Hi,
This series aim to fix multiple issues in how KVM defines & handles
pending vs. injected events. Both in non-nested and nested
scenarios.
The 1st patch removes wrong unnecessary code of clearing pending
NMI/interrupt when injecting event in real-mode when coming from
one of the VMCS event-injection methods (vmx_queue_exception(),
vmx_inject_nmi(), vmx_inject_irq()).
The 2nd patch renames interrupt.pending to interrupt.injected in
order to better represent it's meaning and make it aligned with the
meaning of "pending" & "injected" event for exceptions & NMIs.
A "pending" event should represent an event that it's side-effect
have not been applied yet. In contrast, an "injected" event should
represent an event that it's side-effect have been applied.
The 3rd patch fix some misleading comments in inject_pending_events()
regarding why exception.pending blocks re-injection of NMI/Interrupt
and why it is evaluated first among the pending events.
The 4th patch fixes a critical bug in nVMX which cause L1 to
lose an IPI when it is sent in the exact moment a destination L2 CPU
exits to L0 due to event-delivery.
The last patch removes now irrelevant code which sets
KVM_REQ_EVENT on conditions when it is unnecessary.
Changes from v1 to v2:
* Move bail-out of inject_pending_event() on event re-injection to be
after we evaluated pending exception.
* Removed patch moving call to vmx_complete_nested_posted_interrupt()
in vmx_check_nested_events() to start of function.
Decided to handle that in another patch series.
Changes from v2 to v3:
* Removed patches related to user-mode management of
{interrupt,exception}.{pending.injected}.
Handling them requires delicate backwards-compatabile work.
Therefore, decided to handle that in another patch series.
* Modified patch renaming interrupt.pending to interrupt.injected to
clarify issues with interrupt.pending usage in case
!lapic_in_kernel(vcpu). Added relevant FIXME comments.
Will be handled in future patches.
Regards,
-Liran Alon
