On Fri, Apr 16, 2021 at 07:47:32PM +0200, Miguel Ojeda wrote:

On Fri, Apr 16, 2021 at 7:05 PM Linus Torvalds
[off-list ref] wrote:

quoted

Typical Rust error handling should match the regular kernel
IS_ERR/ERR_PTR/PTR_ERR model fairly well, although the syntax is
fairly different (and it's not limited to pointers).

Yeah, exactly. We already have a `KernelResult<T>` type which is a
`Result<T, Error>`, where `Error` is a wrapper for the usual kernel
int errors.

So, for instance, a function that can either fail or return `Data`
would have a declaration like:

    pub fn foo() -> KernelResult<Data>

A caller that needs to handle the error can use pattern matching or
one of the methods in `Result`. And if they only need to bubble the
error up, they can use the ? operator:

    pub fn bar() -> KernelResult<Data> {
        let data = foo()?;

        // `data` is already a `Data` here, not a `KernelResult<Data>`
    }

Umm...  A fairly common situation is

foo() returns a pointer to struct foo instance or ERR_PTR()
bar() returns a pointer to struct bar instance of ERR_PTR()

bar()
{
	struct foo *p;
	struct bar *res;
	.... // do some work, grab a mutex, etc.
	p = foo();
	if (IS_ERR(p))
		res = ERR_CAST(p); // (void *)p, internally; conceptually it's
				   // ERR_PTR(PTR_ERR(p));
	else
		res = blah();
	.... // matching cleanup
	return res;
}

How well would ? operator fit that pattern?  _If_ it's just a syntax sugar
along the lines of "if argument matches Err(_), return Err(_)", the types
shouldn't be an issue, but that might need some fun with releasing resources,
etc.  If it's something more elaborate... details, please.