[PATCH v2 0/2] KVM: x86: Wake blocked vCPUs before offlining a CPU

[Jinmeng Zhou <jinmengzhou22@gmail.com>]

From: Jinmeng Zhou <zhoujinmeng@bytedance.com>

A vCPU blocked with posted interrupts enabled stays on the wakeup list of
the physical CPU on which it blocked, and its posted-interrupt descriptor
keeps targeting that CPU's APIC.  If that CPU is offlined, a later device
interrupt sends the wakeup notification to a CPU that is gone, and the vCPU
is never woken.

Patch 1 turns CPUHP_AP_X86_KVM_CLK_ONLINE into a generic x86 KVM CPU
hotplug state that is always registered.  Today it is registered only on
hosts without a constant TSC.

Patch 2 adds an optional vendor cpu_offline() callback that runs from that
state's teardown, after the CPU has been deactivated.  VMX uses it to wake
the vCPUs on the dying CPU's wakeup list, so that they move to an active
CPU and rebuild their posted-interrupt state.

Patch 1 is a prerequisite for patch 2, so both are tagged for stable and
should be backported together.

Changes in v2:
- Reuse and generalize CPUHP_AP_X86_KVM_CLK_ONLINE instead of adding an
  arch hook to the generic CPUHP_AP_KVM_ONLINE path, which depends on
  KVM_GENERIC_HARDWARE_ENABLING=y.
- Register the x86 KVM CPU hotplug state unconditionally, so that the
  callback also runs on hosts with a constant TSC.
- Rename the vendor callback from prepare_cpu_offline() to cpu_offline().
- Use a bare __kvm_vcpu_wake_up() instead of KVM_REQ_UNBLOCK, so that the
  guest does not get spurious wakeups.
- Fix the changelog: the migration is started by CPUHP_AP_ACTIVE
  (sched_cpu_deactivate()), not CPUHP_AP_SCHED_WAIT_EMPTY.
- Initialize the per-CPU posted-interrupt wakeup lists before registering
  the CPU hotplug callbacks that walk them.
- Fix the trailers: add Co-developed-by and put the sender's SoB last, and
  add an in-body From: line for the author address.

v1: https://lore.kernel.org/kvm/20260924080648.51014-1-zhoujinmeng@bytedance.com/

Jinmeng Zhou (2):
  KVM: x86: Make the kvmclock CPU hotplug state generic
  KVM: x86: Wake blocked vCPUs before offlining a CPU

 arch/x86/include/asm/kvm-x86-ops.h |  1 +
 arch/x86/include/asm/kvm_host.h    |  1 +
 arch/x86/kvm/vmx/main.c            |  1 +
 arch/x86/kvm/vmx/posted_intr.c     | 18 ++++++++++++++++++
 arch/x86/kvm/vmx/posted_intr.h     |  1 +
 arch/x86/kvm/vmx/vmx.c             | 16 ++++++++++------
 arch/x86/kvm/x86.c                 | 30 ++++++++++++++++++++++++++----
 include/linux/cpuhotplug.h         |  2 +-
 8 files changed, 59 insertions(+), 11 deletions(-)


base-commit: a500db7819c50db59e55f1b4fa1c3baa5a2616f3
-- 
2.39.5