nested_get_vmcs12_pages() processes the posted_intr address in vmcs12. It
caches the kmap()ed page object and pointer, however, it doesn't handle
errors correctly: it's possible to cache a valid pointer, then release
the page and later dereference the dangling pointer.

I was able to reproduce with the following steps:

1. Call vmlaunch with valid posted_intr_desc_addr but an invalid
MSR_EFER. This causes nested_get_vmcs12_pages() to cache the kmap()ed
pi_desc_page and pi_desc. Later the invalid EFER value fails
check_vmentry_postreqs() which fails the first vmlaunch.

2. Call vmlanuch with a valid EFER but an invalid posted_intr_desc_addr
(I set it to 2G - 0x80). The second time we call nested_get_vmcs12_pages
pi_desc_page is unmapped and released and pi_desc_page is set to NULL
(the "shouldn't happen" clause). Due to the invalid
posted_intr_desc_addr, kvm_vcpu_gpa_to_page() fails and
nested_get_vmcs12_pages() returns. It doesn't return an error value so
vmlaunch proceeds. Note that at this time we have a dangling pointer in
vmx->nested.pi_desc and POSTED_INTR_DESC_ADDR in L0's vmcs.

3. Issue an IPI in L2 guest code. This triggers a call to
vmx_complete_nested_posted_interrupt() and pi_test_and_clear_on() which
dereferences the dangling pointer.

Vulnerable code requires nested and enable_apicv variables to be set to
true. The host CPU must also support posted interrupts.

Fixes: 5e2f30b7 "KVM: nVMX: get rid of nested_get_page()"
Cc: stable@vger.kernel.org
Reviewed-by: Andy Honig <ahonig@google.com>
Signed-off-by: Cfir Cohen <cfir@google.com>
Reviewed-by: Liran Alon <liran.alon@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>