Re: [PATCH v3 2/4] rust: print: document safety of _printk FFI calls

[Gary Guo]

On Sun Jul 26, 2026 at 12:59 PM BST, Cian McGuire wrote:
> Two calls in this file had their SAFETY comments left as "TODO.": the
> cast back to `fmt::Arguments` in `rust_fmt_argument`, and the call to
> `bindings::_printk` in `call_printk`.
>
> `rust_fmt_argument` is only reachable through the `%pA` format
> specifier. Every in-tree Rust caller that uses `%pA` -- `call_printk`,
> `device::printk`, `kunit::printk`, and `seq_file::call_printf` -- invokes
> it synchronously with a pointer derived from its own live
> `fmt::Arguments`, as one call in a `switch` dispatch in
> `lib/vsprintf.c`'s pointer formatter that returns its result immediately;
> there is no deferred use or storage of the pointer on the C side. So the
> `fmt::Arguments` value stays alive on the caller's stack for the whole
> time `rust_fmt_argument` is running with `ptr` pointing at it.
>
> The `_printk` call itself is justified by this function's `# Safety`
> section, which this patch also tightens: `format_string` must be one of
> the non-continuation `format_strings` statics, not
> `format_strings::CONT`. `call_printk` always supplies two extra
> arguments -- `module_name` for `%s` and `args` for `%pA` -- which only
> matches the `"%s: %pA"` shape shared by every non-continuation string;
> `CONT`'s format string is `"%pA"` only, so passing it here would make
> `_printk` consume `module_name`'s pointer for `%pA` instead, and
> `rust_fmt_argument` would then reinterpret that `*const c_char` as a
> `*const fmt::Arguments` and dereference it. No caller does this today --
> `call_printk` is only reached from `print_macro!`'s non-continuation arm;
> the continuation arm calls the separate, safe `call_printk_cont`, which
> hardcodes `CONT` together with the single argument it expects -- but the
> previous, broader contract permitted it. With `format_string` guaranteed
> non-`CONT`, and `module_name` guaranteed null-terminated by both of its
> possible origins (`kernel::__LOG_PREFIX` and the one generated by the
> `module!` proc macro), the specifiers always match the supplied
> arguments.
>
> Suggested-by: Miguel Ojeda <ojeda@xxxxxxxxxx>
> Link: https://github.com/Rust-for-Linux/linux/issues/351
> Signed-off-by: Cian McGuire <cianmcguire167@xxxxxxxxx>
> ---
> v3:
> - Tightened `call_printk`'s `# Safety` contract to require a
>   non-continuation format string (i.e. not `format_strings::CONT`),
>   since the function always supplies both a `%s` and a `%pA` argument,
>   which only matches the non-continuation strings' `"%s: %pA"` shape.
>   Passing `CONT` under the old, broader contract would make `_printk`
>   consume `module_name`'s pointer for `%pA` instead of `args`, which
>   `rust_fmt_argument` would then wrongly reinterpret as a
>   `*const fmt::Arguments` and dereference. Updated the inline SAFETY
>   comment to match.
> v2:
> - Reworded the SAFETY comment on `rust_fmt_argument`'s cast to argue
>   from the property that makes every `%pA` call site safe (synchronous
>   invocation with a live `&fmt::Arguments`), instead of enumerating
>   callers by name, so it doesn't go stale if a new caller is added.
>
>  rust/kernel/print.rs | 18 ++++++++++++++----
>  1 file changed, 14 insertions(+), 4 deletions(-)
>
> diff --git a/rust/kernel/print.rs b/rust/kernel/print.rs
> index 6fd84389a858..5a0ee103b54e 100644
> --- a/rust/kernel/print.rs
> +++ b/rust/kernel/print.rs
> @@ -29,7 +29,10 @@
>      use fmt::Write;
>      // SAFETY: The C contract guarantees that `buf` is valid if it's less than `end`.
>      let mut w = unsafe { RawFormatter::from_ptrs(buf.cast(), end.cast()) };
> -    // SAFETY: TODO.
> +    // SAFETY: This function is only reachable via the `%pA` format specifier; using `%pA` from
> +    // Rust requires the caller to invoke it synchronously with a `ptr` derived from a live
> +    // `&fmt::Arguments`, before that reference's scope can end. `ptr` is therefore valid and
> +    // properly aligned for a `fmt::Arguments<'_>` for the duration of this call.

Bulk of this should be safety precondition of the function, not safety comment.

Best,
Gary

>      let _ = w.write_fmt(unsafe { *ptr.cast::<fmt::Arguments<'_>>() });
>      w.pos().cast()
>  }
> @@ -96,8 +99,11 @@ pub mod format_strings {
>  ///
>  /// # Safety
>  ///
> -/// The format string must be one of the ones in [`format_strings`], and
> -/// the module name must be null-terminated.
> +/// The format string must be one of the non-continuation strings in
> +/// [`format_strings`] (i.e. not [`format_strings::CONT`]): this function
> +/// always supplies both `module_name` and `args` as arguments, which only
> +/// matches the `"%s: %pA"` shape used by the non-continuation strings, not
> +/// `CONT`'s `"%pA"`-only shape. The module name must be null-terminated.
>  ///
>  /// [`_printk`]: srctree/include/linux/_printk.h
>  #[doc(hidden)]
> @@ -109,7 +115,11 @@ pub unsafe fn call_printk(
>  ) {
>      // `_printk` does not seem to fail in any path.
>      #[cfg(CONFIG_PRINTK)]
> -    // SAFETY: TODO.
> +    // SAFETY: By this function's safety requirements, `format_string` is one of the
> +    // non-continuation `format_strings` statics (never `CONT`), so its fixed `"%s: %pA"` shape
> +    // matches the two arguments supplied below: `module_name` is null-terminated as `%s`
> +    // requires, and the third argument points at `args`, which remains valid for the duration
> +    // of this call and is only read back synchronously, through `%pA`, by `rust_fmt_argument`.
>      unsafe {
>          bindings::_printk(
>              format_string.as_ptr(),