Re: [PATCH 1/2] rust: sync: refactor static_lock_class!() macro

[Boqun Feng]

On Wed, Jul 23, 2025 at 09:46:03PM +0200, Benno Lossin wrote:
> On Wed Jul 23, 2025 at 6:20 PM CEST, Boqun Feng wrote:
> > On Wed, Jul 23, 2025 at 05:01:39PM +0200, Alice Ryhl wrote:
> >> On Wed, Jul 23, 2025 at 4:36 PM Benno Lossin <lossin@xxxxxxxxxx> wrote:
> >> > On Wed Jul 23, 2025 at 1:49 PM CEST, Alice Ryhl wrote:
> >> > >  impl LockClassKey {
> >> > > +    /// Initializes a statically allocated lock class key.
> >> > > +    ///
> >> > > +    /// This is usually used indirectly through the [`static_lock_class!`] macro.
> >> > > +    ///
> >> > > +    /// # Safety
> >> > > +    ///
> >> > > +    /// The destructor must never run on the returned `LockClassKey`.
> >> >
> >> > I don't know how lockdep works, but Boqun mentioned in the other thread
> >> > that it uses the address of static keys. But AFAIK there is no mechanism
> >> > to differentiate them, so does lockdep just check the address and if it
> >
> > In lockdep, we use `static_obj()` to tell whether it's a static obj or a
> > dynamic allocated one.
> 
> So the code below will go in the non-static code path. Why doesn't it
> need to be initialized/registered? (but other cases need it?)
> 

Becasue all the dynamic lock class keys are put in a hash list (using an
intrusive single linked list), so you have to register it before use and
unregister after use.

> >> > is in a static segment it uses different behavior?
> >> >
> >> > Because from the safety requirements on this function, I could just do
> >> > this:
> >> >
> >> >     // SAFETY: we leak the box below, so the destructor never runs.
> >> >     let class = KBox::new(unsafe { LockClassKey::new_static() });
> >> >     let class = Pin::static_ref(KBox::leak(class));
> >> >     let lock = SpinLock::new(42, c_str!("test"), class);
> >
> > This will trigger a runtime error because `class` is not static, but
> > technically, it won't trigger UB, at least lockdep should be able to
> > handle this case.
> 
> Could you go into more details? What is the "technically it won't
> trigger UB" part about?
> 

If a dynamic key is not registered, lockdep will simply just skip the
initialization of locks, report an error and disable itself entirely. So
it won't cause UB.

Regards,
Boqun

> ---
> Cheers,
> Benno