Message-ID: <87cyf6xv7g.fsf@kernel.org>
Date: Tue, 25 Feb 2025 09:52:19 +0100
From: Andreas Hindborg <a.hindborg@...nel.org>
To: "Tamir Duberstein" <tamird@...il.com>
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>,  "Benno Lossin" <benno.lossin@...ton.me>,
  "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>,  <rust-for-linux@...r.kernel.org>,
  <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v9 01/13] rust: hrtimer: introduce hrtimer support

"Tamir Duberstein" <tamird@...il.com> writes:

> Hi Andreas, mostly grammar and prose clarity comments below.
>
> I still think HasHrTimer::OFFSET is less clear and more fragile than
> just generating compiler-checked implementations in the macro (you're
> already generating OFFSET and one method implementation rather than
> generating 2 method implementations).

I don't agree with you assessment. My argument is that I would rather
generate as little code as possible in the macro, and the trait would in
practice never be implemented by hand.

>
> On Mon, Feb 24, 2025 at 7:06 AM Andreas Hindborg <a.hindborg@...nel.org> wrote:
>>

[...]

>> +//! # Vocabulary
>> +//!
>> +//! States:
>> +//!
>> +//! - Stopped: initialized but not started, or cancelled, or not restarted.
>> +//! - Started: initialized and started or restarted.
>> +//! - Running: executing the callback.
>> +//!
>> +//! Operations:
>> +//!
>> +//! * Start
>> +//! * Cancel
>> +//! * Restart
>> +//!
>> +//! Events:
>> +//!
>> +//! * Expire
>> +//!
>> +//! ## State Diagram
>> +//!
>> +//! ```text
>> +//!                                                   Return NoRestart
>> +//!                       +---------------------------------------------------------------------+
>> +//!                       |                                                                     |
>> +//!                       |                                                                     |
>> +//!                       |                                                                     |
>> +//!                       |                                         Return Restart              |
>> +//!                       |                                      +------------------------+     |
>> +//!                       |                                      |                        |     |
>> +//!                       |                                      |                        |     |
>> +//!                       v                                      v                        |     |
>> +//!           +-----------------+      Start      +------------------+           +--------+-----+--+
>> +//!           |                 +---------------->|                  |           |                 |
>> +//! Init      |                 |                 |                  |  Expire   |                 |
>> +//! --------->|    Stopped      |                 |      Started     +---------->|     Running     |
>> +//!           |                 |     Cancel      |                  |           |                 |
>> +//!           |                 |<----------------+                  |           |                 |
>> +//!           +-----------------+                 +---------------+--+           +-----------------+
>> +//!                                                     ^         |
>> +//!                                                     |         |
>> +//!                                                     +---------+
>> +//!                                                      Restart
>> +//! ```
>> +//!
>> +//!
>> +//! A timer is initialized in the **stopped** state. A stopped timer can be
>> +//! **started** by the `start` operation, with an **expiry** time. After the
>> +//! `start` operation, the timer is in the **started** state. When the timer
>> +//! **expires**, the timer enters the **running** state and the handler is
>> +//! executed. After the handler has finished executing, the timer may enter the
>> +//! **started* or **stopped** state, depending on the return value of the
>> +//! handler. A running timer can be **canceled** by the `cancel` operation. A
>> +//! timer that is cancelled enters the **stopped** state.
>
> This is a bit confusing because it sounds like you're describing a
> *started* timer. After reading the next paragraph I think this wording
> applies to both *started* and *running*, but it isn't unambiguous.

Right, I think I understand. It's a mistake. Last sentence should be:

  A timer in the **started** or **running** state be **canceled** by the
  `cancel` operation. A timer that is cancelled enters the **stopped**
  state.

>
>> +//!
>> +//! A `cancel` or `restart` operation on a timer in the **running** state takes
>> +//! effect after the handler has finished executing and the timer has transitioned
>> +//! out of the **running** state.
>
> There's no external restart, right?

There will be, eventually. Conceptually there is, because the state
diagram and this text describe the operation.

> I think this wording is confused
> by the unification of cancel and restart under operations, though they
> are not isomorphic.

Hmm, I am not following. Can you elaborate? The set of operations is
start, cancel, restart.

> Restart (as I understand it) can only happen from
> the handler, and cancel can only happen via a call to hrtimer_cancel.

This text introduces the restart operation. There is no code path to
reach it from rust at the moment, but I am inclined to add the function
due to this confusion. It would be dead code for now though.

> It's also a bit strange that start isn't mentioned whenever cancel and
> restart are mentioned.

Why is that?

>
>> +//!
>> +//! A `restart` operation on a timer in the **stopped** state is equivalent to a
>> +//! `start` operation.
>> +
>> +use crate::{init::PinInit, prelude::*, time::Ktime, types::Opaque};
>> +use core::marker::PhantomData;
>> +
>> +/// A timer backed by a C `struct hrtimer`.
>> +///
>> +/// # Invariants
>> +///
>> +/// * `self.timer` is initialized by `bindings::hrtimer_setup`.
>> +#[pin_data]
>> +#[repr(C)]
>> +pub struct HrTimer<T> {
>> +    #[pin]
>> +    timer: Opaque<bindings::hrtimer>,
>> +    _t: PhantomData<T>,
>> +}
>> +
>> +// SAFETY: Ownership of an `HrTimer` can be moved to other threads and
>> +// used/dropped from there.
>> +unsafe impl<T> Send for HrTimer<T> {}
>> +
>> +// SAFETY: Timer operations are locked on C side, so it is safe to operate on a
>> +// timer from multiple threads
>
> nit: missing article ("the" C side) and missing period.

Thanks.

>
>> +unsafe impl<T> Sync for HrTimer<T> {}
>> +
>> +impl<T> HrTimer<T> {
>> +    /// Return an initializer for a new timer instance.
>> +    pub fn new() -> impl PinInit<Self>
>> +    where
>> +        T: HrTimerCallback,
>> +    {
>> +        pin_init!(Self {
>> +            // INVARIANTS: We initialize `timer` with `hrtimer_setup` below.
>
> Why plural INVARIANTS?

Mistake, will fix.

>
>> +            timer <- Opaque::ffi_init(move |place: *mut bindings::hrtimer| {
>> +                // SAFETY: By design of `pin_init!`, `place` is a pointer to a
>> +                // live allocation. hrtimer_setup will initialize `place` and
>> +                // does not require `place` to be initialized prior to the call.
>> +                unsafe {
>> +                    bindings::hrtimer_setup(
>> +                        place,
>> +                        Some(T::Pointer::run),
>> +                        bindings::CLOCK_MONOTONIC as i32,
>> +                        bindings::hrtimer_mode_HRTIMER_MODE_REL,
>> +                    );
>> +                }
>> +            }),
>> +            _t: PhantomData,
>> +        })
>> +    }
>> +
>> +    /// Get a pointer to the contained `bindings::hrtimer`.
>> +    ///
>> +    /// This function do not create any references.
>
> s/do/does/

Thanks.

>
> But maybe this should use the same wording from Opaque::raw_get?
>
>     /// This function is useful to get access to the value without
> creating intermediate
>     /// references.

To me those two wordings have the same effect. I don't mind changing the
wording if you feel strongly about it.

>
> If so, consider also naming the argument "this" for consistency. Same
> for other methods below.

Sure.

>
>> +    ///
>> +    /// # Safety
>> +    ///
>> +    /// `ptr` must point to a live allocation of at least the size of `Self`.
>> +    unsafe fn raw_get(ptr: *const Self) -> *mut bindings::hrtimer {
>> +        // SAFETY: The field projection to `timer` does not go out of bounds,
>> +        // because the caller of this function promises that `ptr` points to an
>> +        // allocation of at least the size of `Self`.
>> +        unsafe { Opaque::raw_get(core::ptr::addr_of!((*ptr).timer)) }
>> +    }
>> +
>> +    /// Cancel an initialized and potentially running timer.
>> +    ///
>> +    /// If the timer handler is running, this will block until the handler is
>> +    /// finished.
>
> nit: s/is finished/returns/ and maybe clarify the ordering, namely
> that the timer is definitely in a stopped state after this returns.

    /// If the timer handler is running, this function will block until the
    /// handler returns. Before this function returns, the timer will be in the
    /// stopped state.

If we have a concurrent call to start, the timer might actually be in
the started state when this function returns. But this function will
transition the timer to the stopped state.

>
>> +    ///
>> +    /// Users of the `HrTimer` API would not usually call this method directly.
>> +    /// Instead they would use the safe [`HrTimerHandle::cancel`] on the handle
>> +    /// returned when the timer was started.
>> +    ///
>> +    /// This function does not create any references.
>> +    ///
>> +    /// # Safety
>> +    ///
>> +    /// `self_ptr` must point to a valid `Self`.
>
> Why use different phrasing here than on raw_get? The parameter name is
> also different. Would be nice to be consistent.

They are different requirements, one is stronger than the other. I
construct safety requirements based on the unsafe operations in the
function. The unsafe operations in these two functions have different
requirements. I would not impose a stronger requirement than I have to.

>
>> +    #[allow(dead_code)]
>> +    pub(crate) unsafe fn raw_cancel(self_ptr: *const Self) -> bool {
>> +        // SAFETY: timer_ptr points to an allocation of at least `HrTimer` size.
>> +        let c_timer_ptr = unsafe { HrTimer::raw_get(self_ptr) };
>> +
>> +        // If the handler is running, this will wait for the handler to finish
>> +        // before returning.
>> +        // SAFETY: `c_timer_ptr` is initialized and valid. Synchronization is
>> +        // handled on C side.
>
> missing article here.

👍

>
>> +        unsafe { bindings::hrtimer_cancel(c_timer_ptr) != 0 }
>> +    }
>> +}
>> +
>> +/// Implemented by pointer types that point to structs that embed a [`HrTimer`].
>> +///
>> +/// Target (pointee) must be [`Sync`] because timer callbacks happen in another
>> +/// thread of execution (hard or soft interrupt context).
>
> Is this explaining the bound on the trait, or something that exists
> outside the type system? If it's the former, isn't the Sync bound on
> the trait going to apply to the pointer rather than the pointee?

It is explaining the bound on the trait, and as you say it is not
correct. Pointer types that do not apply synchronization internally can
only be `Sync` when their target is `Sync`, which was probably my line
of thought.

I will rephrase:

  `Self` must be [`Sync`] because timer callbacks happen in another
  thread of execution (hard or soft interrupt context).

>
>> +///
>> +/// Starting a timer returns a [`HrTimerHandle`] that can be used to manipulate
>> +/// the timer. Note that it is OK to call the start function repeatedly, and
>> +/// that more than one [`HrTimerHandle`] associated with a [`HrTimerPointer`] may
>> +/// exist. A timer can be manipulated through any of the handles, and a handle
>> +/// may represent a cancelled timer.
>> +pub trait HrTimerPointer: Sync + Sized {
>> +    /// A handle representing a started or restarted timer.
>> +    ///
>> +    /// If the timer is running or if the timer callback is executing when the
>> +    /// handle is dropped, the drop method of [`HrTimerHandle`] should not return
>> +    /// until the timer is stopped and the callback has completed.
>> +    ///
>> +    /// Note: When implementing this trait, consider that it is not unsafe to
>> +    /// leak the handle.
>
> What does leak mean in this context?

The same as in all other contexts (I think?). Leave the object alive for
'static and forget the address. Thus never drop it and thus never run
the drop method.

>
>> +    type TimerHandle: HrTimerHandle;
>> +
>> +    /// Start the timer with expiry after `expires` time units. If the timer was
>> +    /// already running, it is restarted with the new expiry time.
>> +    fn start(self, expires: Ktime) -> Self::TimerHandle;
>> +}
>> +
>> +/// Implemented by [`HrTimerPointer`] implementers to give the C timer callback a
>> +/// function to call.
>> +// This is split from `HrTimerPointer` to make it easier to specify trait bounds.
>> +pub trait RawHrTimerCallback {
>> +    /// This type is passed to [`HrTimerCallback::run`]. It may be a borrow of
>> +    /// [`Self::CallbackTarget`], or it may be `Self::CallbackTarget` if the
>> +    /// implementation can guarantee correct access (exclusive or shared
>> +    /// depending on the type) to the target during timer handler execution.
>> +    type CallbackTarget<'a>;
>> +
>> +    /// Callback to be called from C when timer fires.
>> +    ///
>> +    /// # Safety
>> +    ///
>> +    /// Only to be called by C code in `hrtimer` subsystem. `ptr` must point to
>
> missing article, should be "...in the `hrtimer`..."

English is difficult 😅

>
>> +    /// the `bindings::hrtimer` structure that was used to start the timer.
>> +    unsafe extern "C" fn run(ptr: *mut bindings::hrtimer) -> bindings::hrtimer_restart;
>> +}
>> +
>> +/// Implemented by structs that can be the target of a timer callback.
>> +pub trait HrTimerCallback {
>> +    /// The type whose [`RawHrTimerCallback::run`] method will be invoked when
>> +    /// the timer expires.
>> +    type Pointer<'a>: RawHrTimerCallback;
>> +
>> +    /// Called by the timer logic when the timer fires.
>> +    fn run(this: <Self::Pointer<'_> as RawHrTimerCallback>::CallbackTarget<'_>)
>> +    where
>> +        Self: Sized;
>> +}
>> +
>> +/// A handle representing a potentially running timer.
>> +///
>> +/// More than one handle representing the same timer might exist.
>> +///
>> +/// # Safety
>> +///
>> +/// When dropped, the timer represented by this handle must be cancelled, if it
>> +/// is running. If the timer handler is running when the handle is dropped, the
>> +/// drop method must wait for the handler to finish before returning.
>> +///
>> +/// Note: One way to satisfy the safety requirement is to call `Self::cancel` in
>> +/// the drop implementation for `Self.`
>> +pub unsafe trait HrTimerHandle {
>> +    /// Cancel the timer, if it is running. If the timer handler is running, block
>> +    /// till the handler has finished.
>
> Here's another case where "running" is confusingly used to refer to
> the timer being in the state before the handler has begun to execute,
> and also to the state after the handler has begun to execute.

Thanks for catching this. How is this:

    /// Cancel the timer, if it is in the started or running state. If the timer
    /// is in the running state, block till the handler has finished executing.



Thanks for the comments!


Best regards,
Andreas Hindborg

Powered by blists - more mailing lists