CI random failure on x86_64-apple-2 InvalidInput in create_dir_all while running test/rustdoc/primitive/no_std.rs

[ehuss]

CI has failed a few times with the following error on x86_64-apple-2:

---- [rustdoc] tests/rustdoc/primitive/no_std.rs stdout ----
thread '[rustdoc] tests/rustdoc/primitive/no_std.rs' panicked at 'called `Result::unwrap()` on an `Err` value: Os { code: 22, kind: InvalidInput, message: "Invalid argument" }', src/tools/compiletest/src/runtest.rs:130:43

or:

---- [rustdoc] src/test/rustdoc/primitive/primitive-generic-impl.rs stdout ----
thread '[rustdoc] src/test/rustdoc/primitive/primitive-generic-impl.rs' panicked at 'called `Result::unwrap()` on an `Err` value: Os { code: 22, kind: InvalidInput, message: "Invalid argument" }', src/tools/compiletest/src/runtest.rs:125:43
Some tests failed in compiletest suite=rustdoc mode=rustdoc host=x86_64-apple-darwin target=x86_64-apple-darwin

I don't have any immediate theories as to the cause. It is conspicuous that it almost always happens on the same test.

The particular line in question is a call to the standard create_dir_all:

rust/src/tools/compiletest/src/runtest.rs

Line 130 in 13b7aa4

create_dir_all(&cx.output_base_dir()).unwrap();

I don't see anything particularly unusual in the implementation: std::fs::create_dir_all. The impl does a fairly standard libc call to mkdir. The macOS man page doesn't mention EINVAL, and I'm not sure under what circumstances that could happen with mkdir.

• 2022-05-26: macros: introduce fluent_messages macro  #97327 (comment)
• 2022-12-05: Rollup of 8 pull requests #105308 (comment)
• 2022-12-07: Start emitting labels even if their pointed to file is not available locally #104449 (comment)
• 2022-12-09: Rollup of 8 pull requests #105448 (comment)
• 2022-12-14: Rollup of 6 pull requests #105649 (comment)
• 2022-12-15: Use impl's def id when calculating type to specify in UFCS #104334 (comment)
• 2022-12-16: always use anonymize_bound_vars #105717 (comment)
• 2022-12-26: Remove unreasonable help message for auto trait #105817 (comment)
• 2022-12-26: Suggest Pin::as_mut when encountering borrow error #106095 (comment)
• 2023-01-18: Do not filter substs in remap_generic_params_to_declaration_params. #106503 (comment)
• 2023-01-23: Compute generator saved locals on MIR #101692 (comment)
• 2023-02-01: [beta] Cargo 1.68.0 backports #107504 (comment)
• 2023-03-13: remove obsolete givens from regionck #107376 (comment)
• 2023-03-20: Revert "Auto merge of #107224 - nikic:llvm-16, r=cuviper" #109326 (comment)

[jyn514]

the test in question:

rust/tests/rustdoc/primitive/no_std.rs

Lines 1 to 16 in a01b4cc

	#![no_std]
	#![deny(warnings)]
	#![deny(rustdoc::broken_intra_doc_links)]
	// @has no_std/fn.foo.html '//a/[@href="{{channel}}/core/primitive.u8.html"]' 'u8'
	// @has no_std/fn.foo.html '//a/[@href="{{channel}}/core/primitive.u8.html"]' 'primitive link'
	/// Link to [primitive link][u8]
	pub fn foo() -> u8 {}
	// Test that all primitives can be linked to.
	/// [isize] [i8] [i16] [i32] [i64] [i128]
	/// [usize] [u8] [u16] [u32] [u64] [u128]
	/// [f32] [f64]
	/// [char] [bool] [str] [slice] [array] [tuple] [unit]
	/// [pointer] [reference] [fn] [never]
	pub fn bar() {}

[jyn514]

Hmm, it stands out to me that this is one of the only tests where there's a directory with the same name as another test (tests/rustdoc/primitive.rs). Maybe mkdir returns an error if the directory already exists, and there's some race condition because the tests are run in parallel?

[ehuss]

I sort of figured out what is going on. The following is a reproduction:

fn main() {
    use std::fs::{create_dir_all, remove_dir_all};
    use std::thread::spawn;
    for _ in 0..10000 {
        let _ = remove_dir_all("rustdoc");
        let t1 = spawn(|| {
            create_dir_all("rustdoc/primitive").unwrap();
            let _ = remove_dir_all("rustdoc/primitive");
            create_dir_all("rustdoc/primitive").unwrap();
        });
        let t2 = spawn(|| {
            create_dir_all("rustdoc/primitive/no_std").unwrap();
            let _ = remove_dir_all("rustdoc/primitive/no_std");
            create_dir_all("rustdoc/primitive/no_std").unwrap();
        });
        t1.join().unwrap();
        t2.join().unwrap();
    }
}

Essentially two tests are trying to create and delete the same directory at the same time. For some reason, macOS returns EINVAL in that situation. I don't think this is necessarily a flaw in Rust's create_dir_all implementation, more in the realm of "don't do that" (or a bug in macOS itself).

I have opened #109509 to fix this for the tests.