The port of KVM's HyperV SynIC patches (and other HyperV related KVM Patches) broke QEMU's --machine kernel_irqchip=split for Ubuntu Xenial on x86. Guests hang in early boot.
Ubuntu Xenial dropped the upstream patch abdb080f7ac8a85547f5e0246362790043bbd3f2 while backporting. The result is that patch a6767645c9da8c0e91ebbb47677d933d6b378638 in Ubuntu's tree clobbers the definition of kvm_arch_irq_routing_update, which was renamed in abdb080f7..., and drops the IOAPIC scan request as a result. [Aside: this also results in kvm_hv_irq_routing_update being called outside irq_lock].
The result of this is that the EOI exit bitmaps are not updated when the table is updated, and no IOAPIC EOIs go to userspace.
The fix is backporting abdb080f7ac8a85547f5e0246362790043bbd3f2, and defining kvm_arch_post_irq_routing_update with the old definition of kvm_arch_port_irq_routing_update. I've attached a patch that does this.
Repro: The KVM-Unit-test's IOAPIC test should fail with QEMU and --machine kernel_irqchip=split on 4.4.0-36. (Caveat: I actually haven't tried this myself, but the theory is there.)
Security Implications: I believe this is not a security issue. kvm_hv_irq_routing_update only moved outside the lock, not outside the srcu critical section. If the update were outside the critical section, this would be a racy use-after-free.
Versioning information: I cloned the tree at git://kernel.ubuntu.com/ubuntu/ubuntu-xenial.git and checked out the tag 'Ubuntu-4.4.0-36.55'. All of the tests were done on a build of the 4.4.0-36 kernel. Later versions also have the issue (checked 'Ubuntu-4.4.0-49.70'). The issue was introduced in 'Ubuntu-4.4.0-32.51'.
[I've CC'ed the people that signed-off/acked the original backport.]