Regression in 1.61 error message for crate not found in local registry

[jonhoo]

Problem

The error message that's presented when a crate doesn't exist in a local registry has regressed from 1.60 to 1.61.

Steps

cargo new local-reg-regression
cd local-reg-regression
echo 'itoa = "1"' >> Cargo.toml
cargo check
mkdir .cargo
cargo local-registry -s Cargo.lock local-registry > .cargo/config.toml
$EDITOR .cargo/config.toml # make it valid TOML
sed -i 's/itoa/gxxy/' Cargo.toml
cargo check

I expected to see (and did in 1.60):

error: no matching package named `gxxy` found
location searched: registry `crates-io`
required by package `local-reg-regression v0.1.0 (/local/home/jongje/dev/tmp/local-reg-regression)`

Instead, with 1.61, I see:

error: failed to get `gxxy` as a dependency of package `local-reg-regression v0.1.0 (/local/home/jongje/dev/tmp/local-reg-regression)`

Caused by:
  failed to query replaced source registry `crates-io`

Caused by:
  failed to read `/local/home/jongje/dev/tmp/local-reg-regression/local-registry/index/gx/xy/gxxy`

Caused by:
  No such file or directory (os error 2)

Possible Solution(s)

I'm not entirely sure how this change in behavior arose. I think it may be related to #10482 and how we don't translate a ErrorKind::NotFound from read_bytes into LoadResponse::NotFound here:

cargo/src/cargo/sources/registry/local.rs

Lines 57 to 60 in f624095

	Poll::Ready(Ok(LoadResponse::Data {
	raw_data: paths::read_bytes(&root.join(path))?,
	index_version: None,
	}))

cc @arlosi

Notes

No response

Version

cargo 1.61.0 (a028ae4 2022-04-29)
release: 1.61.0
commit-hash: a028ae42fc1376571de836be702e840ca8e060c2
commit-date: 2022-04-29
host: aarch64-unknown-linux-gnu
libgit2: 1.4.2 (sys:0.14.2 vendored)
libcurl: 7.80.0-DEV (sys:0.4.51+curl-7.80.0 vendored ssl:OpenSSL/1.1.1m)
os: Amazon Linux AMI 2.0.0 [64-bit]

[jonhoo]

Bisection suggests regression happened in the 2022-03-20 nightly, and specifically with rust-lang/rust#95103.

[jonhoo]

Cargo bisection of the relevant commit range points at f12f025 (part of #10064) being the first bad commit

[jonhoo]

I think I found it. That commit changed (in part) this:

cargo/src/cargo/sources/registry/index.rs

Lines 631 to 645 in 82093ad

	});
	if matches!(err, Poll::Pending) {
	assert!(!hit_closure);
	return Poll::Pending;
	}
	// We ignore lookup failures as those are just crates which don't exist
	// or we haven't updated the registry yet. If we actually ran the
	// closure though then we care about those errors.
	if !hit_closure {
	debug_assert!(cache_contents.is_none());
	return Poll::Ready(Ok(None));
	}
	let _ = err?;

to this

cargo/src/cargo/sources/registry/index.rs

Lines 634 to 647 in f12f025

	})?;
	if matches!(err, Poll::Pending) {
	assert!(!hit_closure);
	return Poll::Pending;
	}
	// We ignore lookup failures as those are just crates which don't exist
	// or we haven't updated the registry yet. If we actually ran the
	// closure though then we care about those errors.
	if !hit_closure {
	debug_assert!(cache_contents.is_none());
	return Poll::Ready(Ok(None));
	}

Note how we now error immediately if .load() fails (the added ?), whereas previously we would ignore the error if the closure was never hit.

[jonhoo]

And that change carried over through #10482, which doesn't map a local registry NotFound error into LoadResponse::NotFound, and thus continues to take the early-exit path.

[jonhoo]

Fix in #10683