`SpinLockIrq` and `SpinLock` use the exact same underlying C structure,
with the only real difference being that the former uses the
irq_disable() and irq_enable() variants for locking/unlocking. These
variants can introduce some minor overhead in contexts where we already
know that local processor interrupts are disabled, and as such we want a
way to be able to skip modifying processor interrupt state in said
contexts in order to avoid some overhead - just like the current C API
allows us to do.
In order to do this, we add some special functions for SpinLockIrq:
lock_with() and try_lock_with(), which allow acquiring the lock without
changing the interrupt state - as long as the caller can provide a
LocalInterruptDisabled reference to prove that local processor
interrupts have been disabled.
In some hacked-together benchmarks we ran, most of the time this did
actually seem to lead to a noticeable difference in overhead:
From an aarch64 VM running on a MacBook M4:
lock() when irq is disabled, 100 times cost Delta { nanos: 500 }
lock_with() when irq is disabled, 100 times cost Delta { nanos: 292 }
lock() when irq is enabled, 100 times cost Delta { nanos: 834 }
lock() when irq is disabled, 100 times cost Delta { nanos: 459 }
lock_with() when irq is disabled, 100 times cost Delta { nanos: 291 }
lock() when irq is enabled, 100 times cost Delta { nanos: 709 }
From an x86_64 VM (qemu/kvm) running on a i7-13700H
lock() when irq is disabled, 100 times cost Delta { nanos: 1002 }
lock_with() when irq is disabled, 100 times cost Delta { nanos: 729 }
lock() when irq is enabled, 100 times cost Delta { nanos: 1516 }
lock() when irq is disabled, 100 times cost Delta { nanos: 754 }
lock_with() when irq is disabled, 100 times cost Delta { nanos: 966 }
lock() when irq is enabled, 100 times cost Delta { nanos: 1227 }
(note that there were some runs on x86_64 where lock() on irq
disabled vs. lock_with() on irq disabled had equivalent benchmarks,
but it very much appeared to be a minority of test runs.)
While it's not clear how this affects real-world workloads yet, let's
add this for the time being so we can find out.
This makes it so that a `SpinLockIrq` will work like a `SpinLock` if
interrupts are disabled. So a function:
(&'a SpinLockIrq, &'a LocalInterruptDisabled) -> Guard<'a, .., SpinLockBackend>
makes sense. Note that due to `Guard` and `LocalInterruptDisabled`
having the same lifetime, interrupts cannot be enabled while the Guard
exists.
Signed-off-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Gary Guo <gary@garyguo.net>
Link: https://patch.msgid.link/20260807070218.27144-19-boqun@kernel.org
ec4fad7c2b