Re: [PATCH] rust: macros: fix soundness issue in `module!` macro

[Wedson Almeida Filho]

On Sun, 31 Mar 2024 at 07:27, Benno Lossin <benno.lossin@xxxxxxxxx> wrote:
>
> On 31.03.24 03:00, Wedson Almeida Filho wrote:
> > On Wed, 27 Mar 2024 at 13:04, Benno Lossin <benno.lossin@xxxxxxxxx> wrote:
> >> +                    #[cfg(not(MODULE))]
> >> +                    #[doc(hidden)]
> >> +                    #[no_mangle]
> >> +                    pub extern \"C\" fn __{name}_exit() {{
> >> +                        __exit()
>
> I just noticed this should be wrapped in an `unsafe` block with a SAFETY
> comment. Will fix this in v2.
>
> >> +                    }}
> >>
> >> -            #[cfg(not(MODULE))]
> >> -            #[doc(hidden)]
> >> -            #[no_mangle]
> >> -            pub extern \"C\" fn __{name}_exit() {{
> >> -                __exit()
> >> -            }}
> >> +                    /// # Safety
> >> +                    ///
> >> +                    /// This function must
> >> +                    /// - only be called once,
> >> +                    /// - not be called concurrently with `__exit`.
> >
> > I don't think the second item is needed here, it really is a
> > requirement on `__exit`.
>
> Fixed.
>
> >
> >> +                    unsafe fn __init() -> core::ffi::c_int {{
> >> +                        match <{type_} as kernel::Module>::init(&THIS_MODULE) {{
> >> +                            Ok(m) => {{
> >> +                                // SAFETY:
> >> +                                // no data race, since `__MOD` can only be accessed by this module and
> >> +                                // there only `__init` and `__exit` access it. These functions are only
> >> +                                // called once and `__exit` cannot be called before or during `__init`.
> >> +                                unsafe {{
> >> +                                    __MOD = Some(m);
> >> +                                }}
> >> +                                return 0;
> >> +                            }}
> >> +                            Err(e) => {{
> >> +                                return e.to_errno();
> >> +                            }}
> >> +                        }}
> >> +                    }}
> >>
> >> -            fn __init() -> core::ffi::c_int {{
> >> -                match <{type_} as kernel::Module>::init(&THIS_MODULE) {{
> >> -                    Ok(m) => {{
> >> +                    /// # Safety
> >> +                    ///
> >> +                    /// This function must
> >> +                    /// - only be called once,
> >> +                    /// - be called after `__init`,
> >> +                    /// - not be called concurrently with `__init`.
> >
> > The second item is incomplete: it must be called after `__init` *succeeds*.
>
> Indeed.
>
> >
> > With that added (which is a different precondition), I think the third
> > item can be dropped because if you have to wait to see whether
> > `__init` succeeded or failed before you can call `__exit`, then
> > certainly you cannot call it concurrently with `__init`.
>
> I would love to drop that requirement, but I am not sure we can. With
> that requirement, I wanted to ensure that no data race on `__MOD` can
> happen. If you need to verify that `__init` succeeded, one might think
> that it is not possible to call `__exit` such that a data race occurs,
> but I think it could theoretically be done if the concrete `Module`
> implementation never failed.

I see. If you're concerned about compiler reordering, then we need
compiler barriers.

> Do you have any suggestion for what I could add to the "be called after
> `__init` was called and returned `0`" requirement to make a data race
> impossible?

If you're concerned with reordering from the processor as well, then
we need cpu barriers. You'd have to say that the cpu/thread executing
`__init` must have a release barrier after `__init` completes, and the
thread/cpu doing `__exit` must have an acquire barrier before starting
`__exit`.

But I'm not sure we need to go that far. Mostly because C is going to
guarantee that ordering for us, so I'd say we can just omit this or
perhaps say "This function must only be called from the exit module
implementation".

> --
> Cheers,
> Benno
>
> >
> >> +                    unsafe fn __exit() {{
> >> +                        // SAFETY:
> >> +                        // no data race, since `__MOD` can only be accessed by this module and there
> >> +                        // only `__init` and `__exit` access it. These functions are only called once
> >> +                        // and `__init` was already called.
> >>                           unsafe {{
> >> -                            __MOD = Some(m);
> >> +                            // Invokes `drop()` on `__MOD`, which should be used for cleanup.
> >> +                            __MOD = None;
> >>                           }}
> >> -                        return 0;
> >>                       }}
> >> -                    Err(e) => {{
> >> -                        return e.to_errno();
> >> -                    }}
> >> -                }}
> >> -            }}
> >>
> >> -            fn __exit() {{
> >> -                unsafe {{
> >> -                    // Invokes `drop()` on `__MOD`, which should be used for cleanup.
> >> -                    __MOD = None;
> >> +                    {modinfo}
> >>                   }}
> >>               }}
> >> -
> >> -            {modinfo}
> >>           ",
> >>           type_ = info.type_,
> >>           name = info.name,
> >>
> >> base-commit: 4cece764965020c22cff7665b18a012006359095
> >> --
> >> 2.44.0
> >>
> >>
>