Bindgen should be configured to use `union` + `ManuallyDrop`

[workingjubilee]

re: Richardhongyu/rfl_empirical_tools#1
re: https://www.usenix.org/system/files/atc24-li-hongyu.pdf

Researchers see that bindgen currently uses the __BindgenUnionField and then make logically incorrect accusations about the (in)capabilities of rustc. Accusations that nonetheless survive peer review because they were not reviewed by Rust ABI experts, precious few of those there are, because most of those people are not considered acclaimed academes. At least, not in the halls of "computer science".

These researchers should be rebutted by adjusting the configuration accordingly, or else new subissues should be filed reflecting the challenges involved in doing so against rust-bindgen and/or rustc, until this issue is resolved.

[ojeda]

Leaving aside the paper, as long as something is an improvement for the kernel, then we should consider it.

If I understand correctly, is the suggestion to use --default-non-copy-union-style manually_drop? (--rust-target is already there, and even when it wasn't, bindgen's default is the highest, so that bit should be fine already).

Enabling that indeed makes bindgen use union with ManuallyDrop for some cases, though not for all, including bpf_array:

struct bpf_array {
    union {
        struct {
            struct {} __empty_value;
            char value[];
        };
    };
};

Cc @pvdrz

[Richardhongyu]

Leaving aside the paper, as long as something is an improvement for the kernel, then we should consider it.

If I understand correctly, is the suggestion to use --default-non-copy-union-style manually_drop? (--rust-target is already there, and even when it wasn't, bindgen's default is the highest, so that bit should be fine already).

Enabling that indeed makes bindgen use union with ManuallyDrop for some cases, though not for all, including bpf_array:

struct bpf_array {
union {
struct {
struct {} __empty_value;
char value[];
};
};
};
Cc @pvdrz

I found this is related to the rule of zero length.

Any union with this pattern will not use a rust union.

	union {
		DECLARE_FLEX_ARRAY(struct foo, foo_name);
		DECLARE_FLEX_ARRAY(struct bar, bar_name);
	};

[tamird]

I filed rust-lang/rust-bindgen#3130 upstream.

[workingjubilee]

If I understand correctly, is the suggestion to use --default-non-copy-union-style manually_drop? (--rust-target is already there, and even when it wasn't, bindgen's default is the highest, so that bit should be fine already).

Yes.

Obviously if there's a bug in bindgen's support that doesn't work out that can be fixed.

[tamird]

The upstream issue was fixed in rust-lang/rust-bindgen@937b8d0. It's not yet released and there's hasn't been a bindgen release since 2025-08-31 ~8 months ago.