Bug 1062368 (CVE-2014-0049)

The problem occurs when the guest performs a pusha with the stack address
pointing to an mmio address (or an invalid guest physical address) to
start with, but then extending into an ordinary guest physical address.
When doing repeated emulated pushes emulator_read_write sets mmio_needed
to 1 on the first one.  On a later push when the stack points to regular
memory, mmio_nr_fragments is set to 0, but mmio_is_needed is not set
to 0.

As a result, KVM exits to userspace, and then returns to
complete_emulated_mmio.  In complete_emulated_mmio
vcpu->mmio_cur_fragment is incremented.  The termination condition of
vcpu->mmio_cur_fragment == vcpu->mmio_nr_fragments is never achieved.
The code bounces back and fourth to userspace incrementing
mmio_cur_fragment past it's buffer.  If the guest does nothing else it
eventually leads to a a crash on a memcpy from invalid memory address.

However if a guest code can cause the vm to be destoryed in another
vcpu with excellent timing, then kvm_clear_async_pf_completion_queue
can be used by the guest to control the data that's pointed to by the
call to cancel_work_item, which can be used to gain execution.

Introduced by:
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=f78146b0f

Acknowledgements:

Red Hat would like to thank Lars Bull of Google for reporting this issue.