[PATCH v10 2/5] rust: support formatting of foreign types

Tamir Duberstein tamird at gmail.com
Tue May 27 15:02:05 UTC 2025


On Mon, May 26, 2025 at 7:01 PM Benno Lossin <lossin at kernel.org> wrote:
>
> On Tue May 27, 2025 at 12:17 AM CEST, Tamir Duberstein wrote:
> > On Mon, May 26, 2025 at 10:48 AM Benno Lossin <lossin at kernel.org> wrote:
> >> On Sat May 24, 2025 at 10:33 PM CEST, Tamir Duberstein wrote:
> >> > Introduce a `fmt!` macro which wraps all arguments in
> >> > `kernel::fmt::Adapter` This enables formatting of foreign types (like
> >> > `core::ffi::CStr`) that do not implement `fmt::Display` due to concerns
> >> > around lossy conversions which do not apply in the kernel.
> >> >
> >> > Replace all direct calls to `format_args!` with `fmt!`.
> >> >
> >> > In preparation for replacing our `CStr` with `core::ffi::CStr`, move its
> >> > `fmt::Display` implementation to `kernel::fmt::Adapter<&CStr>`.
> >> >
> >> > Suggested-by: Alice Ryhl <aliceryhl at google.com>
> >> > Link: https://rust-for-linux.zulipchat.com/#narrow/channel/288089-General/topic/Custom.20formatting/with/516476467
> >> > Signed-off-by: Tamir Duberstein <tamird at gmail.com>
> >> > ---
> >> >  drivers/block/rnull.rs      |   2 +-
> >> >  rust/kernel/block/mq.rs     |   2 +-
> >> >  rust/kernel/device.rs       |   2 +-
> >> >  rust/kernel/fmt.rs          |  77 +++++++++++++++++++++++++++++
> >> >  rust/kernel/kunit.rs        |   6 +--
> >> >  rust/kernel/lib.rs          |   1 +
> >> >  rust/kernel/prelude.rs      |   3 +-
> >> >  rust/kernel/print.rs        |   4 +-
> >> >  rust/kernel/seq_file.rs     |   2 +-
> >> >  rust/kernel/str.rs          |  23 ++++-----
> >> >  rust/macros/fmt.rs          | 118 ++++++++++++++++++++++++++++++++++++++++++++
> >> >  rust/macros/lib.rs          |  19 +++++++
> >> >  scripts/rustdoc_test_gen.rs |   2 +-
> >> >  13 files changed, 235 insertions(+), 26 deletions(-)
> >>
> >> Can you split this into creating the proc-macro, forwarding the display
> >> impls and replacing all the uses with the proc macro?
> >
> > Can you help me understand why that's better?
>
> It makes reviewing significantly easier.
>
> >> > +macro_rules! impl_display_forward {
> >> > +    ($(
> >> > +        $( { $($generics:tt)* } )? $ty:ty $( { where $($where:tt)* } )?
> >>
> >> You don't need `{}` around the `where` clause, as a `where` keyword can
> >> follow a `ty` fragment.
> >
> > This doesn't work:
> > ```
> > error: local ambiguity when calling macro `impl_display_forward`:
> > multiple parsing options: built-in NTs tt ('r#where') or 2 other
> > options.
> >   --> rust/kernel/fmt.rs:75:78
> >    |
> > 75 |     {<T: ?Sized>} crate::sync::Arc<T> where crate::sync::Arc<T>:
> > fmt::Display,
> >    |
> >            ^
> > ```
>
> Ah right that's a shame, forgot about the `tt`s at the end...
>
> >> > +impl_display_forward!(
> >> > +    bool,
> >> > +    char,
> >> > +    core::panic::PanicInfo<'_>,
> >> > +    crate::str::BStr,
> >> > +    fmt::Arguments<'_>,
> >> > +    i128,
> >> > +    i16,
> >> > +    i32,
> >> > +    i64,
> >> > +    i8,
> >> > +    isize,
> >> > +    str,
> >> > +    u128,
> >> > +    u16,
> >> > +    u32,
> >> > +    u64,
> >> > +    u8,
> >> > +    usize,
> >> > +    {<T: ?Sized>} crate::sync::Arc<T> {where crate::sync::Arc<T>: fmt::Display},
> >> > +    {<T: ?Sized>} crate::sync::UniqueArc<T> {where crate::sync::UniqueArc<T>: fmt::Display},
> >> > +);
> >>
> >> If we use `{}` instead of `()`, then we can format the contents
> >> differently:
> >>
> >>     impl_display_forward! {
> >>         i8, i16, i32, i64, i128, isize,
> >>         u8, u16, u32, u64, u128, usize,
> >>         bool, char, str,
> >>         crate::str::BStr,
> >>         fmt::Arguments<'_>,
> >>         core::panic::PanicInfo<'_>,
> >>         {<T: ?Sized>} crate::sync::Arc<T> {where Self: fmt::Display},
> >>         {<T: ?Sized>} crate::sync::UniqueArc<T> {where Self: fmt::Display},
> >>     }
> >
> > Is that formatting better? rustfmt refuses to touch it either way.
>
> Yeah rustfmt doesn't touch macro parameters enclosed in `{}`. I think
> it's better.

OK, but why? This seems entirely subjective.

> >> > +/// Please see [`crate::fmt`] for documentation.
> >> > +pub(crate) fn fmt(input: TokenStream) -> TokenStream {
> >> > +    let mut input = input.into_iter();
> >> > +
> >> > +    let first_opt = input.next();
> >> > +    let first_owned_str;
> >> > +    let mut names = BTreeSet::new();
> >> > +    let first_lit = {
> >> > +        let Some((mut first_str, first_lit)) = (match first_opt.as_ref() {
> >> > +            Some(TokenTree::Literal(first_lit)) => {
> >> > +                first_owned_str = first_lit.to_string();
> >> > +                Some(first_owned_str.as_str()).and_then(|first| {
> >> > +                    let first = first.strip_prefix('"')?;
> >> > +                    let first = first.strip_suffix('"')?;
> >> > +                    Some((first, first_lit))
> >> > +                })
> >> > +            }
> >> > +            _ => None,
> >> > +        }) else {
> >> > +            return first_opt.into_iter().chain(input).collect();
> >> > +        };
> >>
> >> This usage of let-else + match is pretty confusing and could just be a
> >> single match statement.
> >
> > I don't think so. Can you try rewriting it into the form you like?
>
>     let (mut first_str, first_lit) match first_opt.as_ref() {
>         Some(TokenTree::Literal(lit)) if lit.to_string().starts_with('"') => {
>             let contents = lit.to_string();
>             let contents = contents.strip_prefix('"').unwrap().strip_suffix('"').unwrap();
>             ((contents, lit))
>         }
>         _ => return first_opt.into_iter().chain(input).collect(),
>     };

What happens if the invocation is utterly malformed, e.g.
`fmt!("hello)`? You're unwrapping here, which I intentionally avoid.

>
> >> > +        while let Some((_, rest)) = first_str.split_once('{') {
> >> > +            first_str = rest;
> >> > +            if let Some(rest) = first_str.strip_prefix('{') {
> >> > +                first_str = rest;
> >> > +                continue;
> >> > +            }
> >> > +            while let Some((name, rest)) = first_str.split_once('}') {
> >> > +                first_str = rest;
> >> > +                if let Some(rest) = first_str.strip_prefix('}') {
> >>
> >> This doesn't make sense, we've matched a `{`, some text and a `}`. You
> >> can't escape a `}` that is associated to a `{`.
> >
> > Sure, but such input would be malformed, so I don't think it's
> > necessary to handle it "perfectly". We'll get a nice error from
> > format_args anyhow.
>
> My suggestion in this case would be to just remove this if-let. The
> search for `{` above would skip the `}` if it's correct.
>
> > https://play.rust-lang.org/?version=stable&mode=debug&edition=2024&gist=5f529d93da7cf46b3c99ba7772623e33

Makes sense to me.

>
> Yes it will error like that, but if we do the replacement only when the
> syntax is correct, there also will be compile errors because of a
> missing `Display` impl, or is that not the case?

I'm not sure - I would guess syntax errors "mask" typeck errors.

>
> I'm a bit concerned about the ergonomics that this change will
> introduce, but I guess there really isn't anything that we can do about
> except not do it.
>
> >> > +                    first_str = rest;
> >> > +                    continue;
> >> > +                }
> >> > +                let name = name.split_once(':').map_or(name, |(name, _)| name);
> >> > +                if !name.is_empty() && !name.chars().all(|c| c.is_ascii_digit()) {
> >> > +                    names.insert(name);
> >> > +                }
> >> > +                break;
> >> > +            }
> >> > +        }
> >> > +        first_lit
> >>
> >> `first_lit` is not modified, so could we just the code above it into a
> >> block instead of keeping it in the expr for `first_lit`?
> >
> > As above, can you suggest the alternate form you like better? The
> > gymnastics here are all in service of being able to let malformed
> > input fall through to core::format_args which will do the hard work of
> > producing good diagnostics.
>
> I don't see how this is hard, just do:
>
>     let (first_str, first_lit) = ...;

It requires you to unwrap, like you did above, which is what I'm
trying to avoid.

>
>     while ...
>
> >> > +    };
> >> > +
> >> > +    let first_span = first_lit.span();
> >> > +    let adapt = |expr| {
> >> > +        let mut borrow =
> >> > +            TokenStream::from_iter([TokenTree::Punct(Punct::new('&', Spacing::Alone))]);
> >> > +        borrow.extend(expr);
> >> > +        make_ident(first_span, ["kernel", "fmt", "Adapter"])
> >> > +            .chain([TokenTree::Group(Group::new(Delimiter::Parenthesis, borrow))])
> >>
> >> This should be fine with using `quote!`:
> >>
> >>     quote!(::kernel::fmt::Adapter(&#expr))
> >
> > Yeah, I have a local commit that uses quote_spanned to remove all the
> > manual constructions.
>
> I don't think that you need `quote_spanned` here at all. If you do, then
> let me know, something weird with spans is going on then.

You need to give idents a span, so each of `kernel`, `fmt`, and
`adapter` need a span. I *could* use `quote!` and get whatever span it
uses (mixed_site) but I'd rather retain control.

>
> >> > +    };
> >> > +
> >> > +    let flush = |args: &mut TokenStream, current: &mut TokenStream| {
> >> > +        let current = std::mem::take(current);
> >> > +        if !current.is_empty() {
> >> > +            args.extend(adapt(current));
> >> > +        }
> >> > +    };
> >> > +
> >> > +    let mut args = TokenStream::from_iter(first_opt);
> >> > +    {
> >> > +        let mut current = TokenStream::new();
> >> > +        for tt in input {
> >> > +            match &tt {
> >> > +                TokenTree::Punct(p) => match p.as_char() {
> >> > +                    ',' => {
> >> > +                        flush(&mut args, &mut current);
> >> > +                        &mut args
> >> > +                    }
> >> > +                    '=' => {
> >> > +                        names.remove(current.to_string().as_str());
> >> > +                        args.extend(std::mem::take(&mut current));
> >> > +                        &mut args
> >> > +                    }
> >> > +                    _ => &mut current,
> >> > +                },
> >> > +                _ => &mut current,
> >> > +            }
> >> > +            .extend([tt]);
> >> > +        }
> >>
> >> This doesn't handle the following code correctly ):
> >>
> >>     let mut a = 0;
> >>     pr_info!("{a:?}", a = a = a);
> >>
> >> Looks like we'll have to remember what "kind" of an equals we parsed...
> >
> > Hmm, good point. Maybe we can just avoid dealing with `=` at all until
> > we hit the `,` and just split on the leftmost `=`. WDYT? I'll have
> > that in v11.
>
> Sounds good, if there is no `=`, then ignore it.
>
> >> > +/// Like [`core::format_args!`], but automatically wraps arguments in [`kernel::fmt::Adapter`].
> >> > +///
> >> > +/// This macro allows generating `core::fmt::Arguments` while ensuring that each argument is wrapped
> >> > +/// with `::kernel::fmt::Adapter`, which customizes formatting behavior for kernel logging.
> >> > +///
> >> > +/// Named arguments used in the format string (e.g. `{foo}`) are detected and resolved from local
> >> > +/// bindings. All positional and named arguments are automatically wrapped.
> >> > +///
> >> > +/// This macro is an implementation detail of other kernel logging macros like [`pr_info!`] and
> >> > +/// should not typically be used directly.
> >> > +///
> >> > +/// [`kernel::fmt::Adapter`]: ../kernel/fmt/struct.Adapter.html
> >> > +/// [`pr_info!`]: ../kernel/macro.pr_info.html
> >> > +#[proc_macro]
> >> > +pub fn fmt(input: TokenStream) -> TokenStream {
> >>
> >> I'm wondering if we should name this `format_args` instead in order to
> >> better communicate that it's a replacement for `core::format_args!`.
> >
> > Unfortunately that introduces ambiguity in cases where
> > kernel::prelude::* is imported because core::format_args is in core's
> > prelude.
>
> Ahh that's unfortunate.
>
> ---
> Cheers,
> Benno


More information about the dri-devel mailing list