std.process.Child.run fails to kill child on error condition

[adrsimon]

Zig Version

0.14.0-dev.2596+e6879e99e

Steps to Reproduce and Observed Behavior

Hey, I've been tinkering around with Zig and found a strange bug while solving Ziglings. Basically my CPU went to 89°C for no clear reason. Here are the steps to reproduce.

• Clone the ziglings repo
• Open problem 049_quiz6.zig on your favorite editor
• Implement this solution (that doesn't work), modifications are where the //HERE comments are.

// ...

const Elephant = struct {
    letter: u8,
    tail: ?*Elephant = null,
    trunk: ?*Elephant = null,
    visited: bool = false,

    // Elephant tail methods!
    pub fn getTail(self: *Elephant) *Elephant {
        return self.tail.?; // Remember, this means "orelse unreachable"
    }

    pub fn hasTail(self: *Elephant) bool {
        return (self.tail != null);
    }

    // Your Elephant trunk methods go here!
    // --------------------------------------------------- 
    //HERE

    pub fn getTrunk(self: *Elephant) *Elephant {
        return self.trunk.?;
    }

    pub fn hasTrunk(self: *Elephant) bool {
        return (self.trunk != null);
    }

    // ---------------------------------------------------

    pub fn visit(self: *Elephant) void {
        self.visited = true;
    }

    pub fn print(self: *Elephant) void {
        // Prints elephant letter and [v]isited
        const v: u8 = if (self.visited) 'v' else ' ';
        std.debug.print("{u}{u} ", .{ self.letter, v });
    }
};

pub fn main() void {
    var elephantA = Elephant{ .letter = 'A' };
    var elephantB = Elephant{ .letter = 'B' };
    var elephantC = Elephant{ .letter = 'C' };

    // We link the elephants so that each tail "points" to the next.
    elephantA.tail = &elephantB;
    elephantB.tail = &elephantC;
    elephantC.tail = &elephantA; // HERE

    // And link the elephants so that each trunk "points" to the previous.
    elephantB.trunk = &elephantA;
    elephantC.trunk = &elephantB;
    elephantA.trunk = &elephantC; // HERE

    visitElephants(&elephantA);

    std.debug.print("\n", .{});
}

// ...

• Run the specific program : zig build -Dn=49
• Rush to killing the phantom tasks that have been created.

Expected Behavior

Here are three screenshots of the situation.

1. The error message.
2. The phantom processes after running the program 4 times.
3. A btop screenshot after launching the program 4 times in a row.

I basically ran it 4 times in a row after seeing it not working, and it ate half of my CPU cores, while making my CPU heat skyrocket.

I'm not quite sure from where it comes from, either from a Zig compilation problem, or from the ziglings exercise itself. However, I don't have this problem with other Ziglings problems.

I would appreciate any insights on the situation ! Thanks :)

NB:

• I know that this is not the intended solution for the exercise, however, I expect my program not to let open random processes after it terminates.

[mlugg]

These aren't "phantom tasks", they're your compiled program running. By linking these 3 values in a loop, you have set up the code so that visitElephants will busy-loop forever, hence pinning a CPU core at ~100%.

[adrsimon]

@mlugg thanks for your help ! I knew I was doing a loop, and that was kind of intended.

I'm not quite sure I get why doesn't the program stops when it encounters the error to be honest. I might not be enough used to Zig's compilation flow to understand what's going on.

The error kinda says that it was unable to spawn the program and to make it run, but then I still have a program going on ?

[mlugg]

Apologies, I didn't read the issue closely enough. It's a bug that the zig build process is terminating.

[mlugg]

Ahh, okay, std.Build.Step.Run has a mechanism to give up and kill the child process if it emits too much stdio. That explains the error.

But... somehow the child process isn't being killed? Even though there's an errdefer for it:

zig/lib/std/Build/Step/Run.zig

Lines 1352 to 1362 in 7aa95bc

	try child.spawn();
	errdefer {
	_ = child.kill() catch {};
	}
	var timer = try std.time.Timer.start();
	const result = if (run.stdio == .zig_test)
	try evalZigTest(run, &child, prog_node, fuzz_context)
	else
	try evalGeneric(run, &child);

I'm guessing we're hitting an error condition on std.process.Child.kill?

[ianprime0509]

Ziglings doesn't use the standard run step, it has its own custom step which uses std.process.Child.run: https://codeberg.org/ziglings/exercises/src/commit/468e3d7eed7cfbb68067afb34593dbc1109396d7/build.zig#L342-L352

However, given there's no way for the user of std.process.Child.run to do anything about the child process when an error is returned (since the child process is created and spawned all within run), maybe the same errdefer should be added after spawning the child process?

zig/lib/std/process/Child.zig

Line 398 in 7aa95bc

try child.spawn();

(with some extra care taken to ensure the errdefer is only scoped up through try child.wait() and not, say, if try stdout.toOwnedSlice() fails after the process has already terminated)

[alexrp]

cc @matklad (#21144)

[mlugg]

Oh, wow, I didn't realise Ziglings had a custom step for this. That's... interesting.

But yeah, this should be handled by std.process.Child.run.