commit 8f254105da5ca54d48bc3ed6e42e3925654f973e
author Venkatesh Srinivas <venkateshs@google.com> Wed Oct 28 00:30:12 2020 +0000
committer CQ Bot <commit-bot@chromium.org> Wed Oct 28 00:30:12 2020 +0000
tree e041962b80416addc899df106b2b7c3eba6a7c09
parent f3a2aae397f3abe9fb0ec898e82ac7058ddfe5cc

[kernel][x86] Convert halt_interlock from legacy int -> ktl::atomic

|halt_interlock| is a per-CPU flag used in the HLT-version of the x86
idle loop; it is used in the adaptive spin before HLT, so that remote
wakeups can avoid an IPI while a target CPU is spinning.

Specifically:
. When a CPU is idle, it executes the idle thread; the thread executes
  the HLT instruction in a loop until there is something else to do.
. HLT may have a long entry / exit latency
. Before we HLT, we spin a bit, checking for work; while we're spinning
  remote cores can wake us with a pure store. After we enter halt, we need
  to use an IPI to wake a sleeping core.

1. A core that's about to sleep sets a per-cpu variable,
   'halt_interlock', to 1.
2. It then spins a bit, checking that halt_interlock is still == 1
3. It then uses CMPXCHG to switch halt_interlock -> 2.
4. Then it halts

A wakeup does:
1. CMPXCHG switches halt_interlock from 1->0
2. If it fails, we use an IPI to do a heavyweight wake

Convert the flag from a volatile int to a ktl::atomic<>.

Bug: 47117 Use only ktl::atomic in the kernel

Change-Id: I63a6b57e5d906e7d22e8b19919c4d176b8d372c2
Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/441925
Commit-Queue: Venkatesh Srinivas <venkateshs@google.com>
Reviewed-by: Nick Maniscalco <maniscalco@google.com>
Testability-Review: Nick Maniscalco <maniscalco@google.com>