On a second thought ...
I think we could eliminate the kfree() instead, and keep the
atomicity of IPI cross-calls. Linus expressed doubts about
the IPI kmalloc()/kfree() pair we do in the generic SMP IPI
code, suggesting that it probably does not help performance
all that much - so such a change might be upstream-able as well
and would keep -rt closer to mainline.