p.cargo("check -v")
.with_status(101)
.with_stderr_contains(&expected_update)
.with_stderr_contains(expected_ssh_message)
.run();
t.join().ok().unwrap();
}
@ -294,7 +317,7 @@ fn net_err_suggests_fetch_with_cli() {
p.cargo("check -v")
.with_status(101)
.with_stderr(
.with_stderr(format!(
"\
[UPDATING] git repository `ssh://needs-proxy.invalid/git`
warning: spurious network error[..]
@ -316,9 +339,14 @@ Caused by:
https://[..]
Caused by:
failed to resolve address for needs-proxy.invalid[..]
{trailer}
",
)
trailer = if cargo_uses_gitoxide() {
"An IO error occurred when talking to the server\n\nCaused by:\n ssh: Could not resolve hostname needs-proxy.invalid[..]"
} else {
"failed to resolve address for needs-proxy.invalid[..]"
}
))
.run();
p.change_file(
@ -389,8 +417,8 @@ Caused by:
Caused by:
[..]
",
addr = addr
{trailer}",
trailer = if cargo_uses_gitoxide() { "\nCaused by:\n [..]" } else { "" }
))
.run();

View file

@ -5,6 +5,7 @@ use std::ffi::OsStr;
use std::path::PathBuf;
use cargo_test_support::git;
use cargo_test_support::git::cargo_uses_gitoxide;
use cargo_test_support::paths;
use cargo_test_support::project;
use cargo_test_support::registry::Package;
@ -96,6 +97,11 @@ fn use_git_gc() {
#[cargo_test]
fn avoid_using_git() {
if cargo_uses_gitoxide() {
// file protocol without git binary is currently not possible - needs built-in upload-pack.
// See https://github.com/Byron/gitoxide/issues/734 (support for the file protocol) progress updates.
return;
}
let path = env::var_os("PATH").unwrap_or_default();
let mut paths = env::split_paths(&path).collect::<Vec<_>>();
let idx = paths

View file

@ -6,6 +6,7 @@
//! NOTE: The container tests almost certainly won't work on Windows.
use cargo_test_support::containers::{Container, ContainerHandle, MkFile};
use cargo_test_support::git::cargo_uses_gitoxide;
use cargo_test_support::{paths, process, project, Project};
use std::fs;
use std::io::Write;
@ -415,7 +416,11 @@ fn invalid_github_key() {
.build();
p.cargo("fetch")
.with_status(101)
.with_stderr_contains(" error: SSH host key has changed for `github.com`")
.with_stderr_contains(if cargo_uses_gitoxide() {
" git@github.com: Permission denied (publickey)."
} else {
" error: SSH host key has changed for `github.com`"
})
.run();
}
@ -447,16 +452,7 @@ fn bundled_github_works() {
)
.file("src/lib.rs", "")
.build();
let err = if cfg!(windows) {
"error authenticating: unable to connect to agent pipe; class=Ssh (23)"
} else {
"error authenticating: failed connecting with agent; class=Ssh (23)"
};
p.cargo("fetch")
.env("SSH_AUTH_SOCK", &bogus_auth_sock)
.with_status(101)
.with_stderr(&format!(
"\
let shared_stderr = "\
[UPDATING] git repository `ssh://git@github.com/rust-lang/bitflags.git`
error: failed to get `bitflags` as a dependency of package `foo v0.1.0 ([ROOT]/foo)`
@ -472,7 +468,29 @@ Caused by:
Caused by:
failed to authenticate when downloading repository
* attempted ssh-agent authentication, but no usernames succeeded: `git`
*";
let err = if cfg!(windows) {
"error authenticating: unable to connect to agent pipe; class=Ssh (23)"
} else {
"error authenticating: failed connecting with agent; class=Ssh (23)"
};
let expected = if cargo_uses_gitoxide() {
format!(
"{shared_stderr} attempted to find username/password via `credential.helper`, but maybe the found credentials were incorrect
if the git CLI succeeds then `net.git-fetch-with-cli` may help here
https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli
Caused by:
Credentials provided for \"ssh://git@github.com/rust-lang/bitflags.git\" were not accepted by the remote
Caused by:
git@github.com: Permission denied (publickey).
"
)
} else {
format!(
"{shared_stderr} attempted ssh-agent authentication, but no usernames succeeded: `git`
if the git CLI succeeds then `net.git-fetch-with-cli` may help here
https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli
@ -480,9 +498,59 @@ Caused by:
Caused by:
{err}
"
))
)
};
p.cargo("fetch")
.env("SSH_AUTH_SOCK", &bogus_auth_sock)
.with_status(101)
.with_stderr(&expected)
.run();
let shared_stderr = "\
[UPDATING] git repository `ssh://git@github.com:22/rust-lang/bitflags.git`
error: failed to get `bitflags` as a dependency of package `foo v0.1.0 ([ROOT]/foo)`
Caused by:
failed to load source for dependency `bitflags`
Caused by:
Unable to update ssh://git@github.com:22/rust-lang/bitflags.git?tag=1.3.2
Caused by:
failed to clone into: [ROOT]/home/.cargo/git/db/bitflags-[..]
Caused by:
failed to authenticate when downloading repository
*";
let expected = if cargo_uses_gitoxide() {
format!(
"{shared_stderr} attempted to find username/password via `credential.helper`, but maybe the found credentials were incorrect
if the git CLI succeeds then `net.git-fetch-with-cli` may help here
https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli
Caused by:
Credentials provided for \"ssh://git@github.com:22/rust-lang/bitflags.git\" were not accepted by the remote
Caused by:
git@github.com: Permission denied (publickey).
"
)
} else {
format!(
"{shared_stderr} attempted ssh-agent authentication, but no usernames succeeded: `git`
if the git CLI succeeds then `net.git-fetch-with-cli` may help here
https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli
Caused by:
{err}
"
)
};
// Explicit :22 should also work with bundled.
p.change_file(
"Cargo.toml",
@ -498,32 +566,7 @@ Caused by:
p.cargo("fetch")
.env("SSH_AUTH_SOCK", &bogus_auth_sock)
.with_status(101)
.with_stderr(&format!(
"\
[UPDATING] git repository `ssh://git@github.com:22/rust-lang/bitflags.git`
error: failed to get `bitflags` as a dependency of package `foo v0.1.0 ([ROOT]/foo)`
Caused by:
failed to load source for dependency `bitflags`
Caused by:
Unable to update ssh://git@github.com:22/rust-lang/bitflags.git?tag=1.3.2
Caused by:
failed to clone into: [ROOT]/home/.cargo/git/db/bitflags-[..]
Caused by:
failed to authenticate when downloading repository
* attempted ssh-agent authentication, but no usernames succeeded: `git`
if the git CLI succeeds then `net.git-fetch-with-cli` may help here
https://doc.rust-lang.org/cargo/reference/config.html#netgit-fetch-with-cli
Caused by:
{err}
"
))
.with_stderr(&expected)
.run();
}