time: delay the Arc::clone of the scheduler::Handle until registering timer

[ADD-SP]

blocks #7467

Background

This improvement was found while working on the delayed cancellation (#7384),
Since I don't like to include a un-relevant change into a big patch, I made it a separate commit

This might be a low-hanging fruit.

Motivation

tokio/tokio/src/time/sleep.rs

Lines 250 to 256 in 0a3fe46

	pub(crate) fn new_timeout(
	deadline: Instant,
	location: Option<&'static Location<'static>>,
	) -> Sleep {
	use crate::runtime::scheduler;
	let handle = scheduler::Handle::current();
	let entry = TimerEntry::new(handle, deadline);

The current implementation always clone the scheduler::Handle for each timer, even this timer is not registered.

There are two usage of this handle for timer:

1. Ensure the time driver is enabled.
   

   tokio/tokio/src/runtime/time/entry.rs

   Lines 480 to 484 in 0a3fe46

   	impl TimerEntry {
   	#[track_caller]
   	pub(crate) fn new(handle: scheduler::Handle, deadline: Instant) -> Self {
   	// Panic if the time driver is not enabled
   	let _ = handle.driver().time();

2. Registering or clear the entry from the global wheel.
   

   tokio/tokio/src/runtime/time/entry.rs

   Lines 590 to 595 in 0a3fe46

   	if reregister {
   	unsafe {
   	self.driver()
   	.reregister(&self.driver.driver().io, tick, inner.into());
   	}
   	}

For (1), A &Handle is enough, no need to make a clone.

For (2), we can delay the .clone() until we are about to register the entry.

Delaying the Arc::clone improves the performance on multi-core machine.

Solution

• Storing the &schedule:::Handle in the TimerShared.

Benchmark (AMD64 16 cores)

[ADD-SP]

For reviewers

Hiding whitespace reduces diffs.

[ADD-SP]

For reviewers

Triggering panic inside the closure doesn't report the panic location of the caller, so we need to panic outside the closure.

[ADD-SP]

superseded by #7473