Message-ID: <D8A3THTBK4SK.3LI57W1VI580J@proton.me>
Date: Fri, 07 Mar 2025 14:29:58 +0000
From: Benno Lossin <benno.lossin@...ton.me>
To: Andreas Hindborg <a.hindborg@...nel.org>
Cc: Miguel Ojeda <ojeda@...nel.org>, Anna-Maria Behnsen <anna-maria@...utronix.de>, Frederic Weisbecker <frederic@...nel.org>, Thomas Gleixner <tglx@...utronix.de>, Danilo Krummrich <dakr@...nel.org>, Alex Gaynor <alex.gaynor@...il.com>, Boqun Feng <boqun.feng@...il.com>, Gary Guo <gary@...yguo.net>, Björn Roy Baron <bjorn3_gh@...tonmail.com>, Alice Ryhl <aliceryhl@...gle.com>, Trevor Gross <tmgross@...ch.edu>, Lyude Paul <lyude@...hat.com>, Guangbo Cui <2407018371@...com>, Dirk Behme <dirk.behme@...il.com>, Daniel Almeida <daniel.almeida@...labora.com>, Tamir Duberstein <tamird@...il.com>, Markus Elfring <Markus.Elfring@....de>, rust-for-linux@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v10 10/13] rust: hrtimer: implement `HrTimerPointer` for `Pin<Box<T>>`

On Fri Mar 7, 2025 at 3:01 PM CET, Andreas Hindborg wrote:
> "Benno Lossin" <benno.lossin@...ton.me> writes:
>
>> On Fri Mar 7, 2025 at 11:11 AM CET, Andreas Hindborg wrote:
>>> Allow `Pin<Box<T>>` to be the target of a timer callback.
>>>
>>> Acked-by: Frederic Weisbecker <frederic@...nel.org>
>>> Reviewed-by: Lyude Paul <lyude@...hat.com>
>>> Signed-off-by: Andreas Hindborg <a.hindborg@...nel.org>
>>> ---
>>>  rust/kernel/time/hrtimer.rs      |   3 ++
>>>  rust/kernel/time/hrtimer/tbox.rs | 109 +++++++++++++++++++++++++++++++++++++++
>>>  2 files changed, 112 insertions(+)
>>>
>>> diff --git a/rust/kernel/time/hrtimer.rs b/rust/kernel/time/hrtimer.rs
>>> index d2791fd624b7..991d37b0524a 100644
>>> --- a/rust/kernel/time/hrtimer.rs
>>> +++ b/rust/kernel/time/hrtimer.rs
>>> @@ -443,3 +443,6 @@ unsafe fn timer_container_of(ptr: *mut $crate::time::hrtimer::HrTimer<$timer_typ
>>>  pub use pin::PinHrTimerHandle;
>>>  mod pin_mut;
>>>  pub use pin_mut::PinMutHrTimerHandle;
>>> +// `box` is a reserved keyword, so prefix with `t` for timer
>>> +mod tbox;
>>> +pub use tbox::BoxHrTimerHandle;
>>> diff --git a/rust/kernel/time/hrtimer/tbox.rs b/rust/kernel/time/hrtimer/tbox.rs
>>> new file mode 100644
>>> index 000000000000..a3b2ed849050
>>> --- /dev/null
>>> +++ b/rust/kernel/time/hrtimer/tbox.rs
>>> @@ -0,0 +1,109 @@
>>> +// SPDX-License-Identifier: GPL-2.0
>>> +
>>> +use super::HasHrTimer;
>>> +use super::HrTimer;
>>> +use super::HrTimerCallback;
>>> +use super::HrTimerHandle;
>>> +use super::HrTimerPointer;
>>> +use super::RawHrTimerCallback;
>>> +use crate::prelude::*;
>>> +use crate::time::Ktime;
>>> +use core::mem::ManuallyDrop;
>>> +use core::ptr::NonNull;
>>> +
>>> +/// A handle for a [`Box<HasHrTimer<T>>`] returned by a call to
>>> +/// [`HrTimerPointer::start`].
>>> +pub struct BoxHrTimerHandle<T, A>
>>
>> Should this type implement `Send` and `Sync` depending on `T`?
>
> Yes. In practice `T` will always be `Send` and `Sync` because of bounds
> on other traits.
>
> I don't think we have to require `T: Sync`, because the handle does not ever
> create shared references to the underlying `T`?

Oh I meant to do:

    unsafe impl<T: Send + Sync, A> Send for BoxHrTimerHandle<T, A> {}

But since you don't have it, it might be unnecessary.

>>> +where
>>> +    T: HasHrTimer<T>,
>>> +    A: crate::alloc::Allocator,
>>> +{
>>> +    pub(crate) inner: NonNull<T>,
>>> +    _p: core::marker::PhantomData<A>,
>>> +}
>>> +
>>> +// SAFETY: We implement drop below, and we cancel the timer in the drop
>>> +// implementation.
>>> +unsafe impl<T, A> HrTimerHandle for BoxHrTimerHandle<T, A>
>>> +where
>>> +    T: HasHrTimer<T>,
>>> +    A: crate::alloc::Allocator,
>>> +{
>>> +    fn cancel(&mut self) -> bool {
>>> +        // SAFETY: As we obtained `self.inner` from a valid reference when we
>>> +        // created `self`, it must point to a valid `T`.
>>> +        let timer_ptr = unsafe { <T as HasHrTimer<T>>::raw_get_timer(self.inner.as_ptr()) };
>>> +
>>> +        // SAFETY: As `timer_ptr` points into `T` and `T` is valid, `timer_ptr`
>>> +        // must point to a valid `HrTimer` instance.
>>> +        unsafe { HrTimer::<T>::raw_cancel(timer_ptr) }
>>> +    }
>>> +}
>>> +
>>> +impl<T, A> Drop for BoxHrTimerHandle<T, A>
>>> +where
>>> +    T: HasHrTimer<T>,
>>> +    A: crate::alloc::Allocator,
>>> +{
>>> +    fn drop(&mut self) {
>>> +        self.cancel();
>>> +        // SAFETY: `self.inner` came from a `Box::into_raw` call
>>
>> Please add this as an invariant to `Self`.
>
> OK.
>
>>
>>> +        drop(unsafe { Box::<T, A>::from_raw(self.inner.as_ptr()) })
>>> +    }
>>> +}
>>> +
>>> +impl<T, A> HrTimerPointer for Pin<Box<T, A>>
>>> +where
>>> +    T: 'static,
>>> +    T: Send + Sync,
>>> +    T: HasHrTimer<T>,
>>> +    T: for<'a> HrTimerCallback<Pointer<'a> = Pin<Box<T, A>>>,
>>> +    Pin<Box<T, A>>: for<'a> RawHrTimerCallback<CallbackTarget<'a> = Pin<&'a T>>,
>>
>> I don't think this is necessary.
>
> Should I remove it? I feel like it communicates intent.

What intent?

>>> +    A: crate::alloc::Allocator,
>>> +{
>>> +    type TimerHandle = BoxHrTimerHandle<T, A>;
>>> +
>>> +    fn start(self, expires: Ktime) -> Self::TimerHandle {
>>> +        // SAFETY:
>>> +        //  - We will not move out of this box during timer callback (we pass an
>>> +        //    immutable reference to the callback).
>>> +        //  - `Box::into_raw` is guaranteed to return a valid pointer.
>>> +        let inner =
>>> +            unsafe { NonNull::new_unchecked(Box::into_raw(Pin::into_inner_unchecked(self))) };
>>> +
>>> +        // SAFETY:
>>> +        //  - We keep `self` alive by wrapping it in a handle below.
>>> +        //  - Since we generate the pointer passed to `start` from a valid
>>> +        //    reference, it is a valid pointer.
>>> +        unsafe { T::start(inner.as_ptr(), expires) };
>>> +
>>> +        BoxHrTimerHandle {
>>> +            inner,
>>> +            _p: core::marker::PhantomData,
>>> +        }
>>> +    }
>>> +}
>>> +
>>> +impl<T, A> RawHrTimerCallback for Pin<Box<T, A>>
>>> +where
>>> +    T: 'static,
>>> +    T: HasHrTimer<T>,
>>> +    T: for<'a> HrTimerCallback<Pointer<'a> = Pin<Box<T, A>>>,
>>> +    A: crate::alloc::Allocator,
>>> +{
>>> +    type CallbackTarget<'a> = Pin<&'a T>;
>>
>> Why isn't this `Pin<&'a mut T>`?
>
> I don't think it matters much? There can be no other mutable references
> while the callback is running, so why not a shared ref?

IIUC there can be no references to the value, since the user used a
`Pin<Box<T>>` to schedule the timer.

I thought it might make sense to give a pinned mutable reference, since
you explicitly implement the `RawHrTimerCallback` for `Pin<&mut T>`.
Which made me believe one sometimes needs to modify the `T` from the
callback.

Since we're able to do that when the user used a `Box`, I think we
should just do it.

>>> +
>>> +    unsafe extern "C" fn run(ptr: *mut bindings::hrtimer) -> bindings::hrtimer_restart {
>>> +        // `HrTimer` is `repr(C)`
>>> +        let timer_ptr = ptr.cast::<super::HrTimer<T>>();
>>> +
>>> +        // SAFETY: By C API contract `ptr` is the pointer we passed when
>>> +        // queuing the timer, so it is a `HrTimer<T>` embedded in a `T`.
>>> +        let data_ptr = unsafe { T::timer_container_of(timer_ptr) };
>>> +
>>> +        // SAFETY: We called `Box::into_raw` when we queued the timer.
>>> +        let tbox = ManuallyDrop::new(Box::into_pin(unsafe { Box::<T, A>::from_raw(data_ptr) }));
>>
>> Since you turn this into a reference below and never run the drop, why
>> not turn the pointer directly into a reference?
>
> You mean replace with `unsafe {&*data_ptr};`? I guess that could work,
> but it hinges on `Box` being transparent which is more subtle than going
> through the API.

I think that's cleaner. Also why does that rely on `Box` being
transparent?

---
Cheers,
Benno

Powered by blists - more mailing lists