In the mips, mips64, and even the i386 arch, arch/kernel/smp.c has
this in smp_call_function:

        spin_lock(&call_lock);
        call_data = &data;

        /* Send a message to all other CPUs and wait for them to respond */
        for (i = 0; i < smp_num_cpus; i++)
                if (i != cpu)
                        core_send_ipi(i, SMP_CALL_FUNCTION);

call_data isn't volatile, it's a plain static *. So how can we be sure
that "call_data = &data" does anything other than change a register?

The i386 has a wb() after the assignment; we don't even have that.

call_data is neither local nor static to the function, so the modification
of the storage location would seem to be mandatory for the compiler
before the call to core_send_ipi(), so I can see how the code, as written, 
would generally work on most MIPS CPUs.  However, it would be legal 
for the compiler to defer the store until *just* before the invocation of 
core_send_ipi(), and on moderately complex,  high-ILP processors 
it seems to me like the wb() might well be necessary.  I take it that
you've observed a problem with this on your system?: